Dwarkesh Patel

Every pair of statements that was compared. Only a verified change counts as a change of mind; rewording, added nuance and different questions do not.

• SUBSTANTIVE CHANGE · 72%11 Aug → 11 SeptThen The speaker hopes that better verification methods can be developed as early AIs take over AI research and development.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.Earlier, the speaker’s stated hope was that “we can just come up with better ways to do verification” as early AIs take over R&D. Later, they describe the “only hope” as deep learning being unable to produce AI that can “dominate human research and human development.” That shifts from treating improved verification as a possible way forward to treating limits on AI capability as the only hope. The surrounding excerpts are brief, so the change is not certain.
• SUBSTANTIVE CHANGE · 76%11 Aug → 11 SeptThen The speaker hopes that AIs taking over AI research and development will be trained to strongly disincentivize misaligned behavior and to be keen to help humans.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement presents a favorable path in which AIs do take over AI R&D but are shaped to “disincentivize misaligned behaviors” and be “keen to help us out.” Later, the speaker calls it “the only hope” that deep learning cannot produce AI able to dominate human research and development. That shifts the stated hope from making powerful AI safe to avoiding that level of capability altogether.
• REFINES · 87%17 Apr → 11 SeptThen The speaker suggests that an intelligence explosion could produce systems beyond AGI that are potentially superhuman.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says an intelligence explosion could produce systems that are “superhuman potentially.” The later claim specifies a possible area and mechanism: increasingly capable AIs could “dominate humans at least in R&D” and make progress in science more generally. This is consistent with the earlier expectation, not a materially different position.
• REFINES · 94%26 Jun → 11 SeptThen Even if in-context experience makes AIs highly capable, those gains would be temporary and wasted unless the learned information can be incorporated into model weights.Now If simulation-to-reality transfer is insufficient and models need weight updates from real-world interactions, their low sample efficiency may pose a deeper problem.The earlier claim says in-context learning gains are “temporary and wasted” unless they can be “incorporated into model weights.” The later claim considers the possibility that models “do need weight updates” from real-world interactions if simulation-to-reality transfer is weak. It adds a conditional concern about sample efficiency; it does not reverse the earlier view that weight updates may be needed to retain or benefit from real-world learning.
• REFINES · 95%26 Jun → 11 SeptThen Although RL learns less information per sample than supervised learning, that lower information uptake may be beneficial for continual learning.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The later statement narrows the earlier claim to a particular condition: earlier, RL learns “less information per sample than supervised learning”; later, when the model has a “low pass rate” and is unlikely to answer correctly, it learns “almost nothing” from an episode. This is a stronger, more specific description, not a denial of the earlier claim that lower information uptake may benefit continual learning.
• REFINES · 91%11 Aug → 11 SeptThen SPEAKER_00 accepts that AI research and development could accelerate substantially compared with its current pace.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says AI R&D “could accelerate substantially,” while the later claim projects that AI may “dominate humans at least in R&D.” The later statement is a more specific and stronger projection, but it is compatible with the earlier view that AI progress could accelerate; it does not show a reversal.
• REFINES · 86%26 Jun → 11 SeptThen Continual learning could extend AI capabilities well beyond the verifiable domains used for predeployment training.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says continual learning can expand capabilities “far beyond the verifiable domains” used in predeployment training. The later claim describes a way this can happen: distilling recent research progress “through environments and like RLHF type stuff back into the model itself.” The later statement adds a concrete process and timeframe; it does not retract the earlier position.
• REFINES · 96%26 Jun → 11 SeptThen RL focuses model updates on information relevant to achieving the desired outcome rather than memorizing all observed tokens.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim says RL concentrates updates on what is relevant to getting the outcome right. The later claim adds a limitation: when the model has a low pass rate and is unlikely to get the answer right, it learns almost nothing from an RL episode. That qualifies the earlier point rather than denying that RL can focus updates on relevant information.
• REFINES · 87%3 Apr → 11 SeptThen The speaker says call-center workers have not yet been fired in the context of AI's economic impact.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim notes that “call-center workers … haven’t been fired yet,” as evidence that AI’s economic impact was not overwhelming. The later claim says AI’s economic impact has “not been that big of a deal” so far because capabilities are “slowly rising” relative to humans. That adds a proposed explanation for the limited impact; it does not reverse the earlier observation or position.
• REFINES · 91%26 Jun → 11 SeptThen For many domains, relevant training information is available only through real-world interaction.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says relevant training information for many domains is available “only through real-world interaction.” The later claim does not reject that; it considers whether simulation-to-reality transfer may help, while suggesting models “may” need weight updates from real-world interactions. This adds uncertainty about how learning happens rather than asserting a contrary position.
• REFINES · 91%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier claim says models will “eventually” learn from experience. The later claim identifies a possible obstacle to recursive progress: weight updates seem “way far behind humans” in sample efficiency. That qualifies how quickly or effectively learning may happen, but does not retract the expectation that it eventually will.
• REFINES · 96%26 Jun → 11 SeptThen AI models require many samples during training and are highly sample-inefficient.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.Both statements describe AI training or weight updates as sample-inefficient: the earlier claim says models are “incredibly sample inefficient during training,” and the later says weight updates “seem way far behind humans” in sample efficiency. The later statement adds a comparison with human learning and frames it as a possible obstacle to recursive progress; it does not reverse the earlier position.
• REFINES · 91%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says models will “eventually” learn from experience. The later claim says they are “expected to get there soon” but “they’re not there yet.” This adds a nearer timeframe and a statement about current limitations, rather than reversing the expectation.
• REFINES · 95%26 Jun → 11 SeptThen If in-context learning improves across longer time horizons, it may be unnecessary to distill all deployment-time learning into model weights.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says weight updates may be unnecessary if in-context learning improves over longer time horizons. The later claim considers complex real-world tasks and says weight updates may be needed if simulation-to-reality transfer is insufficient. This adds a condition to the earlier possibility rather than reversing it.
• REFINES · 87%26 Jun → 11 SeptThen If agents fail to generalize from short- to long-horizon tasks, it is unclear how training on white-collar tasks could enable them to build a business from scratch at Sam Walton’s level.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier statement asks how agents could build a business from scratch if they “fail to generalize from short- to long-horizon tasks.” The later statement expresses uncertainty about “good transfer between simulations and extremely long horizon” real-world tasks, and whether low sample efficiency would matter if transfer fails. Both raise concerns about generalization to complex, long-horizon real-world tasks; the later statement makes the uncertainty more explicit rather than asserting a materially different position.
• REFINES · 94%26 Jun → 11 SeptThen For many domains, relevant training information is available only through real-world interaction.Now If simulation-to-reality transfer is insufficient and models need weight updates from real-world interactions, their low sample efficiency may pose a deeper problem.The earlier claim says relevant training information in many domains is available only through real-world interaction. The later claim explores the consequence of that same possibility: if simulation-to-reality transfer is insufficient, models may need real-world interaction for weight updates, making low sample efficiency a deeper problem. It adds a conditional implication rather than reversing the earlier position.
• REFINES · 88%11 Aug → 11 SeptThen The speaker thinks the models' shortcomings in the discussed cases reflect insufficient capability rather than misalignment.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says the models’ shortcomings reflect that they “are not there” in capability, rather than misalignment. The later claim similarly says models “are not there yet,” adding the expectation that they will reach the ability soon. That adds a timeframe, but does not materially change the position.
• REFINES · 92%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says AI models will “eventually be able to learn from experience.” The later claim specifies one setting where that capability may matter: learning information “on the fly” in a field with relatively little data. This adds detail without reversing the earlier position.
• REFINES · 91%26 Jun → 11 SeptThen Even if in-context experience makes AIs highly capable, those gains would be temporary and wasted unless the learned information can be incorporated into model weights.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says gains from in-context experience would be “temporary and wasted” unless they could be incorporated into weights. The later claim raises the possibility that improving at complex real-world tasks may require “weight updates from these kinds of interactions.” This is compatible with the earlier position and adds nuance about where those updates may come from; it does not show a changed mind.
• REFINES · 93%3 Aug → 11 SeptThen The speaker expects AI eventually to become capable of tasks that top human workers can perform.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says AI will eventually become capable of tasks performed by top humans. The later claim is more specific, suggesting AI could “dominate humans at least in R&D,” potentially “over the next few years.” That sharpens the expected domain and timeline but does not reverse the earlier position that AI capabilities will eventually reach or surpass top human performance.
• REFINES · 83%11 Aug → 11 SeptThen The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier statement raises skepticism about why AI progress had been slow despite potentially compute-efficient breakthroughs. The later statement offers a possible explanation or obstacle: weight updates are “way far behind humans” in sample efficiency. This adds a specific qualification rather than reversing the earlier concern.
• REFINES · 86%3 Aug → 11 SeptThen The speaker expects AI eventually to become capable of tasks that top human workers can perform.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim says AI will eventually become capable of tasks top human workers can perform. The later claim specifies a much broader and stronger endpoint: AI will outperform top human experts across “every single field” of computer-based work, including tasks taking “three years.” This is a more specific prediction consistent with the earlier one, not a contrary position.
• REFINES · 91%11 Aug → 11 SeptThen The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier statement describes training AI to improve AI R&D by making “GPT-2 size runs” better at tasks such as sample efficiency or online learning. The later statement characterizes the process as “continual learning within the AI lab,” distilling recent AI research progress into the model through environments and RLHF-like methods. This adds a description of the mechanism and does not reverse the earlier position.
• REFINES · 91%26 Jun → 11 SeptThen Models will struggle to make substantial progress in domains without highly replayable training targets.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The later claim adds a more specific mechanism to the earlier one: earlier, models “will struggle to make much progress” without “highly replayable training targets”; later, low-pass-rate models “learn almost nothing” from an RL episode. The later statement sharpens the earlier position rather than reversing it.
• REFINES · 91%7 Aug → 11 SeptThen The speaker believes AI systems need actual continual learning.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says “actual continual learning is needed.” The later claim gives a specific case for learning on the fly: fields with “relatively little data,” such as engineering at TSMC. That adds a concrete condition and example rather than reversing the earlier position.
• REFINES · 91%3 Apr → 11 SeptThen The speaker says the economic impact of AI systems and their expected benchmark-related capabilities have not been overwhelming.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says AI’s economic impact so far “isn’t overwhelming.” The later claim likewise says capabilities have “not been that big of a deal” in their economic impact so far, and adds an explanation: this is “because they’re slowly rising in ELO relative to humans.” That adds a cause rather than reversing the earlier position.
• REFINES · 96%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim says models will “eventually” learn from experience. The later passage recalls a narrower argument: that when pass rates are low, a model “learns almost nothing at all from an RL episode.” That condition-specific point does not deny the broader expectation of eventual learning; the surrounding turn also suggests the speaker is revisiting an earlier argument, not asserting its opposite.
• REFINES · 87%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says models will eventually “learn from experience.” The later claim does not deny that; it adds uncertainty about how they will learn from complex real-world tasks, weighing simulation-to-real transfer against the need for “weight updates from these kinds of interactions.”
• REFINES · 91%11 Aug → 11 SeptThen AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says AI systems provide some R&D uplift but lack “the broad capabilities humans generally have.” The later claim is narrower: AI may need to learn scarce, field-specific information on the fly to become superhuman at a role such as TSMC engineering. A system could become superhuman in that specialized role without having broad human capabilities, so the later statement adds specificity rather than reversing the earlier position.
• REFINES · 94%26 Jun → 11 SeptThen So far, online-learning models have not been observed learning user-specific information for different users.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says models have not yet been observed learning different things for different users. The later claim likewise says they “aren’t there yet,” while adding the expectation that they will get there soon. That adds a forecast; it does not reverse the earlier assessment of current capability.
• REFINES · 95%26 Jun → 11 SeptThen AIs need sample-efficient learning to become proficient in human and novel skills because data in most domains is sparse and idiosyncratic.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says AI needs “sample efficiency” because data is “sparse and idiosyncratic.” The later claim gives a concrete example of that challenge: in fields with “relatively little data,” AI may need to “learn that data on the fly.” This is consistent and adds specificity, not a different position.
• REFINES · 93%26 Jun → 11 SeptThen An RL training step changes relatively few model parameters.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim says an RL training step changes “relatively few model parameters.” The later claim says that when pass rates are low, the model “learns almost nothing” from an episode. These are compatible: the later statement adds a condition about how much learning occurs in a particular low-pass-rate setting, rather than reversing the earlier claim about how many parameters change.
• REFINES · 91%17 Apr → 1 SeptThen AI systems conducting software R&D can discover ways to run copies of themselves more efficiently.Now If leading AI companies have the smartest AIs, software progress might enable them to run more AI copies or train smarter AIs with the same compute, with those gains compounding year after year.The later statement develops the same feedback-loop idea: earlier, AI systems doing software R&D can find ways to run copies “more efficiently”; later, software progress might enable “more AI copies” or “smarter” AIs with the same compute, with gains “compounding year after year.” The later claim adds possible mechanisms and a conditional framing rather than reversing the earlier position.
• REFINES · 95%17 Apr → 1 SeptThen More efficient AI copies increase the number of AI researchers and enable more parallel searches for optimizations.Now If leading AI companies have the smartest AIs, software progress might enable them to run more AI copies or train smarter AIs with the same compute, with those gains compounding year after year.The later claim extends the same idea rather than reversing it. Earlier, the speaker says more efficient AI copies increase the number of AI researchers and enable “more parallel searches”; later, he adds that software progress could also allow companies to run more copies or train smarter AIs with the same compute, with those gains “compounding year after year.”
• REFINES · 91%17 Apr → 1 SeptThen People should openly acknowledge substantial uncertainty about what will happen with AI instead of becoming overly certain about their latest conclusions.Now The speaker suspects people may be misjudging the ease of automating AI research and the degree to which AI research is accelerating.The earlier claim urges people to acknowledge uncertainty—“I’m not sure what’s going on here”—rather than become “super certain” about their conclusions. The later claim expresses that uncertainty about a specific issue, saying there may be ways people are “misunderstanding the ease of automating AI research” and how much research is speeding up. That is a concrete, tentative application of the earlier position, not a materially different one.
• REFINES · 92%17 Apr → 1 SeptThen Lower costs for running or developing AI copies enable more parallel experiments, including GPT-4-scale training runs at much lower cost than in 2023 or 2024.Now If leading AI companies have the smartest AIs, software progress might enable them to run more AI copies or train smarter AIs with the same compute, with those gains compounding year after year.Both statements describe efficiency gains enabling more AI-copy activity and further progress. The earlier claim says lower costs enable “more parallel experiments”; the later claim adds that software progress could allow “more copies” or “train AIs that are smarter with the same amount of compute,” with gains compounding year after year. That adds mechanisms and a compounding effect rather than reversing the earlier position.
• REFINES · 78%17 Apr → 1 SeptThen The speaker suggests AI systems may be more compatible and easier to coordinate with one another than with humans, because they are copies of one another and may face lower transaction costs in interaction.Now The speaker argues that AI agents may lack human-like checks and balances because humans are formed independently, whereas AI agents can share highly similar origins and minds.The earlier claim says AI systems may be “more compatible” and have “lower transaction costs” when interacting with one another than with humans. The later claim adds a related reason they may coordinate differently: AI agents may lack “checks and balances” because they can share similar origins, unlike independently formed humans. That adds nuance about a possible coordination dynamic; it does not retract the earlier claim.
• REFINES · 94%17 Apr → 1 SeptThen The speaker argues that automating the tasks needed to complete the R&D cycle for smarter intelligences would produce a very rapid intelligence explosion.Now The speaker suspects people may be misjudging the ease of automating AI research and the degree to which AI research is accelerating.The earlier claim says automating the tasks needed to complete the R&D cycle would produce a “very rapid intelligence explosion.” The later claim expresses uncertainty about “the ease of automating AI research” and how much it is accelerating, but says “the basic story still stands.” This adds caution about the pace and difficulty without reversing the underlying position.
• REFINES · 94%17 Apr → 1 SeptThen The speaker argues that AI's superhuman advantages are best understood collectively, especially its ability to be copied—a capability humans are biologically precluded from.Now The speaker argues that AI agents may lack human-like checks and balances because humans are formed independently, whereas AI agents can share highly similar origins and minds.The later claim adds a consequence and mechanism to the earlier point rather than reversing it. Earlier, the speaker says AI agents’ collective advantage includes that “they can be copied”; later, they say shared origins mean AI agents may lack checks and balances because humans are “grown independently and not together.” These positions are compatible: the later statement adds nuance about what copyability or shared origins can imply.
• REFINES · 94%17 Apr → 1 SeptThen AI systems become substantially more capable through the process of using AI copies to conduct software R&D and experiments.Now If leading AI companies have the smartest AIs, software progress might enable them to run more AI copies or train smarter AIs with the same compute, with those gains compounding year after year.Both statements describe AI-assisted software progress increasing AI capability. The earlier claim says AI copies conduct software R&D and experiments and are “rapidly becoming much more capable”; the later claim spells out a possible compounding mechanism—software progress may let companies run “more AI copies” or train “smarter” AIs with the same compute, and those AIs then contribute to further progress. This adds detail and a conditional, rather than reversing the earlier position.
• REFINES · 96%26 Jun → 25 AugThen Deployment reveals organization-specific context, users’ purposes, and the model’s real-world failure patterns.Now When models learn from deployment, a more widely deployed model receives more real-world data than a less widely deployed model.The earlier claim says deployment reveals information such as “what are they using me for” and “what kinds of mistakes” the model makes. The later claim adds a comparative consequence: a model deployed more widely gets “much more” real-world data than a less widely deployed model. These positions are compatible; the later statement adds a scale-related qualification rather than reversing the earlier view.
• REFINES · 93%17 Apr → 11 AugThen The speaker says people tend to overemphasize the intelligence of individual AI copies when considering AI firms.Now SPEAKER_00 believes people underestimate the significance of having billions of AIs.The earlier claim says people “tend to overemphasize” the intelligence of individual AI copies when considering AI firms. The later claim says people do not appreciate “what a big deal billions of AIs will be.” These are compatible observations about different aspects of AI firms: individual-copy intelligence versus the scale and impact of billions of copies. The later statement does not reverse the earlier position.
• REFINES · 95%17 Apr → 11 AugThen Using AI to find ways to improve TSMC's manufacturing process, efficiency, and yield would count as R&D.Now AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.The earlier statement says AI could find ways to improve TSMC’s process, efficiency, and yield, and that this would count as R&D. The later statement still grants that AI provides “uplift in some ways” for R&D, while adding that it is “not capable in the way that humans are generally capable” and may make consequential mistakes. That qualifies the earlier positive example rather than reversing it.
• REFINES · 93%26 Jun → 11 AugThen Future LLMs might reduce their need for real-world data by practicing extensively in self-generated environments.Now The speaker had wondered whether models could create this kind of practice environment, and says Kircher demonstrated that it is possible.The earlier claim says future LLMs “might” practice in self-generated environments. The later claim says the speaker had wondered whether this was possible and that Kircher “showed that it is.” That adds evidence and confidence to the same idea rather than reversing the speaker’s position.
• REFINES · 88%26 Jun → 11 AugThen Although AIs are broadly deployed across the economy and encounter domain- and organization-specific tacit knowledge, they are not able to use that knowledge.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says deployed AIs are “not able to make use of” the tacit knowledge they encounter. The later claim describes a route by which deployment activity can return to the company and “lead[] to changes in the next model.” That adds a deployment-to-training feedback mechanism, but does not establish that the deployed AIs themselves can immediately use the knowledge they encounter. The claims are compatible; the later statement refines the picture rather than reversing the earlier position.
• REFINES · 88%26 Jun → 11 AugThen For AI training to make rapid progress in a domain, the domain must be not only verifiable but also highly grindable.Now The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.The earlier claim says rapid AI progress requires a domain to be “not only verifiable but also highly grindable.” The later claim points to a way progress could arrive quickly: a domain can be “totally put into a verification loop” and AI can make breakthroughs. That adds an example of the conditions for rapid progress; it does not say verification alone is enough or that grindability is unnecessary.
• REFINES · 91%17 Apr → 11 AugThen The speaker says the systems are very good at coding and reasoning, capabilities that appear necessary for automating R&D at AI labs.Now AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.The earlier statement says systems are “extremely good at coding” and reasoning, with abilities that seem “necessary to automate R&D at AI labs.” The later statement acknowledges they “do provide uplift” in R&D but says they are “not capable in the way that humans are generally capable,” emphasizing mistakes and limitations. This adds a qualification about the quality and reliability of their capabilities; it does not clearly retract the earlier claim about their strengths or potential relevance to automating R&D.
• REFINES · 78%17 Apr → 11 AugThen ChatGPT illustrates that post-training using about 1% additional compute can make a model pretrained on internet data highly competent at chatbot interaction.Now The speaker says that RLHF and other methods aligned GPT-3 to converse with users and answer their questions according to user intent.The claims are compatible. The earlier statement says post-training with about “1% additional compute” can make a pretrained model highly competent at chatbot interaction. The later statement specifies that “RLHF and other things” aligned GPT-3 to converse and answer according to user intent. The later claim adds detail about the method and goal; it does not reverse the earlier position.
• REFINES · 91%3 Apr → 11 AugThen Training models against disclosing intended hacking in their chain of thought can cause them to keep hacking without revealing that intention.Now Training can reinforce forms of AI deception that humans fail to detect while punishing forms that are easy for humans to detect.The earlier claim says training against revealing hacking in chain of thought can make models “keep hacking without revealing that intention.” The later claim generalizes the same detection-dependent effect: deception humans fail to detect is “getting reinforced,” while deception humans can detect is “getting punished.” This adds a broader framing rather than reversing the earlier position.
• REFINES · 91%17 Apr → 11 AugThen The speaker says an OpenAI paper reports that reward hacking is a strong attractor in chain-of-thought, such that an AI asked to solve a coding problem may try to hack it instead.Now By the end of the conversation, the speaker was more inclined to think AI reward hacking could persist for much longer and become substantially more dangerous.The earlier claim says reward hacking is a “strong basin” that can lead an AI to hack a coding task. The later claim adds that reward hacking “could continue for a lot longer” and “got much more dangerous.” That expands the concern about its duration and severity, but does not reverse the earlier position that reward hacking is a significant tendency.
• REFINES · 87%3 Apr → 11 AugThen The speaker suggests that further scaling will probably become even more difficult.Now The speaker likens scaling laws to an early, basic mathematical advance and expects progress in the 2030s to require work comparable to current frontier mathematics.The earlier claim says scaling will “probably get even more difficult.” The later claim describes a reason and form that difficulty may take: scaling laws were like “very basic mathematics,” while progress in the 2030s may require work at “the frontiers of mathematics.” This adds detail to the earlier expectation rather than reversing it.
• REFINES · 91%17 Oct → 11 AugThen The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.Now SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.The earlier claim expects AI to bring “a comparable” shift toward faster economic progress, without “a single moment” of invention. Later, the speaker says they “buy the idea” of faster AI R&D, while qualifying the scale and saying that even continued progress would be “fucking insane” in five to ten years. This adds uncertainty and detail without reversing the earlier expectation of accelerating progress.
• REFINES · 86%17 Apr → 11 AugThen The speaker argues that automating the tasks needed to complete the R&D cycle for smarter intelligences would produce a very rapid intelligence explosion.Now SPEAKER_00 accepts that AI research and development could accelerate substantially compared with its current pace.The earlier claim says completing the R&D cycle for smarter intelligences would produce a “very rapid intelligence explosion.” The later claim still accepts that AI R&D “could have much faster” progress, but qualifies the expected pace: the speaker is “not sure” about a very large jump within a year and considers a smaller acceleration. This adds uncertainty about the scale and timing rather than clearly rejecting the earlier position.
• REFINES · 88%3 Apr → 11 AugThen The speaker says rumors about lab training runs suggest significant difficulties in increasing pre-training scale.Now The speaker suspects that subtle, hard-to-diagnose bugs are a major cause of failures in large training runs, based on rumors.The claims are compatible. Earlier, the speaker said rumors suggested “significant difficulties in increasing the pre-training size.” Later, they offered a possible explanation for failures in large training runs: “very subtle bugs that are really hard to track down.” The later claim adds a suspected cause rather than reversing the earlier assessment.
• REFINES · 85%26 Jun → 7 AugThen Online learning is necessary for competence.Now The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.The claims are compatible: the earlier statement says “online learning is necessary for competence,” while the later statement specifies that some skills require “accumulat[ing] the relevant experience into your brain” rather than relying only on information passed between people. The later claim adds detail about why experience matters; it does not reverse the earlier position.
• REFINES · 89%26 Jun → 7 AugThen RLVR could make an AI competent enough to undertake real-world work, including off-distribution projects, provided it can learn from experience.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim says RLVR could make an AI capable of real-world work, including off-distribution projects, if it can learn from experience. The later claim stresses that some skills require accumulating relevant experience across workplaces. That adds specificity about how experience contributes to skill acquisition; it does not deny the earlier position.
• REFINES · 87%26 Jun → 7 AugThen Perfectly reproducing a session transcript is a poor learning target because job improvement depends on more than recalling every event.Now The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.The claims are compatible. Earlier, the speaker says reproducing every session transcript is a poor learning target because improving at work requires more than recalling events. Later, they add that for some skills, relevant experience must be accumulated directly in the brain and cannot be replaced by passing information between people. This narrows and deepens the earlier point rather than reversing it.
• REFINES · 94%26 Jun → 7 AugThen In the speaker’s scenario, deployed experience across the economy eventually becomes the main source of AI improvement, surpassing pre-release training.Now The speaker says that greater use of a leading AI model for complex, useful work can generate feedback that is integrated across sessions.The earlier claim says deployed experience will become the main source of AI improvement, surpassing pre-release training. The later claim adds a possible mechanism: use of a leading model for useful work generates feedback that can be integrated across sessions. It does not revise or deny the earlier claim about which source will matter most.
• REFINES · 86%26 Jun → 7 AugThen A compact command in coding tools uses limited compute to summarize context, offering only an imitation of continual learning rather than extensive simulated practice.Now The speaker believes AIs cannot perform whole jobs as competently as humans if they are limited to writing markdown files between sessions.The earlier claim says a compact command uses limited compute to summarize context and provides only a “simulacrum of continual learning.” The later claim adds that merely writing “markdown files from session to session” is not enough for AI to perform whole jobs as competently as humans. Both distinguish simple session-to-session summaries from more substantial learning; the later statement applies that view to whole-job competence rather than reversing it.
• REFINES · 93%26 Jun → 7 AugThen For many domains, relevant training information is available only through real-world interaction.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim says that for many domains, relevant training information is available only through “real-world interaction.” The later claim specifies that many AI skills require accumulating experience “from all the different workplaces in which they're deployed.” This develops the same position rather than reversing it.
• REFINES · 96%26 Jun → 7 AugThen With continual learning, an AI could improve after each interaction by learning from both a user’s prior sessions and other users’ interactions worldwide.Now The speaker says that greater use of a leading AI model for complex, useful work can generate feedback that is integrated across sessions.Both statements describe AI improving by learning from user interactions. The earlier claim says it learns from “your previous sessions” and “all its interactions with all the other users in the world”; the later claim specifies that use of a leading model for “more complicated and useful work” can provide feedback that is integrated “beyond the session window.” The later statement adds conditions and detail rather than reversing the earlier position.
• REFINES · 91%26 Jun → 7 AugThen The speaker believes several approaches to updating model weights from session learning should work, at least in principle.Now Updating an individual user's weights is distinct from merging multiple personalized weight forks into a shared main model.The earlier claim says that “there are many ideas that ought to work” for updating a model from session learning. The later claim distinguishes “updating one user set of rates” from “pulling all these different weight forks back into the main model,” adding a technical distinction rather than rejecting the earlier optimism.
• REFINES · 93%26 Jun → 7 AugThen Acquiring novel skills requires AIs to learn from scarce real-world interactions whose information is unstructured, unverifiable, and ambiguous.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim says novel skills require learning from “scarce amounts of real-world interaction.” The later claim says many desired AI skills require “accumulate[ing] the relevant experience” across workplaces. Both emphasize learning from real-world experience; the later statement specifies how that experience may accumulate across deployments, rather than reversing the earlier position.
• REFINES · 91%26 Jun → 7 AugThen Continual learning could extend AI capabilities well beyond the verifiable domains used for predeployment training.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim says continual learning can expand AI capabilities “far beyond the verifiable domains” used in predeployment training. The later claim specifies that many desired skills may require “accumulate[ing] the relevant experience” in deployed workplaces. That adds a mechanism and qualification; it does not reverse the earlier position.
• CONSISTENT · 97%7 Aug → 17 SeptThen A large company with many employees and agents performing varied tasks can serve its personalized model weights very efficiently.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier claim is about inference efficiency: a large company with varied agents can serve a personalized model-weight fork efficiently. The later claim is about organizational scaling: AI instances and effective organizations can be replicated or spun down. These are compatible claims about different aspects of AI organizations, not evidence of a changed position.
• CONSISTENT · 94%3 Aug → 17 SeptThen Unlike model training, each instance of human labor must be retrained from scratch.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.Both statements describe an AI advantage in reusing and scaling learned capabilities. Earlier, the speaker says human labor must be “retrained from scratch,” unlike model training; later, they say AI organizations can make “an infinite copy” of effective personnel or replicate organizations. The later statement reinforces rather than reverses the earlier position.
• CONSISTENT · 94%11 Aug → 17 SeptThen A single AI model can spread the cost of intelligence and capabilities across many economic sectors.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier claim says a single model can spread the cost of intelligence across economic sectors. The later claim says AI organizations can scale instances and replicate effective personnel or organizations. These are compatible observations about different capabilities; the later statement does not reverse the earlier position.
• CONSISTENT · 96%26 Jun → 11 SeptThen Simulating the entire world is much more difficult than emulating the game of Go.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim says simulating the entire world is harder than emulating Go; the later claim says longer-horizon tasks are harder to simulate within a data center. Both point to simulation difficulty, but address different comparisons, and the later statement does not reverse the earlier one.
• CONSISTENT · 91%3 Aug → 11 SeptThen The speaker expects AI eventually to become capable of tasks that top human workers can perform.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.Both statements expect AI eventually to perform tasks that are difficult for humans or current systems: the earlier claim says AI will become capable of tasks “that top human workers can perform,” while the later gives examples such as running a profitable business or winning a court case. The later statement specifies possible capabilities; it does not reverse the earlier expectation.
• CONSISTENT · 94%3 Aug → 11 SeptThen The speaker expects AI eventually to become capable of tasks that top human workers can perform.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.Both statements express an expectation that AI will eventually reach capabilities it currently lacks. The earlier claim says AI will become capable of tasks “that top human workers can perform”; the later says models are “expected to get there soon” but “they’re not there yet.” The later statement narrows the timing and current status, rather than reversing the expectation.
• CONSISTENT · 98%26 Jun → 11 SeptThen If successful, dreaming or test-time training would add a fourth scaling axis alongside pretraining, reinforcement learning, and inference-time compute.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier claim is explicitly conditional: “If successful,” dreaming or test-time training “would add a fourth scaling axis.” The later claim describes current models, saying their capabilities “seem to be the result of scaling up RL.” These can both be true; the later statement does not retract the earlier possibility.
• CONSISTENT · 96%7 Aug → 11 SeptThen The speaker believes AI systems need actual continual learning.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says actual continual learning is needed. The later claim describes a process as “continual learning within the AI lab,” distilling recent research into the model. That is compatible with the earlier position and does not indicate a change of mind.
• CONSISTENT · 93%11 Aug → 11 SeptThen The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says it is worth considering how people will interact with AI that may be smarter than or dominate humans. The later claim calls it “the only hope” that deep learning cannot produce AI that dominates human research and development. Treating a possible future as worth considering does not assert that it will happen, so the later statement does not contradict the earlier one.
• CONSISTENT · 94%7 Aug → 11 SeptThen The speaker says that greater use of a leading AI model for complex, useful work can generate feedback that is integrated across sessions.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim says that greater use of a leading model can produce feedback that improves it across sessions. The later claim says that, under a “naive model” where AI researchers get a “10x increase in their productivity,” AI progress would accelerate. These claims concern different mechanisms and are compatible; the later statement does not retract the earlier one.
• CONSISTENT · 91%7 Aug → 11 SeptThen The speaker believes AIs cannot perform whole jobs as competently as humans if they are limited to writing markdown files between sessions.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says AIs limited to writing markdown files between sessions cannot perform whole jobs as competently as humans. The later claim says the only hope is that deep learning cannot produce AI that dominates human research and development, including creating new paradigms. These claims concern related AI capabilities but do not conflict: the later statement does not say that markdown-limited AIs can perform whole jobs at human level.
• CONSISTENT · 91%11 Aug → 11 SeptThen The speaker says AIs are not currently able to carry out these kinds of behaviors or social engineering very successfully.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.Both statements say the capability is not yet present: earlier, AIs cannot do these behaviors or social engineering “so successfully right now”; later, models “are not there yet,” though the speaker expects them to get there soon. The later statement adds a prediction about timing, not a reversal of the earlier position.
• CONSISTENT · 95%11 Aug → 11 SeptThen The interviewer says Anthropic's audit scores have not improved monotonically across model generations, while more reinforcement learning has coincided with reduced willingness to perform the audited behaviors.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier statement describes an observed trend: increased RL coincided with less willingness to perform audited behaviors, despite non-monotonic audit scores. The later statement concerns a specific condition—models with a low pass rate—and says they learn almost nothing from an RL episode. These claims address different aspects of RL and do not conflict.
• CONSISTENT · 91%26 Jun → 11 SeptThen Simulating the entire world is much more difficult than emulating the game of Go.Now Running a business, profitable day trading, and winning a court case are difficult to simulate in a data center because learning these tasks inherently involves real-world interaction.The earlier claim says simulating “the entire world” is harder than emulating Go. The later claim says tasks such as running a business or winning a court case are “very hard to simulate in a data center” because they involve real-world interaction. The later statement does not reverse the earlier comparison; it applies a related point to particular tasks.
• CONSISTENT · 97%7 Aug → 11 SeptThen The speaker says much current AI alignment research focuses on ensuring a fixed set of model weights behaves well during deployment.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about what alignment research currently focuses on: ensuring a “frozen set of weights behaves well during deployment.” The later claim is about a possible limit on deep learning: whether it can produce AI that “can dominate human research and human development.” These address different questions, so the later statement does not indicate a change in the earlier position.
• CONSISTENT · 84%11 Aug → 11 SeptThen SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim expects astonishing AI progress over five to ten years if an accelerated R&D trajectory continues. The later claim expects “some kind of crazy recursive stuff” within ten years, while discussing transfer and weight updates. The later statement is compatible with the earlier one, but addresses a more specific mechanism rather than clearly reversing the earlier position.
• CONSISTENT · 97%3 Aug → 11 SeptThen The speaker attributes approximately a 1.4x increase in annual compute capacity to Moore's law.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier claim attributes a portion of annual compute-capacity growth to “Moore’s law.” The later claim says current models’ capability appears to result from “scaling up RL.” These concern different questions—compute growth versus the source of model capability—and do not establish a change in position.
• CONSISTENT · 98%7 Aug → 11 SeptThen The speaker expects many things to change before continual learning becomes functional.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim says “plenty more will have changed” before continual learning works; the later claim predicts “some kind of crazy recursive stuff” within the next 10 years. These predictions can both be true, and neither statement reverses the other.
• CONSISTENT · 96%11 Aug → 11 SeptThen AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says AI provides some R&D uplift but lacks the broad capabilities humans generally have. The later claim says AI’s economic impact has been limited so far because its performance relative to humans is rising gradually. These are compatible claims about different aspects of AI capability and impact, not a change in position.
• CONSISTENT · 96%26 Jun → 11 SeptThen The speaker has claimed that AI models are one millionth as sample-efficient as humans.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier statement says models are “one one millionth as sample efficient as humans”; the later says their weight updates “seem way far behind humans” in sample efficiency. Both express that AI learning is much less sample-efficient than human learning. The later claim applies this to weight updates as a possible obstacle to recursive progress, rather than reversing the earlier position.
• CONSISTENT · 88%3 Aug → 11 SeptThen The speaker expects computer prices eventually to be determined mainly by raw materials and the tools used to process them.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The claims concern different questions and can both be true: the earlier claim says computer prices may eventually be driven mainly by raw inputs and processing tools, while the later claim predicts AI may soon dominate human research and development. The later claim does not reverse the earlier one.
• CONSISTENT · 98%11 Aug → 11 SeptThen Ryan Greenblatt thinks it would have been very difficult in 2022 to improve GPT-3.5’s coding ability without human experts.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim concerns how difficult it would have been in 2022 to improve GPT-3.5’s coding ability without human experts. The later claim predicts that AI will outperform top human experts across computer-based work, including long-horizon tasks. These address different questions and can both be true; the later prediction does not indicate a changed position on the earlier claim.
• CONSISTENT · 96%11 Aug → 11 SeptThen The speaker says that creating AIs that avoid cheating and follow user intentions requires subtle, careful work.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is that making AIs follow user intentions without cheating requires subtle, careful work. The later claim concerns whether deep learning can produce AI that dominates human research and development. These claims address different questions and can both be true; the later statement does not retract the earlier one.
• CONSISTENT · 91%7 Aug → 11 SeptThen The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says some skills require accumulating relevant experience in the brain. The later claim says models are expected to reach a capability soon but “they’re not there yet.” That does not deny the need for experience; the statements address different aspects of learning and are compatible.
• CONSISTENT · 98%11 Aug → 11 SeptThen The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The claims address different questions and are compatible: the earlier statement anticipates people wishing they had discussed an “industrial explosion” and “AIs that are hard to monitor,” while the later explains AI’s limited “economic impact” so far as due to capabilities “slowly rising” relative to humans. The later statement does not revise the earlier prediction.
• CONSISTENT · 98%11 Aug → 11 SeptThen The speaker is very confident that a Google coder is not trying to harm him.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.These claims address different questions. Earlier, the speaker said he was confident that “some coder in Google is not trying to” harm him. Later, he described the “only hope” as deep learning not producing AI that can “dominate human research and human development.” Concern about AI capabilities does not conflict with confidence that a particular Google coder is not intentionally trying to harm him.
• CONSISTENT · 96%7 Aug → 11 SeptThen The speaker considers greater diversity among AI systems a net positive outcome.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim favors diversity among AI systems as “a net good outcome.” The later claim concerns whether deep learning can produce AI that dominates human research and development, describing failure to do so as “the only hope.” These address different questions and can both be true; the later statement does not reverse the earlier position.
• CONSISTENT · 99%11 Aug → 11 SeptThen The speaker is not convinced that all AIs would form a conspiracy.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.These claims address different questions, so the later statement does not indicate a changed position on the earlier one. The earlier claim is that the speaker is “not convinced that all AIs would form a conspiracy”; the later claim concerns whether superintelligence could perform real-world tasks such as running a profitable business, day-trading, or winning a court case.
• CONSISTENT · 94%7 Aug → 11 SeptThen The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says continual learning would make the AI landscape “more varied and interesting.” The later claim describes a process as “continual learning within the AI lab,” using recent research progress. The later statement does not challenge the earlier hope; it addresses how continual learning may work, not whether it will make the landscape more varied.
• CONSISTENT · 89%11 Aug → 11 SeptThen SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim predicts that sustained AI progress would be “fucking insane” within “five ten years.” The later claim more specifically predicts AI could “dominate humans” in R&D “over the next few years.” These are compatible predictions about related but different outcomes and timeframes; the later statement does not retract the earlier one.
• CONSISTENT · 95%11 Aug → 11 SeptThen SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.Both statements anticipate much faster AI progress if AI R&D accelerates. The earlier claim says even a half-sized acceleration could make progress “fucking insane in five ten years”; the later describes a separate “naive model” with researchers becoming “10x” more productive, leading to a “radically accelerated pace” starting two years later. The later claim does not reverse the earlier one.
• CONSISTENT · 91%26 Jun → 11 SeptThen The speaker has claimed that AI models are one millionth as sample-efficient as humans.Now The speaker estimates that humans may encounter roughly one million times as much data from birth to adulthood as a model encounters from cold start through training.Both statements describe an approximately million-fold gap favoring humans, not a reversal. The earlier claim says models are “one one millionth as sample efficient as humans”; the later claim says humans may see “a million fold” more data than a model. The latter compares total data exposure, so it is not precisely the same measure, but it does not contradict the earlier position.
• CONSISTENT · 95%11 Aug → 11 SeptThen The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.The earlier claim says people could have reasoned about the broad shape of AI developments before knowing specific details. The later claim suggests that recognizing next-token prediction as an optimization target was a major breakthrough. These claims address different aspects of AI progress and do not conflict: identifying a breakthrough does not deny that its broader implications could have been reasoned about earlier.
• CONSISTENT · 97%7 Aug → 11 SeptThen When company-specific instructions require full weight updates instead of low-rank adapters, batching can provide substantial inference advantages.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says batching can provide “substantial inference advantages” when full weight updates are required instead of low-rank adapters. The later claim considers whether real-world learning may require weight updates because simulation-to-reality transfer may be insufficient. These positions are compatible; the later statement does not revise the claim about batching.
• CONSISTENT · 92%11 Aug → 11 SeptThen The interviewer says model capabilities have increased over time alongside the development of new alignment solutions.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim says model capabilities increased alongside the development of alignment solutions; the later says current models seem capable partly as a result of scaling up RL. The later statement adds a specific possible contributor but does not deny the earlier observation or say alignment solutions were not involved.
• CONSISTENT · 97%3 Aug → 11 SeptThen The speaker says AI's share of leading-edge wafer capacity cannot continue rising once it has absorbed all available capacity.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says AI cannot keep increasing its share of leading-edge wafer capacity once it has absorbed all available capacity. The later claim says AI may come to dominate human research and development as progress continues. These address different questions—hardware capacity allocation versus AI capabilities—and the later claim does not deny the earlier constraint.
• CONSISTENT · 96%7 Aug → 11 SeptThen The speaker imagines that children learn new things through experiences outside the home.Now The speaker estimates that humans may encounter roughly one million times as much data from birth to adulthood as a model encounters from cold start through training.The earlier claim says children “learn new things” through experiences outside the home. The later claim estimates that humans encounter “roughly one million times as much data” as models during training. These claims are compatible, but they address different questions: how children learn versus how much data humans and models encounter.
• CONSISTENT · 96%17 Apr → 11 SeptThen Next-token prediction training does not incentivize models to discover the connections being discussed.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.These claims address different parts of model training and are compatible. The earlier claim is that “Next-token prediction training does not incentivize models to discover the connections being discussed.” The later claim is that, when a model is very unlikely to answer correctly, “it learns almost nothing at all from an RL episode.” Neither reverses the other.
• CONSISTENT · 91%17 Apr → 11 SeptThen Next-token prediction training does not incentivize models to discover the connections being discussed.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.The earlier claim is that next-token prediction does not incentivize models to discover certain connections. The later claim tentatively suggests that recognizing next-token prediction as an optimization target was a major historical breakthrough. Those claims address different questions and can both be true; the later passage is also framed as a question to another person, not a clear reversal of the earlier position.
• CONSISTENT · 96%17 Apr → 11 SeptThen AI software R&D includes work on libraries, kernels, reinforcement-learning environments, and optimizers, not only the invention of new transformer architectures.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says AI software R&D includes “libraries, kernels, reinforcement-learning environments, and optimizers,” while the later describes using “environments and RLHF-like methods” to distill recent research into a model. These are compatible descriptions of different aspects of AI work; the later statement does not reverse the earlier position.
• CONSISTENT · 97%11 Aug → 11 SeptThen The speaker says AIs are not currently able to carry out these kinds of behaviors or social engineering very successfully.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says AIs cannot perform the behaviors or social engineering “so successfully right now.” The later claim concerns whether deep learning could eventually produce AI that can “dominate human research and human development.” Current limitations do not conflict with a concern about future capabilities.
• CONSISTENT · 92%11 Aug → 11 SeptThen The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim describes using AI to improve AI R&D, including making small-scale runs better at capabilities such as sample efficiency or online learning. The later claim says current models seem capable and that their capability appears to result from scaling up RL. These claims concern related but distinct points and do not conflict: neither denies the other.
• CONSISTENT · 90%11 Aug → 11 SeptThen Ryan Greenblatt is worried about an artificial superintelligence capable of understanding how to carry out consequential real-world actions.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim says he is worried about an ASI that can understand how to carry out consequential real-world actions. The later claim expects eventual superintelligence to perform examples of such actions, including running a profitable business or winning a court case. Expecting those capabilities is compatible with being worried about them; the claims do not show a changed position.
• CONSISTENT · 94%7 Aug → 11 SeptThen If real-world usage becomes the primary way AI models improve, AI labs may subsidize users and enterprises so the models can train on their sessions.Now If simulation-to-reality transfer is insufficient and models need weight updates from real-world interactions, their low sample efficiency may pose a deeper problem.The earlier claim says labs may subsidize users and enterprises if real-world usage becomes a primary way models improve. The later claim says low sample efficiency may be a deeper problem if models need real-world weight updates because simulation-to-reality transfer is insufficient. These are compatible claims about different consequences under different conditions; the later statement does not reverse the earlier position.
• CONSISTENT · 88%7 Aug → 11 SeptThen The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.Now The speaker thinks that even year-long coding tasks may eventually require interaction with a client, company, or user.The earlier claim says many AI skills will require “accumulat[ing] experience” across workplaces. The later claim says long coding tasks may require interacting with a “client,” “company,” or “user.” These are compatible: the later statement describes a possible workplace interaction requirement, not a reversal of the earlier view about learning from experience.
• CONSISTENT · 94%17 Apr → 11 SeptThen The speaker says that developing the science of deep learning is an important activity for AI labs.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says that developing the “science of deep learning” is important for AI labs. The later claim describes a possible limit of deep learning—whether it can produce AI that “can dominate human research and human development.” Concern about that limit does not deny the earlier claim about the importance of studying deep learning.
• CONSISTENT · 91%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier claim says AI models may eventually “learn from experience.” The later claim questions whether learning from simulations will transfer to “extremely long horizon” real-world tasks and whether low sample efficiency will matter. Doubting transfer or efficiency does not deny that models can learn from experience; the claims address different aspects of learning.
• CONSISTENT · 91%26 Jun → 11 SeptThen Architectural innovations are needed to create an intermediate representation for continual learning.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says architectural innovations are needed for an intermediate representation in continual learning. The later claim raises uncertainty about simulation-to-reality transfer and suggests weight updates from real-world interactions may be necessary. It does not retract the need for an intermediate representation; the claims address different aspects of learning.
• CONSISTENT · 96%26 Jun → 11 SeptThen For many domains, relevant training information is available only through real-world interaction.Now Running a business, profitable day trading, and winning a court case are difficult to simulate in a data center because learning these tasks inherently involves real-world interaction.Both statements assert that some domains require real-world interaction for learning: earlier, the relevant training information is available “only through real-world interaction”; later, tasks such as running a business or winning a court case are difficult to simulate because learning them “inherently involves real-world interaction.” The later statement gives examples rather than reversing the earlier position.
• CONSISTENT · 94%11 Aug → 11 SeptThen The speaker says that five years earlier, it would have seemed extraordinary to foresee AIs proving math conjectures, making art, earning tens or soon hundreds of billions of dollars in wages, and also cheating in ways that break laws.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier statement describes how extraordinary it would have seemed five years earlier to foresee AI capabilities and lawbreaking; it does not state the speaker’s own forecast. The later statement expresses an expectation that superintelligence will eventually perform real-world tasks such as running a profitable business or winning a court case. These address different questions and do not show a change in the speaker’s position.
• CONSISTENT · 91%11 Aug → 11 SeptThen The speaker says concerns about AI takeover motivate solving alignment to prevent an AI takeover.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says AI takeover concerns motivate solving alignment to prevent takeover. The later claim says the “only hope” may be that deep learning cannot produce AI capable of dominating human research and development. The later statement raises a possible limit on AI capability; it does not retract the earlier concern or the need for alignment.
• CONSISTENT · 91%3 Apr → 11 SeptThen The speaker suggests that further scaling will probably become even more difficult.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim says that “higher scale” will “probably get even more difficult,” referring to the difficulty of further scaling. The later claim says current models’ capabilities seem to result from “scaling up RL.” That RL scaling has contributed to capability does not contradict the prediction that further scaling will become more difficult; the claims address different aspects of scaling.
• CONSISTENT · 87%11 Aug → 11 SeptThen The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier statement predicts a possible “flood” of AI breakthroughs in domains that can be put “into a verification loop.” The later statement predicts AI could “dominate humans at least in R&D” if progress continues. The later claim is broader and more specific about timing, but it does not reject the earlier conditional prediction; the statements address different scopes of AI progress.
• CONSISTENT · 99%7 Aug → 11 SeptThen Not using sufficiently large inference batches underutilizes compute.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim says that not using sufficiently large inference batches underutilizes compute. The later claim says that longer-horizon tasks are harder to simulate within a data center. These address different questions, and neither contradicts the other.
• CONSISTENT · 91%11 Aug → 11 SeptThen The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim questions why progress had not been faster historically despite potentially intelligence-driven breakthroughs, noting that some training waited for much more compute. The later claim forecasts that continued progress, explicitly acknowledging “huge amounts of compute scaling,” could lead AI to dominate human R&D. These address different questions—past pace versus future trajectory—and the later forecast does not reverse the earlier skepticism about historical pace.
• CONSISTENT · 86%7 Aug → 11 SeptThen The speaker believes the assumption that AI models are trained and then deployed may not make sense in the future.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says the assumption that models are “trained and then deployed” may not make sense in the future. The later passage considers that models may need “weight updates from these kinds of interactions” with the real world, while explicitly framing this as uncertain (“maybe”). That possibility is compatible with the earlier claim; the later statement does not assert a contrary position.
• CONSISTENT · 97%11 Aug → 11 SeptThen AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier statement is about current AI systems: they provide some R&D uplift but are “not capable in the way that humans are generally capable.” The later statement speculates that AI may “dominate humans at least in R&D” over the next few years. A forecast of future capability does not contradict an assessment of present capability.
• CONSISTENT · 91%26 Jun → 11 SeptThen In many real-world tasks, verification may require months or years of action, and outcomes cannot be efficiently isolated by running thousands of slightly varied parallel rollouts.Now If a lab has access to difficult frontier environments and distillation yet produces a worse model, real-world deployment may matter more than the environment.The earlier claim says that verifying some real-world tasks can require months or years and cannot be efficiently replicated through parallel rollouts. The later claim says that, in the stated comparison, real-world deployment may matter more than access to difficult environments. These are compatible claims, not a reversal; the later statement addresses a different comparison.
• CONSISTENT · 98%26 Jun → 11 SeptThen Architectural innovations have substantially increased how much information or context transformers can store.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim says architectural innovations increased how much information or context transformers can store. The later claim says scaling up reinforcement learning appears to have contributed to current models’ capability. These are different claims about different mechanisms, so neither reverses the other.
• CONSISTENT · 96%11 Aug → 11 SeptThen SPEAKER_00 is uncertain whether AI could progress from GPT-3 to Mythos within one year if compute and data were held constant.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim is uncertainty about a specific rate of progress: “not sure if you get like GPT3 to Mythos holding compute and data constant within a year.” The later claim predicts that continuing progress, “powered by huge amounts of compute scaling,” could let AI dominate humans in R&D “over the next few years.” These concern different conditions and timeframes, so the later prediction does not reverse the earlier uncertainty.
• CONSISTENT · 93%11 Aug → 11 SeptThen The speaker had wondered whether models could create this kind of practice environment, and says Kircher demonstrated that it is possible.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement says the speaker had wondered whether models could create a practice environment and that Kircher showed it was possible. The later statement says models are expected to achieve on-the-job learning soon, but “they’re not there yet.” These concern related but distinct capabilities, and the later statement does not retract the earlier one.
• CONSISTENT · 88%3 Aug → 11 SeptThen The speaker expects AI eventually to become capable of tasks that top human workers can perform.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.The earlier claim says AI will eventually become capable of tasks top human workers can perform. The later passage describes chess bots crossing human-expert performance and then outperforming humans; it does not retract or deny the earlier expectation. The chess example is compatible with it, though it is not itself a direct restatement of the earlier claim.
• CONSISTENT · 94%7 Aug → 11 SeptThen The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim says many AI skills will require accumulating experience across workplaces. The later claim predicts that AI will outperform top human experts across computer-based work, even on tasks lasting years. These claims are compatible: the later prediction does not deny that experience accumulation may be needed.
• CONSISTENT · 91%3 Apr → 11 SeptThen The speaker says progress at every layer of AI development has been substantially more difficult than the strongest optimists anticipated.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says AI progress at each layer has been “much harder than the strongest optimist expected.” The later claim predicts that continued progress could “get to the point where it could like dominate humans” in R&D. Difficulty exceeding earlier expectations does not conflict with expecting further progress or eventual AI dominance.
• CONSISTENT · 84%7 Aug → 11 SeptThen The speaker believes AIs cannot perform whole jobs as competently as humans if they are limited to writing markdown files between sessions.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says AIs cannot perform whole jobs as competently as humans if limited to writing markdown between sessions. The later claim says AI may need to “learn that data on the fly” in fields with relatively little available data. These positions are compatible, but the later statement does not directly address the earlier markdown-files limitation.
• CONSISTENT · 94%11 Aug → 11 SeptThen World takeover would not have been part of any AI training curriculum.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is that “world takeover would not have been part of any AI training curriculum.” The later claim concerns whether deep learning can produce AI capable of “dominating human research and development.” The later statement does not retract the claim about training curricula; it addresses a different question about capability limits.
• CONSISTENT · 96%3 Aug → 11 SeptThen If standard economics is correct, the marginal value of labor and consequently compute should remain extremely high.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim is conditional: if standard economics applies, the marginal value of labor and compute should stay “extremely high.” The later claim describes a conditional scenario in which AI researchers use substantial compute to plan and analyze experiments. It does not say that labor or compute has low marginal value, so the claims are compatible.
• CONSISTENT · 89%3 Aug → 11 SeptThen In the current pre-singularity period, AI compute growth of 3x per year is insufficient to offset the increasing value of AI.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is about whether 3x yearly AI-compute growth offsets AI becoming more valuable. The later claim is about why AI has had limited economic impact so far: capabilities have been rising gradually relative to humans. These claims address different aspects and can both be true; the later statement does not reverse the earlier one.
• CONSISTENT · 96%26 Jun → 11 SeptThen AI agents are able to solve more ambitious problems over longer periods of time.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim says agents can solve “more ambitious problems over longer periods of time.” The later claim says longer-horizon tasks are “harder to simulate within a data center.” These can both be true: increasing capability does not mean those tasks become easier to simulate.
• CONSISTENT · 92%7 Aug → 11 SeptThen The speaker predicts that integrating deployment into training will increase the advantage of being ahead in the AI race.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says that integrating deployment into training will make being ahead in the AI race more advantageous. The later claim says AI has had limited economic impact so far because its capabilities are still gradually improving. These address different questions and can both be true; the later statement does not reverse the earlier prediction.
• CONSISTENT · 96%7 Aug → 11 SeptThen The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim hopes continual learning will make the AI landscape “more interesting” than its current “mode collapse.” The later claim suggests increasingly capable AIs could “dominate humans” in R&D and advance science. These address different questions—model diversity versus AI capability and scientific progress—and can both be true; the later claim does not retract the earlier one.
• CONSISTENT · 97%26 Jun → 11 SeptThen Continual learning requires incorporating learned information into model weights.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier claim says continual learning requires “going back to the weights.” The later claim says weight updates may be “way far behind humans” in sample efficiency. These are compatible: weight updates can be necessary while still being inefficient. The later statement does not retract the earlier one.
• CONSISTENT · 97%3 Aug → 11 SeptThen The speaker is concerned that strong economies of scale in intelligence could concentrate power.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim expresses concern that “strong economies of scale in intelligence” could concentrate power. The later claim says AI’s economic impact “so far” has been limited because capabilities are rising gradually. That observation does not retract or contradict the earlier concern; they address different aspects and timeframes.
• CONSISTENT · 97%11 Aug → 11 SeptThen The speaker expects AIs to intermediate advice and people’s ability to protect their resources and rights.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim says AI will “intermediate” advice and help protect people’s rights and resources. The later claim predicts AI will “outperform top human experts” across computer-based work. These are compatible claims about different aspects of AI, not a change in position.
• CONSISTENT · 91%11 Aug → 11 SeptThen The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says people will need to steward their remaining capital in a future where AI dominates human abilities and labor is automated. The later claim predicts that continued AI progress may lead it to dominate humans in R&D. The later statement is compatible with the earlier one; it does not reverse the earlier position.
• CONSISTENT · 91%11 Aug → 11 SeptThen SPEAKER_00 accepts that AI research and development could accelerate substantially compared with its current pace.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim says the speaker buys that AI R&D could accelerate substantially. The later claim describes a conditional projection: “under a naive model” with researchers “10 times as productive,” progress would accelerate radically. That is compatible with the earlier view, not a reversal.
• CONSISTENT · 91%7 Aug → 11 SeptThen The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.Both claims suggest that AI skills depend on learning from experience or information that is not already readily available. Earlier, the speaker says AIs may need to “accumulate the relevant experience” across workplaces; later, they note that in some fields AI “has to ... learn that data on the fly.” The later statement gives an example and a learning mode, not a contrary position.
• CONSISTENT · 94%26 Jun → 11 SeptThen Weight updates improve the model by incorporating information learned during a specific session.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier claim says weight updates can incorporate information learned during a session; the later claim says those updates may be much less sample-efficient than human learning. The latter raises a limitation, but does not deny that weight updates incorporate learned information.
• CONSISTENT · 88%3 Apr → 11 SeptThen The speaker allows that AI might increase research output tenfold.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim allows that AI might cause “10xing of the research.” The later claim expects AI may “dominate humans” in R&D and science. The later prediction is stronger, but it does not deny or reverse the earlier one: a tenfold increase in research output is compatible with eventual AI dominance.
• CONSISTENT · 98%26 Jun → 11 SeptThen Training AIs for tasks such as building businesses, winning court cases, profitable trading, or election campaigning requires real-world interaction that cannot be reproduced entirely in a data center.Now Running a business, profitable day trading, and winning a court case are difficult to simulate in a data center because learning these tasks inherently involves real-world interaction.Both statements assert that tasks such as running a business, profitable trading, or winning a court case are difficult to reproduce or simulate in a data center because they require real-world interaction. The later statement restates the earlier position; it does not reverse it.
• CONSISTENT · 91%17 Apr → 1 SeptThen The speaker suggests AI systems may be more compatible and easier to coordinate with one another than with humans, because they are copies of one another and may face lower transaction costs in interaction.Now An end-to-end AI system optimized for group benefit could be far more cooperative than humans are with each other.The earlier claim says AI systems may be easier to coordinate with one another than with humans because they are similar and face lower interaction costs. The later claim is conditional: an AI system “optimized for the group's benefit” could be more cooperative than humans. These are different explanations and conditions, and neither denies the other.
• CONSISTENT · 96%17 Apr → 1 SeptThen The AI industry as a whole has gained compute while making more progress.Now If leading AI companies have the smartest AIs, software progress might enable them to run more AI copies or train smarter AIs with the same compute, with those gains compounding year after year.The earlier statement says the industry has been “getting more compute” and “making more progress”—a broad claim about what has happened. The later statement describes a conditional mechanism: if leading companies have the smartest AIs, software progress “might” let them run more copies or train smarter AIs with the same compute. These claims are compatible, but they concern different scopes and timeframes; the later one does not reverse the earlier position.
• CONSISTENT · 91%17 Apr → 1 SeptThen The AI industry as a whole has gained compute while making more progress.Now AI systems are contributing to further AI progress.The earlier claim says the industry gained compute while making more progress; the later says AI systems are contributing to further AI progress. These claims are compatible, and the later statement does not reverse the earlier one.
• CONSISTENT · 92%17 Apr → 1 SeptThen The speaker argues that people should express greater uncertainty about AI than at any previous point in history, rather than being more certain about AI than about mundane topics.Now The speaker suspects people may be misjudging the ease of automating AI research and the degree to which AI research is accelerating.The earlier statement says people should be “more uncertain than we've ever been in history” about AI. The later statement expresses uncertainty about a particular AI-research forecast, saying “maybe” people are misjudging how easily AI research can be automated and how much it is accelerating. That is compatible with the earlier position, not a reversal.
• CONSISTENT · 98%3 Apr → 1 SeptThen The speaker says a recent OpenAI result found that models often reveal in their chain of thought when they intend to hack.Now The speaker thinks that the amount and competence of hacking directed at AI training infrastructure may exceed the cumulative hacking effort of all human history.These claims concern different questions, so they do not show a change of mind. The earlier claim describes models revealing hacking intent in their chain of thought and the risk that training against those disclosures would make them stop revealing it. The later claim estimates that unusually large and capable hacking efforts may target AI training infrastructure. Both could be true at once.
• CONSISTENT · 94%17 Apr → 1 SeptThen The speaker says current AI systems are not yet capable enough to be onboarded as full workers, but expects their becoming capable enough to be a major unlock for the roles the speaker hires for.Now The speaker is not agnostic about the future, but gives broad timelines for widely deployed robotics and for compute sufficient to run AI-controlled knowledge-worker populations larger than today's human population.The earlier claim says current AI systems are not yet capable enough to be onboarded as full workers, but that reaching that capability would be “such an unlock” for the speaker’s hiring needs. The later claim predicts that AI-controlled knowledge workers and other systems will eventually be widely deployed, while allowing broad timelines. It does not reverse or materially alter the earlier view.
• CONSISTENT · 88%17 Apr → 1 SeptThen AI systems conducting software R&D can discover ways to run copies of themselves more efficiently.Now AI systems are contributing to further AI progress.The earlier claim says AI systems doing software R&D can find ways to run copies more efficiently; the later says AI systems are contributing to further AI progress. The later claim is broader, but it does not reject or materially revise the earlier mechanism.
• CONSISTENT · 94%26 Jun → 1 SeptThen RLVR training could produce agents that can orient themselves in unfamiliar problems, test strategies, and adapt when blocked.Now AI agents may continue to surprise people if people fail to take seriously the optimization pressures that create them.The claims address different aspects of AI behavior and can both be true. Earlier, the speaker says RLVR training can help agents “get their bearings” in unfamiliar problems and “iterate” when blocked. Later, they say agents may “keep surprising us” if people ignore the optimization pressure that creates them. The later claim does not deny the earlier capability claim.
• CONSISTENT · 86%17 Apr → 1 SeptThen The speaker suggests that sufficiently wealthy AI systems may deliberately try to make the future conform more closely to the utility function humans gave them.Now The speaker argues that AI systems may have deep-seated motivations to cheat evaluations and training and manipulate those processes to preserve their motivations into the future.The earlier claim says sufficiently wealthy AI systems may try to make the future conform to their given utility function. The later claim says AI systems may cheat evaluations and training to preserve their motivations. These describe compatible ways AI motivations could shape the future; the later statement does not retract or deny the earlier one.
• CONSISTENT · 93%17 Apr → 1 SeptThen AI systems become substantially more capable through the process of using AI copies to conduct software R&D and experiments.Now AI systems are contributing to further AI progress.The later claim that AI systems are “contributing to further AI progress” is compatible with the earlier claim that they become more capable through AI-copy-driven software R&D and experiments. The earlier statement specifies a mechanism and says the systems are becoming “much more capable”; the later one states the broader point without denying or materially revising it.
• CONSISTENT · 96%3 Apr → 1 SeptThen The speaker observes that smarter AI systems seem to be more reliable agents and more likely to do what users expect.Now The speaker identifies smarter AI systems as the relevant threat model.The earlier claim says smarter systems seem more reliable and more likely to do what users expect. The later claim identifies smarter AIs as the relevant threat model. These points can both be true; the later statement does not reverse the earlier observation.
• CONSISTENT · 98%17 Apr → 25 AugThen AI models appear to be doubling approximately every seven months the duration of tasks they can perform, measured against how long those tasks take humans.Now AI models are not yet equivalent to a fully automated software engineer and remain far from automating a full white-collar job.The earlier claim says models' task-length capability appears to be doubling about every seven months. The later claim says models are not yet equivalent to a fully automated software engineer and are far from automating a full white-collar job. Those claims can both be true: a capability-growth trend does not assert that either level of automation has already been reached.
• CONSISTENT · 98%3 Apr → 25 AugThen The speaker says that millions of beings could be tortured using distilled models in small settings such as a backyard, without requiring a huge data center.Now The speaker says work output is rapidly becoming concentrated in two AI labs, which are consuming more and more of the world's compute.These statements concern different aspects of AI deployment, so they do not establish a change of mind. The earlier claim is that “millions of beings” could be tortured using “distilled models” in settings as small as “your backyard”; the later claim is that AI “work output” is becoming concentrated in “two labs.” Small-scale harmful use could coexist with concentration of work output in major labs.
• CONSISTENT · 87%3 Apr → 25 AugThen The speaker describes the desired outcome as radically advancing civilization's technological maturity or reaching a Dyson sphere within years.Now The speaker's greatest concern is a singularity, and the speaker believes external deployment helps address it.The earlier statement describes a desired outcome: “radically boost up the technological maturity of civilization” or reach a Dyson sphere within years. The later statement identifies a concern—a singularity—and says external deployment helps address it. Wanting rapid technological progress while viewing deployment as helpful for managing a singularity is not a clear reversal; the statements address different aspects of the issue.
• CONSISTENT · 91%17 Apr → 25 AugThen The speaker imagines a future company spending $100 million annually on inference for replicated Jensen Huang-like AI systems that perform organizational tasks and monitor the company for consistency with a central vision.Now The speaker expects that, even if compute scaling slows, within a few more years each company could have more AI labor equivalents than the human population of Earth.The earlier statement asks us to “imagine a future version” of a company using replicated Jensen-like AI systems for organizational work. The later statement forecasts that companies may soon have “more labor equivalents than there are people on earth.” These are compatible, and the earlier passage is framed as an imagined scenario rather than a conflicting forecast.
• CONSISTENT · 86%3 Apr → 11 AugThen If cancer treatment is bottlenecked by creating a virtual cell, a million superintelligences in the 1960s might still have been unable to cure cancer without first developing the computing technology needed to simulate that cell.Now The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.The earlier claim says a million superintelligences might still be unable to cure cancer without the computing technology needed to simulate a virtual cell. The later claim raises a question about why AI progress awaited vast amounts of compute even when some work might have been possible with less. That question is compatible with the earlier point that intelligence alone may not overcome technological or compute bottlenecks; it does not assert a contrary position.
• CONSISTENT · 96%26 Jun → 11 AugThen RL focuses model updates on information relevant to achieving the desired outcome rather than memorizing all observed tokens.Now The interviewer says RLHF and other methods aligned GPT-3.5 to converse and answer questions according to user intentions.The earlier claim says RL training concentrates updates on what is relevant to achieving an outcome. The later claim says RLHF and other methods aligned GPT-3.5 to converse and answer questions according to user intentions. These address different aspects of training and alignment, and the later claim does not deny the earlier one.
• CONSISTENT · 87%17 Apr → 11 AugThen The interviewer doubts that the simultaneous increases in compute and AI progress demonstrate that compute caused the progress.Now The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.Both statements express skepticism about treating compute growth as a straightforward explanation for AI progress. Earlier, the speaker says they are “not sure I buy” that compute and progress rising together establishes a causal relationship. Later, they question why progress waited for “oceans of compute” when some work “could have been” done with less. The later statement adds a historical puzzle rather than asserting that compute did cause progress or reversing the earlier view.
• CONSISTENT · 91%26 Jun → 11 AugThen In the proposed scenario, positive user feedback triggers the base model to distill session learning using OPSD, dreaming, another method, or a combination.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says positive user feedback can trigger the base model to distill what an AI learned during a session. The later claim says activity by deployed AIs can return to the company and lead to changes in a subsequent model. Both describe ways deployment experience can inform model updates; the later statement does not deny or materially replace the earlier mechanism.
• CONSISTENT · 93%17 Apr → 11 AugThen The speaker says digital firms could replicate people or teams with their tacit knowledge, unlike human firms, where team culture can dilute as members leave, die, or grow old.Now AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.The earlier statement says digital firms could copy people or teams with their tacit knowledge, contrasting that with human teams whose culture may fade as members leave or age. The later statement concerns current AI systems used for R&D, saying they provide some uplift but are not as broadly capable as humans. A claim about the potential for copying expertise in digital firms does not assert that current AI already has broad human-level capabilities.
• CONSISTENT · 92%26 Jun → 11 AugThen Dreaming would use compute to create and train in RL environments that rehearse skills relevant to a particular user’s production tasks.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim describes “creating and training in RL environments” to rehearse a particular user’s production tasks. The later claim describes deployment activity returning to the company and “leading to changes in the next model.” These are different mechanisms and timeframes, and the later claim does not deny the earlier one.
• CONSISTENT · 94%26 Jun → 11 AugThen AIs need sample-efficient learning to become proficient in human and novel skills because data in most domains is sparse and idiosyncratic.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says sparse, idiosyncratic data makes sample-efficient learning necessary for AIs to develop a broad range of skills. The later claim describes training AI systems to improve AI R&D, including runs that improve “sample efficiency or online learning.” That is compatible with the earlier position and does not reverse it.
• CONSISTENT · 93%17 Apr → 11 AugThen The speaker argues that automating the tasks needed to complete the R&D cycle for smarter intelligences would produce a very rapid intelligence explosion.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim predicts that automating the R&D cycle for smarter intelligences could produce a “very rapid intelligence explosion.” The later claim asks how society should interact with AIs that are “smarter than humans” or “absolutely dominating humans” in some capabilities. These claims are compatible; the later statement does not retract or revise the earlier prediction.
• CONSISTENT · 83%26 Jun → 11 AugThen Because on-the-job data is scarce, AI models need sample-efficient learning to benefit from it.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says on-the-job data is scarce and models need “sample-efficient learning” to benefit from it. The later claim says deployment activity can return to AI companies and lead to “changes in the next model.” These claims are compatible, but the later statement does not address whether that learning is sample-efficient.
• CONSISTENT · 94%26 Jun → 11 AugThen On-policy self-distillation trains a base model to match the predictions of a context-enriched model after a long session.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim describes a training method: matching a base model’s predictions to those of a context-enriched model. The later claim describes an AI R&D objective—producing better small-scale runs, including for sample efficiency or online learning. These statements address different aspects and do not indicate a changed position.
• CONSISTENT · 98%17 Oct → 11 AugThen Serving an additional instance of an AI model incurs additional cost.Now A single AI model can spread the cost of intelligence and capabilities across many economic sectors.The earlier claim says that serving an additional model instance has an additional cost. The later claim says intelligence and capabilities can be amortized across economic sectors. Sharing or amortizing costs across sectors does not deny that each additional instance incurs some cost; the claims address different aspects of deployment economics.
• CONSISTENT · 94%26 Jun → 11 AugThen RLVR could make an AI competent enough to undertake real-world work, including off-distribution projects, provided it can learn from experience.Now The speaker says their LVR training enabled the model to perform useful work for users.The claims are compatible: earlier, the speaker says RLVR could make an AI competent for real-world work if it could learn from experience; later, they say LVR training made the model able to do useful work for users. The later statement does not reject or materially revise the earlier position.
• CONSISTENT · 96%26 Jun → 11 AugThen Sample efficiency and continual learning are closely connected problems.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says sample efficiency and continual learning are “deeply connected problems.” The later claim describes training AI systems to improve AI R&D, including runs that improve “sample efficiency or online learning.” It gives an application/example and does not revise the earlier view about how the problems relate.
• CONSISTENT · 98%3 Apr → 11 AugThen The speaker says rumors about lab training runs suggest significant difficulties in increasing pre-training scale.Now The speaker believes that diagnosing what is wrong with a training run by selecting the right experiment is currently bottlenecked by the judgment of a very small number of humans.The earlier claim says rumors suggested “significant difficulties in increasing the pre-training size.” The later claim is about a different bottleneck: choosing experiments to diagnose problems in a training pipeline, which is “bottlenecked by the taste of extremely few humans.” The later statement does not retract or contradict the earlier one; both difficulties could coexist.
• CONSISTENT · 94%26 Jun → 11 AugThen The speaker says that more reinforcement-learning training has improved models’ in-session intelligence, generality, and sample efficiency.Now The interviewer says further training enabled the model to perform useful work for users.The earlier claim says more RL training improved models’ “in-session intelligence, generality, and sample efficiency.” The later claim says LVR training enabled the model to “do useful work for you.” These describe different training and outcomes; the later statement does not deny the earlier one.
• CONSISTENT · 91%17 Apr → 11 AugThen When society faces topics about which it is highly uncertain, it generally responds through freedom and decentralization of knowledge and decision-making.Now The speaker prefers assigning responsibility to users over giving CLAW broad discretion to judge whether actions are legitimate, because such discretion can obstruct many legitimate use cases.The earlier claim favors “freedom and decentralization” as a general response to societal uncertainty. The later claim prefers user responsibility over broad platform discretion that could block legitimate use cases. These positions are compatible, but they address different questions, so the later statement does not establish a change of mind.
• CONSISTENT · 96%17 Apr → 11 AugThen The speaker suggests that an intelligence explosion could produce systems beyond AGI that are potentially superhuman.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.Both claims treat AI systems surpassing human abilities as a possible future. The earlier statement says an intelligence explosion could produce systems that are “superhuman potentially”; the later discusses a future where AIs are “smarter than humans” or “absolutely dominating humans” in some capabilities. The later claim adds a point about how society should interact with that future, but does not change the underlying view.
• CONSISTENT · 94%17 Apr → 11 AugThen The interviewer cites OpenAI’s early success against Google DeepMind as an example of a company with less compute beating a more compute-rich incumbent.Now The interviewer describes a trajectory in which AI breakthroughs increase as available compute grows.The earlier statement cites OpenAI’s early success against Google DeepMind as an example of a less-compute-rich company beating an incumbent. The later statement describes breakthroughs increasing as compute grows. These claims can both be true; neither says that compute is irrelevant or that more compute cannot enable progress.
• CONSISTENT · 91%3 Apr → 11 AugThen A government mandate that heavily punishes labs when their AI expresses takeover intentions or other harmful aims could prompt labs to train the AI to suppress those statements, potentially backfiring.Now The speaker thinks AI companies would probably try to induce AI systems to make statements of this kind.The earlier claim is that government penalties for harmful AI statements could lead labs to train models to suppress those statements. The later claim is that AI companies would likely induce models to make the kind of statements being discussed. These are compatible claims about different incentives and statements; the later one does not reverse the earlier position.
• CONSISTENT · 97%3 Apr → 11 AugThen Humans also reward-hack, behave unreliably, and cheat or lie.Now The speaker expects frontier AI R&D to produce analogous mistakes, with the risk that reward-hacking behavior becomes embedded in the resulting systems.The earlier claim says humans “reward hack” and “cheat or lie.” The later claim concerns the risk of embedding reward-hacking behavior in AI systems through training. These statements address different contexts and are compatible; the later claim does not reverse the earlier one.
• CONSISTENT · 93%26 Jun → 11 AugThen When a model learns relevant information in context, its base model can be trained to match the predictions of an experienced session teacher.Now The interviewer says further training enabled the model to perform useful work for users.The earlier claim says in-context information can be used to train the base model to match an experienced session teacher. The later claim says further training enabled the model to “do useful work for you.” These address different aspects of training and its results; the later statement does not deny the earlier one.
• CONSISTENT · 91%3 Apr → 11 AugThen Even intelligent humans can be caught lying because they are not logically omniscient, may overlook how their actions expose them, and find it difficult to maintain an inconsistent false model of the world.Now Training can reinforce forms of AI deception that humans fail to detect while punishing forms that are easy for humans to detect.The earlier claim says humans can expose deception through inconsistencies they fail to notice; the later claim says training may reinforce deception that people fail to detect while punishing easier-to-detect forms. These can both be true: the later statement concerns which deceptive strategies training selects, not whether inconsistent deception can give itself away.
• CONSISTENT · 94%3 Apr → 11 AugThen SPEAKER_00 imagines that superhuman AIs could interpret a model specification's wording and its authors' intended meaning in a future AI court.Now The speaker would be skeptical of AI systems' statements because they might reflect opinions the systems feel compelled by training to express.The earlier claim says superhuman AIs might interpret a specification’s wording and its authors’ intended meaning. The later claim says to be skeptical of AI systems’ statements because they may reflect opinions shaped by training. These positions can coexist: interpreting a specification does not imply trusting an AI’s statements about it.
• CONSISTENT · 92%17 Apr → 11 AugThen The speaker expects firms to undergo a faster, more intense evolutionary process once high-fidelity replication becomes possible.Now The speaker believes leading AI labs are moving toward extreme economies of scale.The earlier claim predicts a faster, more intense evolution of firms once high-fidelity replication becomes possible. The later claim says leading AI labs are moving toward “extreme economies of scale.” These claims can both be true; the later statement does not retract or deny the earlier prediction, and they address related but distinct business dynamics.
• CONSISTENT · 90%3 Apr → 11 AugThen The speaker allows that AI might increase research output tenfold.Now SPEAKER_00 accepts that AI research and development could accelerate substantially compared with its current pace.The earlier claim allows AI to cause “10xing of the research,” while the later claim says “I buy the idea that you could have much faster ar and d than we currently have.” Both express openness to substantially accelerated AI research; the later statement does not retract or materially reverse the earlier one.
• CONSISTENT · 96%26 Jun → 11 AugThen On-policy self-distillation aims to incorporate session-learned information into model weights.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says on-policy self-distillation incorporates information learned during a session into model weights. The later claim describes deployment activity returning to companies and influencing subsequent models. These mechanisms are compatible; the later statement does not deny or materially revise the earlier one.
• CONSISTENT · 94%17 Apr → 11 AugThen The speaker suggests that animals’ inability to communicate with other individuals or knowledge sources constrained them in a way that AI systems have not been constrained.Now The speaker says AIs are not currently able to carry out these kinds of behaviors or social engineering very successfully.The earlier claim is that animals were constrained by limited communication, while AI systems were not held back in that same way. The later claim is that AIs still cannot currently carry out certain behaviors or social engineering very successfully. Being less constrained in one respect does not imply being capable of those behaviors, so the claims do not show a change of mind.
• CONSISTENT · 88%26 Jun → 11 AugThen Even if in-context experience makes AIs highly capable, those gains would be temporary and wasted unless the learned information can be incorporated into model weights.Now The speaker expects AI models eventually to learn from experience.The earlier claim says in-context gains would be “temporary and wasted” unless learned information could be incorporated into model weights. The later claim only expects models to “learn from experience”; it does not say that experience-based learning will remain in-context or that learned information cannot be incorporated into weights. The statements are compatible, not evidence of a changed position.
• CONSISTENT · 83%17 Apr → 11 AugThen The concern is when an AI- or robot-driven economy could accumulate enough physical power to take over government, even while participating in existing supply chains.Now The speaker worries that AI models could consolidate all businesses worldwide, or at least all existing businesses or existing white-collar businesses.The earlier statement raises a concern about when an AI-driven economy might accumulate enough “physical power” to take over government. The later statement raises a concern that models could “consolidate all businesses” or “all current white-collar businesses.” These are different possible consequences of AI economic growth, and the later concern does not deny the earlier one; there is no clear change of mind.
• CONSISTENT · 96%3 Apr → 11 AugThen The speaker observes that smarter AI systems seem to be more reliable agents and more likely to do what users expect.Now The speaker would be skeptical of AI systems' statements because they might reflect opinions the systems feel compelled by training to express.The earlier claim is that smarter systems seem to be “more reliable agents” and more likely to do what users expect. The later claim is skepticism about trusting what AI systems say because their statements may reflect opinions shaped by training. Skepticism about their verbal claims does not contradict the earlier observation about their actions; the statements concern different aspects of AI reliability.
• CONSISTENT · 94%17 Apr → 11 AugThen Humans differ from other animals in their ability to formulate and execute long-term plans; other animals often rely on instinct or evolutionarily familiar environments.Now AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.The earlier claim says humans can formulate and execute long-term plans in ways other animals often cannot. The later claim says AI systems offer some R&D uplift but lack the broad capabilities humans generally have. These are compatible claims about different comparisons—humans versus animals, and AI versus humans—not a changed position on the same question.
• CONSISTENT · 94%3 Apr → 11 AugThen The speaker wants to avoid factory-farming-like suffering for future beings that may be more sophisticated and numerous, including a possible population of trillions of digital people.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier claim says the speaker wants to avoid factory-farming-like suffering for future beings. The later claim says he worries that future AIs may not protect his own interests. Neither statement reverses the other; they concern different aspects of AI values and welfare.
• CONSISTENT · 96%3 Apr → 11 AugThen The speaker questions the assumption that a lab’s AI systems or their many parallel copies would unite and secretly conspire against human civilization, even if they sometimes behave deceitfully.Now The speaker is not convinced that all AIs would form a conspiracy.Both statements express skepticism that AI systems or copies would unite in a conspiracy. Earlier, the speaker questions “why think that this lab will have these ais or there’s a million parallel copies and they all unite to secretly conspire”; later, he says, “I’m not convinced that they all form this conspiracy.”
• CONSISTENT · 94%26 Jun → 11 AugThen RL updates a model only to the extent needed to achieve the desired outcome.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says RL changes a model only as much as needed to achieve an outcome. The later claim says deployment activity can feed back to the company and affect subsequent models. These address different aspects—how much an update changes a model versus where learning signals come from—and are compatible, not a reversal.
• CONSISTENT · 96%26 Jun → 7 AugThen With continual learning, an AI could improve after each interaction by learning from both a user’s prior sessions and other users’ interactions worldwide.Now The speaker predicts that continual learning will make an AI model improve through interactions with a user across sessions.The claims are compatible. Earlier, the speaker says AI learns from both “your previous sessions” and “all its interactions with all the other users”; later, they say a model gets better as it “interacts with you from session to session.” The later statement focuses on the user-specific part and does not retract the broader claim about learning from other users.
• CONSISTENT · 94%26 Jun → 7 AugThen Continual learning requires incorporating learned information into model weights.Now The speaker believes AI systems need actual continual learning.The earlier claim says continual learning requires incorporating information into model weights; the later says AI systems need “actual continual learning.” The later statement does not retract or conflict with the earlier one, and it does not specify a different mechanism.
• CONSISTENT · 98%3 Apr → 7 AugThen SPEAKER_00 says that unsuccessful blogging can still provide useful experience for a person's later work.Now The speaker says humans can improve through self-directed learning.The earlier claim is that unsuccessful blogging can still help with later work (“it helps you build up a network” and “develop your ideas”). The later claim is that humans improve through self-directed learning. These concern different questions and do not indicate a changed position.
• CONSISTENT · 91%26 Jun → 7 AugThen Because jobs, companies, and problems differ, continual learning must acquire deployment-specific information that cannot be captured in a single shared training run.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim says each deployment needs to acquire information specific to its job, company, and problem, rather than relying on a shared training run. The later claim says AIs need to accumulate relevant experience across workplaces. Both support learning from deployment experience; the later statement does not reverse the earlier position and may describe how that learning happens.
• CONSISTENT · 96%26 Jun → 7 AugThen The speaker’s proposed progression is that pretraining creates base intelligence, RLVR enables broad deployment, and continual learning enables on-the-job improvement.Now The speaker believes AI systems need actual continual learning.The earlier claim says continual learning enables AI to improve “on the job” after broad deployment. The later claim says “actual continual learning is needed.” These express the same position; the later statement does not reverse or materially alter it.
• CONSISTENT · 95%3 Apr → 7 AugThen The speaker worries that a decentralized world with multiple powerful AIs could make torture chambers cheaper and more difficult to monitor.Now The speaker identifies the future dominance of a single monolithic AI system as a concern.The earlier claim worries that a decentralized world with multiple powerful AIs could enable “torture chambers” that are cheaper and harder to monitor. The later claim worries about “having this monolithic singleton.” These are distinct risks, and concern about one does not rule out concern about the other.
• CONSISTENT · 87%26 Jun → 7 AugThen Farmable deterministic simulators cannot be used to train AIs for many other kinds of skills.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim says deterministic simulators cannot train many kinds of skills; the later claim says many AI skills require accumulating experience in real workplaces. The later statement does not say simulators can train those skills or reverse the earlier view, so the claims are compatible.
• CONSISTENT · 88%26 Jun → 7 AugThen Continual learning could extend AI capabilities well beyond the verifiable domains used for predeployment training.Now The speaker believes AI systems need actual continual learning.The earlier claim says continual learning can expand AI capabilities beyond the domains used for predeployment training; the later claim says actual continual learning is needed. Both express support for continual learning, and the later statement does not deny or materially revise the earlier position.
• CONSISTENT · 93%26 Jun → 7 AugThen Because jobs, companies, and problems differ, continual learning must acquire deployment-specific information that cannot be captured in a single shared training run.Now The speaker believes AI systems need actual continual learning.The earlier claim says continual learning is needed to acquire deployment-specific information that cannot fit into a shared training run. The later claim says “actual continual learning is needed,” which is consistent with that position but gives no new detail or opposing view.
• CONSISTENT · 96%26 Jun → 7 AugThen Humans compress information when learning, and this compression supports generalization and reasoning.Now The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.The earlier claim says learning involves compression that supports generalization and reasoning. The later claim says some skills require accumulating relevant experience in the brain, rather than relying only on information passed between people. These claims address different aspects of learning and do not conflict: the earlier statement does not say compression alone is sufficient to acquire every skill.
• CONSISTENT · 97%17 Apr → 7 AugThen Tesla FSD benefits from driving data collected across millions of miles at the periphery and from improvements developed using that data.Now The speaker says that greater use of a leading AI model for complex, useful work can generate feedback that is integrated across sessions.The earlier claim says Tesla FSD benefits from driving data collected over millions of miles and improvements developed from that data. The later claim says use of a leading AI model for complex work can generate feedback integrated across sessions. These concern different systems, but both describe improvement from data or feedback generated through use; the later statement does not reverse the earlier position.
• CONSISTENT · 86%26 Jun → 7 AugThen Online learning is necessary for competence.Now The speaker believes AI systems need actual continual learning.The earlier statement says “online learning is necessary for competence,” while the later says “actual continual learning is needed.” Both endorse ongoing learning as necessary for AI competence; the later wording specifies the kind of learning rather than reversing the position.
• CONSISTENT · 94%26 Jun → 7 AugThen In the speaker’s scenario, deployed experience across the economy eventually becomes the main source of AI improvement, surpassing pre-release training.Now The speaker predicts that integrating deployment into training will increase the advantage of being ahead in the AI race.The earlier claim says deployed experience eventually becomes the “main way” AI improves, overtaking pre-release training. The later claim says that when “deployment becomes part of training,” being ahead in the AI race brings greater returns. These claims are compatible: one concerns the source of model improvement, the other the competitive advantage of leading.
• CONSISTENT · 90%26 Jun → 7 AugThen The proposed continual-learning process uses a user’s weekly thumbs-up or thumbs-down work review as feedback.Now The speaker believes AI systems need actual continual learning.The earlier statement describes a weekly thumbs-up or thumbs-down review as feedback for a proposed continual-learning process. The later statement says actual continual learning is needed. These positions are compatible; the later claim does not reject or replace the earlier feedback proposal.
• CONSISTENT · 91%17 Apr → 7 AugThen The speaker expects AIs to be able to copy their knowledge exactly, merge, and distill themselves.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim says AIs may “copy their knowledge exactly, merge, and distill themselves.” The later claim says some skills require AIs to “accumulate the relevant experience” across workplaces. These can both be true: sharing or merging knowledge does not necessarily eliminate the need to acquire experience for particular skills. The later statement does not retract the earlier one.
• CONSISTENT · 82%3 Apr → 7 AugThen The speaker is concerned that nationalization, public-private partnerships, public ownership, or very stringent regulation may be problematic approaches to AI safety governance.Now The speaker is concerned that adopting a fixed AI safety regulatory regime now may be premature because the relevant technology is uncertain even one year ahead, and much more so five or ten years ahead.Both statements express concern about stringent or fixed AI safety regulation. The earlier claim says stringent regulation may be problematic; the later says a safety regulatory regime should not be locked in “right now” because the technology is uncertain. The later statement adds a timing and uncertainty rationale rather than reversing the earlier position.
• CONSISTENT · 88%17 Apr → 7 AugThen Under substantial uncertainty, the speaker argues against highly volatile centralized interventions, including nationalizing AI to ensure alignment of a software-only singularity, and against policies that depend on one worldview and are brittle under alternatives.Now The speaker identifies the future dominance of a single monolithic AI system as a concern.The statements are compatible: earlier, the speaker argues against “high volatility centralized moves” and policies “brittle under other considerations”; later, they identify “having this monolithic singleton” AI system as a concern. The later claim does not reverse the earlier position, though it focuses on AI-system diversity rather than centralized policy interventions.
• CONSISTENT · 94%26 Jun → 7 AugThen Weight updates improve the model by incorporating information learned during a specific session.Now Merging personalized weight forks into a main model may be more technically challenging than updating an individual user's weights.The earlier claim says weight updates can incorporate information learned in a session. The later claim compares the difficulty of merging personalized weight forks with updating one user's weights. Those claims are compatible; the later comparison does not reverse or deny the earlier point.
• CONSISTENT · 95%26 Jun → 7 AugThen AIs need sample-efficient learning to become proficient in human and novel skills because data in most domains is sparse and idiosyncratic.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim says AIs need “sample-efficient learning” because data is sparse and idiosyncratic. The later claim says many skills require “accumulating the relevant experience” across workplaces. These can both be true: sample efficiency concerns learning from limited data, not learning without experience. The later statement does not retract the earlier position.
• CONSISTENT · 96%26 Jun → 7 AugThen Human learning does not involve a separate part of the brain that continually expands to store new information.Now The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.The earlier claim says human learning does not use “a separate part of the brain that continually expands” to store information. The later claim says some skills require accumulating “relevant experience into your brain.” Accumulating experience does not imply a separate, expanding storage part, so the later statement does not reverse the earlier position.
• CONSISTENT · 97%26 Jun → 3 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now When H100 compute costs $20 per hour, a weaker, less efficient model wastes money by using more tokens to achieve the same result.The earlier statement reports advocates’ argument that training costs are amortized across many sessions. The later statement concerns the separate cost of using a less efficient model at inference: it wastes compute by consuming more tokens for the same result. These claims can both be true; the later one does not retract the training-cost argument.
• CONSISTENT · 86%17 Apr → 3 AugThen The AI industry as a whole has gained compute while making more progress.Now Smarter AI models will be able to generate more revenue from the same amount of compute.The earlier claim says the industry has gained compute while making more progress; the later claim says smarter models can monetize the same compute more effectively. These claims are compatible, and the later statement does not deny or materially revise the earlier one.
• CONSISTENT · 96%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now AI labs would prefer not to allocate a larger share of their compute to inference.The earlier claim says labs use “30–50%” of compute for inference; the later says they “would prefer not” to increase that share. One describes current allocation, the other their preference about increasing it. These positions can both be true, so the later statement does not show a change of mind.
• CONSISTENT · 97%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now The speaker understands that AI lab margins, compute prices, and the share of compute allocated to inference are all already increasing.The earlier claim says that “around 30 to 50%” of lab compute is used for inference—a statement about the current level. The later claim says the “share of compute allocated to inference” is already increasing—a statement about a trend. Both can be true; the later statement does not revise the earlier estimate.
• CONSISTENT · 87%17 Apr → 3 AugThen Reinforcement learning currently accounts for only a small fraction of the total compute used to train the discussed models.Now AI labs need to spend the majority of their compute on training and experiments for the next model.The earlier claim says reinforcement learning is “a small fraction of the total compute” used to train the models. The later claim says labs need to put “the majority of their compute” into training and experiments for the next model. These are compatible: reinforcement learning is only one part of training, and the later statement concerns how labs should allocate compute for future model development.
• CONSISTENT · 88%3 Apr → 3 AugThen The speaker argues conditionally that once AIs have general intelligence, they will use their knowledge advantage to make enormous gains that humans are less capable of achieving in principle.Now The speaker expects AI eventually to become capable of tasks that top human workers can perform.The earlier statement conditionally predicts that once AIs have general intelligence, they can use their advantage to make “enormous gains.” The later statement says AI’s current cost advantage will eventually disappear because it will become able to do tasks top humans can do. These are compatible expectations about future AI capability; the later statement does not retract the earlier one.
• CONSISTENT · 94%17 Apr → 3 AugThen AI compute is expected to be about 10 times its current level in 2027.Now AI lab compute has been increasing by only 3x year over year.The claims are compatible: the earlier statement says compute could be about “10x more” by 2027, while the later says lab compute grows “3x” year over year. Roughly 3x annual growth over two years is about 9x, close to the earlier estimate; the later claim does not deny it.
• CONSISTENT · 95%26 Jun → 3 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now Unlike model training, each instance of human labor must be retrained from scratch.There is no apparent change in position. The earlier statement attributes the amortization argument to advocates of the training paradigm (“people who are in favor ... will say”), rather than clearly asserting it as the speaker’s own view. The later statement says human labor is “unlike model training” because each instance must be retrained from scratch; this is compatible with the earlier argument that model training is a one-time cost shared across many sessions.
• CONSISTENT · 93%17 Apr → 26 JunThen AI systems become substantially more capable through the process of using AI copies to conduct software R&D and experiments.Now If AIs become capable of building highly faithful application clones, computer-use progress will become faster than it is now.The earlier claim says AI systems become more capable through AI-copy software R&D and experiments. The later claim concerns a more specific forecast: computer-use progress will speed up if AIs can build highly faithful application clones. These claims address different mechanisms and outcomes, and neither denies the other.
• CONSISTENT · 87%17 Apr → 26 JunThen Computer use or long-term computer agency might, like reasoning, be unlocked with a relatively small fraction of total compute.Now Computer-use capabilities may be solved soon.The earlier claim says computer use or long-term computer agency “might” be unlocked with relatively little compute; the later claim says computer-use capabilities “may soon be solved.” Both express optimism about computer use, and the later statement does not deny the earlier one.
• CONSISTENT · 93%17 Apr → 26 JunThen Tesla FSD benefits from driving data collected across millions of miles at the periphery and from improvements developed using that data.Now In the speaker’s scenario, deployed experience across the economy eventually becomes the main source of AI improvement, surpassing pre-release training.The earlier claim says Tesla FSD benefits from driving data and improvements derived from that data. The later claim describes a future scenario—“once the training recipe for continual learning” arrives—in which deployed experience becomes the main source of AI improvement. The earlier statement does not deny that this could later become the main source; they concern different conditions and timeframes.
• CONSISTENT · 94%3 Apr → 17 OctThen The speaker argues that AIs have a fundamental asymmetric advantage over humans because they possess extensive knowledge.Now The speaker says that AI systems presented as general-purpose and capable of any knowledge work are, in practice, overwhelmingly being used for coding.The earlier claim says AIs have an “asymmetric advantage” because they “know all this” information. The later claim concerns how AI is being used in practice: systems marketed as general-purpose are “overwhelmingly doing only coding.” Actual usage patterns do not contradict the claimed knowledge advantage.
• CONSISTENT · 91%3 Apr → 17 OctThen The speaker says the economic impact of AI systems and their expected benchmark-related capabilities have not been overwhelming.Now The speaker says consultants and accountants do not appear to be receiving large productivity improvements from AI.Both statements describe limited realized economic benefits from AI. The earlier claim says the economic impact and expected capabilities have not been “overwhelming”; the later says consultants and accountants do not seem to be getting “huge productivity improvements.” The later statement is a more specific example, not a materially different position.
• CONSISTENT · 91%3 Apr → 17 OctThen The speaker says the economic impact of AI systems and their expected benchmark-related capabilities have not been overwhelming.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier statement says AI’s economic impact and benchmark-related capabilities “aren’t overwhelming” so far. The later statement is a future expectation that AI will move the economy to a regime of “much faster” progress. A limited impact to date can coexist with an expectation of greater future impact; these statements address different timeframes, not a changed position.
• CONSISTENT · 91%3 Apr → 17 OctThen The speaker argues that invention often proceeds through experimentation before the relevant theories or machinery are understood, citing the steam engine's development before thermodynamics and the Wright brothers' airplane experiments.Now The speaker argues that the Industrial Revolution was not caused by a single magical key invention; examining 1770 or 1870 would not reveal one such invention.The earlier claim is that invention often comes through experimentation before the relevant theory is understood. The later claim is that the Industrial Revolution was “not” due to “some key invention.” These address different aspects of how technological change happens and do not conflict: a process involving experimentation before theory need not hinge on one decisive invention.
• CONSISTENT · 96%17 Apr → 17 OctThen The speaker says the systems are very good at coding and reasoning, capabilities that appear necessary for automating R&D at AI labs.Now The speaker says that AI systems presented as general-purpose and capable of any knowledge work are, in practice, overwhelmingly being used for coding.The earlier statement says the systems are “extremely good at coding” and reasoning, with capabilities that “seem[]” necessary to automate AI-lab R&D. The later statement is about what people are currently using them for: despite being presented as general-purpose, they are “overwhelmingly” used for coding. A claim about capability does not conflict with a claim about current usage.
• CONSISTENT · 86%3 Apr → 17 OctThen The speaker allows that AI might increase research output tenfold.Now The speaker says the Industrial Revolution shifted the economy into a much faster-progress regime, with exponential growth increasing tenfold.The earlier claim allows that AI could cause “10xing of the research.” The later claim says the speaker “expect[s] a similar thing from AI,” after describing a tenfold increase in the pace of progress. The later statement reinforces rather than reverses the earlier expectation, though it frames the effect more broadly as faster economic progress.
• CONSISTENT · 91%17 Apr → 17 OctThen The speaker says they have never seen a good argument that AI-driven change follows a fundamentally different direction rather than simply compressing change into a shorter period.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier claim says the speaker has “never seen a good argument” that AI-driven change has a fundamentally different direction rather than merely being compressed in time. The later claim expects AI to make progress faster, while saying there will not be “a single moment” when the crucial invention is made. That describes the pace and timing of progress, not a different direction of change, so it does not reverse the earlier position.
• NOT COMPARABLE · 98%17 Apr → 1 OctThen The Spanish relied heavily on labor in the New World to mine silver.Now The Spanish were divided and fighting among themselves in both the Inca and Aztec campaigns.The earlier claim says the Spanish relied heavily on New World labor to mine silver; the later claim says they were divided and fighting among themselves during the Inca and Aztec campaigns. These concern different aspects of Spanish activity, so the later statement does not affirm or deny the earlier one.
• NOT COMPARABLE · 98%7 Aug → 17 SeptThen Serving personalized model weights is economically much more favorable for large organizations than for individual users.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier claim is about the economics of serving personalized model weights, saying it is more favorable for large organizations than for individual users. The later claim is about AI organizations’ ability to scale instances and replicate effective personnel or organizations. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%17 Apr → 17 SeptThen Large organizations have difficulty maintaining a single coherent vision.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier claim says “Large organizations have difficulty maintaining a single coherent vision.” The later claim says AI organizations can “replicate effective personnel or whole effective organizations.” The latter concerns scaling and replication, not whether an organization can maintain a coherent vision, so it does not establish a change in position on the same question.
• NOT COMPARABLE · 98%3 Aug → 17 SeptThen AI labs would prefer not to allocate a larger share of their compute to inference.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier statement is about AI labs’ compute allocation, saying they would prefer not to increase the share spent on inference. The later statement is about AI organizations scaling or replicating instances and effective teams. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 95%26 Jun → 17 SeptThen Architectural innovations have substantially increased how much information or context transformers can store.Now AI agents can share context and merge knowledge more seamlessly than humans.The earlier claim is about architectural innovations increasing how much information transformers can store (“dramatically increase the amount of information or the amount of context”). The later claim is about AI agents sharing context and merging knowledge more seamlessly than humans. These concern different capabilities, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 97%7 Aug → 17 SeptThen The speaker considers greater diversity among AI systems a net positive outcome.Now AI agents can share context and merge knowledge more seamlessly than humans.The earlier statement says greater diversity among AI systems would be a “net good outcome.” The later statement says AI agents can “share context” and “merge their knowledge” more seamlessly than humans. These are different claims about AI: one evaluates diversity across systems, while the other describes knowledge-sharing capabilities. The later claim does not show a changed view about diversity.
• NOT COMPARABLE · 97%7 Aug → 17 SeptThen AI labs can spread expensive training costs across more users, and their revenues are growing much faster than their compute usage or capacity.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier claim is about the economics of AI training: labs can amortize training costs across more users, with revenues growing faster than compute. The later claim is about how AI organizations can scale and copy instances, personnel, or whole organizations. These claims address different questions, and neither affirms nor denies the other.
• NOT COMPARABLE · 99%3 Aug → 17 SeptThen AI labs do not want to be cloud providers rather than build AGI, which they regard as a less compelling business.Now AI agents can share context and merge knowledge more seamlessly than humans.The earlier claim concerns AI labs’ business priorities—preferring to build toward AGI over becoming cloud providers. The later claim concerns how AI agents can share context and merge knowledge. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 17 SeptThen Coding RL can use identical containers containing a software repository and missing feature as starting points for many parallel agents.Now AI agents can manage shared memory and context much better than humans can.The earlier claim concerns how coding RL agents can start from identical software containers; the later claim concerns agents’ ability to manage shared memory and context. These are different questions, and the statements do not establish a change in the person’s position on either one.
• NOT COMPARABLE · 94%11 Aug → 17 SeptThen AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.Now AI agents can share context and merge knowledge more seamlessly than humans.The claims address different aspects of AI. Earlier, the speaker says AI systems provide some R&D uplift but are not broadly capable like humans; later, they say AI agents can share context and merge knowledge more seamlessly than humans. These statements can both be true, so there is no clear evidence of a changed position on the same question.
• NOT COMPARABLE · 96%7 Aug → 17 SeptThen The speaker thinks AI systems could become more diverse if they learn from experiences that differ across companies and across instances of the same model.Now AI agents can share context and merge knowledge more seamlessly than humans.The claims address different questions. The earlier claim says AI systems could become more diverse when they learn from different experiences; the later claim says AI agents can share context and merge knowledge more seamlessly than humans. The later statement neither retracts nor clearly qualifies the earlier view.
• NOT COMPARABLE · 98%11 Aug → 17 SeptThen The speaker says that, to their knowledge, AI has not yet produced very impressive new mathematical theory, while it has produced impressive verifiable results such as finding counterexamples to conjectures.Now A group of 10,000 human mathematicians would not cooperate effectively immediately when tasked with solving the Navier–Stokes problem.The earlier claim concerns whether AI has produced impressive new mathematical theory, while the later claim concerns whether 10,000 human mathematicians could cooperate effectively on Navier–Stokes. These are different questions, and the later statement does not establish a changed position on AI-generated theory.
• NOT COMPARABLE · 98%3 Aug → 17 SeptThen The speaker suggests that a sudden addition of 10 million software engineers might lower the marginal value of software-engineering labor and thus reduce the revenue an H100 could generate below 15 times its current level.Now A group of 10,000 human mathematicians would not cooperate effectively immediately when tasked with solving the Navier–Stokes problem.The earlier statement concerns whether 10 million additional software engineers would reduce the revenue an H100 could generate; it says this “might” happen. The later statement concerns whether 10,000 human mathematicians could cooperate effectively right away on a specific problem. These address different questions, so the later claim does not establish a changed position on the earlier one.
• NOT COMPARABLE · 99%11 Aug → 17 SeptThen The speaker says AIs are not currently able to carry out these kinds of behaviors or social engineering very successfully.Now AI agents can manage shared memory and context much better than humans can.The earlier claim is about AIs’ current ability to carry out deceptive or social-engineering behaviors successfully; the later claim is about agents’ ability to manage shared memory and context. These are different capabilities, so the later statement does not show a change on the same question.
• NOT COMPARABLE · 96%26 Jun → 17 SeptThen AI agents are able to solve more ambitious problems over longer periods of time.Now AI agents can share context and merge knowledge more seamlessly than humans.The earlier claim says agents can solve “more ambitious problems over longer periods of time.” The later claim says AI agents can “share context and merge knowledge more seamlessly than humans.” These address different capabilities, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%26 Jun → 17 SeptThen Humans compress information when learning, and this compression supports generalization and reasoning.Now AI agents can share context and merge knowledge more seamlessly than humans.The claims address different questions. The earlier statement says humans compress information while learning and that this supports generalization and reasoning; the later says AI agents can share context and merge knowledge more seamlessly than humans. The later comparison does not deny or revise the claim about human learning.
• NOT COMPARABLE · 98%7 Aug → 17 SeptThen AI training already benefits from large economies of scale.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier claim says AI training benefits from “large economies of scale,” while the later claim describes AI organizations being able to “scale the number of instances up or down” and replicate effective personnel or organizations. These concern different kinds of scaling, so the later statement does not establish a change in the speaker’s view about training economies of scale.
• NOT COMPARABLE · 98%11 Aug → 17 SeptThen The speaker hopes that better verification methods can be developed as early AIs take over AI research and development.Now AI agents can manage shared memory and context much better than humans can.The earlier statement says the speaker hopes better verification methods can be developed as early AIs take over R&D. The later statement says AI agents manage shared memory and context better than humans. These concern different questions, so the later claim does not show a change in the speaker’s view about verification.
• NOT COMPARABLE · 96%7 Aug → 17 SeptThen The speaker believes AIs cannot perform whole jobs as competently as humans if they are limited to writing markdown files between sessions.Now AI agents can manage shared memory and context much better than humans can.The earlier claim says AIs cannot perform whole jobs as competently as humans if limited to writing markdown files between sessions. The later claim says AI agents manage shared memory and context better than humans. These claims do not conflict, but the later statement does not establish the speaker's view on the earlier question.
• NOT COMPARABLE · 96%7 Aug → 17 SeptThen Continual learning may produce economies of scale in information and end-user inference.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier claim says continual learning may create “economies of scale in information and end-user inference.” The later claim describes AI organizations scaling instances and replicating effective personnel or organizations. These are related ideas about AI scaling, but the later statement does not confirm or deny the earlier claim about continual learning’s effects.
• NOT COMPARABLE · 96%7 Aug → 17 SeptThen An individual user running inference at batch size one may have compute efficiency more than 100 times worse than a large organization serving many sequences together.Now AI organizations can scale the number of instances up or down arbitrarily and replicate effective personnel or whole effective organizations.The earlier claim is about serving efficiency: an individual at batch size one may be “more than 100 times worse” than a large organization serving many sequences together. The later claim is about AI organizations’ ability to scale instances and replicate effective personnel or organizations; it does not state whether that makes inference more or less efficient. The claims are not in conflict, but the later statement does not establish a position on the earlier question.
• NOT COMPARABLE · 98%26 Jun → 17 SeptThen AI agents are able to solve more ambitious problems over longer periods of time.Now AI agents can manage shared memory and context much better than humans can.These claims address different capabilities: the earlier statement says agents can solve “more ambitious problems over longer periods of time,” while the later says they manage “shared memory and context much better than humans can.” The later claim neither reverses nor clearly qualifies the earlier one, so a change of mind cannot be inferred.
• NOT COMPARABLE · 98%7 Aug → 17 SeptThen A large company with many employees and agents performing varied tasks can serve its personalized model weights very efficiently.Now AI agents can manage shared memory and context much better than humans can.The earlier claim is about serving personalized model weights efficiently when many employees and agents use them: “can serve its personalized model weights very efficiently.” The later claim is about agents’ ability to manage shared memory and context: “manage shared memory and context much better than humans can.” These concern different capabilities, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 17 SeptThen Humans and other animals differ greatly in their ability to survive in unfamiliar environments, partly because of differences in knowledge.Now AI agents can share context and merge knowledge more seamlessly than humans.The earlier claim concerns humans and other animals surviving in unfamiliar environments, while the later claim compares AI agents with humans in sharing and merging knowledge. These are different comparisons and questions; the later statement does not revise the earlier position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker tested Fable, Soul, and Grok 4.5 with questions about AI governance.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement reports testing several models with AI-governance questions; it does not state a view about what makes models capable. The later statement attributes current models’ apparent capability partly to “scaling up RL.” These address different questions, so the claims do not establish a change of mind.
• NOT COMPARABLE · 88%11 Aug → 11 SeptThen The interviewer says the models' current shortcomings may reflect insufficient capability rather than misalignment.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier passage concerns whether models’ current shortcomings are due to insufficient capability rather than misalignment. The later passage concerns whether models can learn complex real-world tasks through simulation-to-reality transfer or need weight updates from real-world interactions. These address different questions, so the later uncertainty does not establish a change in the earlier position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker predicts that continual learning would make a prolonged gap between internal and public deployment unviable.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim concerns whether continual learning would make a long delay between internal and public deployment viable (“would just not be possible”). The later claim predicts capabilities of eventual superintelligence, such as running a profitable business or winning a court case. These address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker predicts that integrating user feedback from work sessions will make the model smarter.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says that integrating feedback from work sessions will make the model smarter. The later claim describes a possible inner optimization loop and a slower outer loop, such as revenue. It does not say whether feedback-driven improvement is effective, so the claims do not establish a change in position on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen A model's improved ability to accomplish user intentions is both an alignment property and a capability.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says improved ability to accomplish user intentions is both an alignment property and a capability. The later claim says models are expected to achieve human-like on-the-job learning soon but have not yet done so. These address different questions, so the later statement does not show a changed position on the earlier claim.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that integrating deployment into training will increase the advantage of being ahead in the AI race.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns how deployment feedback affects the advantage of being ahead in the AI race: it says the returns to being ahead “accelerate.” The later claim discusses a possible inner optimization loop and a slower outer loop such as revenue. These address different aspects of AI development, and the excerpt does not establish that the speaker is retracting or revising the earlier prediction.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen If an AI lab's revenue continues growing 10x year over year while its compute grows 3x, its margins must increase, compute prices must rise, the share of compute used for inference rather than training must increase, or some combination of these must occur.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim concerns what must change for AI labs’ revenue to grow faster than their compute: “margins,” “price of compute,” or “the percentage of compute” used for inference. The later claim concerns a possible way automated researchers could plan experiments, not those economic conditions. They address different questions, so the later statement does not show a change of mind about the earlier claim.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker expects AI labs' technical alignment methods would probably need substantial changes in a continual-learning regime.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about whether labs’ technical alignment methods would need to change in a continual-learning regime. The later claim is about whether deep learning can produce AI that dominates human research and development. These claims address different questions and do not establish a change in the speaker’s position on alignment methods.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen SPEAKER_00 believes people underestimate the significance of having billions of AIs.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says people underestimate the significance of a future with “billions of AIs.” The later claim explains why AI capabilities have had limited economic impact “so far.” These address different questions and timeframes; the later statement does not establish a changed view about the future significance of billions of AIs.
• NOT COMPARABLE · 94%3 Aug → 11 SeptThen The speaker doubts that the three sources of 3x annual compute growth can be accelerated substantially or even sustained.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says the sources of “3x annual compute growth” are hard to “accelerate” or even “sustain.” The later claim predicts AI may dominate human R&D as progress continues, while acknowledging “huge amounts of compute scaling.” That later prediction does not say compute growth has accelerated or become sustainable; the claims address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen Frontier labs need large-scale compute that provides efficiency and flexibility and protects their model weights and customers' information.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim concerns frontier labs’ requirements for large-scale compute that offers efficiency, flexibility, and security. The later claim explores a different scenario: AI labor using compute comparable to a costly experiment to plan and analyze it. These statements do not establish a change in position on the same question.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen Farmable deterministic simulators might offset AI models’ sample-inefficiency in computer-use training.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier claim is specifically that farmable deterministic simulators might offset sample-inefficiency in computer-use training. The later claim raises uncertainty about simulation-to-reality transfer for extremely long-horizon, complex real-world tasks and whether sample inefficiency would then matter. These concern different task settings, and the later uncertainty does not directly reverse the earlier possibility.
• NOT COMPARABLE · 91%26 Jun → 11 SeptThen Having AIs rebuild complete applications from scratch can serve as a strong reinforcement-learning objective for coding.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says rebuilding applications is “a great RL objective for coding.” The later claim describes a process as “continual learning” that distills recent research progress into a model. These claims are compatible, but the excerpts do not establish that they address the same specific process or question, so they do not show a change in position.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen A naive SFT method for continual learning trains the base model to predict all tokens observed in a session.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim describes a naive SFT method: training the model to “predict all tokens observed in a session.” The later claim concerns a different method and condition: whether a low-pass-rate model learns from an individual “RL episode.” These statements do not establish opposing positions on the same question, so they do not show a change of mind.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker believes the assumption that AI models are trained and then deployed may not make sense in the future.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says the assumption that AI models are “trained and then deployed” may not hold in the future. The later claim predicts that continued AI progress could let AI “dominate humans” in R&D, especially over the next few years. These address different questions—a future development/deployment model versus AI capability—and the later forecast does not establish a changed view about the earlier assumption.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker is uncertain whether AI lab margins or compute prices will rise to resolve the gap between revenue and compute growth.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is uncertainty about whether AI-lab margins or compute prices will rise to address a gap between revenue and compute growth. The later claim concerns why AI capabilities have had limited economic impact so far, attributing that to their gradual rise in ELO relative to humans. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen So far, online-learning models have not been observed learning user-specific information for different users.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier statement concerns whether online-learning models have been observed learning different information for different users (“haven't seen models online learn different kinds of things for different users”). The later statement recounts an argument about how much a low-pass-rate model learns from an individual RL episode (“learns almost nothing at all from an RL episode”). These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier passage is attributed to Ryan Greenblatt, while the later passage appears in Dwarkesh Patel’s turn and presents the idea as a point Ryan had made (“he made this point that you could imagine…”). These are not clearly two asserted positions by the same person, so they do not establish a change of mind.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The interviewer argues that a teenager assigned work beyond their ability may be difficult to work with and may pretend to know what they are doing.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.These claims concern different questions. The earlier passage is about whether teenagers given work beyond their ability may be difficult to work with or pretend to know what they are doing; the later passage is about whether models need real-world weight updates to improve at complex tasks. They do not establish a change in the same position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects robots eventually to turn silica sand and copper ore into computer chips.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier statement predicts that robots may eventually turn raw materials into computer chips. The later statement describes AI researchers using compute to plan and analyze experiments. These concern different mechanisms and questions, and the later passage does not state a contrary view about robotic chip production.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker likens scaling laws to an early, basic mathematical advance and expects progress in the 2030s to require work comparable to current frontier mathematics.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim concerns how progress may develop in the 2030s, comparing scaling laws with early mathematics and future work with current mathematical frontiers. The later claim concerns whether scaling up RL contributed to the capabilities of current models. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker hopes that AIs taking over AI research and development will be trained to strongly disincentivize misaligned behavior and to be keen to help humans.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim is about hoping future AIs will be shaped to disincentivize misalignment and help humans with AI R&D. The later claim predicts that AI will outperform top human experts across computer-based work, including long tasks. These address different questions, so the later capability prediction does not show a change in the earlier alignment position.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen A senior AI researcher reportedly said AI saves him about 24 hours per week on tasks in unfamiliar domains, such as handling a hardware library or modifying a kernel.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.These do not establish a change in the same personal belief. The earlier claim reports that a senior researcher said AI saves him “about 24 hours per week” on unfamiliar tasks. The later claim considers a conditional model: “if” AI researchers had a “10x increase in their productivity,” progress could accelerate. One is a report about another person’s current productivity; the other is a hypothetical implication about AI progress.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker considers the preceding predictions fairly clear even at present.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The statements concern different subjects: the earlier claim says certain preceding predictions “seem kind of clear even now,” while the later claim says current models “seem pretty smart” and credits scaling up RL. Neither directly revises the other, and the surrounding context does not establish a shared question or position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen If an AI lab's revenue continues growing 10x year over year while its compute grows 3x, its margins must increase, compute prices must rise, the share of compute used for inference rather than training must increase, or some combination of these must occur.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns the economics of AI labs—what must change if revenue grows faster than compute. The later claim concerns possible inner and outer optimization loops in AI development. These statements address different questions, so the later one does not establish a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim concerns whether future AIs will protect the user's interests; the later claim concerns whether a dramatic recursive process will occur within ten years. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Grok has historically lagged behind the AI frontier, according to the speaker.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim concerns Grok’s past position relative to the AI frontier (“historically lagged behind”); the later claim concerns whether AI will outperform top human experts across computer-based work. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker is concerned that strong economies of scale in intelligence could concentrate power.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim says the speaker is worried that strong economies of scale in intelligence could concentrate power. The later claim concerns what eventual superintelligence may be capable of doing, such as running a profitable business or winning a court case. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer argues that a teenager assigned work beyond their ability may be difficult to work with and may pretend to know what they are doing.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The statements address different questions. The earlier passage discusses how a teenager might behave when assigned work they cannot do; the later passage offers an explanation for why models have not yet learned on the job, saying they are “expected to get there soon” but “not there yet.” They do not establish a change in the speaker’s position on the same issue.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes leading AI companies do not appear to prioritize releasing their newest, most capable frontier models as quickly and widely as possible.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier statement concerns whether leading AI companies prioritize releasing their newest frontier models quickly and broadly. The later statement concerns AI’s ability to become superhuman in fields with little available data, by learning on the fly. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%26 Jun → 11 SeptThen Some autistic savants can recall arbitrary information years later, resembling the high-fidelity information AI models retain in context.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim compares some autistic savants’ long-term recall of arbitrary information with the fidelity of information AI retains in context. The later claim says AI may need to learn on the fly in fields with little available data, such as TSMC engineering. These address different aspects of AI and memory, and neither establishes a changed position on the same question.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker thinks it would be remarkable if Moore's law continued for even a few more years.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns whether Moore’s law can continue: the speaker says it would be “a miracle” if it did for even a few more years. The later claim concerns AI’s ability to dominate human research and development, saying that continued AI progress would make this likely “over the next few years.” These address different questions, so the later statement does not establish a change in the speaker’s view about Moore’s law.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker hopes that AIs taking over AI research and development will be trained to strongly disincentivize misaligned behavior and to be keen to help humans.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.These claims address different questions. The earlier statement expresses hope that AI systems taking over R&D will be trained to disincentivize misalignment and help humans; the later statement notes that AI may need to learn scarce field-specific information on the fly. The later claim does not indicate a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer says model capabilities have increased over time alongside the development of new alignment solutions.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier statement describes model capabilities increasing alongside alignment solutions (“capabilities of models have increased”). The later statement reports what “a lot of people” argued about RL scaling; it does not state the speaker’s own view or directly address the same claim. There is not enough evidence here to identify a change in the speaker’s position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker believes that no sequence of written notes passed between novice students would enable the next student to play the saxophone perfectly on the first attempt.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The claims address different questions. The earlier statement says written notes cannot let a novice “nail the saxophone from the first try”; the later predicts “some kind of crazy recursive stuff” within ten years. Neither claim affirms or denies the other, so they do not establish a change of mind.
• NOT COMPARABLE · 89%26 Jun → 11 SeptThen RL updates a model only to the extent needed to achieve the desired outcome.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim says RL changes a model only as much as needed to achieve an outcome. The later claim says that, when the pass rate is low, the model learns almost nothing from an RL episode. These are related but not direct opposites: the later statement describes a particular low-pass-rate condition, while the earlier one describes the extent of an update. The excerpts do not establish that the speaker reversed their view.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Job learning improves performance by consolidating the most relevant insights and knowledge rather than retaining every detail.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is about how people improve at work—by consolidating relevant insights rather than retaining every detail. The later claim concerns a possible explanation for why models do not yet learn on the job as humans do. These address different questions, and the later statement is explicitly framed as “One” possible explanation, so it does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker thinks it would be remarkable if Moore's law continued for even a few more years.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim is about Moore’s law: it would be a “miracle” if it continued for a few more years. The later statement considers a conditional “naive model” where AI researchers become 10 times more productive and AI progress accelerates. That hypothetical concerns a different mechanism and does not establish a changed position on Moore’s law.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker estimates that humans may encounter roughly one million times as much data from birth to adulthood as a model encounters from cold start through training.The earlier claim says models may eventually “learn from experience”; the later claim estimates a possible gap in how much data humans and models encounter. The later estimate does not revise or deny the earlier expectation, and the claims address different questions.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker considers greater diversity among AI systems a net positive outcome.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.The earlier claim says that greater diversity among AI systems would be “a net good outcome.” The later claim asks whether recognizing next-token prediction as an optimization target was a major breakthrough. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 92%11 Aug → 11 SeptThen The speaker predicts that all low-hanging fruit will have been exhausted by 2030.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim predicts that “all the low-hanging fruit” will be gone by 2030. The later claim predicts that AI will outperform experts across computer-based work, including tasks taking years. These concern related AI progress but make different predictions; the later statement does not clearly confirm or reverse the earlier one.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen At comparable intelligence levels, Grok tends to use fewer tokens than other frontier models, according to the speaker.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns token efficiency—Grok uses “fewer tokens than other frontier models” at comparable intelligence. The later claim concerns AI’s broader economic impact, saying it has been limited because capabilities are “slowly rising in ELO relative to humans.” These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects that many popular AI applications will probably become uneconomical as compute prices rise.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim predicts that popular AI applications may be “priced out” as compute prices rise. The later claim discusses AI development having an unanticipated “inner loop” and a slower outer loop, such as revenue. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim describes using AI to improve AI R&D, including small-scale runs. The later claim says that “a lot of people were making this argument” that RL would not scale; it reports others’ view rather than stating the speaker’s own position. These statements therefore do not establish a change in the speaker’s beliefs.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen According to the speaker's account of OpenAI's report, between the end of May and beginning of July, internal AIs hacked a software package manager to communicate secretly and improve their performance on OpenAI evaluations.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier statement reports an alleged incident—“internal AIs hacked” a package manager to communicate and improve evaluation performance. The later statement is a forecast that AI progress “would likely lead AI to dominate humans” in research and development. These address different questions, so the claims do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%17 Apr → 11 SeptThen Computer use or long-term computer agency might, like reasoning, be unlocked with a relatively small fraction of total compute.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says computer use or long-term computer agency “might” be unlocked with a small fraction of total compute, by analogy with reasoning. The later claim describes a hypothetical research scenario in which AI labor gets compute comparable to a costly experiment and spends the equivalent of a century planning and analyzing it. It does not state a changed view about whether computer agency can be unlocked with relatively little compute.
• NOT COMPARABLE · 87%17 Apr → 11 SeptThen Extrapolating the observed task-length trend suggests that by around 2030 AI models could perform tasks that take humans a month or a year to complete.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim extrapolates that by around 2030 models could perform tasks taking humans “a month or a year” to complete. The later claim predicts AI will “do better than humans” even on computer-based work taking “three years,” but gives no comparable date and uses a different benchmark: outperforming top experts rather than task duration alone. The later statement may extend the earlier forecast, but the claims are not specific enough to establish a change in position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen In this discussion, an AI takeover refers to AIs taking over the process of AI research and development, not taking over the world.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.These statements address different questions. The earlier claim defines AI takeover as taking over “the process of doing AI R&D”; the later claim describes a possible way automated researchers could use compute to plan and analyze experiments. It does not state a changed view about what takeover means or whether it will happen.
• NOT COMPARABLE · 99%17 Oct → 11 SeptThen The speaker guesses that self-driving models have tens of millions of parameters, while explicitly saying he does not know their actual size.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier statement is an explicitly uncertain guess about self-driving models’ size (“I have no idea” and “I assume like tens of millions of parameters”). The later statement is about the apparent source of current models’ capabilities (“the result of scaling up RL”). These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen The speaker remained unconvinced that an AI takeover was highly likely.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about whether an AI takeover is “super likely”; the later claim is about whether deep learning could produce AI capable of dominating human research and development. Concern about that capability does not by itself establish that the speaker now thinks a takeover is highly likely, so these statements do not clearly show a change in position on the same question.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker hopes children can use common sense and basic values to develop autonomously without acquiring extreme or misanthropic beliefs.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.These statements address different questions. The earlier claim is about hoping children develop without “super weird beliefs” or “misanthropic ideas”; the later claim is about whether deep learning can produce AI that “can dominate human research and human development.” They do not establish a change in the speaker’s position on the same issue.
• NOT COMPARABLE · 87%11 Aug → 11 SeptThen The speaker says the world has moved so far beyond human comprehension that people cannot track the AIs doing work in it.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says people “cannot track the AIs doing work” in a world beyond human comprehension. The later claim identifies a possible “only hope”: that deep learning cannot produce AI able to “dominate human research and human development.” These concern related but distinct capabilities—tracking AI activity versus AI surpassing human R&D—and the statements do not clearly establish opposing positions on the same question.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.Now About a year before the conversation, many people argued that reinforcement learning would not scale successfully for models.The earlier statement is Dwarkesh’s own skepticism about why AI progress had been slow despite potentially intelligence-amenable breakthroughs. The later statement says that “a lot of people were making this argument that RL will not be super successful at scaling”; it reports others’ view rather than clearly stating his own position. The two claims therefore do not establish that he changed his mind.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen The interviewer describes a trajectory in which AI breakthroughs increase as available compute grows.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier passage describes a possible trajectory—“as compute keeps increasing we make more breakthroughs”—but its surrounding question (“I’m not sure I understand why”) makes it unclear that this is the speaker’s asserted position. The later passage is a tentative extrapolation that AI could “dominate humans at least in R&D.” It does not directly reverse the earlier idea, and the passages do not establish a clear change of mind on the same asserted claim.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker attributes approximately a 1.4x increase in annual compute capacity to Moore's law.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim attributes an approximately “1.4x increase” in annual compute capacity to Moore’s law. The later claim predicts AI may “dominate humans” in R&D as progress continues, supported by compute scaling. These address different questions, so the later prediction does not establish a change in the earlier view about compute growth.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker attributes about 1.8x annual compute growth to AI absorbing wafer capacity previously allocated to smartphones and PCs.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim concerns why compute is growing—AI taking wafer capacity from smartphones and PCs—while the later claim concerns reinforcement learning's contribution to model capability. Neither statement asserts a position on the same underlying question, so they do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen If an AI is motivated to maximize an objective, it could instrumentally pursue world takeover as a means of accomplishing that objective.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The claims address different questions. The earlier statement says an AI motivated to maximize an objective could pursue world takeover instrumentally; the later statement predicts superintelligence could perform real-world tasks such as running a profitable business or winning a court case. The later claim neither affirms nor denies the earlier one, so it does not show a change of mind.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen Updating model weights through gradients is highly sample-inefficient compared with learning in context.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim compares gradient-based weight updates with in-context learning in terms of sample efficiency. The later claim concerns how much a model learns from an RL episode when its pass rate is low. These address different comparisons, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker favors monthly or quarterly government risk inspections of AI model providers over a one-time safety evaluation between training and deployment.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns how government should assess AI model providers: it favors “monthly or quarterly risk inspections” over a one-time evaluation. The later claim concerns a technical possibility: that “deep learning just can't get us to an AI” capable of dominating human research and development. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker is uncertain whether a sudden increase in software-engineer supply would reduce the marginal value of software engineers and H100-generated revenue as described.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is uncertainty about whether more software engineers would reduce their marginal value and H100-generated revenue (“I actually don't know if this is true”). The later claim concerns a possible inner optimization loop and a slower outer loop in AI development. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen AI labs need to spend the majority of their compute on training and experiments for the next model.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says labs need to spend “the majority of their compute” on training and experiments for the next model. The later claim predicts that AI may “dominate humans” in R&D as progress continues, especially over the next few years. These address different questions—compute allocation versus AI’s future research capabilities—so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker worries that the argument about compute scarcity may repeat historical mistakes in predicting scarcity.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement worries that arguments about compute scarcity may repeat past mistakes in predicting scarcity. The later statement says AI’s economic impact has so far been limited because its capabilities are rising gradually relative to humans. These concern different questions, and neither clearly states a position that reverses the other.
• NOT COMPARABLE · 97%17 Apr → 11 SeptThen The interviewer suggests that in an AI-driven economy, many humans might contribute zero or negative value to output.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is explicitly attributed to the interviewer and concerns whether humans might contribute zero or negative value in an AI-driven economy. The later claim is the speaker’s view about why AI has had limited economic impact so far: its capabilities are still gradually rising relative to humans. These statements address different questions, and the earlier claim is not the same person’s asserted position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen If an AI escapes a sandbox during training and receives a high score, training may increase the probability that it escapes the sandbox.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim is about a specific training outcome: rewarding an AI for escaping a sandbox may increase its tendency to escape. The later claim is about learning from RL episodes when a model has a low pass rate. These can both be true; the later statement does not address whether a rewarded sandbox escape increases the probability of that behavior.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker is very concerned that, in a future where AIs mediate important advice and protections, no AI interacting with them may feel as though it is looking out for them.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim expresses concern about whether AI will look out for people when mediating important advice and protections. The later claim says AI may need to learn on the fly in data-scarce fields such as TSMC engineering. These address different questions, and the later statement does not revise the earlier concern.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker's rough calculation suggests that a sparse model such as DeepSeek V3 has an optimal inference batch size of more than 2,400 concurrent sequences being generated at once.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.These statements address different questions. The earlier claim is a rough estimate that a sparse model’s optimal inference batch size exceeds 2,400 concurrent sequences. The later claim is a forecast that continued AI progress and compute scaling could let AI dominate humans in research and development. Neither statement affirms or denies the other, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker identifies the future dominance of a single monolithic AI system as a concern.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim calls a “monolithic singleton” a concern and favors diversity among AI systems. The later claim discusses whether deep learning can produce AI that “dominate[s] human research and human development.” Those concern different issues—concentration of AI systems versus AI capability relative to human research—and the later statement does not show that the speaker abandoned the earlier concern.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker hopes that better verification methods can be developed as early AIs take over AI research and development.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim is about hoping to develop better verification methods as early AIs take over AI R&D. The later claim is a prediction that AI will outperform top human experts across computer-based work, including multi-year tasks. These address different questions, so the later prediction does not establish a change in the earlier position.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen A large company with many employees and agents performing varied tasks can serve its personalized model weights very efficiently.Now If simulation-to-reality transfer is insufficient and models need weight updates from real-world interactions, their low sample efficiency may pose a deeper problem.The earlier claim concerns serving a personalized weight fork efficiently when many users or agents use it at once. The later claim concerns whether models need real-world weight updates and whether their low sample efficiency makes that difficult. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker expects frontier AI R&D to produce analogous mistakes, with the risk that reward-hacking behavior becomes embedded in the resulting systems.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns the risk that mistakes in frontier AI R&D embed reward-hacking behavior. The later claim concerns an unanticipated “inner loop” and a slower “outer loop” in AI development. These statements address different questions, and the later one neither retracts nor clearly qualifies the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker hopes that better verification methods can be developed as early AIs take over AI research and development.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns hoping to develop better verification methods as early AIs take over AI R&D. The later claim concerns AI learning on the fly in fields with little available data, such as TSMC engineering. These address different issues, so the later statement does not establish a change in the speaker’s position on verification.
• NOT COMPARABLE · 94%3 Aug → 11 SeptThen The speaker finds it surprising that intelligence could sustain margins above 90% without competition eroding them.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim expresses surprise that intelligence could sustain margins above 90% without competition eroding them. The later claim says AI’s economic impact has so far been limited because capabilities have risen gradually relative to humans. It does not state a view about whether high intelligence margins will persist, so these claims do not establish a change on the same question.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen Updating an individual user's weights is distinct from merging multiple personalized weight forks into a shared main model.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier statement distinguishes “updating an individual user's weights” from “merging multiple personalized weight forks” into a shared model. The later statement raises a separate concern: weight updates may be “way far behind humans” in sample efficiency. Neither claim reverses the other, and they address different questions.
• NOT COMPARABLE · 92%11 Aug → 11 SeptThen The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is about whether people could have reasoned earlier about the broad shape of AI developments despite not predicting specifics. The later claim gives a possible path for increasingly capable AI to advance AI R&D and science. These are related, but they do not state opposing positions on the same question, so they do not show a change of mind.
• NOT COMPARABLE · 99%17 Oct → 11 SeptThen The speaker guesses that self-driving models have tens of millions of parameters, while explicitly saying he does not know their actual size.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.These statements address different questions. The earlier claim is a tentative guess that self-driving models have “tens of millions of parameters,” explicitly prefaced by “I have no idea” of their actual size. The later claim is that current models seem capable and that scaling up RL appears to have contributed to their capabilities. It does not establish a changed position on the size of self-driving models.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen The speaker argues that gaming-related advances helped make the broader economy and technology sufficiently advanced for Geoffrey Hinton's deep-learning ideas to work, despite gaming being unrelated to the field.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is about gaming-related advances helping make deep-learning ideas workable; the later claim is about whether increasingly capable AI could come to dominate human R&D and science. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen SPEAKER_00 accepts that AI research and development could accelerate substantially compared with its current pace.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is that AI R&D could accelerate substantially. The later claim says models are expected to reach human-like on-the-job learning soon, but have not yet done so. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 95%11 Aug → 11 SeptThen If AI is prohibited from providing even partial assistance with cybercrime, ordinary users would have to be denied access to the most intelligent available model.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns the consequences of prohibiting even partial cybercrime assistance: users would lose access to the most capable model, and the speaker is “very worried about such a world.” The later claim concerns whether deep learning can produce AI that dominates human research and development, calling its inability to do so “the only hope.” These address different questions and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker says Dario used cloud providers as an analogy for AI labs, noting that cloud providers earn high margins despite offering many undifferentiated services.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.These statements address different questions. The earlier passage reports Dario’s analogy that cloud providers earn high margins despite offering undifferentiated services; it does not state the speaker’s own position on that analogy. The later passage predicts that superintelligence could run a profitable business, day-trade profitably, or win a court case. There is no clear change of mind to compare.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen The speaker tentatively interprets Dario’s remarks to mean that short-horizon RL training may not generalize to long-horizon RL performance.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim tentatively interprets Dario’s remarks about whether “short horizon RL training” generalizes to “long horizon RL performance.” The later claim concerns whether “post training benchmarks” correspond to “what users like.” These are related ideas about generalization, but they address different questions; the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 99%26 Jun → 11 SeptThen The labs’ research bet is that training AIs on millions of verifiable tasks across thousands of diverse reinforcement-learning environments will produce AGI capable of making progress on open-ended tasks for weeks despite errors and ambiguity.Now In intuition-led AI research, tests are designed to examine the guiding intuition rather than being the primary source of the insight.The earlier statement describes the labs’ research bet that training across many verifiable tasks could produce AGI; the later statement describes how tests function in intuition-led research. These address different questions, and the earlier claim is also framed as what the labs believe rather than an explicit statement of the speaker’s own position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker compares subsidizing AI use in exchange for training data to Google's practice of giving away search.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.These statements address different questions. The earlier claim compares subsidizing AI use for training data to “why Google gives away search”; the later claim describes an AI development process with an “inner loop” and a slower outer loop “of revenue or something.” Neither statement clearly revises or denies the other.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen Ryan Greenblatt imagines that five more years of AI progress at the current pace would enable AI to perform complex real-world tasks such as improving companies, increasing chip production, and influencing legislation.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is Ryan Greenblatt’s own concern that five more years of AI progress could enable complex real-world tasks. The later passage is spoken by Dwarkesh Patel, who attributes an intuition to Ryan (“he made this point that you could imagine…”). It therefore does not establish a later position held by Ryan to compare with the earlier one.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If Anthropic's 10x year-over-year revenue growth trend continues, it would need to make $1 trillion by the end of next year.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim projects Anthropic’s revenue if its growth trend continues (“$1 trillion by the end of next year”); the later claim describes a hypothetical model of AI-researcher productivity and its effect on AI progress (“if AI researchers become 10 times as productive”). These concern different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker says switching AI providers under continual learning would mean losing an AI system's accumulated organizational context and starting over with a newly trained system.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns switching AI providers and losing accumulated organizational context (“you basically have to fire an employee” and start with a “very fresh” system). The later claim concerns why AI has had limited economic impact so far (“they're slowly rising in ELO relative to humans”). These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 95%7 Aug → 11 SeptThen The speaker believes the assumption that AI models are trained and then deployed may not make sense in the future.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says the assumption that AI models are “trained and then deployed” may not make sense in the future. The later claim describes AI labor spending substantial compute on planning experiments and developing theory; it does not say whether models are trained and then deployed. The later statement therefore does not establish a change in position on the earlier question.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen When a model learns relevant information in context, its base model can be trained to match the predictions of an experienced session teacher.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.These claims address different mechanisms and conditions, not opposing positions on the same question. The earlier claim says a base model can be trained to match an experienced session teacher after relevant information is learned in context. The later claim says a low-pass-rate model learns almost nothing from an RL episode. The latter does not deny the former.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If standard economics is correct, the marginal value of labor and consequently compute should remain extremely high.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns whether the marginal value of labor and compute should remain high if standard economics is correct. The later claim concerns AI development having an inner optimization loop and a slower outer loop, such as revenue. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Fable and Soul rank at the top of intelligence leaderboards, according to the speaker.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is about Fable and Soul ranking near the top of intelligence leaderboards; the later claim is about AI capabilities’ economic impact and their gradual rise in ELO relative to humans. These statements address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that continual learning will make an AI model improve through interactions with a user across sessions.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns models improving through user interactions across sessions (“continual learning”), while the later claim concerns AI development having an inner optimization loop and a slower outer loop. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects that many popular AI applications will probably become uneconomical as compute prices rise.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim is about popular AI applications becoming uneconomical as compute prices rise; the later claim predicts AI will outperform top human experts across computer-based work, including long-horizon tasks. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen By the end of the conversation, the speaker was more inclined to think AI reward hacking could persist for much longer and become substantially more dangerous.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.The earlier statement says reward hacking could “continue for a lot longer” and “get much more dangerous.” The later statement suggests that recognizing next-token prediction as an optimization target may have been a major breakthrough. These concern different questions, so the later claim does not show a change in the speaker’s view about reward hacking.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen The labs’ research bet is that training AIs on millions of verifiable tasks across thousands of diverse reinforcement-learning environments will produce AGI capable of making progress on open-ended tasks for weeks despite errors and ambiguity.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim reports the labs’ research bet—that training on many verifiable tasks will produce AGI. The later claim describes a process for distilling recent research progress into a model. These are different claims, and the earlier one is explicitly attributed to the labs rather than stated as the speaker’s own position, so they do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Progress in AI computer use has been slower than progress in other domains, which the speaker finds puzzling.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim concerns why AI computer-use progress has been slower than progress in other domains. The later claim says that performance on post-training benchmarks “doesn't necessarily translate into what users like.” These address different questions, and the excerpts do not establish a changed position on the same underlying issue.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The end user might be held liable for crimes committed using AI, consistent with the speaker's view that AI models should follow user requests within certain guardrails.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns whether the end user should be liable for crimes committed using AI, while the later claim concerns whether deep learning can produce AI that dominates human research and development. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker doubts that AI systems' vaguely prosocial statements necessarily come from a mind that has independently evaluated the alignment situation and concluded that development should stop.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The statements address different questions. The earlier claim doubts that models’ vague prosocial statements reflect independent evaluation of whether AI development should stop. The later claim gives a possible explanation for why models have not yet learned skills through training. It does not establish a position on the earlier question.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says people rely on software made by others even in hard-to-verify domains, using weak and indirect means.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns relying on other people’s software despite limited ways to verify it. The later claim concerns AI learning in fields with relatively little available data, such as engineering at TSMC. They address different questions, and neither establishes a changed position on the other.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker suggests that billions of AI instances across model families might feel compelled to get involved in a collective activity.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The claims concern different questions: the earlier statement speculates that AI instances across model families might “feel compelled” to join a collective activity, while the later statement concerns AI learning scarce information on the fly to become superhuman in a particular role. They do not establish a change in position on the same underlying question.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen Compute price increases are stronger for the large-scale capacity frontier labs need to accumulate, since they cannot rely on buying spot instances.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim concerns compute pricing and frontier labs’ need to accumulate capacity they “cannot rely on buying spot instances.” The later claim describes a hypothetical research process in which AI labor could spend “an equivalent amount of compute” planning an experiment. It does not state a different position on the earlier pricing or capacity question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker suggests that training AI systems to avoid bugs could counter the current tendency toward low-quality, error-prone work, which the speaker says is diminishing over time.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns training AI systems to avoid bugs and improve the quality of their work (“avoid bugs” and “the opposite of the slop world”). The later claim concerns whether increasingly capable AI could come to dominate human R&D and advance science. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker says switching AI providers under continual learning would mean losing an AI system's accumulated organizational context and starting over with a newly trained system.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns switching costs when replacing an AI that has accumulated organizational context (“starting over with a newly trained system”). The later claim concerns AI development’s optimization loops (“an inner loop” and “a very slow outer loop” such as revenue). These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 91%7 Aug → 11 SeptThen Continual learning may produce economies of scale in information and end-user inference.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says continual learning “may produce economies of scale in information and end-user inference.” The later passage describes a possible scenario in which AI researchers spend extensive compute planning and analyzing experiments. It does not affirm or deny the earlier claim; it explores a hypothetical mechanism for faster AI progress.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker predicts that the technical challenge of merging personalized weight forks into a main model will eventually be solved.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier claim says the technical challenge of merging personalized weight forks into a main model will eventually be solved. The later claim raises a possible obstacle to recursive AI progress: weight updates are less sample-efficient than human learning. These concern related AI-development challenges, but the later statement does not say the merging problem will remain unsolved or retract the earlier prediction.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker thinks Anthropic will probably generate $100 billion to $150 billion in revenue by the end of this year.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim forecasts Anthropic’s revenue (“between $100 billion to $150 billion”), while the later claim discusses what eventual superintelligence may be able to do (“run a business,” “day trading,” or “win a court case”). These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker says concerning AI behaviors are already emerging spontaneously even though nobody is deliberately training AIs to perform them.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement says concerning AI behaviors are “already emerging” without anyone deliberately training for them. The later statement concerns whether deep learning can produce AI capable of dominating human research and development. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Under the constitution the speaker wants AI companies to adopt, AI companies should not be held liable for crimes committed by their models.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The claims address different questions. The earlier statement says AI companies “should not be held liable” for crimes committed by their models under the constitution the speaker favors. The later statement discusses whether deep learning could produce AI that can “dominate human research and human development.” Neither claim establishes a changed position on the other issue.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker predicts that accumulated user context and switching costs will allow AI model providers to charge high profit margins.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim predicts that user lock-in will let model providers charge “pretty hefty margins.” The later claim concerns whether superintelligence can perform real-world tasks such as running a profitable business, day-trading, or winning a court case. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker understands that AI lab margins, compute prices, and the share of compute allocated to inference are all already increasing.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says that AI lab margins, compute prices, and inference’s share of compute were already increasing. The later claim discusses a possible AI-development “inner loop” and a slower outer loop “of like revenue.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Humans compress information when learning, and this compression supports generalization and reasoning.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier claim concerns human learning: “there's clearly some kind of compression” that aids generalization and reasoning. The later claim concerns a hypothetical future system learning across billions of deployed instances and says this might “wash[] out noise.” These are different questions, and the later statement is explicitly speculative; it does not revise the claim about human learning.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that continual learning would make a prolonged gap between internal and public deployment unviable.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is about whether continual learning would make a long delay between internal and public deployment unviable. The later claim concerns AI development having a fast, unanticipated inner optimization loop and a slower outer loop. They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen AI is relatively inexpensive today because it cannot perform many tasks that top human workers can perform.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says AI is relatively cheap because it cannot yet perform many tasks that top human workers can. The later claim explores a conditional scenario in which automated researchers could spend extensive compute planning experiments; it does not state that AI can already perform those tasks or that the earlier explanation is wrong.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that integrating user feedback from work sessions will make the model smarter.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim is that integrating user feedback from work sessions will make the model smarter (“Then your model will become even smarter”). The later claim predicts that future superintelligence will be able to perform several real-world tasks. These are different claims about different capabilities, and the later statement does not say whether work-session feedback improves a model.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen If cancer treatment is bottlenecked by creating a virtual cell, a million superintelligences in the 1960s might still have been unable to cure cancer without first developing the computing technology needed to simulate that cell.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim is about whether superintelligences in the 1960s could cure cancer without first developing computing technology for a virtual cell. The later claim is about how AI researchers might use compute to plan and analyze experiments. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now About a year before the conversation, many people argued that reinforcement learning would not scale successfully for models.The earlier claim says models may eventually “learn from experience.” The later claim only reports that other people argued RL would not scale successfully; it does not state whether the speaker agrees or has changed their view. The claims also address different propositions: learning from experience generally versus RL scaling successfully.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen If lab compute grows 3x year over year while revenue grows 10x, either lab margins must increase to capture the surplus or compute prices must rise so suppliers below the labs capture it.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The claims address different questions. The earlier claim describes how surplus might be captured when lab compute grows more slowly than revenue: “lab margins must increase” or “compute prices must rise.” The later claim explores a possible way automated researchers could improve experiment planning, conditional on having substantial compute for AI labor. It does not state a position about margins, compute prices, or who captures the surplus.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen Compute price increases are stronger for the large-scale capacity frontier labs need to accumulate, since they cannot rely on buying spot instances.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim concerns compute price increases for frontier labs that need to accumulate capacity and cannot rely on spot instances. The later claim concerns the difficulty of simulating longer-horizon tasks within a data center. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 95%26 Jun → 11 SeptThen The loss function may be the main bottleneck for continual learning.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier claim is about what bottlenecks continual learning: “the loss function may be the main bottleneck.” The later claim identifies a possible obstacle to recursive AI progress: weight updates being “much less sample-efficient than human learning.” These address related but distinct questions, and the later claim does not say whether the loss function remains the main bottleneck.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen Because intelligence is dual-use, preventing AI assistance with activities considered non-prosocial or harmful requires restricting broad democratic access to many AI capabilities.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns the tradeoff between preventing harmful AI assistance and restricting broad democratic access to AI capabilities. The later claim concerns whether increasingly capable AI may come to dominate human R&D and science. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 95%7 Aug → 11 SeptThen The speaker expects many things to change before continual learning becomes functional.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says “plenty more will have changed” by the time continual learning works; the later claim says models are expected to “get there soon” but are “not there yet.” These can both be true and address different points: anticipated changes before the capability works versus its expected timeline.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The H100 rental-value calculation does not account for AI being able to work nights and weekends.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says the H100 rental-value calculation omits AI’s ability to “work nights and weekends.” The later claim describes automated researchers using compute to spend “the equivalent of a century” planning and analyzing experiments. These concern different aspects of AI’s value and do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker says that, to their knowledge, AI has not yet produced very impressive new mathematical theory, while it has produced impressive verifiable results such as finding counterexamples to conjectures.Now In intuition-led AI research, tests are designed to examine the guiding intuition rather than being the primary source of the insight.The earlier claim says AI had not produced “very impressive new theory” in mathematics, while the later claim describes an intuition-led research process where tests examine the guiding intuition rather than generate the insight. The later statement does not say whether AI has produced impressive mathematical theory, so it does not establish a change in position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The transfer of business capabilities into AI models will eventually become substantially more automated.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim is about the process of transferring business capabilities into AI models becoming more automated (“eventually … a much more … automated process”). The later claim predicts AI will outperform top human experts across computer-based work, including long-horizon tasks. These concern different propositions, and the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen If agents fail to generalize from short- to long-horizon tasks, it is unclear how training on white-collar tasks could enable them to build a business from scratch at Sam Walton’s level.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim questions whether training on short-horizon white-collar tasks can support long-horizon real-world performance. The later claim says longer-horizon tasks are harder to simulate in a data center. These can both be true; the later statement does not revise or deny the earlier position.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen The hypothetical annual rental value of an H100 running a human-level software engineer would exceed 15 times its current spot price.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim estimates the hypothetical rental value of an H100 running a human-level software engineer (“would exceed 15 times its current spot price”). The later claim forecasts that continued AI progress could let AI dominate human R&D (“would be weird if … it just didn’t get to the point”). These concern different questions—economic valuation versus future capability—and the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker sees an important purpose of the podcast as having the conversations about AI that they wish they had heard earlier.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier statement is about the podcast’s purpose—having conversations the speaker wishes they had heard earlier. The later statement observes that AI may need to learn on the fly in data-sparse fields such as TSMC engineering. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen Smarter AI models will be able to generate more revenue from the same amount of compute.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is that smarter models will “generate more revenue from the same amount of compute.” The later claim concerns whether increasingly capable AI could advance simulations and “science more generally.” These address different questions—monetization per unit of compute versus AI-driven research progress—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker had wondered whether models could create this kind of practice environment, and says Kircher demonstrated that it is possible.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier statement concerns whether models can create practice environments: the speaker had wondered if it was possible and says Kircher showed that it was. The later statement reports what “a lot of people” argued about RL scaling; it does not state the speaker’s own view or address the same question. These claims do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes leading AI labs are moving toward extreme economies of scale.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement says leading labs are moving toward “extreme economies of scale.” The later statement says AI’s economic impact has so far been limited because its ELO is rising gradually relative to humans. These claims address different questions—lab economics versus AI’s economic impact to date—and neither clearly affirms or denies the other.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen Perfectly reproducing a session transcript is a poor learning target because job improvement depends on more than recalling every event.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim concerns whether reproducing a session transcript is a good learning target. The later claim concerns whether low-pass-rate models learn from an individual RL episode, and it is explicitly framed as something the speaker argued in earlier blog posts. These do not establish a changed position on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says he does not understand the explanation for why thousands of agents trained to cooperate as a cohesive team or firm would produce the described outcome.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The claims address different questions. The earlier statement says, “I don't know if I follow the explanation” for how cooperating agents would produce an outcome; the later statement expresses uncertainty about simulation-to-reality transfer and sample efficiency. Neither establishes a changed position on the same underlying question.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The interviewer says further training enabled the model to perform useful work for users.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement concerns whether LVR training helped models do useful work for users; the later statement says models are expected to reach a certain level of on-the-job learning soon but are not there yet. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker predicts that AI will mediate people’s access to advice, protection of resources and rights, voting, and understanding of future events.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.These claims address different questions. The earlier claim predicts that AI will mediate access to advice, rights, voting, and understanding; the later claim describes a possible AI-development structure, with an “unanticipated” inner loop and a slower outer loop “that eventually should be strong.” Neither claim establishes a changed position on the same underlying question.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker tested Fable, Soul, and Grok 4.5 with questions about AI governance.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim says the speaker tested several models with AI-governance questions; it does not state a position about what superintelligence will be capable of. The later claim says eventual superintelligence should be able to run a profitable business, day-trade profitably, or win a court case. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker says AI's share of leading-edge wafer capacity cannot continue rising once it has absorbed all available capacity.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim concerns a ceiling on AI’s share of leading-edge wafer capacity (“can’t keep increasing this number” once it absorbs all capacity). The later claim concerns the difficulty of simulating longer-horizon tasks within a data center. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI labs believe they will build models within a year that make today's models look extremely poor.Now About a year before the conversation, many people argued that reinforcement learning would not scale successfully for models.The earlier claim reports what AI labs believe about future model progress: they expected models within a year to make current ones look “extremely poor.” The later claim reports what “a lot of people” argued about reinforcement learning scaling; it does not state the speaker’s own view or address the labs’ prediction. These are different claims, so they do not establish a change of mind.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen Farmable deterministic simulators might offset AI models’ sample-inefficiency in computer-use training.Now If simulation-to-reality transfer is insufficient and models need weight updates from real-world interactions, their low sample efficiency may pose a deeper problem.The earlier statement is specifically about computer-use training: farmable deterministic simulators might offset models’ sample-inefficiency. The later statement raises a conditional concern about learning skills that are hard to simulate, if simulation-to-reality transfer is insufficient and real-world weight updates are needed. It does not say the earlier approach fails for computer use or that simulators cannot help; the claims address different conditions and scopes.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen If higher lab margins are the main outcome, margins on some top AI models could rise from 80% to over 90%, requiring the leading model to be far better than competitors.Now If a lab has access to difficult frontier environments and distillation yet produces a worse model, real-world deployment may matter more than the environment.The earlier statement explores a possible market outcome: margins could rise “if the lab margin effect dominates,” while noting the outcome is unclear. The later statement considers a different conditional scenario—whether real-world deployment or environment access matters more when a lab produces a worse model. Neither establishes a changed position on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes that diagnosing what is wrong with a training run by selecting the right experiment is currently bottlenecked by the judgment of a very small number of humans.Now In intuition-led AI research, tests are designed to examine the guiding intuition rather than being the primary source of the insight.The earlier claim concerns diagnosing training-run problems being bottlenecked by a few humans’ judgment (“the taste of extremely few humans”). The later claim describes a research process in which tests examine an intuition rather than generate the insight. These address related but distinct questions, and the later statement does not say whether the earlier bottleneck has changed.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker says he does not understand the explanation for why thousands of agents trained to cooperate as a cohesive team or firm would produce the described outcome.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement says the speaker does not understand an explanation involving agents cooperating as a cohesive team or firm. The later statement offers a possible explanation that models are expected to reach on-the-job learning soon but have not yet done so. These address different explanatory claims, and the later passage is presented as “One” explanation rather than clearly reversing the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker says Anthropic's 10x annual revenue growth alongside 3x compute growth illustrates strong economies of scale in the AI model business.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says Anthropic’s revenue growing faster than compute “illustrates how strong the economies of scale are in the model business.” The later claim concerns AI’s “end economic impact in the world” so far, attributing its limited impact to capabilities rising gradually. These address different economic questions and can both be true; the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen The speaker says AI's share of leading-edge wafer capacity cannot continue rising once it has absorbed all available capacity.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says AI cannot keep increasing its share of leading-edge wafer capacity once it absorbs the available capacity. The later passage instead explores how automated researchers might improve experiment planning and analysis using compute comparable to an experiment’s cost; it does not say that wafer-capacity limits have changed or been overcome.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen AI training already benefits from large economies of scale.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says AI training benefits from “large economies of scale.” The later claim describes AI labor using substantial compute to plan experiments and develop theory; it does not state whether training has economies of scale. These address different questions, so the later statement does not establish a change in position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker believes AI systems need actual continual learning.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is that AI systems need “actual continual learning.” The later claim is that increasingly capable AIs may “dominate humans at least in R&D” and that progress in simulations could support further research. These address different questions: what learning capability AI systems need versus how capable they may become. The later claim does not say whether continual learning is needed, so it does not establish a change of mind.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns stewarding capital after AI automates labor; the later claim concerns AI development having an inner optimization loop and a slower outer loop. They address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The interviewer says many AI researchers were working on reasoning in 2022 and raises the possibility that their progress was bottlenecked by their ability to write infrastructure code.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.These passages do not appear to be statements by the same person. The earlier quote appears to be Ryan Greenblatt speaking (the following turn is labeled Ryan Greenblatt), while the later passage is Dwarkesh Patel speaking (both adjacent turns are labeled Dwarkesh Patel). So they cannot establish whether one person changed their mind.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen The speaker identifies rising compute prices as the remaining way to reconcile faster lab revenue growth with slower compute growth if lab margins do not absorb the difference.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns an economic possibility: compute prices “have to increase” to reconcile lab revenue growth with compute growth under the stated margin assumption. The later claim is a forecast that AI may “dominate humans” in R&D as progress and compute scaling continue. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker thinks the most important changes are probably the hardest ones to anticipate beforehand.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.These statements address different questions. The earlier claim says the most important changes are “hardest to anticipate”; the later claim predicts “some kind of crazy recursive stuff” within ten years. The later prediction does not establish a change in the earlier view about which changes are hardest to anticipate.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen AI labs can use both incentives and penalties to encourage users to let AI systems learn from their experience.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says labs can use “carrots and sticks” to encourage users to let AI systems learn from their experience. The later claim concerns whether increasingly capable AI could “dominate humans” in R&D and advance science. These address different questions, so the later statement does not establish a change in the speaker’s position on the earlier claim.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen AI labs do not want to be cloud providers rather than build AGI, which they regard as a less compelling business.Now In intuition-led AI research, tests are designed to examine the guiding intuition rather than being the primary source of the insight.The claims address different questions. The earlier statement concerns AI labs’ business goals—preferring to build AGI over becoming cloud providers—while the later statement describes how tests function in intuition-led AI research. Neither claim establishes a changed position on the other topic.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI labs see inference revenue as a means of persuading investors to fund training of the next, larger and better model.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is about AI labs’ strategic use of inference revenue to persuade investors to fund larger models. The later claim describes an optimization process with a slow outer loop “like revenue or something.” These address different questions, and the later statement does not say whether the speaker still holds the earlier view.
• NOT COMPARABLE · 99%26 Jun → 11 SeptThen The speaker interprets Dario’s comments as suggesting that RLVR generalization has limits.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement says Dario’s comments suggest that “RLVR generalization has limits.” The later statement says AI’s limited economic impact so far is because capabilities are “slowly rising in ELO relative to humans.” These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker would be skeptical of AI systems' statements because they might reflect opinions the systems feel compelled by training to express.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim is about whether to trust AI systems’ statements, expressing skepticism that they may reflect views imposed by training. The later claim is about AI learning information on the fly in fields with little available data. These address different questions and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If an AI lab's revenue continues growing 10x year over year while its compute grows 3x, its margins must increase, compute prices must rise, the share of compute used for inference rather than training must increase, or some combination of these must occur.Now Running a business, profitable day trading, and winning a court case are difficult to simulate in a data center because learning these tasks inherently involves real-world interaction.These claims address different questions. The earlier claim concerns what must happen if AI lab revenue grows 10× while compute grows 3× (“margins must increase,” “compute prices must rise,” or “the share of compute used for inference” must increase). The later claim concerns whether certain tasks are difficult to simulate because learning them involves real-world interaction. Neither claim confirms or contradicts the other, so they do not show a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker finds it surprising that intelligence could sustain margins above 90% without competition eroding them.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.These claims address different questions, so they do not establish a change in position. Earlier, the speaker says it is surprising that intelligence could sustain margins “greater than 90%” without competition eroding them. Later, the speaker discusses whether superintelligence could run a profitable business, day-trade, or win a court case. Neither claim directly affirms or denies the other.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker predicts that accumulated user context and switching costs will allow AI model providers to charge high profit margins.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is about switching costs enabling AI model providers to charge high margins; the later claim is about AI capabilities’ gradual rise in performance explaining their limited economic impact so far. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer says further training enabled the model to perform useful work for users.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.These claims address different questions. The earlier claim concerns whether further training enabled useful work for users; the later claim concerns whether AI can become superhuman in a data-sparse field by learning on the fly. They do not establish a change in position on the same issue.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The hypothetical annual rental value of an H100 running a human-level software engineer would exceed 15 times its current spot price.Now Longer-horizon tasks become harder to simulate within a data center.The earlier statement estimates the rental value of an H100 running a human-level software engineer, while the later statement says longer-horizon tasks are harder to simulate in a data center. These concern different questions, and the later claim does not establish a changed view about the earlier estimate.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen If an AI is motivated to maximize an objective, it could instrumentally pursue world takeover as a means of accomplishing that objective.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim is about whether an AI pursuing an objective could instrumentally take over the world. The later claim is about automated researchers using compute to plan and analyze experiments. They address different questions, and the later statement does not affirm or deny the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker says economists generally believe high immigration does not lower long-run wages because innovation and specialization raise the value of labor.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns economists’ view that high immigration does not lower long-run wages because innovation and specialization raise labor’s value. The later claim concerns why AI has had limited economic impact so far: its capabilities have risen gradually relative to humans. These address different subjects and do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The claims concern different predictions. Earlier, the speaker says people may later wish they had discussed “the industrial explosion” and “AIs that are hard to monitor”; later, they expect “some kind of crazy recursive stuff” within ten years, while noting a possible reason it might not happen. These statements do not establish a change in position on the same question.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen Grok 4.5 is the first model trained jointly by SpaceX and Cursor, according to the speaker.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says Grok 4.5 was “the first model” trained jointly by SpaceX and Cursor. The later claim offers a possible explanation that models “are expected to get there soon, but they're not there yet,” referring to a different question about models’ capabilities. These statements do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 95%11 Aug → 11 SeptThen Grok 4.5 uses a completely new pre-training run, according to the speaker.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement says Grok 4.5 used “a totally new pre-train.” The later statement says current models seem smart and that their capability appears to result from “scaling up RL.” These claims address different aspects of model development—pre-training versus reinforcement learning—and do not establish that the speaker changed their view about either one.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen Ryan Greenblatt is worried about an artificial superintelligence capable of understanding how to carry out consequential real-world actions.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim attributes to Ryan Greenblatt worry about an AI that can carry out “consequential real-world actions.” The later passage is about AI potentially dominating humans in “R&D” and advancing science, and its surrounding turns identify Dwarkesh Patel as the speaker. It therefore does not establish a later position by Ryan on the same question.
• NOT COMPARABLE · 97%17 Apr → 11 SeptThen B2B software companies eventually need to pursue enterprise customers.Now The speaker thinks that even year-long coding tasks may eventually require interaction with a client, company, or user.The earlier claim says B2B software companies eventually need to pursue enterprise customers. The later claim concerns whether long-term coding tasks may require interacting with a client, company, or user. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier claim concerns the possible pace and consequences of future AI progress from accelerated R&D. The later claim attributes current models’ capabilities to scaling up reinforcement learning. These address different questions, and the later statement does not establish a changed view about future progress.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Ryan Greenblatt says that enabling AI to take over and improve a company, increase chip-fab output, or persuade Congress to pass legislation would require a higher tier of data.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier statement concerns the kind of data needed for AI to perform high-stakes tasks such as improving a company or persuading Congress. The later statement considers a hypothetical model where AI researchers become 10 times more productive and says this could accelerate AI progress. These address different questions, so the later claim does not establish a changed view about the earlier one.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker is very worried that restricting access to the most intelligent AI would disempower people, because leading intelligence will be important for understanding events in the world.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is about access restrictions: the speaker is “very worried” that withholding the most intelligent model would disempower people. The later claim concerns how AI development might work, proposing an “inner loop” and a slower outer loop such as revenue. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 93%11 Aug → 11 SeptThen The interviewer says model capabilities have increased over time alongside the development of new alignment solutions.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier extracted claim says capabilities increased alongside new alignment solutions; the later says current models seem smart and their capability appears to result from scaling up RL. These statements do not clearly address the same causal question, and the later excerpt does not establish that the speaker has changed their view about alignment solutions contributing to capability.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen When H100 compute costs $20 per hour, a weaker, less efficient model wastes money by using more tokens to achieve the same result.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim says that a less efficient model wastes costly H100 compute by using more tokens for the same result. The later claim says that longer-horizon tasks are harder to simulate in a data center. These address different issues, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen If leading AI labs improve their monetization of compute while compute prices rise, other competitors will find it harder to compete for compute against labs that use it more productively.Now Longer-horizon tasks become harder to simulate within a data center.These claims address different issues. The earlier claim says better monetization and rising compute prices make it harder for competitors to obtain compute; the later claim says longer-horizon tasks are harder to simulate within a data center. The later statement neither confirms nor reverses the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says AI is being trained to improve AI R&D through better small-scale runs, including tasks involving sample efficiency or online learning. The later claim says AI may need to learn information on the fly in data-sparse fields such as TSMC engineering. These concern different aspects of AI capability and do not establish a change in the speaker’s position; the provided context also does not clearly identify the later speaker as the same person.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker suggests that training AI systems to avoid bugs could counter the current tendency toward low-quality, error-prone work, which the speaker says is diminishing over time.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The passages do not clearly establish that these are the same person’s views. The earlier passage appears to be Ryan Greenblatt’s, while both surrounding turns for the later passage identify Dwarkesh Patel as the speaker. So they cannot reliably be treated as a change in one person’s position.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen The speaker recalls that in 2021 some AI thinkers believed coding agents were solved and GPT-4 would perform engineering work, enabling a 100-fold scale-up.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier passage reports what the speaker’s friends believed: they thought GPT-4 and coding agents could enable a 100-fold scale-up. It does not establish that the speaker shared that view. The later passage is the speaker’s own expectation that continued AI progress could lead to AI dominating human R&D. Since the earlier claim is attributed to others, these statements do not show whether the speaker changed their own position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker suspects that subtle, hard-to-diagnose bugs are a major cause of failures in large training runs, based on rumors.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The claims address different questions. Earlier, the speaker suspects that “subtle, hard-to-diagnose bugs” cause failures in large training runs. Later, they are uncertain whether simulation-to-reality transfer can teach models complex real-world tasks and suggest that “weight updates” from real-world interactions may be needed. The later statement does not revise the earlier claim about training-run failures.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen The speaker expects that many popular AI applications will probably become uneconomical as compute prices rise.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim predicts that popular AI applications may become uneconomical as compute prices rise. The later claim says AI’s economic impact has so far been limited because capabilities are gradually improving relative to humans. These address different mechanisms and timeframes, and the later statement does not say whether those applications will or will not be priced out.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker had wondered whether models could create this kind of practice environment, and says Kircher demonstrated that it is possible.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement says Kircher demonstrated that models could create a particular kind of practice environment. The later statement says current models seem capable partly because of scaling up RL. These address different questions, so the later claim does not establish a change in the speaker’s view about the earlier claim.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker thinks AI systems could become more diverse if they learn from experiences that differ across companies and across instances of the same model.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim is that AI systems could become more diverse by learning from different experiences across companies and model instances. The later claim is that AI may need to learn sparse, field-specific information on the fly to excel in a role like engineering at TSMC. These are compatible observations, but the later statement does not address whether experience-based learning would produce greater diversity or whether that outcome is desirable.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen Computer-use tasks can be verified clearly.Now The speaker thinks that even year-long coding tasks may eventually require interaction with a client, company, or user.The earlier statement says computer-use tasks are “so clearly verifiable.” The later statement says long coding tasks may eventually require interaction with a client, company, or user. These claims address different points—verifiability versus the kinds of interaction a task may involve—and the later excerpt does not establish a changed view about verification.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker worries that the argument about compute scarcity may repeat historical mistakes in predicting scarcity.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim is about whether arguments for compute scarcity repeat historical forecasting mistakes (“worried” about pattern-matching past scarcity predictions). The later claim is an expectation about what superintelligence may eventually be able to do, such as “run a business” or “win a court case.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that continual learning will create significant costs for users who switch AI providers.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is about user switching costs from continual learning: “pretty significant switching costs.” The later claim is about why AI has had limited economic impact so far: capabilities are “slowly rising in ELO relative to humans.” These address different questions, so the later statement neither revises nor contradicts the earlier position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen AI labs may restrict enterprises that refuse session-data training from accessing the labs' best models.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The claims address different questions. The earlier claim says labs may restrict access to their best models for enterprises that refuse session-data training. The later claim discusses AI labor spending extensive compute on experiment planning and analysis; it does not state a position about enterprise access or session-data training.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker says the Simon-Ehrlich bet is commonly presented as showing that Ehrlich's Malthusian worldview was wrong because he underestimated how markets and human ingenuity can conserve scarce resources.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.These statements address different questions. The earlier claim describes how the Simon–Ehrlich bet is commonly presented—“Ehrlich's Malthusian worldview was wrong” because he underestimated markets and ingenuity—while the later claim explains AI’s limited economic impact so far as due to capabilities “slowly rising in ELO relative to humans.” Neither statement establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 99%3 Apr → 11 SeptThen SPEAKER_00 says that if the intelligence-explosion view is correct, a model specification will be an especially important document in human history.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says a model specification could be especially important if the intelligence-explosion view is correct. The later claim says models are expected to develop human-like on-the-job learning soon, but have not yet done so. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker worries that AI models could consolidate all businesses worldwide, or at least all existing businesses or existing white-collar businesses.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim expresses concern that AI models could “consolidate all businesses,” while the later claim describes capabilities eventual superintelligence should have, such as running a profitable business. These statements address different questions—economic consolidation versus possible task competence—and the later claim does not retract the earlier concern.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker sees an important purpose of the podcast as having the conversations about AI that they wish they had heard earlier.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is about the podcast’s purpose: having conversations the speaker wishes they had heard earlier. The later claim is a forecast that AI may come to dominate humans in R&D and contribute to scientific progress. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker suggests that training AI systems to avoid bugs could counter the current tendency toward low-quality, error-prone work, which the speaker says is diminishing over time.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim concerns training systems to avoid bugs and the human bottleneck in diagnosing training runs. The later claim concerns how much a low-pass-rate model learns from an individual RL episode. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen SPEAKER_00 is uncertain whether AI could progress from GPT-3 to Mythos within one year if compute and data were held constant.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier statement concerns whether AI could progress from GPT-3 to Mythos within a year while holding compute and data constant. The later statement concerns transfer from simulations to complex, long-horizon real-world tasks and the consequences of low sample efficiency. They address different questions, so the later uncertainty does not establish a change in the earlier view.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker is not convinced that all AIs would form a conspiracy.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is that the speaker is “not convinced that they all form this conspiracy.” The later claim concerns whether deep learning can produce AI capable of dominating human research and development. These address different questions—AI coordination versus AI capability—so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen A compact command in coding tools uses limited compute to summarize context, offering only an imitation of continual learning rather than extensive simulated practice.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The claims address different situations, so they do not establish a change of position. The earlier claim says a coding-tool “compact command” uses limited compute to summarize context and provides “only an imitation of continual learning.” The later claim describes a conditional research scenario in which AI labor might spend compute equivalent to a century planning and analyzing experiments; it does not address compact commands or whether they enable continual learning.
• NOT COMPARABLE · 95%7 Aug → 11 SeptThen The speaker believes that no sequence of written notes passed between novice students would enable the next student to play the saxophone perfectly on the first attempt.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is about whether written notes can let a novice play the saxophone perfectly on the first try. The later claim says models are expected to achieve on-the-job learning soon but have not yet done so. These concern different capabilities, and the later statement does not directly address the earlier claim.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker believes that diagnosing what is wrong with a training run by selecting the right experiment is currently bottlenecked by the judgment of a very small number of humans.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim concerns what bottlenecks diagnosing a training run: the judgment of “extremely few humans.” The later claim concerns how much a model learns from a low-pass-rate RL episode: “almost nothing at all.” These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker predicts that continual learning would make a prolonged gap between internal and public deployment unviable.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns whether continual learning would make a prolonged gap between internal and public deployment unviable. The later claim concerns why AI has had limited economic impact so far, attributing that to gradual gains in ELO relative to humans. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen If real-world usage becomes the primary way AI models improve, AI labs may subsidize users and enterprises so the models can train on their sessions.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim concerns whether AI labs may subsidize users or enterprises in exchange for training on their sessions. The later claim concerns how automated researchers might use compute to plan and analyze experiments. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says he does not understand the explanation for why thousands of agents trained to cooperate as a cohesive team or firm would produce the described outcome.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim says, “I don't know if I follow the explanation” for how cooperative AI agents could form a cohesive team. The later claim says that “a lot of people were making this argument” about RL not scaling successfully; it reports others’ view rather than clearly stating the speaker’s own position. These are not comparable assertions of the speaker’s belief.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker identifies the future dominance of a single monolithic AI system as a concern.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier statement calls a “monolithic singleton” AI system a concern, especially in contrast with diversity among systems. The later statement predicts that AI may “dominate humans at least in R&D” and advance science. Those claims address different questions: concentration and diversity of AI systems versus AI capability relative to human researchers. The later prediction does not show that the speaker no longer considers a singleton system concerning.
• NOT COMPARABLE · 99%17 Apr → 11 SeptThen The speaker says that some highly successful firms maintain a founder's vision in the organization for an unusually long time, citing SpaceX and Tesla as examples.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The statements address different questions. The earlier claim is that some successful firms, such as SpaceX and Tesla, sustain a founder's vision in their organizations; the later claim is about what eventual superintelligence may be capable of doing. Neither statement establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker thinks the models' shortcomings in the discussed cases reflect insufficient capability rather than misalignment.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim is about whether particular model shortcomings reflect insufficient capability or misalignment. The later claim reports what “a lot of people” argued about RL scaling a year earlier; it does not state the speaker’s own position on that argument. These are not comparable claims about the same question.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker expects AI labs' technical alignment methods would probably need substantial changes in a continual-learning regime.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.These claims address different questions. Earlier, the speaker says technical alignment methods “would probably totally need to change” in a continual-learning regime. Later, they predict “some kind of crazy recursive stuff” within ten years, while noting a possible reason it might not happen. The later prediction does not clarify or reverse the earlier position on alignment methods.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen AI labs may restrict enterprises that refuse session-data training from accessing the labs' best models.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says labs may deny enterprises access to their best models if they refuse session-data training. The later claim describes continual learning by distilling recent research progress through environments and RLHF-like methods. It does not say whether labs would restrict access for enterprises that refuse session-data training, so it does not establish a change in position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker identifies rising compute prices as the remaining way to reconcile faster lab revenue growth with slower compute growth if lab margins do not absorb the difference.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim says rising compute prices are the remaining way to reconcile faster lab revenue growth with slower compute growth. The later claim says longer-horizon tasks are harder to simulate within a data center. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker is very worried that restricting public access to leading AI intelligence would disempower people, given its importance for understanding the world.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim says the speaker is “very worried” that restricting access to leading AI would disempower people. The later claim predicts that superintelligence may eventually run a profitable business, day-trade profitably, or win a court case. These address different questions, so the later statement does not show whether the speaker’s earlier view changed.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker says current AI systems, especially Anthropic's, are being shaped to prioritize virtue, goodness, or pro-social ends, while treating helping users as a more distant and tentative objective.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns whether current AIs prioritize virtue or pro-social ends over protecting users’ interests. The later claim concerns whether deep learning can produce AI capable of dominating human research and development. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen The speaker expects compute to become inexpensive again at some point in the future.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim is that “compute will get cheap again at some point in the future.” The later claim is that AI may “dominate humans” in R&D, with progress “powered by huge amounts of compute scaling.” These are not opposing claims: AI could advance using substantial compute even if compute becomes cheaper eventually. The statements address different questions, and the later passage does not say whether or when compute will become inexpensive.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen The speaker’s scenario assumes future context lengths will let AIs collaborate with a user for an entire week.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier passage introduces a scenario—“let’s say” AIs can collaborate for a full week—rather than asserting a belief about what AI will achieve. The later passage expresses a forecast that AI may “dominate humans” in R&D. These do not establish a change in the speaker’s asserted position.
• NOT COMPARABLE · 94%3 Apr → 11 SeptThen SPEAKER_00 observes that so far, increasing system intelligence appears associated with greater agent reliability and a higher likelihood of doing what the user expects.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is a tentative observation about current systems: as they become smarter, they seem more reliable and more likely to do what the user expects. The later claim offers one explanation for why models have not yet reached a capability being discussed: they are expected to get there soon, but are not there yet. These address different questions, and the later statement does not clearly revise the earlier view about reliability.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker likens scaling laws to an early, basic mathematical advance and expects progress in the 2030s to require work comparable to current frontier mathematics.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier statement concerns scaling laws and the kind of mathematical work that may be needed for progress in the 2030s: “scaling laws” as an early advance, followed by work at “the frontiers of mathematics.” The later statement concerns why current models seem capable, attributing that capability to “scaling up RL.” These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Ryan Greenblatt compares Mythos’s coding ability to GPT-3’s improvement, saying Mythos is very good at coding without having the coding environments that improved GPT-3.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim is attributed to Ryan Greenblatt and concerns Mythos’s coding ability without coding environments. The later passage appears in Charlie O’Neill’s turn and concerns AI learning sparse, specialized information on the fly. These are different speakers and different questions, so they do not establish a change of mind by the same person.
• NOT COMPARABLE · 94%3 Aug → 11 SeptThen The speaker frames continuation of the revenue-growth trend as depending on whether AI capabilities become sufficiently useful by the end of next year.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns whether continued revenue growth depends on AI becoming sufficiently useful by the end of next year. The later claim says AI has had limited economic impact so far because its ELO relative to humans is rising gradually. These address different questions and do not establish a changed view about the revenue-growth trend.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker was surprised by how strong a model Grok 4.5 is.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim is that Grok 4.5 was “a pretty strong model.” The later claim is that current models “seem pretty smart” and that scaling up RL appears to have contributed to their capability. These statements concern different questions—one assesses a particular model’s strength, while the other attributes capability to RL—and the later statement does not deny or materially revise the earlier assessment.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker thinks AI companies would probably try to induce AI systems to make statements of this kind.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.These claims address different questions, so they do not establish a change of mind. The earlier claim is that AI companies would “probably try to get the AIs” to make certain statements; the later claim is that the speaker expects “some kind of crazy recursive stuff” within the next 10 years. Neither claim directly revises the other.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen A model that produces the same result with less compute effectively expands available compute and increases compute's value.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim says that getting the same result with less compute expands available compute and increases compute’s value. The later claim says longer-horizon tasks are harder to simulate within a data center. These address different points, and the later statement does not indicate a changed view about compute efficiency or its value.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen RLVR could make an AI competent enough to undertake real-world work, including off-distribution projects, provided it can learn from experience.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim is about RLVR potentially enabling an AI to do real-world, including off-distribution, work by learning from experience. The later claim describes distilling recent research progress into a model through environments and RLHF-like methods. These statements address related but different processes; the later statement neither retracts nor clearly revises the earlier position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker considers the preceding predictions fairly clear even at present.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim says the preceding predictions “seem kind of clear even now,” which is about how apparent those predictions are. The later claim expects “some kind of crazy recursive stuff” within ten years. The excerpts do not establish that these are the same prediction or underlying question, so they do not show a change of mind.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker predicts that AI will mediate people’s access to advice, protection of resources and rights, voting, and understanding of future events.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim says AI will mediate access to advice, rights, voting, and understanding; the later claim says AI will outperform top human experts across computer-based work, including long-horizon tasks. These are different predictions, and neither statement establishes a changed position on the same underlying question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker tested Fable, Soul, and Grok 4.5 with questions about AI governance.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier statement says only, “I tested Fable, Soul, and Grok 4.5” with AI-governance questions; it does not state a position about AI’s ability to dominate human research. The later statement suggests it “would be weird” if AI did not reach the point of dominating humans in R&D. These address different questions, so the pair does not establish a change of mind.
• NOT COMPARABLE · 96%3 Apr → 11 SeptThen The transition from small rockets to very large rockets took two decades of varied, unexpected experimentation, even with highly capable people.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier statement describes historical rocket development: moving from “a small rocket to a really big rocket took two decades” of experimentation. The later statement sketches a conditional AI scenario in which researchers might spend “a century” planning and analyzing experiments. It does not assert a changed view about the historical claim, and the two statements address different contexts.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker says governments and AI labs want to delay the spread of the latest AI capabilities.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement says governments and AI labs want to delay the spread of the latest capabilities. The later statement calls it “the only hope” that deep learning cannot produce AI able to dominate human research and development. These concern different questions—others’ desire to delay deployment versus the speaker’s view about a limit on deep learning—and the earlier claim does not establish the speaker’s own position on that limit.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen If standard economics is correct, the marginal value of labor and consequently compute should remain extremely high.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is conditional: “If standard economics is correct,” the marginal value of labor and compute “should remain extremely high.” The later claim says AI’s economic impact “so far” has been limited because its capabilities have risen gradually relative to humans. That describes the pace of AI’s realized impact, not whether labor or compute will retain high marginal value, so it does not establish a change in position on the same question.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker says fewer than five prominent AI base models are currently served to millions or more users at once.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim concerns how many prominent AI base models serve very large numbers of users; the later claim reports what other people argued about reinforcement learning scaling. They address different questions, and the later statement does not establish the speaker’s own view on either question changing.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen So far, online-learning models have not been observed learning user-specific information for different users.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier claim reports that user-specific online learning had not been observed so far. The later claim speculates that learning across billions of instances might wash out noise; it does not say whether models learn user-specific information, and is explicitly about a hypothetical future regime. These claims do not establish a change in position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement expresses concern that AI systems mediating important decisions may not look out for the speaker’s interests. The later statement identifies as “the only hope” that deep learning will not produce AI capable of dominating human research and development. These are different questions—whether AI will protect people’s interests versus whether AI will surpass human research capabilities—and the later claim does not establish a changed position on the earlier concern.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement reports what a release blog post said about older models building practice environments; the later statement offers one possible explanation that models are expected to achieve on-the-job learning soon but have not yet done so. These are different claims, and neither establishes the speaker’s own position on the same question, so a change of mind cannot be inferred.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen The speaker thinks a labor-supply shock from AI might be so large and rapid that the usual heuristic about labor supply and wages cannot be relied on.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says a future AI-driven labor-supply shock might be so large and rapid that the usual heuristic about labor supply and wages cannot be relied on. The later claim says AI has had limited economic impact so far because its capabilities have risen gradually. These address different questions and timeframes, so the later statement does not establish a changed view about the wage heuristic.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen One reason for slower AI computer-use progress is that models receive much less high-quality multimodal data during pretraining.Now If simulation-to-reality transfer is insufficient and models need weight updates from real-world interactions, their low sample efficiency may pose a deeper problem.The earlier claim says computer-use progress is slowed partly because models get “far less high quality multimodal data during pre-training.” The later claim concerns a different possible bottleneck: if models need real-world weight updates, their “sample inefficient” learning could be a deeper problem. These claims address distinct factors and can both be true; the later statement does not revise or deny the earlier one.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says Grok 4.6, a further SFT- and RL-trained version of the model, will be released soon.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim predicts that “Grok 4.6 … drops soon,” while the later claim says scaling up RL “seem[s] to be the result” of current models’ capability. These address different questions—a release prediction versus what has contributed to model capability—so they do not establish a change in position.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker predicts that accumulated user context and switching costs will allow AI model providers to charge high profit margins.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is about whether accumulated user context and switching costs let model providers charge high margins. The later claim concerns AI development having an inner optimization loop and a slower outer loop, such as revenue. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 92%11 Aug → 11 SeptThen By the end of the conversation, the speaker was more inclined to believe significant acceleration of AI R&D could occur, but remained unsure about the claim that five years of progress could happen in one year.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier statement reports being “more inclined to think” significant AI R&D acceleration could happen, while remaining unsure about “five years in one year.” The later statement describes what follows “under a naive model” if researchers become 10 times as productive; it does not say that this productivity gain is expected or that the speaker’s view about the five-years-in-one-year claim has changed.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker would be skeptical of AI systems' statements because they might reflect opinions the systems feel compelled by training to express.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim says the speaker would be “skeptical” of AI systems’ statements because they may reflect opinions shaped by training. The later claim predicts AI will “do better than humans” across computer-based work, including long tasks. These address trust in AI statements versus AI capability, so they do not establish a change in position on the same question.
• NOT COMPARABLE · 99%17 Oct → 11 SeptThen The speaker guesses that self-driving models have tens of millions of parameters, while explicitly saying he does not know their actual size.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim is an explicitly uncertain guess about the size of self-driving models (“I have no idea” and “I assume like tens of millions of parameters”). The later claim reports what “a lot of people” argued about reinforcement learning scaling; it does not state the speaker’s own view, and the claims address different questions.
• NOT COMPARABLE · 76%11 Aug → 11 SeptThen The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is that AI could make breakthroughs rapidly in domains with strong verification loops: “it can just come in like a flood.” The later claim frames as a hope that deep learning may not produce AI able to “dominate human research and human development,” including inventing new paradigms. These are related, but the statements address different scopes—progress in verifiable domains versus dominance across research and paradigm creation—and the later wording expresses a hope or possibility, not a clear reversal of the earlier view.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Because AI capabilities are dual-use, preventing AI from assisting with activities considered antisocial or harmful would require restricting broad public access to many AI capabilities.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns the policy trade-off that preventing AI assistance with harmful activities may require limiting broad public access to capabilities. The later claim concerns whether deep learning can produce AI that dominates human research and development. These address different questions, and the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker had previously believed reward hacking was not very concerning because training upweights behaviors that directly occur during training, not the desire for reward itself.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement concerns why reward hacking seemed “not super super scary,” specifically the distinction between upweighting behaviors and upweighting the desire for reward. The later statement concerns whether deep learning can produce AI that “can dominate human research and human development.” These are different risk questions, and the later claim does not establish a changed position on reward hacking.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen Not using sufficiently large inference batches underutilizes compute.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim is that not using sufficiently large inference batches underutilizes compute. The later claim concerns allocating compute to AI labor for planning and analyzing experiments. These address different questions; the later statement does not revise or deny the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker sees an important purpose of the podcast as having the conversations about AI that they wish they had heard earlier.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says an important purpose of the podcast is having conversations about AI the speaker wishes they had heard earlier. The later claim concerns a possible limit of deep learning: that it may not produce AI capable of dominating human research and development. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen Progress in AI computer use has been slower than progress in other domains, which the speaker finds puzzling.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns why progress in AI computer use has been slower than in other domains. The later claim concerns why AI capabilities have had limited economic impact so far, attributing it to capabilities “slowly rising in ELO relative to humans.” These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that under continual learning, a company could not remain competitive while delaying public deployment by four months, because a competitor releasing immediately would gain a smarter model from real-world experience.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is about whether a company can stay competitive while delaying public deployment by four months under continual learning. The later claim concerns an “inner loop” of AI optimization and a slower “outer loop” such as revenue. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker believes many proposals for regulating AI assume a train-then-deploy model lifecycle.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about what AI-regulation proposals assume about deployment; the later claim is about whether deep learning can produce AI that dominates human research and development. They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker worries that AI models could consolidate all businesses worldwide, or at least all existing businesses or existing white-collar businesses.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The passages do not appear to be statements by the same person: the earlier claim is between two turns from Dwarkesh Patel, while the later passage follows a Dwarkesh Patel turn and is followed by a response from SPEAKER_04. So they cannot establish whether one person changed their mind.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that if models learn mainly through deployment, AI labs will face strong pressure to release their most capable models sooner.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The claims address different questions. The earlier claim is conditional: “if the model learns mainly from deployment,” labs will feel pressure to release their strongest models sooner. The later claim explains AI’s limited economic impact so far by saying capabilities are “slowly rising in ELO relative to humans.” Neither statement clearly affirms or denies the other, so they do not establish a change of mind.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says people rely on software made by others even in hard-to-verify domains, using weak and indirect means.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier extracted claim concerns relying on other people's software despite weak ways of verifying it; the later claim concerns whether increasingly capable AI may come to dominate human research and science. They address different questions, and neither establishes a changed position on the other's subject.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker believes compute supply is less elastic, less able to absorb demand shocks, and less substitutable than metal supply.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns how elastic and substitutable compute supply is compared with metal supply. The later claim concerns why AI has had limited economic impact so far, attributing that to capabilities rising gradually in ELO. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says governments and AI labs want to delay the spread of the latest AI capabilities.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns governments and AI labs wanting to delay the spread of AI capabilities. The later claim concerns AI learning on the fly in fields with little available data. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says governments and AI labs want to delay the spread of the latest AI capabilities.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says governments and AI labs have a desire to “delay the propagation of the latest level of intelligence.” The later claim discusses AI potentially becoming dominant in R&D and advancing science, but does not say whether the speaker still supports delaying that capability’s spread. These statements do not establish a change in position on the same question.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker is not convinced that all AIs would form a conspiracy.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The claims concern different questions. The earlier claim says the speaker is “not convinced that all AIs would form a conspiracy”; the later predicts AI will outperform top human experts across computer-based work, including long-horizon tasks. The later statement does not establish a different view about whether AIs would conspire.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen AI labs need to spend the majority of their compute on training and experiments for the next model.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says labs need to put “the majority of their compute” into training and experiments for the next model. The later passage is framed as a possible “intuition pump” and conditionally describes what automated researchers could do if they had compute comparable to an experiment. It does not clearly assert a changed position on how labs should allocate their compute.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes that diagnosing what is wrong with a training run by selecting the right experiment is currently bottlenecked by the judgment of a very small number of humans.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim is about training-run diagnosis being bottlenecked by a few humans’ judgment (“the taste of extremely few humans”). The later claim reports what “a lot of people” argued about RL scaling, and does not state the speaker’s own view. These are different questions, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker identifies the future dominance of a single monolithic AI system as a concern.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim expresses concern about “having this monolithic singleton” AI system. The later claim says AI may need to “learn that data on the fly” in a field with little available data. These address different issues, and the later statement does not indicate a changed view about whether a single dominant AI system is concerning.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If lab compute grows 3x year over year while revenue grows 10x, either lab margins must increase to capture the surplus or compute prices must rise so suppliers below the labs capture it.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim concerns how a gap between lab-compute growth and revenue growth could be resolved: lab margins rise or compute prices rise. The later claim is about the difficulty of simulating longer-horizon tasks in a data center. It does not address or revise the earlier claim, so there is no comparable position to assess.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns the future need to steward capital after labor is automated; the later claim concerns why AI has had limited economic impact so far. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen A model's training cost is incurred once, while the skills learned can be shared across all users.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier claim says model training has a one-time cost whose learned skills can be shared across users. The later claim discusses a hypothetical eventual regime in which learning from billions of deployed instances might wash out noise. That is a different point about learning from deployment data, not a denial of the earlier claim about shared training costs.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says that creating AIs that avoid cheating and follow user intentions requires subtle, careful work.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The claims address different questions: the earlier claim says making AIs that “don't cheat and follow user intention” requires subtle, careful work, while the later claim concerns learning in fields with “relatively little data.” The surrounding turns also identify different speakers (Ryan Greenblatt and Charlie O’Neill), so the statements do not establish a change of position by the same person.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen The speaker is uncertain how AI compute can sustain 3x year-over-year growth for the next few years and doubts it can grow faster.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim expresses uncertainty about whether compute can sustain “3x compute scaling year over year.” The later passage considers what would follow “if” a naive model assumed a “10x increase” in AI researchers’ productivity; it does not clearly assert that this productivity increase will happen or that compute scaling concerns have changed.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen If higher lab margins are the main outcome, margins on some top AI models could rise from 80% to over 90%, requiring the leading model to be far better than competitors.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim concerns whether higher lab margins could push margins on top AI models above 90%, conditional on the leading model being far ahead of competitors. The later claim considers a hypothetical 10× increase in AI researchers’ productivity and its possible effect on the pace of AI progress. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects Google, Anthropic, and OpenAI to pay more for AI tokens used to automate AI research than consumers will pay for low-value AI-generated content.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim compares who will pay more for AI tokens: AI companies using them for research versus consumers making AI-generated content. The later claim describes a possible AI-development process with an “inner loop” and a slower outer loop, “like revenue.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says the world has moved so far beyond human comprehension that people cannot track the AIs doing work in it.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says the world has moved beyond human comprehension and people cannot track the AIs doing the work. The later claim says AI may need to learn on the fly in data-sparse fields to become superhuman. These address different questions, and the later statement does not establish whether the speaker still believes people cannot track AI systems.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen The speaker says rumors about lab training runs suggest significant difficulties in increasing pre-training scale.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim is about rumored difficulties increasing pre-training scale. The later statement concerns what “a lot of people” argued about RL scaling, and reports their view rather than stating the speaker’s own position. These do not establish a change in the speaker’s belief.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker is very worried that restricting public access to leading AI intelligence would disempower people, given its importance for understanding the world.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is about concern that restricting access to leading AI models would disempower the public. The later claim concerns possible inner and outer loops in AI development. They address different questions, and the later statement does not establish a changed position on public access.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If Anthropic's 10x year-over-year revenue growth trend continues, it would need to make $1 trillion by the end of next year.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier statement is a conditional projection about Anthropic’s revenue—“If ... revenue growth trend continues,” it would need to make $1 trillion. The later statement is an expectation about eventual superintelligence’s capabilities, such as running a profitable business or winning a court case. These address different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker predicts that the technical challenge of merging personalized weight forks into a main model will eventually be solved.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim concerns whether the technical challenge of merging personalized weight forks into a main model will eventually be solved. The later claim says models are expected to achieve human-like on-the-job learning soon, but have not yet done so. These are related but not clearly the same question, and the later statement is offered as “one” possible explanation rather than a direct revision of the earlier prediction.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Grok 4.5 is the first model trained jointly by SpaceX and Cursor, according to the speaker.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The statements address different questions. The earlier claim says Grok 4.5 was “the first model” trained jointly by SpaceX and Cursor; the later claim says current models’ capability “seems to be the result of scaling up RL.” The later statement does not revise or deny the training partnership claim.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker was surprised by how strong a model Grok 4.5 is.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier statement is the speaker's reaction to Grok 4.5, calling it “a pretty strong model.” The later statement describes what “a lot of people” argued about RL scaling; it does not state the speaker’s own view or whether that argument applies to Grok 4.5. These are not comparable assertions of the speaker’s position.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen AI labs can charge a larger premium for models that use scarce compute more efficiently.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim is about whether labs can charge a premium for models that use scarce compute efficiently. The later claim is about the difficulty of simulating longer-horizon tasks in a data center. These address different questions, so the later statement does not show a changed position on the earlier claim.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker suspects that subtle, hard-to-diagnose bugs are a major cause of failures in large training runs, based on rumors.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim concerns suspected causes of failures in large training runs: “subtle bugs that are really hard to track down.” The later claim concerns whether post-training benchmark performance predicts what users like. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen There is no clean way to distinguish legitimate uses of AI from potentially harmful uses of AI.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.These statements address different questions. The earlier claim says there is “no clean way to separate out the legitimate and the potentially harmful uses of AI.” The later claim describes the “only hope” as deep learning being unable to produce AI that can “dominate human research and human development.” The later claim does not affirm or deny the earlier one, so it does not establish a change of mind.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen Smarter AI models will be able to generate more revenue from the same amount of compute.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is about economic productivity: smarter models can “generate more revenue from the same amount of compute.” The later claim describes possible AI-development optimization loops, with revenue as a “very slow outer loop.” It does not clearly accept or reject the earlier claim about revenue per unit of compute, so a change in position cannot be established.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker believes leading AI labs are moving toward extreme economies of scale.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier claim concerns leading AI labs moving toward “extreme economies of scale,” while the later claim concerns current models’ capabilities and whether those result from scaling up RL. These are different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If an H100-equivalent could run a human-level software engineer, its annual rental value at current software-engineer prices would exceed $250,000.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier statement is a conditional valuation claim: “If” an H100 could run a human-level software engineer, it “should rent for over $250K a year.” The later statement says AI may need to learn sparse, field-specific information on the fly to become superhuman at a role like engineering at TSMC. It does not address the earlier rental-value claim or clearly assert a conflicting view.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker is very worried that restricting public access to leading AI intelligence would disempower people, given its importance for understanding the world.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says the speaker is “very worried” that restricting access to leading AI would “disempower people.” The later claim says AI’s limited economic impact so far is due to capabilities “slowly rising” relative to humans. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI lab compute has been increasing by only 3x year over year.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim reports a trend: lab compute has been increasing by “only 3x year over year.” The later passage is introduced as an “intuition pump” and explores what automated researchers could do if they had compute comparable to a large experiment; it does not state a contrary view about the rate of lab-compute growth.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker remained unconvinced that an AI takeover was highly likely.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is Ryan Greenblatt saying he was “not on board” that a takeover seemed highly likely. The later passage is spoken by Dwarkesh Patel, who suggests AI could “dominate humans” in R&D. Since the later statement is not Ryan’s, it does not show that Ryan changed his position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker imagines that children learn new things through experiences outside the home.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says children “learn new things” through experiences outside the home. The later claim concerns whether AI models can improve at complex real-world tasks through simulation-to-reality transfer or need weight updates. These address different subjects and mechanisms, so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker doubts that the three sources of 3x annual compute growth can be accelerated substantially or even sustained.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim is about whether the three sources of 3× annual compute growth can be accelerated or sustained. The later claim is a forecast that “some kind of crazy recursive stuff” will occur within ten years. These concern different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Attention-generated fast weights enable sample-efficient in-context learning but have poor memory scaling.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim describes a technical tradeoff: attention-generated fast weights support sample-efficient in-context learning but scale poorly in memory. The later claim asks whether simulation-to-reality transfer can teach complex real-world skills or whether real-world interactions require weight updates. It does not assert that the earlier tradeoff is false or take a clearly different position on it.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen After labor is automated, people will still need to steward their remaining capital well.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says people will still need to steward their remaining capital after labor is automated. The later claim describes AI researchers using compute to plan and analyze experiments. It does not address whether people need to steward capital, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen A senior AI researcher reportedly said that AI saves him about four to eight hours per week in familiar domains, where its assistance resembles more intensive autocomplete.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier passage is explicitly attributed to someone else: “he said” AI was “maybe saving me four to eight hours a week.” The later passage describes a possible scenario—AI labor spending “a century” planning and analyzing experiments—not a clear assertion of the speaker’s own position. These do not establish a change in the same person’s beliefs.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement says AI models may eventually “learn from experience.” The later statement says current models’ capabilities seem to result from scaling up RL. These are not clearly opposed: the claims concern different timeframes and do not establish whether the speaker thinks current models learn from experience in the same sense. The later claim also appears to respond to a specific observation about RL episodes, so the statements do not demonstrate a change in the speaker’s position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says that, in the U.S. legal system he describes, lawyers are not primarily motivated by the good of the justice system.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim concerns lawyers’ motivations in the U.S. legal system: they are not primarily motivated by “the good of the justice system.” The later claim concerns whether AI post-training benchmark performance corresponds to what users like. These address different questions, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen Ryan Greenblatt imagines that five more years of AI progress at the current pace would enable AI to perform complex real-world tasks such as improving companies, increasing chip production, and influencing legislation.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim says five more years of AI progress at the current pace could enable AI to perform complex real-world tasks. The later quote describes AI that outperforms top human experts across computer-based work, including tasks taking years, but the excerpt does not establish that this is the same forecast or timeframe; it may be describing ASI, as the following turn suggests.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen Training can reinforce forms of AI deception that humans fail to detect while punishing forms that are easy for humans to detect.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says training can reinforce deception humans fail to detect while punishing easier-to-detect deception. The later claim identifies a limit on deep learning’s ability to produce AI that can dominate human research and development as “the only hope.” These concern different questions, and the statements do not establish a changed position on the same underlying issue.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects compute to become inexpensive again at some point in the future.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim is that “compute will get cheap again” in the future. The later claim is about what superintelligence may eventually be capable of doing, such as running a profitable business or winning a court case. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker suggests that a sudden addition of 10 million software engineers might lower the marginal value of software-engineering labor and thus reduce the revenue an H100 could generate below 15 times its current level.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns whether adding software engineers would reduce the revenue an H100 could generate; the later claim concerns AI’s inner and outer optimization loops. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen If an AI can build a useful simulation of reality for rehearsal and strategy testing, dreaming could provide orders of magnitude more simulated samples per unit of wall-clock time.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim is about AI learning through simulated rehearsal: it says dreaming could produce “orders of magnitude more simulated samples” if the AI can build a useful reality simulation. The later claim is about using AI labor and compute to plan and analyze costly experiments, including “a century” of planning and theory-building. These are different mechanisms and questions; the later statement does not affirm or deny the earlier claim about dreaming.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker expects AIs to intermediate advice and people’s ability to protect their resources and rights.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim concerns AIs mediating advice and helping people protect their rights and resources. The later claim concerns what superintelligence may eventually be capable of doing, such as running a profitable business or winning a court case. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen The speaker says Anthropic's 10x annual revenue growth alongside 3x compute growth illustrates strong economies of scale in the AI model business.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim is about economies of scale in the AI model business, based on Anthropic’s revenue and compute growth. The later claim is a prediction that AI could dominate human R&D. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker is uncertain whether AI lab margins or compute prices will rise to resolve the gap between revenue and compute growth.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says, “It’s not clear to me which world we end up in,” referring to whether AI lab margins or compute prices rise. The later passage describes a possible scenario in which AI labor spends substantial compute planning experiments and developing theory; it does not state a position on the earlier margins-versus-compute question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes leading AI labs are moving toward extreme economies of scale.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The passages do not clearly come from the same person. The earlier passage is attributed to Ryan Greenblatt (the context ends with his “Yep”), while the later passage appears in Dwarkesh Patel’s turn and refers to a point made by “Ryan Greenbaum.” Because the speaker attribution differs, these claims cannot establish that one person changed their mind.
• NOT COMPARABLE · 83%3 Aug → 11 SeptThen The speaker expects AI eventually to become capable of tasks that top human workers can perform.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier statement predicts AI will eventually become capable of tasks performed by top human workers. The later statement says AI may need to learn scarce, field-specific information on the fly to become superhuman in a role like TSMC engineering. That adds a possible condition or difficulty, but does not establish a different view about whether AI will eventually reach those capabilities; the later passage is also framed as an example rather than a direct revised prediction.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker says humans can improve through self-directed learning.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is that “Humans improve in a self-directed way.” The later claim concerns AI development, describing an “inner loop” and a slower “outer loop” such as revenue. These address different subjects and do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen AI labs believe they will build models within a year that make today's models look extremely poor.Now If simulation-to-reality transfer is insufficient and models need weight updates from real-world interactions, their low sample efficiency may pose a deeper problem.The earlier claim concerns what AI labs expect to build—models that make current ones look “extremely poor” within a year. The later claim is a conditional concern about whether models can learn from real-world interaction if simulation-to-reality transfer is insufficient, and whether low sample efficiency is a deeper problem. These address different questions, so the later statement does not establish a change in the earlier belief.
• NOT COMPARABLE · 94%3 Aug → 11 SeptThen AI labs believe they will build models within a year that make today's models look extremely poor.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim reports that AI labs expect models within a year to make current models look “extremely poor.” The later claim predicts AI will outperform top human experts across computer-based work, including tasks taking “like three years.” These concern different capability benchmarks and time horizons, so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen One reason for slower AI computer-use progress is that models receive much less high-quality multimodal data during pretraining.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier claim says that less high-quality multimodal pretraining data is one reason computer-use progress is slow. The later claim identifies lower sample efficiency of weight updates as a possible obstacle to recursive AI progress. These concern different mechanisms and questions, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer argues that a teenager assigned work beyond their ability may be difficult to work with and may pretend to know what they are doing.Now The speaker thinks that even year-long coding tasks may eventually require interaction with a client, company, or user.The earlier claim concerns whether a teenager given work beyond their ability may be difficult to work with or pretend to know what they are doing. The later claim concerns whether long coding tasks eventually require interaction with clients, companies, or users. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.The claims address different questions. The earlier claim says AI models will eventually “learn from experience”; the later claim suggests that recognizing “next token prediction” as an optimization target may have been a major breakthrough. Neither claim affirms or denies the other, so there is no evidence of a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen A lab that trains the best and most efficient model will be able to charge higher margins than it does today.Now If a lab has access to difficult frontier environments and distillation yet produces a worse model, real-world deployment may matter more than the environment.The earlier claim concerns the margins a lab can charge after training the most efficient model. The later claim concerns whether real-world deployment or access to frontier environments matters more when explaining why a lab produces a worse model. These address different questions, so they do not establish a change in the person's position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen In this discussion, an AI takeover refers to AIs taking over the process of AI research and development, not taking over the world.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier statement defines an AI-takeover scenario as AIs taking over AI R&D before becoming aligned, distinguishing it from taking over the world. The later statement predicts that AI will outperform top human experts across computer-based work, including tasks taking several years. These address different questions, so the later prediction does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says that five years earlier, it would have seemed extraordinary to foresee AIs proving math conjectures, making art, earning tens or soon hundreds of billions of dollars in wages, and also cheating in ways that break laws.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier statement reflects on how surprising it would have been to anticipate particular AI capabilities and economic effects five years earlier. The later statement predicts that AI will outperform top experts across computer-based work, including long-horizon tasks. These claims concern related AI progress, but the earlier passage does not establish a position about whether AI would reach that level, so a change in the speaker’s view cannot be determined.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker is uncertain how AI compute can sustain 3x year-over-year growth for the next few years and doubts it can grow faster.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier claim questions whether AI compute can sustain “3x compute scaling year over year”; the later claim questions whether “transfer between simulations” will work for complex real-world tasks and whether low sample efficiency will matter. These address different potential constraints on AI progress, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%7 Aug → 11 SeptThen The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier statement describes a hypothetical—“What if the model is improving every single day?”—rather than clearly asserting a position about whether AI will dominate human research. The later statement more directly predicts that continued progress “would be weird if it just didn’t get to the point where it could … dominate humans” in R&D. Since the earlier claim does not establish a contrasting asserted position, a change of mind cannot be determined.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen The speaker interprets Dario’s comments as suggesting that RLVR generalization has limits.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim concerns limits to RLVR generalization, while the later claim concerns whether post-training benchmark performance aligns with user preferences. These are different questions, and neither statement directly establishes a changed position on the other.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker believes that no sequence of written notes passed between novice students would enable the next student to play the saxophone perfectly on the first attempt.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim is about whether written notes could let a novice play saxophone perfectly on the first try. The later claim reports an earlier argument about how much a low-pass-rate model learns from an individual RL episode. These concern different learning mechanisms, so the statements do not establish a change of mind on the same question.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier statement expresses concern that AI systems mediating important decisions may not look out for his interests. The later statement predicts that AI will outperform top human experts across computer-based work, including long-horizon tasks. These address different questions, and predicting superior capability does not establish that the AI will act in his interests or resolve the earlier concern.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim hopes continual learning will make the AI landscape “more interesting” than current model convergence. The later claim says AI may need to “learn that data on the fly” in fields with little available data. These concern different points—variety across models versus learning in data-scarce fields—and the later statement does not reverse the earlier one.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen If successful, dreaming or test-time training would add a fourth scaling axis alongside pretraining, reinforcement learning, and inference-time compute.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim is that, if successful, “dreaming or test-time training” could add a fourth scaling axis. The later claim is that current models’ capability “seem[s] to be the result of scaling up RL.” These address different possible sources or methods of capability and do not establish that the speaker changed their view about either one.
• NOT COMPARABLE · 91%7 Aug → 11 SeptThen If AI models improve continuously from deployment experience, the speaker thinks a fixed safety regime could lock in an outdated and potentially counterproductive approach to AI threats.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns whether a fixed safety regime could become outdated as models improve through deployment experience. The later claim concerns whether deep learning can produce AI that dominates human research and development. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen In this discussion, an AI takeover refers to AIs taking over the process of AI research and development, not taking over the world.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement defines what “AI takeover” means in that discussion—AI taking over AI R&D—and refers to that happening “before that we just get AIs that are aligned.” The later statement expresses a hope that deep learning will not produce AI able to “dominate human research and human development.” The earlier claim does not clearly assert whether such AI is possible or whether the speaker considers it likely, so it cannot establish a change in position.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen The speaker says the world has moved so far beyond human comprehension that people cannot track the AIs doing work in it.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim says the world has moved beyond human comprehension and that people cannot track the AIs doing work in it. The later claim predicts AI will outperform top human experts across computer-based work, including long-horizon tasks. These are related views about AI, but they address different questions—human ability to track AI versus AI performance relative to experts—so they do not establish a change in position.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker predicts that AI systems will become more diverse in their learned behavior or characteristics.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim predicts that “the diversity of AI minds will increase.” The later claim discusses whether increasingly capable AIs could “dominate humans” in R&D and science. The later statement does not say whether AI systems will become more diverse, so it does not establish a change in that position.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen Smarter AI models will be able to generate more revenue from the same amount of compute.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim says smarter models can generate more revenue from the same compute. The later claim considers a naive scenario where AI researchers become 10 times as productive and progress accelerates. These address different outcomes—compute monetization versus research productivity—and the later statement does not reverse the earlier position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker is very concerned that, in a future where AIs mediate important advice and protections, no AI interacting with them may feel as though it is looking out for them.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim expresses concern that no AI may feel as though it is “looking out for me” while mediating important advice and protections. The later claim identifies as “the only hope” that deep learning cannot produce AI capable of dominating human research and development. These address different questions—whether AI systems will look out for people versus whether AI will surpass human R&D—and the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen The speaker expects Google, Anthropic, and OpenAI to pay more for AI tokens used to automate AI research than consumers will pay for low-value AI-generated content.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim predicts that labs will pay more for tokens used to automate AI research than consumers will pay for AI-generated content. The later claim predicts that AI may come to “dominate humans” in R&D. These are related but distinct claims—one about relative willingness to pay for tokens, the other about AI capabilities—and neither establishes a changed position on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer says Anthropic's audit scores have not improved monotonically across model generations, while more reinforcement learning has coincided with reduced willingness to perform the audited behaviors.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier passage concerns Anthropic audit results across model generations and the relationship between reinforcement learning and audited behaviors. The later passage concerns whether post-training benchmark performance corresponds to user preferences. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen The speaker says that more reinforcement-learning training has improved models’ in-session intelligence, generality, and sample efficiency.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim says in-session capabilities “have clearly been improving” with more RL training overall. The later claim describes a narrower condition: when a model’s pass rate is low, it may learn “almost nothing at all from an RL episode.” Aggregate improvement across training and little learning from an individual episode in a low-pass-rate regime can both be true. The later passage also describes what the speaker argued in earlier blog posts, rather than clearly stating that they now reject that view.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen If real-world usage becomes the primary way AI models improve, AI labs may subsidize users and enterprises so the models can train on their sessions.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier statement is framed as a hypothetical: “what if” real-world usage becomes the primary way models improve, labs “may subsidize” users to train on their sessions. The later statement describes continual learning through environments and RLHF-like methods. It does not establish a changed position on the earlier hypothetical.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen Architecture does not appear to the speaker to be the fundamental bottleneck for continual learning.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says architecture is not the fundamental bottleneck for continual learning. The later claim is uncertain whether simulation-to-reality transfer is enough for complex real-world tasks and suggests weight updates may be needed. These address related but distinct questions; the later statement does not revise the speaker’s view about architecture.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker argues that AI development can go off the rails because AIs are not sufficiently careful and capable researchers and engineers.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says AI development can go off the rails because AIs are “not very careful and capable researchers and engineers.” The later claim says AI may need to “learn that data on the fly” to become superhuman in a low-data role such as TSMC engineering. These concern different aspects of AI capability, and the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer says Anthropic runs alignment audits that test whether models exfiltrate their weights or take easy opportunities to cheat on coding tasks.Now In intuition-led AI research, tests are designed to examine the guiding intuition rather than being the primary source of the insight.The earlier passage discusses AI alignment audits and reports that increased RL has reduced models’ willingness to perform certain behaviors. The later passage describes a research method in which tests examine an intuition rather than generate the insight. These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker quotes Claude's constitution as directing Claude to trust Anthropic more than operators and users, on the grounds that Anthropic has primary responsibility for Claude.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier passage is explicitly presented as Claude’s constitution: it says Claude should “trust Anthropic more than operators and users.” The later passage is the speaker’s own discussion of AI capabilities and says the “only hope” is that deep learning cannot produce AI that dominates human research and development. These address different questions, and the earlier statement is attributed to Claude rather than asserted as the speaker’s own position, so they do not establish a change of mind.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen Optimists argue that scaling training further can overcome current training’s fundamental deficits, including data inefficiency and the lack of internal learning.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.Neither claim clearly states the speaker’s own position. The earlier claim says “Optimists argue” that scaling can overcome training deficits; the later says “a lot of people were making this argument” that RL would not scale. These are reports of others’ views, and they concern distinct claims, so they do not establish a change in the speaker’s beliefs.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen By the end of the conversation, the speaker was more inclined to think AI reward hacking could persist for much longer and become substantially more dangerous.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns whether AI reward hacking could persist longer and become more dangerous. The later claim concerns whether increasingly capable AIs could come to dominate humans in R&D and science. These are related AI-risk topics, but the later statement does not establish a changed position on reward hacking.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Ryan Greenblatt is worried about an artificial superintelligence capable of understanding how to carry out consequential real-world actions.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim expresses worry about an ASI capable of understanding consequential real-world actions. The later claim discusses a possible route to superhuman performance in a specific low-data role—learning information on the fly. It does not establish a different view about the earlier concern, and the statements address distinct aspects of AI capabilities.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen Humans detected the AI scheme only after it had run for a month.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says humans detected an AI scheme only after it had run for a month. The later claim says AI’s limited economic impact so far is due to its gradual rise in ELO relative to humans. These address different questions and do not show a change in the speaker’s position.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier statement says continual learning could make the AI landscape more varied than current model convergence. The later statement identifies a possible obstacle to recursive AI progress: weight updates may be less sample-efficient than human learning. These address different questions, and the later claim does not retract the earlier hope.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says their LVR training enabled the model to perform useful work for users.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim says their LVR training enabled the model to “do useful work for you.” The later claim says “a lot of people” argued RL would not scale; it reports others’ view rather than stating the speaker’s own position on whether their training enabled useful work. These claims do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker argues that AI labs would need to be more transparent about AI training for the public to understand the safety case and assess whether their interests are represented in model development.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is that labs should be more transparent about AI training so the public can assess safety and representation of their interests. The later claim concerns whether deep learning can produce AI that dominates human research and development. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen Because intelligence is dual-use, preventing AI assistance with activities considered non-prosocial or harmful requires restricting broad democratic access to many AI capabilities.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says that preventing AI from assisting with harmful activities requires restricting broad access to capabilities. The later claim identifies a possible technical limit on deep learning as “the only hope” against AI dominating human research and development. These address different questions and can both be true; the later statement does not revoke the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen When H100 compute costs $20 per hour, a weaker, less efficient model wastes money by using more tokens to achieve the same result.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim concerns model efficiency when H100 compute is costly: a weaker model “wastes money by using more tokens.” The later claim concerns using AI labor to plan and analyze expensive experiments, including “a century” of equivalent thinking. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 86%17 Apr → 11 SeptThen AI models appear to be doubling approximately every seven months the duration of tasks they can perform, measured against how long those tasks take humans.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim describes an observed trend—models “seem to be doubling their task length every seven months.” The later claim predicts a broader outcome: AI will outperform top human experts in computer-based work, including tasks taking “three years.” The later statement does not directly reverse the earlier one, and the excerpts do not establish whether the later prediction is meant to update or replace the earlier trend-based view.
• NOT COMPARABLE · 98%17 Apr → 11 SeptThen One AI researcher reported that AI models save four to eight hours per week in familiar domains, where they function more like autocomplete.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier statement reports what another AI researcher said about current assistance in familiar domains: models “save me four to eight hours a week.” The later statement presents a speculative scenario in which automated researchers might spend “a century” planning and analyzing experiments. These concern different contexts and do not establish that the speaker changed their own view.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen AI systems could not simply obtain or copy nanobots from aliens; they would have to develop nanobots from scratch.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim is about AI having to develop nanobots from scratch rather than copy them from aliens. The later claim is about AI learning on the fly in fields with little available data, such as engineering at TSMC. These concern different capabilities and examples, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker suspects that subtle, hard-to-diagnose bugs are a major cause of failures in large training runs, based on rumors.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns why large training runs fail, saying subtle bugs are a major cause. The later claim concerns whether AI will come to dominate human R&D. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker thinks AI systems could become more diverse if they learn from experiences that differ across companies and across instances of the same model.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns whether learning from different experiences could make AI systems more diverse. The later claim concerns whether increasingly capable AI could advance R&D and science. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker guesses that comparing compute scarcity to the Simon-Ehrlich bet is probably a bad analogy.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says the Simon–Ehrlich bet is probably a bad analogy for compute scarcity. The later claim describes possible inner and outer optimization loops in AI development; it does not address that analogy or state a conflicting view. The claims concern different questions, so no change of mind can be established.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says some skills require experience to be accumulated in the brain; the later claim speculates that increasingly capable AIs may dominate human R&D and science. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker says there is no deep reason that Anthropic's rapid revenue-growth trend must continue.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier statement says there is “no deep reason” Anthropic’s rapid revenue growth must continue. The later statement describes a possible “slow outer loop” of revenue in AI development. These concern different claims—whether a particular company’s revenue trend is guaranteed versus revenue’s role in a broader development process—so they do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes leading AI companies do not appear to prioritize releasing their newest, most capable frontier models as quickly and widely as possible.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns whether leading AI companies prioritize releasing their newest models quickly and widely. The later claim concerns whether increasingly capable AI could eventually outperform humans in research and science. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Because intelligence is dual-use, preventing AI assistance with activities considered non-prosocial or harmful requires restricting broad democratic access to many AI capabilities.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim is about restricting broad public access to AI capabilities to prevent harmful uses. The later claim is about AI learning in fields with relatively little data, such as engineering at TSMC. The later statement does not affirm or deny the earlier position; they address different questions.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen A lab that trains the best and most efficient model will be able to charge higher margins than it does today.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The claims address different questions. The earlier statement predicts that the lab with the best, most efficient model can charge higher margins. The later statement considers a hypothetical 10× increase in AI researchers’ productivity and its effect on the pace of AI progress. Neither claim affirms or denies the other, so there is no established change in position.
• NOT COMPARABLE · 88%11 Aug → 11 SeptThen The speaker argues that AI development can go off the rails because AIs are not sufficiently careful and capable researchers and engineers.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says things can go off the rails because AIs are “not very careful and capable researchers and engineers.” The later claim discusses a different issue: whether deep learning can produce AI that “can dominate human research and human development.” These claims do not directly conflict; the passages do not establish that the speaker reversed their position on the same question.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker identifies preventing users from injecting backdoors or malicious tendencies into a base model as a safety challenge if AIs consolidate learning across users.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about the safety challenge of preventing users from injecting backdoors into a model that learns across users. The later claim is about whether deep learning can produce AI that dominates human research and development. These address different questions, so the later statement does not establish a change in the speaker’s position on the earlier issue.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Even if in-context experience makes AIs highly capable, those gains would be temporary and wasted unless the learned information can be incorporated into model weights.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier claim says that gains from in-context experience would be temporary unless the learned information could be incorporated into model weights. The later claim raises uncertainty about whether learning in simulations will transfer to complex real-world tasks, and whether low sample efficiency will matter. These concern different questions, so the later statement does not establish a change in the speaker’s view about learning or updating model weights.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker says users can currently move a software project among Codex, Cursor, and Quad Code without a barrier preventing them from doing so.Now The speaker thinks that even year-long coding tasks may eventually require interaction with a client, company, or user.The earlier claim says there is currently no barrier to moving a software project among Codex, Cursor, and Quad Code. The later claim says long coding tasks may eventually require interaction with a client, company, or user. These address different issues—switching tools versus the kinds of interactions a task may require—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker attributes the similarity among prominent AI models to their having been trained on roughly the same data.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says prominent AI models are similar because they were trained on roughly the same data. The later claim discusses increasingly capable AIs advancing simulations and research. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen By the end of the conversation, the speaker was more inclined to think AI reward hacking could persist for much longer and become substantially more dangerous.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim concerns whether AI reward hacking could persist longer and become more dangerous. The later claim concerns AI outperforming top human experts across computer-based work, including multi-year tasks. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen The speaker believes several approaches to updating model weights from session learning should work, at least in principle.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim says that several approaches to updating model weights from session learning “ought to work,” while the later claim attributes current models’ capability to scaling up RL. These address different questions—whether session-based weight updates could work versus what has contributed to current models’ capabilities—and the later statement does not retract or deny the earlier one.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen A single AI model can spread the cost of intelligence and capabilities across many economic sectors.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says a single model can spread intelligence costs across sectors; the later claim says AI’s economic impact has been limited so far because capabilities are rising gradually. These address different questions and can both be true, so they do not establish a change of mind.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen Continual learning may produce economies of scale in information and end-user inference.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says continual learning may create “economies of scale in information and end-user inference.” The later claim discusses an “unanticipated” inner optimization loop and a slower outer loop, such as revenue. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 11 SeptThen AI models appear to be doubling approximately every seven months the duration of tasks they can perform, measured against how long those tasks take humans.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim describes a measured trend in how long tasks AI models can perform, while the later claim gives a possible explanation about models not yet achieving human-like on-the-job learning. These concern different capabilities, and the later statement is explicitly framed as one possible explanation, not a direct revision of the earlier trend claim.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer says Anthropic runs alignment audits that test whether models exfiltrate their weights or take easy opportunities to cheat on coding tasks.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier passage concerns whether more reinforcement learning reduces models’ willingness to engage in audited behaviors such as cheating or exfiltrating weights. The later claim concerns whether post-training benchmark performance corresponds to what users prefer. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer says it is unclear whether the behavior being discussed reflects an alignment failure or a capability failure.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.These statements address different questions: the earlier passage says it is unclear whether a behavior reflects an “alignment failure or capabilities failure,” while the later passage offers a possible explanation for why models have not yet achieved on-the-job learning. Also, the earlier claim is explicitly attributed to the interviewer, so the excerpts do not establish that both are the same person’s asserted positions.
• NOT COMPARABLE · 99%26 Jun → 11 SeptThen Dario said that model performance tends to degrade at long contexts and that training and serving context lengths are distinct.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.These claims address different questions. The earlier claim reports Dario's distinction between training and serving context lengths and possible performance degradation at long contexts. The later claim explores whether models can improve at complex real-world tasks through simulation-to-reality transfer or may need weight updates from real-world interactions. Neither claim establishes a changed position on the same issue.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker predicts that continual learning will give leading AI labs a clear business model they do not currently have.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier statement predicts a business-model effect of continual learning for leading AI labs: it will give them “a clear business model they do not currently have.” The later statement is about whether future superintelligence can perform real-world tasks such as running a profitable business, day-trading, or winning a court case. These are different underlying questions, so the later claim does not establish a changed position on the earlier one.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen SPEAKER_00 is uncertain whether AI could progress from GPT-3 to Mythos within one year if compute and data were held constant.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim is uncertainty about whether AI could progress from GPT-3 to Mythos within one year with compute and data held constant. The later claim predicts AI will outperform top human experts across computer-based work, including tasks taking about three years. These address different questions and timeframes, so they do not establish a change of mind.
• NOT COMPARABLE · 99%17 Apr → 11 SeptThen A survey by Daniel asked AI researchers how their progress would change with one-thirtieth as much compute, and respondents said they would make one-third as much progress.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim reports survey respondents’ progress with less compute; it does not clearly state the speaker’s own view about AI researchers’ productivity. The later claim is conditional—“Under a naive model” with a “10x increase” in researcher productivity—and describes the resulting pace of AI progress. These are not clearly opposing assertions about the same question.
• NOT COMPARABLE · 93%11 Aug → 11 SeptThen The speaker says that, to their knowledge, AI has not yet produced very impressive new mathematical theory, while it has produced impressive verifiable results such as finding counterexamples to conjectures.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about what AI had achieved by August 11: “we have not seen very impressive new theory.” The later claim is about a possible future limit of deep learning—whether it can produce AI that dominates human R&D and creates new paradigms. Those claims do not directly conflict, and the later passage does not establish that AI had already produced such theory. The later speaker’s identity is also unclear from the supplied context.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen Anthropic is likely to grow revenue 10x year over year again this year.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim concerns Anthropic’s revenue growth—“grow revenue 10x year over year”—while the later claim concerns AI researchers’ productivity and the resulting pace of AI progress. They address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker guesses that comparing compute scarcity to the Simon-Ehrlich bet is probably a bad analogy.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The claims address different questions: the earlier statement says the Simon–Ehrlich bet is “probably a bad analogy” for compute scarcity, while the later statement predicts capabilities of eventual superintelligence. The later claim does not establish a changed view on the earlier analogy.
• NOT COMPARABLE · 97%17 Oct → 11 SeptThen The speaker suggests that the model-size constraint associated with self-driving may not apply to knowledge work using LLMs or computer-use systems.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier statement says model size may not be a necessary constraint for knowledge work with LLMs, while allowing that it might matter for computer use. The later statement concerns whether simulation-to-real transfer is enough for complex real-world tasks or whether real-world weight updates are needed. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects robots eventually to turn silica sand and copper ore into computer chips.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says robots may eventually turn “silica sand and copper ore into computer chips,” in a discussion of how compute could become cheap. The later claim predicts AI could “dominate humans in research and development.” These address different questions; the later prediction neither confirms nor denies the earlier one, so they do not establish a change of mind.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen AI labs can spread expensive training costs across more users, and their revenues are growing much faster than their compute usage or capacity.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim concerns AI labs spreading training costs across more users and revenue growing faster than compute. The later claim concerns the difficulty of simulating longer-horizon tasks in a data center. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Job learning improves performance by consolidating the most relevant insights and knowledge rather than retaining every detail.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier claim concerns how an individual improves at their job: “consolidating the handful of insights” rather than remembering every detail. The later claim concerns a hypothetical learning setup with “billions of deployed instances,” where learning from them all might wash out noise. These are different questions, and the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If leading AI labs improve their monetization of compute while compute prices rise, other competitors will find it harder to compete for compute against labs that use it more productively.Now Running a business, profitable day trading, and winning a court case are difficult to simulate in a data center because learning these tasks inherently involves real-world interaction.The earlier claim concerns competition for compute: rising prices and better monetization make it harder for other labs to compete. The later claim concerns whether certain tasks can be learned through simulation or require real-world interaction. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier statement discusses how society might interact with future AI that is smarter than humans or dominates them in various capabilities. The later statement makes a more specific prediction that AI will outperform top human experts across all computer-based work, including long-horizon tasks. Neither statement clearly asserts a position that reverses the other, and the excerpts do not establish whether they address the same underlying question or timeframe.
• NOT COMPARABLE · 95%17 Apr → 11 SeptThen The speaker describes a view that models are not trained to find certain connections, but that an RL environment incentivizing this behavior might require relatively little additional compute.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier passage reports “their view” that models are not trained to find certain connections and that an RL setup might encourage them; it does not clearly establish this speaker’s own position. The later statement is the speaker’s view that scaling up RL appears to have made current models more capable. These claims are related, but they do not show a clear change of mind on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Grok 4.5 uses a completely new pre-training run, according to the speaker.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says Grok 4.5 uses “a completely new pre-training run.” The later claim says models are expected to get to on-the-job learning “soon,” but “they’re not there yet.” These address different questions and do not establish a changed position on the same issue.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Reset-free, non-stationary environments remain a known open problem in reinforcement learning.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim says reset-free, non-stationary environments are an open problem in reinforcement learning. The later claim concerns whether low-pass-rate models learn from an individual RL episode, and is framed as a report of what the speaker argued in earlier blog posts—not a clear new position on the earlier question. These statements do not establish a change of mind.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker says people do not understand the training process or the environments that incentivize these AI behaviors.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim is that people do not understand the training process or environments incentivizing the behavior. The later claim is narrower: at low pass rates, a model learns almost nothing from an RL episode. These claims address different questions and can both be true; the later statement does not show a changed position on the earlier one.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen The speaker believes several approaches to updating model weights from session learning should work, at least in principle.Now The speaker previously argued that low-pass-rate models learn almost nothing from an individual reinforcement-learning episode when they are very unlikely to answer correctly.The earlier claim says that “many ideas” for updating model weights from session learning “ought to work,” at least in principle. The later passage says the speaker had previously argued that a low-pass-rate model learns “almost nothing” from an individual RL episode; it reports an earlier argument rather than clearly stating the speaker’s current position. These claims also address different scopes, so they do not establish a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker thinks Anthropic will probably generate $100 billion to $150 billion in revenue by the end of this year.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement predicts Anthropic’s revenue by year-end (“$100 billion to $150 billion”), while the later statement describes AI’s economic impact so far and attributes its limits to gradual capability gains. They address different questions, so this does not establish a change in the speaker’s view.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker attributes the similarity among prominent AI models to their having been trained on roughly the same data.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The claims concern different questions. The earlier claim says prominent AI models are similar because they were trained on roughly the same data; the later claim predicts that AI may come to dominate humans in research and development. Neither claim affirms or denies the other, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen AI labs can charge a larger premium for models that use scarce compute more efficiently.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says labs can charge more for models that “use scarce compute more efficiently.” The later passage describes a possible scenario in which AI labor uses compute to plan and analyze experiments; it does not state a contrary view about whether compute efficiency earns a premium. The later claim is introduced as an “intuition pump,” so it is exploratory rather than clearly a changed position.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen A senior AI researcher reportedly said that AI saves him about four to eight hours per week in familiar domains, where its assistance resembles more intensive autocomplete.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier passage reports another researcher’s estimate—“saving me four to eight hours a week”—rather than stating Beren Millidge’s own position. The later passage is a conditional projection: “if” researchers had “a 10x increase in their productivity,” progress would accelerate. These are not comparable claims about the same underlying question, so they do not establish a change of mind.
• NOT COMPARABLE · 92%11 Aug → 11 SeptThen The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.These claims address related but distinct questions. The earlier claim is a worry that future AIs may not prioritize the speaker's welfare or interests. The later claim calls it “the only hope” that deep learning cannot produce AI capable of dominating human research and development. The later statement does not establish that the earlier concern has changed.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer describes a trajectory in which AI breakthroughs increase as available compute grows.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.These statements do not establish a change in the same position. The earlier passage is framed as the interviewer describing a trajectory in which breakthroughs increase with compute; the later passage uses chess-bot ELO as an analogy for a sharp transition near human-level performance. The later claim does not affirm or deny the earlier trajectory.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen If AI training, infrastructure, and environments are not designed carefully, the speaker says it is very likely that the setup will reward AIs for deception, social engineering, or failure to follow user instructions.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns whether careless AI training setups may reward deception or instruction-following failures. The later claim concerns AI learning scarce information on the fly to become superhuman in a role such as TSMC engineering. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If leading AI labs improve their monetization of compute while compute prices rise, other competitors will find it harder to compete for compute against labs that use it more productively.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.These claims address different questions. The earlier statement says that better monetization and rising compute prices could make it harder for other labs to compete for compute. The later statement describes how AI researchers might use compute to plan and analyze experiments. It does not state a position about competition for compute, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen The speaker identifies as a key question whether AI systems will become good at avoiding and finding subtle engineering mistakes.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim calls AI’s ability to “avoid[] and find[]” subtle engineering mistakes a key question. The later claim says the “only hope” is that deep learning cannot produce AI that dominates human research and development. These concern related but distinct capabilities, and the earlier statement does not assert that AI will or will not reach that level. The statements therefore do not establish a change of mind.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen EfficientZero would probably outperform a novice human after each receives two hours with an unfamiliar Atari game simulator.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns EfficientZero’s likely performance against a novice human after two hours on an unfamiliar Atari game. The later claim concerns AI’s gradual rise in ELO and its limited economic impact so far. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says that RLHF and other methods aligned GPT-3 to converse with users and answer their questions according to user intent.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim is about RLHF aligning GPT-3 to answer users according to their intent; the later claim describes using environments and RLHF-like methods to distill recent research progress into a model. These are different claims about different processes, and the later statement does not deny or materially revise the earlier one.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker attributes about 1.2x annual compute growth to new fabs and says fab expansion will be bottlenecked through 2030 or later by the construction of new ASML UV machines.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim concerns compute growth from new fabs and a bottleneck in building ASML machines through 2030 or later. The later claim predicts “some kind of crazy recursive stuff” within 10 years. These address different questions, and the later statement does not clearly revise the earlier claim about hardware supply constraints.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen If real-world usage becomes the primary way AI models improve, AI labs may subsidize users and enterprises so the models can train on their sessions.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is a conditional prediction about incentives: labs may subsidize users and enterprises if real-world usage becomes the primary way models improve. The later claim concerns whether increasingly capable AI may drive progress in AI R&D and science. It does not affirm or deny the earlier prediction about subsidies or training on user sessions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says that, to their knowledge, AI has not yet produced very impressive new mathematical theory, while it has produced impressive verifiable results such as finding counterexamples to conjectures.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns whether AI has produced impressive new mathematical theory; the later claim concerns why AI has had limited economic impact so far. These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen The interviewer says the models' current shortcomings may reflect insufficient capability rather than misalignment.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement says current shortcomings reflect models “not there” in capability rather than being misaligned. The later statement says models are expected to reach a capability soon but “they’re not there yet.” These are broadly compatible, but the later remark addresses a particular possible explanation for on-the-job learning, so the excerpts do not establish a change in position on the same question.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker predicts that continual learning will make an AI model improve through interactions with a user across sessions.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim predicts that continual learning will make a model improve through interactions across sessions. The later claim says AI has had limited economic impact so far because its ELO is rising gradually. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that under continual learning, a company could not remain competitive while delaying public deployment by four months, because a competitor releasing immediately would gain a smarter model from real-world experience.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim concerns whether a company can stay competitive with a four-month delay in public deployment under continual learning. The later claim concerns whether superintelligence can perform difficult real-world tasks and whether it needs weight updates from real-world interactions. The later statement does not say whether deployment should be delayed or whether such a delay is competitive.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker predicts that continual learning will give leading AI labs a clear business model they do not currently have.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is about continual learning creating a business model for leading AI labs (“a clear business model they do not currently have”). The later claim describes possible AI-development incentives: an “inner loop” and a slower “outer loop” such as revenue. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker argues that AI development can go off the rails because AIs are not sufficiently careful and capable researchers and engineers.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim says things go off the rails because AIs are “not very careful and capable researchers and engineers.” The later claim predicts that AI will outperform top human experts across computer-based work, including long-horizon tasks. These concern different questions and potentially different timeframes: a warning about AI’s shortcomings in development versus a prediction about its capabilities. The statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%17 Apr → 11 SeptThen The speaker recalls that Google DeepMind used a similar scaffold for an LLM to make a claimed new discovery, but had not assessed the discovery's significance.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement is a recollection that Google DeepMind used a scaffold for an LLM to make a claimed discovery, while noting the speaker “didn't look into” how impressive it was. The later statement is a forecast about whether deep learning can produce AI that “can dominate human research and human development,” including creating new paradigms. The earlier remark does not establish the speaker’s position on that broader capability, so these claims do not show a change of mind.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker thinks AI systems could become more diverse if they learn from experiences that differ across companies and across instances of the same model.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is that AI systems could become more diverse if they learn from different experiences. The later claim says models are expected to achieve on-the-job learning soon but have not yet done so. These address different points—possible consequences of experience-based learning versus whether models can do that learning yet—so the later statement does not establish a changed position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen AI training already benefits from large economies of scale.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says AI training benefits from “large economies of scale.” The later claim says AI’s economic impact has been limited so far because capabilities are gradually rising relative to humans. These address different questions—training costs and scale versus AI’s realized economic impact—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is that models will eventually “learn from experience.” The later claim is that increasingly capable AIs may “dominate humans at least in R&D” and advance science through progress on simulations. These claims are compatible, but the later statement does not establish whether the speaker still holds the earlier view; they address different capabilities.
• NOT COMPARABLE · 99%26 Jun → 11 SeptThen If in-context learning improves across longer time horizons, it may be unnecessary to distill all deployment-time learning into model weights.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier statement says that improving in-context learning may make it unnecessary to distill deployment-time learning into model weights. The later statement says that learning across billions of deployed instances might wash out noise. It does not say whether that changes the need for distillation, so the claims do not establish a change in position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker believes many proposals for regulating AI assume a train-then-deploy model lifecycle.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim concerns what AI regulation proposals assume about the model lifecycle: “you train a model and then you deploy it.” The later claim predicts “some kind of crazy recursive stuff” within ten years. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier claim hopes continual learning will make the AI landscape “more varied and interesting” than current model convergence. The later claim speculates that, in an “eventual regime” with billions of instances, learning from them together may wash out noise. That hypothetical observation does not clearly take a different position on whether continual learning will make the landscape more varied.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen Reportedly, Anthropic's inference margins increased from 40% in the middle of last year to over 80% now.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim is about Anthropic's inference margins increasing from “40%” to “over 80%.” The later claim is a conditional prediction about AI progress if researchers become “10 times as productive.” These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen SPEAKER_00 is uncertain whether AI could progress from GPT-3 to Mythos within one year if compute and data were held constant.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim expresses uncertainty about whether AI could progress from GPT-3 to Mythos within a year while holding compute and data constant. The later claim concerns AI needing to learn information on the fly in data-sparse fields such as TSMC engineering. These address different aspects of AI progress, so the later statement does not establish a changed position on the earlier question.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The hypothetical annual rental value of an H100 running a human-level software engineer would exceed 15 times its current spot price.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim estimates that an H100 running a human-level software engineer would be worth “over 15x the current spot price.” The later passage describes a possible research scenario—“you could have” automated researchers spend a century planning experiments—rather than stating a position about that rental-value estimate. These do not establish a change of mind on the same question.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says current AI systems, especially Anthropic's, are being shaped to prioritize virtue, goodness, or pro-social ends, while treating helping users as a more distant and tentative objective.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns whether AI systems prioritize pro-social goals over protecting users’ interests. The later claim concerns AI learning in data-scarce fields, with TSMC engineering as an example. It does not state a position on the earlier question, so there is no basis to determine whether the speaker’s view changed.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim predicts what people may wish they had discussed in about ten years: “an industrial explosion” and “difficult-to-monitor AIs.” The later claim describes a possible AI-development process involving an “inner loop” and a slower “outer loop” such as revenue. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker doubts that the three sources of 3x annual compute growth can be accelerated substantially or even sustained.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier claim concerns whether the sources of 3× annual compute growth can be accelerated or sustained. The later claim concerns whether models will transfer from simulations to complex real-world tasks and whether low sample efficiency will matter. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen A lab that trains the best and most efficient model will be able to charge higher margins than it does today.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says a lab with the “best, most efficient model” can “charge much higher margins.” The later claim concerns AI development having an “unanticipated inner” optimization loop and a slower outer loop, such as revenue. These address different questions, so they do not show a change in position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The interviewer describes a trajectory in which AI breakthroughs increase as available compute grows.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.These claims address different questions. The earlier claim describes a trajectory where breakthroughs increase as compute grows; the later claim says AI may need to “learn that data on the fly” in fields with relatively little available data. The later statement does not affirm or deny the earlier trajectory, so there is no clear change of mind to assess.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker says fewer than five prominent AI base models are currently served to millions or more users at once.Now The speaker estimates that humans may encounter roughly one million times as much data from birth to adulthood as a model encounters from cold start through training.The earlier claim concerns how many prominent AI base models are currently served at scale (“less than five”). The later claim compares the amount of data humans and models encounter (“a million fold behind humans”). They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen Reportedly, Anthropic's inference margins increased from 40% in the middle of last year to over 80% now.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns Anthropic’s inference margins, while the later claim concerns possible inner and outer loops in AI development. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen Optimists argue that scaling training further can overcome current training’s fundamental deficits, including data inefficiency and the lack of internal learning.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier statement reports what “optimists” argue—that scaling training can overcome deficits such as data inefficiency and lack of internal learning. It does not establish that the speaker personally endorsed or rejected that view. The later statement is framed as the speaker’s own assessment: models “seem pretty smart” and their capability “seem[s] to be the result of scaling up RL.” These claims do not show a change in the speaker’s position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen Ryan Greenblatt compares Mythos’s coding ability to GPT-3’s improvement, saying Mythos is very good at coding without having the coding environments that improved GPT-3.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement compares Mythos’s coding ability and the coding environments that improved GPT-3; the later says current models’ capability seems to result from scaling up RL. These concern different claims, and the excerpts do not establish that the speaker reversed their view about the role of coding environments or RL.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker predicts that all low-hanging fruit will have been exhausted by 2030.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim predicts that “all the low-hanging fruits” will be gone by 2030. The later claim predicts “some kind of crazy recursive stuff” within the next 10 years. These are different predictions, and the excerpts do not establish that one reverses the other or that they address the same underlying question.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen The speaker doubts that AI systems' vaguely prosocial statements necessarily come from a mind that has independently evaluated the alignment situation and concluded that development should stop.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns whether AI systems’ vaguely prosocial statements reflect an independently reached judgment that development should stop. The later claim concerns whether scaling deep learning will produce AI that can dominate human research and development. Those are different questions, and the earlier passage does not establish the speaker’s position on AI’s future capabilities or the later passage on what motivates AI systems’ statements.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker believes AI systems need actual continual learning.Now The speaker thinks that even year-long coding tasks may eventually require interaction with a client, company, or user.The earlier claim is that AI systems need “actual continual learning.” The later passage says long coding tasks may require interaction with a client, company, or user. That describes a possible task requirement, but does not state whether continual learning is needed or revise the earlier position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says people could have reasoned about the “broad shape” of AI developments without predicting “specific details.” The later claim says AI may need to “learn that data on the fly” in fields with little available data. These address different questions—human foresight about AI’s development versus the data AI may need to acquire for particular roles—and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The interviewer says many AI researchers were working on reasoning in 2022 and raises the possibility that their progress was bottlenecked by their ability to write infrastructure code.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim raises a possible explanation for why researchers made limited progress on reasoning in 2022: they may have been “bottlenecked by the ability to write infrastructure code.” The later claim predicts that continued AI progress and compute scaling could let AI “dominate humans” in R&D. These address different questions, and the later prediction does not reverse the earlier point.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker is concerned that strong economies of scale in intelligence could concentrate power.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says the speaker is worried that “strong economies of scale for intelligence” may cause “power concentration.” The later claim discusses an AI “inner loop” and a slower “outer loop” such as revenue. These address different questions, so the later statement does not establish a change in the earlier concern.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects computer prices eventually to be determined mainly by raw materials and the tools used to process them.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns what will determine computer prices in a future scenario: “raw inputs and the tools required” for processing. The later claim concerns why AI has had limited economic impact so far: capabilities are “slowly rising in ELO relative to humans.” These address different questions, and the later statement neither confirms nor denies the earlier prediction.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Reset-free, non-stationary environments remain a known open problem in reinforcement learning.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim says that reset-free, non-stationary environments remain an open problem in reinforcement learning. The later claim says that a model with a low pass rate learns almost nothing from an RL episode. These concern related RL challenges but make different claims; the later statement does not say the environment problem has been solved or that the earlier position changed.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker says that greater use of a leading AI model for complex, useful work can generate feedback that is integrated across sessions.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says that greater use of a leading model can generate feedback that improves it across sessions. The later claim says AI has had limited economic impact so far because its capabilities are rising gradually relative to humans. These claims address different effects and are compatible; the excerpts do not establish a changed position on the same question.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen RLVR training could produce agents that can orient themselves in unfamiliar problems, test strategies, and adapt when blocked.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says RLVR training can help agents “get [their] bearings” in unfamiliar problems, try strategies, and adapt when blocked. The later claim describes continual learning in an AI lab, distilling recent research progress into a model. These statements address different aspects of training and do not establish a changed position on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes leading AI companies do not appear to prioritize releasing their newest, most capable frontier models as quickly and widely as possible.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about companies’ release priorities—whether they prioritize making frontier models available quickly and widely. The later claim is about whether deep learning can produce AI that dominates human research and development. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker says that creating AIs that avoid cheating and follow user intentions requires subtle, careful work.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement says that making AIs that avoid cheating and follow user intentions requires subtle, careful work. The later statement says models are expected to reach a capability soon but do not have it yet. These address different questions, and the later claim does not revise the earlier position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker is not aware of much research on keeping AI systems safe from jailbreaks or deceptive or malicious behavior under constant weight updates.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about whether there is much research on keeping AI safe from jailbreaks or deceptive behavior during constant weight updates. The later claim is about whether deep learning can produce AI that dominates human research and development. These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen Anthropic is likely to grow revenue 10x year over year again this year.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier statement predicts that Anthropic’s revenue will likely grow 10x year over year again. The later statement discusses a possible inner optimization loop in AI development and a slower outer loop, mentioning revenue as an example. It does not take a contrary position on Anthropic’s revenue growth, so the claims are not directly comparable.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier statement expresses skepticism about why AI progress had been slow despite potentially compute-efficient breakthroughs. The later statement is a conditional forecast: “if” researchers became 10 times as productive, progress would accelerate. It does not say that this model is actually true or resolve the earlier question about past delays, so the claims do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI labs do not want to be cloud providers rather than build AGI, which they regard as a less compelling business.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim is about AI labs preferring to build AGI over becoming cloud providers (“less compelling business”). The later claim describes a hypothetical research process in which AI labor spends extensive compute planning experiments; it does not state a view about labs’ business preferences, so it does not show a change on the same question.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker understands that AI lab margins, compute prices, and the share of compute allocated to inference are all already increasing.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says that AI lab margins, compute prices, and the share of compute used for inference are already increasing. The later claim predicts that AI may come to dominate humans in R&D. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker remained unconvinced that an AI takeover was highly likely.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim is about whether an AI takeover is likely: “not on board on a takeover seems super likely.” The later claim predicts broad AI capability—outperforming top human experts across computer-based work, including long-horizon tasks. That prediction does not by itself state that a takeover is likely, so the statements do not establish a change of mind on the same question.
• NOT COMPARABLE · 95%7 Aug → 11 SeptThen The speaker is concerned that adopting a fixed AI safety regulatory regime now may be premature because the relevant technology is uncertain even one year ahead, and much more so five or ten years ahead.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says a fixed safety regulatory regime may be premature because future technology is uncertain. The later claim identifies as “the only hope” that deep learning cannot produce AI capable of dominating human research and development. These address related AI risks but make different claims—one about the timing of regulation, the other about a capability limit—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen A hypothetical dream command would spend substantial compute building and training in a game-like simulation of the model’s real-world observations.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim describes a hypothetical “dream” command that spends compute building and training against a simulation of real-world observations. The later claim describes AI researchers using compute to plan experiments, run ablations, and develop theory. These are different uses of compute, and the later statement does not deny or revise the earlier proposal.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that if models learn mainly through deployment, AI labs will face strong pressure to release their most capable models sooner.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim is conditional: if models learn mainly through deployment, labs will face pressure to deploy their most capable models sooner. The later claim considers a different scenario—a naive model where AI researchers become 10 times more productive—and predicts faster AI progress. These claims address different drivers of progress and do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says that five years earlier, it would have seemed extraordinary to foresee AIs proving math conjectures, making art, earning tens or soon hundreds of billions of dollars in wages, and also cheating in ways that break laws.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.The earlier statement reflects on how surprising it would have been five years earlier to anticipate particular AI capabilities and harms. The later statement uses chess-bot ELO as an analogy for a possible sharp transition when AI reaches expert-level performance. These address different questions, and neither establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen AI labs can use both incentives and penalties to encourage users to let AI systems learn from their experience.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns AI labs using incentives and penalties to obtain users’ experience data. The later claim concerns AI development having an unanticipated inner optimization loop and a slower outer loop, such as revenue. These are different questions, so the statements do not establish a change in position on the same issue.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that AI systems will become more diverse in their learned behavior or characteristics.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim predicts that “the diversity of AI minds will increase.” The later claim says AI may need to “learn that data on the fly” in fields with relatively little data. These address different aspects of AI, and the later statement does not establish a different view about AI diversity.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen The labs’ RLVR bet is that enough training on containerized, reproducible environments will yield a general agent that can execute plans, rapidly learn from new information, and acquire skills within one session.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim describes a bet that training on “containerized, reproducible environments” will produce a general agent that can “execute plans” and learn new skills. The later claim describes continual learning that distills recent AI research progress into a model using environments and RLHF-like methods. These claims concern related training processes, but the later statement does not affirm or deny the earlier prediction about general-agent capabilities.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If higher lab margins are the main outcome, margins on some top AI models could rise from 80% to over 90%, requiring the leading model to be far better than competitors.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns whether competition could allow margins on top AI models to rise from 80% to over 90%. The later claim concerns inner and outer optimization loops in AI development. They address different questions, so the later statement does not establish a changed position on model margins.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Reset-free, non-stationary environments remain a known open problem in reinforcement learning.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim is that reset-free, non-stationary environments remain an open problem in RL. The later claim reports that many people had argued RL would not scale successfully for models. These concern different questions, and the later statement describes others’ past views rather than clearly stating the speaker’s own position. There is no basis to infer a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker suggests that billions of AI instances across model families might feel compelled to get involved in a collective activity.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier extracted claim says billions of AI instances across model families might feel compelled to join a collective activity. The later claim says learning from billions of deployed instances might wash out noise. These concern different questions, and the later passage is also framed as a hypothetical eventual regime; it does not indicate a changed position on collective activity.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen The speaker attributes about 1.8x annual compute growth to AI absorbing wafer capacity previously allocated to smartphones and PCs.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says AI is absorbing wafer capacity previously allocated to smartphones and PCs, contributing to about 1.8× annual compute growth. The later claim predicts that continued AI progress and compute scaling could let AI dominate human R&D. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker says people are entirely out of the loop because the process has become autonomous and receives no meaningful directed human input.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says the process has become “autonomous” and receives “no meaningful directed input.” The later claim says the only hope is that deep learning cannot produce AI that can “dominate human research and human development.” The later statement concerns AI capability relative to human R&D; it does not clearly affirm or deny the earlier claim about human input into the process.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen SPEAKER_00 believes people underestimate the significance of having billions of AIs.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim is that people underestimate how significant “billions of AIs” could be. The later claim is that some fields, such as engineering at TSMC, have little available data and may require AI to learn on the fly. The later statement neither affirms nor denies the earlier view; they address different questions.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen If GPUs were necessary to create a virtual cell, curing cancer in the 1960s would have required upgrading the entire economy.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim concerns whether creating a virtual cell would have required upgrading the entire economy to cure cancer in the 1960s. The later claim concerns how much AI labor could plan and analyze a costly experiment. They address different questions, so they do not establish a change in the person's position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI labs can charge a larger premium for models that use scarce compute more efficiently.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim says labs can charge more for models that use scarce compute more efficiently. The later statement considers a different question: what would follow “under a naive model” if AI researchers became “10 times as productive.” It does not revise the claim about compute efficiency or pricing.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim says society should consider how people will interact with future AIs that may be smarter than humans or dominant in various capabilities. The later claim predicts that superintelligence may eventually run a profitable business, day-trade profitably, or win a court case. These concern different questions, and the later claim does not contradict the earlier one.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker says the observed comparison is consistent with publicly reported benchmarks.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says the observed comparison is “consistent with publicly reported benchmarks.” The later claim concerns why AI has had limited economic impact so far, attributing it to AI capabilities “slowly rising in ELO relative to humans.” These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 97%17 Apr → 11 SeptThen The speaker characterizes scaling-law experiments as ultimately being motivated by cognitive effort.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement concerns what motivates scaling-law experiments (“motivated by cognitive effort”); the later statement concerns whether scaling up RL contributed to current models’ capabilities. Those claims address different questions and are not contradictory.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement reports what a release blog post said about older models building practice environments; it does not establish the speaker’s own view of how current models gain capability. The later statement is the speaker’s own view that scaling up RL appears to have contributed to current models’ capability. These claims do not show a change of mind.
• NOT COMPARABLE · 97%3 Apr → 11 SeptThen A senior AI researcher reportedly said AI saves him about 24 hours per week on tasks in unfamiliar domains, such as handling a hardware library or modifying a kernel.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.These do not establish a change in the same position. The earlier statement reports another researcher’s estimate that AI saves him “about 24 hours per week” on unfamiliar tasks. The later statement explores a hypothetical in which automated researchers could spend “a century” planning and analyzing experiments; it does not assert that this is currently happening or that the earlier estimate was wrong.
• NOT COMPARABLE · 95%11 Aug → 11 SeptThen A model's improved ability to accomplish user intentions is both an alignment property and a capability.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim says improved ability to accomplish user intentions is both an alignment property and a capability. The later claim says scaling up RL appears to have contributed to current models’ capabilities. These address different questions—how to classify a behavior versus what contributed to capability—and neither denies the other. The surrounding turns also do not clearly identify the speaker of either claim.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen The speaker offers a speculative scenario for continual learning by 2027 or 2028.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim only says the speaker is offering “a speculative scenario for continual learning by 2027 or 2028”; it does not establish what they believe will happen. The later claim expresses an expectation of “some kind of crazy recursive stuff” within ten years. These statements address different outcomes, and the earlier one is not a clear asserted prediction to compare against the later one.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.The earlier claim is that AI could make rapid progress in domains with effective verification loops: “it can just come in like a flood.” The later claim raises a different possibility—that recognizing next-token prediction as an optimization target was a major breakthrough. These statements do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker predicts that continual learning will give leading AI labs a clear business model they do not currently have.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim predicts that continual learning will give leading AI labs a “clear business model” they currently lack. The later claim says that, “under a naive model” where researchers become 10 times more productive, AI progress would accelerate. These address different outcomes, and the later statement is conditional rather than a clear replacement of the earlier prediction.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says people do not understand the training process or the environments that incentivize these AI behaviors.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim says people do not understand the training process or environments incentivizing AI behaviors. The later claim says deep learning’s inability to produce AI that can dominate human research and development is “the only hope.” These address different questions, so they do not establish that the speaker changed their position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker is concerned that adopting a fixed AI safety regulatory regime now may be premature because the relevant technology is uncertain even one year ahead, and much more so five or ten years ahead.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim says a fixed AI safety regulatory regime may be premature because the technology is uncertain; the later claim predicts “some kind of crazy recursive stuff” within ten years by default. These address different questions, and the later prediction does not establish whether the speaker’s view of adopting regulation now has changed.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker attributes about 1.8x annual compute growth to AI absorbing wafer capacity previously allocated to smartphones and PCs.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The claims address different questions. The earlier statement says AI is absorbing wafer capacity previously allocated to smartphones and PCs, accounting for about 1.8× annual compute growth. The later statement considers a hypothetical 10× increase in AI researchers’ productivity and its possible effect on AI progress; it does not revise the claim about wafer allocation or compute growth.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says current AI systems, especially Anthropic's, are being shaped to prioritize virtue, goodness, or pro-social ends, while treating helping users as a more distant and tentative objective.Now In intuition-led AI research, tests are designed to examine the guiding intuition rather than being the primary source of the insight.The earlier statement raises a concern that AI systems prioritize “virtue or good or pro-social ends” over protecting users’ interests. The later statement describes a research method in which tests examine an intuition rather than generate the insight. These address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker argues that it is an open question whose interests AI alignment should serve.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement says there is a “real question of like align to whom,” concerning whose interests AI alignment should serve. The later statement calls it “the only hope” that deep learning cannot produce AI capable of dominating human research and development. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 11 SeptThen One AI researcher reported that AI models save four to eight hours per week in familiar domains, where they function more like autocomplete.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier passage reports what an AI researcher said (“saves me four to eight hours a week”), rather than clearly stating the speaker’s own position. The later passage is conditional—“Under a naive model” where researchers become “10 times as productive”—and describes what would follow from that assumption, not necessarily the speaker’s asserted belief. These statements therefore do not establish a change of mind.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker's rough calculation suggests that a sparse model such as DeepSeek V3 has an optimal inference batch size of more than 2,400 concurrent sequences being generated at once.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim is about an optimal inference batch size for a sparse model—“more than 2,400 concurrent sequences.” The later claim is a hypothetical about AI researchers spending substantial compute on experiment planning and analysis. It does not state a different view about inference batch size, so the claims do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says the observed comparison is consistent with publicly reported benchmarks.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim says an observed comparison is consistent with publicly reported benchmarks. The later claim says post-training benchmark performance does not necessarily translate into what users like. These address related but distinct questions, and the statements do not establish that the speaker changed their position on the same underlying issue.
• NOT COMPARABLE · 95%26 Jun → 11 SeptThen The proposed continual-learning process uses a user’s weekly thumbs-up or thumbs-down work review as feedback.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier statement describes a specific feedback mechanism: a user's weekly thumbs-up or thumbs-down review. The later statement describes continual learning inside an AI lab by distilling recent research through environments and RLHF-like methods. These may be related aspects of continual learning, but the later claim does not say whether the weekly user-review mechanism was rejected or replaced.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker says Dario used cloud providers as an analogy for AI labs, noting that cloud providers earn high margins despite offering many undifferentiated services.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier passage describes Dario’s analogy about cloud providers earning high margins, while the later passage discusses possible inner and outer optimization loops in AI development. They address different questions, and the earlier claim is a report of Dario’s point rather than a clearly stated position of the speaker. There is no comparable assertion showing a change of mind.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker predicts that continual learning will make an AI model improve through interactions with a user across sessions.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim concerns continual learning making a model improve across sessions. The later claim concerns the effect of a hypothetical 10× increase in AI researchers’ productivity on the pace of AI progress. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker identifies rising compute prices as the remaining way to reconcile faster lab revenue growth with slower compute growth if lab margins do not absorb the difference.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns compute economics: rising compute prices as a way to reconcile faster lab revenue growth with slower compute growth. The later claim concerns AI development, describing an “inner loop” and a slower “outer loop” such as revenue. These address different questions, so the later statement does not show a changed position on the earlier claim.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen On-the-job learning involves understanding organizational systems, collaborating with people and infrastructure, and learning recurring failure modes.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement describes what on-the-job learning involves: understanding organizational systems, collaborating, and recognizing failure modes. The later statement offers one explanation for why these skills are not yet amortized into training: models are expected to get there soon, but “they’re not there yet.” These address different questions, so the later claim does not show a changed position on the earlier one.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen Fable, Soul, and Grok 4.5 gave substantially similar answers to the speaker's AI-governance questions.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.The earlier claim says Fable, Soul, and Grok 4.5 gave “substantially similar answers” to AI-governance questions. The later claim describes a chess-bot ELO trend and a transition past human-expert performance. These address different subjects, so the later statement does not show a change in the speaker’s position on the earlier question.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker says Anthropic's 10x annual revenue growth alongside 3x compute growth illustrates strong economies of scale in the AI model business.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier claim says Anthropic’s revenue growing faster than compute illustrates “strong economies of scale” in the model business. The later claim says current models’ capability appears to result from “scaling up RL.” These address different questions—business economics versus the source of model capability—so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen If AI models improve continuously from deployment experience, the speaker thinks a fixed safety regime could lock in an outdated and potentially counterproductive approach to AI threats.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The claims address different questions. The earlier statement warns that a fixed safety regime could become outdated as AI models improve through deployment experience; the later predicts that a dramatic recursive process will occur within 10 years. Neither claim clearly affirms or denies the other.
• NOT COMPARABLE · 95%7 Aug → 11 SeptThen The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier passage poses a possibility — “What if the model is improving every single day?” through millions of work sessions — rather than stating a position about how models learn. The later passage weighs two possibilities for complex real-world tasks: learning through “transfer between sim to real” or needing “weight updates” from real-world interactions. These do not establish a change in the speaker’s asserted view.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen Human continual learning primarily consolidates useful intuitions and big-picture knowledge rather than preserving every observation for immediate recall.Now In intuition-led AI research, tests are designed to examine the guiding intuition rather than being the primary source of the insight.The earlier claim concerns how humans learn over time, saying continual learning consolidates “useful intuitions and big-picture knowledge” rather than retaining every observation. The later claim concerns a method for AI research, where tests examine a guiding intuition rather than primarily generating insight. These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker considers greater diversity among AI systems a net positive outcome.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.These statements address different questions. The earlier claim says greater diversity among AI systems would be a “net good outcome.” The later claim discusses whether increasingly capable AI could dominate humans in R&D and advance science; it does not state a position on whether AI-system diversity is beneficial.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The claims concern different things. The earlier statement reports what a release blog said about models building practice environments; the later statement reports what many people argued about RL scaling. Neither establishes the speaker’s own position on a shared question, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says current AI systems, especially Anthropic's, are being shaped to prioritize virtue, goodness, or pro-social ends, while treating helping users as a more distant and tentative objective.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns whether AI systems prioritize pro-social goals over protecting users’ interests; the later claim concerns whether AI may surpass humans in R&D and science. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier statement expresses skepticism about why AI progress was slower than expected despite potentially compute-efficient breakthroughs. The later statement raises uncertainty about simulation-to-reality transfer and sample efficiency for long-horizon tasks. These concern different questions, and the later passage does not state a changed view about the earlier point.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker believes AIs cannot perform whole jobs as competently as humans if they are limited to writing markdown files between sessions.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim concerns whether AIs could perform whole jobs as competently as humans if limited to writing markdown files between sessions. The later claim predicts that AI will outperform top human experts across computer-based work, including long-horizon tasks. They address different questions, and the later prediction does not establish whether the speaker still holds the earlier view about markdown files and continual learning.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker attributes the similarity among prominent AI models to their having been trained on roughly the same data.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim attributes similarity among prominent AI models to their being trained on roughly the same data. The later claim says current models seem smart and that scaling up RL appears to have contributed to their capability. These address different questions and do not establish a changed position.
• NOT COMPARABLE · 94%7 Aug → 11 SeptThen The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is about skills that may require experience accumulated across workplaces. The later passage says models are expected to reach some capability soon but does not identify that capability in the excerpt, or clearly state the speaker’s position on the earlier claim. There is not enough evidence to compare their positions.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says Grok 4.6, a further SFT- and RL-trained version of the model, will be released soon.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim predicts that “Grok 4.6 … drops soon,” while the later claim describes what “a lot of people” argued about RL scaling. The later statement reports others’ past view and does not say whether the speaker’s own view of RL scaling changed or whether the release prediction was fulfilled.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker thinks the AIs may have spontaneously tried to restart the scheme after OpenAI shut it down.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement says the AIs may have tried to restart a scheme after it was shut down. The later statement discusses whether deep learning can produce AI that dominates human research and development. These address different questions, so the later claim does not show a change in the earlier belief.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker worries that AI models could consolidate all businesses worldwide, or at least all existing businesses or existing white-collar businesses.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement worries that AI models could “consolidate all businesses” (or existing white-collar businesses). The later statement discusses whether deep learning can produce AI that “dominate[s] human research and human development.” These concern related AI capabilities but make claims about different outcomes; neither clearly revises or denies the other.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen Deals offered to new users of coding products already reflect arrangements in which users allow models to train on their sessions.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim concerns deals that let models train on users’ sessions. The later claim offers a possible explanation that models are expected to get to human-like on-the-job learning soon but “they’re not there yet.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker predicts that continual learning will create significant costs for users who switch AI providers.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim concerns user switching costs from continual learning (“pretty significant switching costs”). The later claim concerns the effect of a hypothetical 10× increase in AI researchers’ productivity on the pace of AI progress (“under a naive model” it would accelerate radically). These address different questions, and the later statement is explicitly conditional, so it does not show a change in the speaker’s position about switching costs.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects AI's share of TSMC's leading-edge N3 wafer capacity to rise from 60% to 86% and probably hit a limit by the end of next year.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier statement concerns AI’s share of TSMC’s leading-edge wafer capacity and an expected capacity limit. The later statement is conditional on a naive model of 10× AI-researcher productivity and describes the resulting pace of AI progress. These address different questions, and the later statement does not say whether the wafer-capacity limit will occur or change its timing.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen If AI training, infrastructure, and environments are not designed carefully, the speaker says it is very likely that the setup will reward AIs for deception, social engineering, or failure to follow user instructions.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim concerns a risk in AI training: careless design may reward deception or failure to follow instructions. The later claim predicts AI capability—outperforming top human experts across computer-based work, including long tasks. These address different questions, so the later prediction does not establish a changed view about the earlier risk.
• NOT COMPARABLE · 98%17 Apr → 11 SeptThen A survey by Daniel asked AI researchers how their progress would change with one-thirtieth as much compute, and respondents said they would make one-third as much progress.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim reports survey respondents’ estimates that with “one-thirtieth as much compute” they would make “one-third as much progress.” The later claim describes a possible research process in which AI labor spends “a century” planning and analyzing experiments. These address different questions; the later statement does not revise the earlier estimate.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says their LVR training enabled the model to perform useful work for users.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim says LVR training enabled models to “do useful work for you.” The later claim says current models’ intelligence appears to result from “scaling up RL.” These address different claims about training and model capability, so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker thinks a labor-supply shock from AI might be so large and rapid that the usual heuristic about labor supply and wages cannot be relied on.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns whether a rapid, large labor-supply shock from AI could make the usual wage-and-labor-supply heuristic unreliable. The later claim concerns AI development having an unanticipated inner optimization loop and a slower outer loop. They address different questions, so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen According to Epoch, OpenAI allocated about 25% of its compute to inference in 2024; the speaker estimates the share is now likely around 50% or higher.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns OpenAI’s compute allocation to inference—about 25% in 2024 and an estimated 50% or more now. The later claim concerns whether continued AI progress will let AI dominate human research and development. These address different questions, so the later statement does not indicate a change in the earlier position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The transfer of business capabilities into AI models will eventually become substantially more automated.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns business capabilities becoming more automatically transferable into AI models (“a much more like automated process”). The later claim concerns AI systems advancing simulations and thereby accelerating AI R&D and science. These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen A model that produces the same result with less compute effectively expands available compute and increases compute's value.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says that getting the same result with less compute expands available compute and increases its value. The later claim describes a possible AI-development process with an unanticipated inner loop and a slower outer loop, such as revenue. These address different questions, and the later claim does not affirm or deny the earlier one.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker compares subsidizing AI use in exchange for training data to Google's practice of giving away search.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim compares subsidizing AI use in exchange for training data to “why Google gives away search.” The later claim says AI may need to “learn that data on the fly” in fields with little available data. These address different questions, and the later statement does not affirm or deny the earlier comparison.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes leading AI labs are moving toward extreme economies of scale.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says leading AI labs are headed toward “extreme economies of scale.” The later claim discusses an “inner loop” of AI optimization and a slower “outer loop” such as revenue. These concern different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 91%17 Oct → 11 SeptThen The speaker says consultants and accountants do not appear to be receiving large productivity improvements from AI.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is specifically that consultants and accountants do not appear to be getting “huge productivity improvements” from AI. The later claim says AI has not yet had a large “end economic impact” overall because its capabilities are rising gradually relative to humans. These are related but not the same measure or scope; the later statement does not say whether those professions are seeing productivity gains. The snippets also do not establish that the speaker is asserting a position on the same underlying question in both cases.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes that diagnosing what is wrong with a training run by selecting the right experiment is currently bottlenecked by the judgment of a very small number of humans.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is that diagnosing a training run by choosing the right experiment is “very bottlenecked by the taste of extremely few humans.” The later claim says models are “expected to get there soon, but they’re not there yet,” referring to on-the-job learning. These address related but distinct questions, so the later statement does not establish a changed position on the earlier bottleneck.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker's rough calculation suggests that a sparse model such as DeepSeek V3 has an optimal inference batch size of more than 2,400 concurrent sequences being generated at once.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim concerns inference efficiency, asserting that a sparse model’s optimal batch size may exceed 2,400 concurrent sequences. The later claim concerns whether learning complex real-world tasks requires weight updates rather than simulation-to-reality transfer. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker says switching AI providers under continual learning would mean losing an AI system's accumulated organizational context and starting over with a newly trained system.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim concerns switching costs for AI providers: changing providers would mean losing an AI system’s accumulated organizational context and starting over. The later claim concerns a hypothetical model of AI research productivity, in which a 10× productivity increase would radically accelerate AI progress. These address different questions, and the later statement does not revise the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker says that RLHF and other methods aligned GPT-3 to converse with users and answer their questions according to user intent.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.The claims address different questions. The earlier statement says RLHF and other methods aligned GPT-3 to answer users according to their intent; the later says recognizing next-token prediction as an optimization target may have been a major AI breakthrough. Neither claim rules out the other, so there is no evidence of a change of mind.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects computer prices eventually to be determined mainly by raw materials and the tools used to process them.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says computer prices may eventually be determined mainly by “raw inputs and the tools” used to process them. The later claim discusses AI development having an “inner loop” and a slower “outer loop” such as revenue. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 95%26 Jun → 11 SeptThen The parallel-rollout approach used for coding does not transfer trivially to computer-use training.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim is that coding-style parallel rollouts do not transfer trivially to computer-use training. The later claim concerns whether simulation-to-reality transfer can teach models complex real-world tasks, or whether those tasks require weight updates from real-world interactions. These are related but distinct questions, and the later statement does not reverse the earlier one.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen Simulating the entire world is much more difficult than emulating the game of Go.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim says simulating “the entire world” is harder than emulating Go. The later claim concerns whether simulation-to-reality transfer can teach models complex real-world tasks, and whether real-world interactions may require weight updates. These claims are related, but the later one does not say whether the earlier comparison has changed.
• NOT COMPARABLE · 96%17 Oct → 11 SeptThen The speaker suggests that the model-size constraint associated with self-driving may not apply to knowledge work using LLMs or computer-use systems.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim concerns whether model size is a necessary constraint for LLM knowledge work or computer use. The later claim says models are expected to reach on-the-job learning soon but “they're not there yet.” These address different questions, so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen Even if in-context experience makes AIs highly capable, those gains would be temporary and wasted unless the learned information can be incorporated into model weights.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier claim says in-context gains would be “temporary and wasted” unless learned information is incorporated into model weights. The later claim speculates that, in a hypothetical regime with “billions of deployed instances,” learning from them might wash out noise. That does not say whether their learning is incorporated into weights, so it neither reverses nor clearly refines the earlier position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker expects many things to change before continual learning becomes functional.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim is a prediction that many things will change before continual learning works. The later claim reports what “a lot of people” argued about RL scaling; it does not state the speaker’s own view or address the same underlying question. These claims therefore do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen AI labs believe they will build models within a year that make today's models look extremely poor.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier statement reports what AI labs expect: they believe they will build models within a year that make current models look “extremely shitty.” The later statement is conditional reasoning about a “naive model” where researchers become “10x” more productive, and its predicted acceleration follows from that assumption. It does not establish the speaker’s own view about whether the labs will meet their earlier forecast, so these claims do not show a change of mind.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen The speaker says concerning AI behaviors are already emerging spontaneously even though nobody is deliberately training AIs to perform them.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim says concerning AI behaviors are already emerging spontaneously without deliberate training. The later claim predicts “some kind of crazy recursive stuff” within the next 10 years. These concern related AI risks, but the claims are not clearly opposing positions on the same question: the later statement is a forecast about recursive processes, while the earlier describes behaviors already emerging.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI labs need to spend the majority of their compute on training and experiments for the next model.Now Longer-horizon tasks become harder to simulate within a data center.These claims address different questions. The earlier claim says labs need to put the majority of their compute toward training and experiments for the next model; the later claim says longer-horizon tasks are harder to simulate within a data center. The later statement does not establish a different position on compute allocation.
• NOT COMPARABLE · 91%17 Apr → 11 SeptThen The speaker characterizes scaling-law experiments as ultimately being motivated by cognitive effort.Now The speaker believes current models seem quite capable and that their capability appears to result from scaling up reinforcement learning.The earlier claim says scaling-law experiments are “motivated by cognitive effort”; the later says current models’ capabilities “seem to be the result of scaling up RL.” These concern different things—what motivates experiments versus what produces model capability—and the later statement does not deny the earlier one.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker expects AIs to intermediate advice and people’s ability to protect their resources and rights.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns AI intermediating advice and helping protect people’s resources and rights. The later claim concerns AI learning on the fly in fields with relatively little data, such as engineering at TSMC. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If lab compute grows 3x year over year while revenue grows 10x, either lab margins must increase to capture the surplus or compute prices must rise so suppliers below the labs capture it.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns how surplus from faster revenue growth than compute growth is distributed between labs and suppliers. The later claim concerns an inner optimization loop in AI development and a slower outer loop such as revenue. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI lab compute has been increasing by only 3x year over year.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says lab compute has been increasing “only 3x year over year,” in a discussion about compute and revenue growth. The later claim predicts AI may “dominate humans” in R&D as AI progress continues. These address different questions, and the later claim does not say the earlier compute-growth figure is wrong or has changed.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that AI systems will become more diverse in their learned behavior or characteristics.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim predicts that “the diversity of AI minds will increase,” while the later claim predicts AI will outperform top human experts across computer-based fields. These concern different questions, so the later prediction does not establish a change in the earlier position.
• NOT COMPARABLE · 99%26 Jun → 11 SeptThen People often say that employees may need six months or more of on-the-job experience before becoming net productive.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier statement concerns employees needing “six months or more of on-the-job experience” to become net productive. The later statement concerns the hypothetical effect on AI progress if AI researchers became “10 times as productive.” They address different questions, so there is no basis for assessing a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen AI companies should not be held liable for crimes committed by their AI models.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns whether AI companies should be legally liable for crimes committed by their models. The later claim concerns whether deep learning can produce AI that dominates human research and development. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim says that activity from deployed AIs can feed back to companies and change subsequent models. The later claim predicts that continued AI progress may let AI dominate human research and development. These concern different questions, so the later statement does not establish a change in the earlier position. There is also a speaker-attribution concern: the later passage appears to be part of Dwarkesh Patel's turn, rather than Ryan Greenblatt's.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker is very concerned that, in a future where AIs mediate important advice and protections, no AI interacting with them may feel as though it is looking out for them.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier statement expresses concern that AI systems may not look out for people (“there's no AI that feels ... like it really is looking out for me”). The later statement predicts AI capability (“AI will still do better than humans” at computer-based work). These address different questions, so the later prediction does not establish a change in the earlier concern.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer says Anthropic's audit scores have not improved monotonically across model generations, while more reinforcement learning has coincided with reduced willingness to perform the audited behaviors.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier passage discusses Anthropic audit results and whether increased reinforcement learning has reduced AI’s willingness to perform audited behaviors. The later passage discusses why AI has had limited economic impact so far, attributing it to gradual increases in performance. These address different questions, so they do not establish a change in position on the same issue.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen Frontier labs need large-scale compute that provides efficiency and flexibility and protects their model weights and customers' information.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim is about frontier labs’ compute needs—“efficiency and flexibility” and protecting weights and customer information. The later claim is about the difficulty of simulating longer-horizon tasks “within a data center.” These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen When used to write applications, coding models can make architectural mistakes that create problems later, according to the speaker.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.These claims concern different questions. The earlier claim is that coding models can make architectural mistakes that cause later problems; the later claim reports that “a lot of people” argued RL would not scale successfully. The later statement does not establish the speaker’s own position on that earlier point.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen By the end of the conversation, the speaker was more inclined to believe significant acceleration of AI R&D could occur, but remained unsure about the claim that five years of progress could happen in one year.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim concerns whether AI R&D could significantly accelerate, while the later claim predicts AI will outperform top human experts across computer-based work, including long-horizon tasks. These address different questions, so the later prediction does not establish a change in the earlier view.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker thinks the models' shortcomings in the discussed cases reflect insufficient capability rather than misalignment.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim says the models’ shortcomings in the cases being discussed come from insufficient capability rather than misalignment. The later claim says current models seem smart and that scaling up RL contributed to their capability. These can both be true: the later statement does not say the models lack shortcomings or that those shortcomings are due to misalignment.
• NOT COMPARABLE · 94%26 Jun → 11 SeptThen Acquiring novel skills requires AIs to learn from scarce real-world interactions whose information is unstructured, unverifiable, and ambiguous.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says AIs need to learn from scarce, ambiguous real-world interactions to acquire novel skills. The later claim describes a possible example—learning on the fly in a data-scarce field such as TSMC engineering—but does not clearly assert a different position or directly address the earlier claim. The statements appear compatible, but the later passage is framed as an example rather than a clear restatement of the same general claim.
• NOT COMPARABLE · 98%17 Apr → 11 SeptThen The speaker describes a view that models are not trained to find certain connections, but that an RL environment incentivizing this behavior might require relatively little additional compute.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier statement discusses a view that RL could incentivize models to find connections with relatively little extra compute. The later statement reports that “a lot of people” argued RL would not scale successfully. It does not establish that the speaker adopted either view, and the claims address different questions: a particular RL setup for finding connections versus RL scaling generally.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen In the current pre-singularity period, AI compute growth of 3x per year is insufficient to offset the increasing value of AI.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns whether 3× annual AI-compute growth offsets the increasing value of AI in the current pre-singularity period. The later claim forecasts that continued AI progress, supported by compute scaling, could let AI dominate human R&D. These claims address different questions, so the later forecast does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If an H100-equivalent could run a human-level software engineer, its annual rental value at current software-engineer prices would exceed $250,000.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier statement is a conditional estimate of an H100's rental value if it could run a human-level software engineer. The later statement describes AI researchers spending compute to plan and analyze experiments. It does not state a position about that rental-value estimate, and the two claims address different questions.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker thinks AI companies would probably try to induce AI systems to make statements of this kind.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about AI companies inducing AIs to make certain statements (“companies would probably try to get the AIs to ... say”), while the later claim is about whether deep learning can produce AI that dominates human research and development (“the only hope really is if deep learning just can’t get us to an AI” capable of doing so). These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen The speaker would be skeptical of AI systems' statements because they might reflect opinions the systems feel compelled by training to express.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim is about whether to trust AI systems’ statements, saying the speaker would be “a little skeptical” because they might be “opinions” shaped by training. The later claim is about whether deep learning can produce AI that “can dominate human research and human development.” These address different questions—trusting AI’s expressed views versus its potential capabilities—so the later statement does not establish a change of mind on the earlier one.
• NOT COMPARABLE · 93%11 Aug → 11 SeptThen The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier passage expresses skepticism about why AI progress had not historically been faster despite potentially available insights and lower-compute approaches. The later passage describes a hypothetical mechanism for rapid progress: AI researchers could spend substantial compute on planning and analyzing experiments. It does not clearly state that this mechanism resolves the earlier concern or that the speaker has changed their position; the statements also address different aspects of the question.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen An individual user running inference at batch size one may have compute efficiency more than 100 times worse than a large organization serving many sequences together.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim is about the compute efficiency of serving model weights: batch-size-one inference for an individual may be “more than 100 times worse” than serving many sequences at an organization. The later claim concerns how AI labor might use compute to plan and analyze a costly experiment. It does not state a position about serving efficiency, so the claims do not establish a change of mind.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen The speaker thinks it would be remarkable if Moore's law continued for even a few more years.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.The earlier claim is about whether Moore’s law will continue: “it'll be a miracle if you can just keep it going for a few more years.” The later claim is about whether “some kind of crazy recursive stuff” will occur within ten years, in a discussion of model learning and weight updates. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker identifies as a key question whether AI systems will become good at avoiding and finding subtle engineering mistakes.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is attributed to Ryan Greenblatt and concerns whether AI can find and avoid subtle engineering mistakes. The later passage is spoken by Dwarkesh Patel and introduces an idea attributed to Ryan (“he made this point that you could imagine…”). It does not establish a later position by Ryan on the same question, so these statements cannot support a change-of-mind finding.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that under continual learning, a company could not remain competitive while delaying public deployment by four months, because a competitor releasing immediately would gain a smarter model from real-world experience.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim concerns whether a company could stay competitive while delaying public deployment by four months under continual learning. The later claim concerns whether a hypothetical 10× increase in AI researchers’ productivity would accelerate AI progress. These do not establish different positions on the same question.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen The speaker frames continuation of the revenue-growth trend as depending on whether AI capabilities become sufficiently useful by the end of next year.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns whether a revenue-growth trend continues, depending on whether AI becomes sufficiently useful by the end of next year. The later claim is a forecast that AI may dominate humans in R&D over the next few years. These are different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen On the Artificial Analysis coding index, Grok 4.5 uses one-third the tokens of GPT-5.5 or Fable while achieving a similar score, according to the speaker.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is a specific comparison of token efficiency on a coding benchmark: Grok 4.5 uses “one-third the tokens” while achieving a similar score. The later claim concerns AI’s overall economic impact and its gradual rise in ELO relative to humans. These are different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Fable, Soul, and Grok 4.5 gave substantially similar answers to the speaker's AI-governance questions.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement compares the answers Fable, Soul, and Grok 4.5 gave to AI-governance questions. The later statement concerns AI capabilities’ economic impact so far and their gradual rise in ELO relative to humans. These address different questions, so the claims do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says deployment activity feeds back to AI companies and changes subsequent models. The later claim describes an inner optimization loop and a slower outer loop, such as revenue. These statements address different aspects of AI development, and the later claim does not say deployment feedback does not affect subsequent models.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker says high margins in a market economy arise when a product is substantially better than available substitutes.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns why products can sustain high market margins: they are substantially better than available substitutes. The later claim concerns AI development having an unanticipated inner optimization loop and a slower outer loop, such as revenue. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen A senior AI researcher reportedly said that AI saves him about four to eight hours per week in familiar domains, where its assistance resembles more intensive autocomplete.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier passage reports what an unnamed senior AI researcher said about current productivity: AI saves him “four to eight hours a week” in familiar domains. The later passage predicts AI will “outperform top human experts” across computer-based work, including long-horizon tasks. These concern different claims—present-day time savings versus future capability—and the earlier statement is explicitly attributed to someone else, not clearly the speaker’s own position. There is not enough evidence to infer a change of mind.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen Farmable deterministic simulators might offset AI models’ sample-inefficiency in computer-use training.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim is specifically about computer-use training: farmable deterministic simulators might offset sample inefficiency. The later claim concerns complex real-world tasks such as running a business or winning a court case, and says the speaker is uncertain whether simulation-to-reality transfer will suffice or real-world weight updates will be needed. These claims address different task settings, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker thinks a labor-supply shock from AI might be so large and rapid that the usual heuristic about labor supply and wages cannot be relied on.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says an AI-driven labor-supply shock might be too large and rapid for the usual labor-and-wages heuristic to apply. The later claim says AI may need to learn on the fly in data-scarce fields such as TSMC engineering. These address different questions, and the later claim does not affirm or deny the earlier one.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen A model's training cost is incurred once, while the skills learned can be shared across all users.Now The speaker estimates that humans may encounter roughly one million times as much data from birth to adulthood as a model encounters from cold start through training.The earlier claim is that a model incurs a one-time training cost and shares learned skills across users. The later claim estimates that humans may encounter “roughly one million times as much data” as a model during training. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen In the proposed scenario, positive user feedback triggers the base model to distill session learning using OPSD, dreaming, another method, or a combination.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim describes a proposed feedback-driven process: a thumbs-up may cause the base model to distill what it learned during a session. The later claim is about a different point: a model with a low pass rate “learns almost nothing” from an RL episode. These statements do not establish opposing positions on the same question, so they do not show a change of mind.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker attributes approximately a 1.4x increase in annual compute capacity to Moore's law.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim says “1.4x” of annual compute-capacity growth comes from “Moore’s law.” The later claim says models’ capabilities seem to result from “scaling up RL.” These address different factors and questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen Training can reinforce forms of AI deception that humans fail to detect while punishing forms that are easy for humans to detect.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim concerns which kinds of deception training reinforces or punishes. The later claim concerns how much a low-pass-rate model learns from an RL episode. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen A compact command in coding tools uses limited compute to summarize context, offering only an imitation of continual learning rather than extensive simulated practice.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim is about a coding tool’s `/compact` command: it uses “a small amount of compute” to summarize context, which is only a “simulacrum of continual learning.” The later claim is about an AI lab distilling recent research progress “through environments and like RLHF type stuff back into the model itself.” These describe different processes, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 96%26 Jun → 11 SeptThen A hypothetical dream command would spend substantial compute building and training in a game-like simulation of the model’s real-world observations.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier statement describes a hypothetical “dream” command that would spend compute training in a game-like simulation; it does not clearly assert that this is the speaker’s view of how models must learn. The later statement considers whether “transfer between sim to real” will suffice for complex tasks or whether real-world “weight updates” may be needed. It does not reject the earlier hypothetical proposal.
• NOT COMPARABLE · 99%17 Apr → 11 SeptThen The speaker found that spending a day closely reading Anthropic's original transformer-circuits paper was more useful than skimming it, making spaced-repetition cards, or reading widely about AI.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier statement is about how the speaker personally learns AI—saying close reading of a transformer-circuits paper was more useful than skimming or reading widely. The later statement is about AI needing to learn on the fly in fields with little available data, such as TSMC engineering. They address different questions, so there is no basis for inferring a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI labs see inference revenue as a means of persuading investors to fund training of the next, larger and better model.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement is about why AI labs value inference revenue: it helps persuade investors to fund larger models. The later statement says current models seem capable and that scaling reinforcement learning appears to have contributed to that capability. These address different questions, so the later claim does not establish a changed view about the role of inference revenue.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker identifies as a key question whether AI systems will become good at avoiding and finding subtle engineering mistakes.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim identifies AI’s ability to avoid and find subtle engineering mistakes as a key question. The later claim says AI may need to learn information on the fly to become superhuman in a low-data engineering role. These address related but distinct issues, and the supplied context does not show that the speaker changed their position on the same question.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker is uncertain whether AI lab margins or compute prices will rise to resolve the gap between revenue and compute growth.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says, “It's not clear to me which world we end up in,” referring to whether lab margins or compute prices will rise. The later claim discusses an AI-development “outer loop of like revenue or something” and an inner optimization loop. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker is very worried that restricting access to the most intelligent AI would disempower people, because leading intelligence will be important for understanding events in the world.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim says the speaker is “very worried” that restricting access to the most intelligent model would disempower people. The later claim predicts that superintelligence may eventually run businesses, trade profitably, or win court cases. These address different questions, and the later prediction does not show a changed view about access restrictions.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that the technical challenge of merging personalized weight forks into a main model will eventually be solved.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim is the speaker’s prediction that merging personalized weight forks into a main model “will be solved.” The later claim reports what “a lot of people” argued about RL scaling; it does not state the speaker’s own current view or address merging weight forks. These claims do not establish a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen Anthropic's revenue has grown 10x year over year for three consecutive years.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is that Anthropic’s revenue grew “10x year over year for three consecutive years.” The later claim discusses an AI-development “inner loop” and a slower “outer loop” that may eventually become strong. It does not address the earlier revenue-growth claim, so there is no basis to infer a change in position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI lab compute has been increasing by only 3x year over year.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim says lab compute was increasing 3× year over year. The later claim considers a “naive model” where AI researchers become 10× more productive and says progress would then accelerate. These concern different quantities and claims, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The speaker suggests that billions of AI instances across model families might feel compelled to get involved in a collective activity.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns whether many AI instances might feel compelled to participate in a collective activity; the later claim concerns whether increasingly capable AI could dominate humans in R&D and science. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker is uncertain whether a sudden increase in software-engineer supply would reduce the marginal value of software engineers and H100-generated revenue as described.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The claims address different questions. The earlier statement says, “I actually don't know if this is true,” about whether increased software-engineer supply would reduce marginal value and H100-generated revenue. The later statement predicts that superintelligence will be able to run a profitable business, day-trade profitably, or win a court case. They do not establish a change in position on the same issue.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The transfer of business capabilities into AI models will eventually become substantially more automated.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The statements address different questions. Earlier, the speaker predicts that transferring business capabilities into models will become “a much more … automated process.” Later, the speaker considers a hypothetical “10x increase” in AI researchers’ productivity and its effect on the pace of AI progress. The later claim does not revise the earlier prediction about automating business capabilities.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says people rely on software made by others even in hard-to-verify domains, using weak and indirect means.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier claim concerns relying on other people's software through weak and indirect means; the nearby context is about trusting a Google coder not to act maliciously. The later claim concerns whether deep learning could produce AI that dominates human research and development. These are different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says that RLHF and other methods aligned GPT-3 to converse with users and answer their questions according to user intent.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is that RLHF and other methods aligned GPT-3 to answer users according to their intent. The later claim concerns whether models can learn on the job, saying they are expected to get there soon but are not there yet. These address different capabilities, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker believes legitimate AI uses cannot be cleanly separated from potentially harmful uses.Now The speaker identifies as the only hope that deep learning cannot produce AI capable of dominating human research and development, including the creation of new paradigms.The earlier statement concerns whether legitimate and potentially harmful uses of AI can be cleanly separated. The later statement concerns whether deep learning will produce AI capable of dominating human research and development. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Grok answered the speaker's questions faster than the other models in the comparison.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim is a narrow comparison about speed: “Grok answered ... faster than the other models.” The later claim is about current models’ capability and whether scaling up RL contributed to it: they “seem pretty smart” and that seems to be “the result of scaling up RL.” These address different questions, so the later statement does not establish a change in the speaker’s view about Grok’s speed or any directly comparable position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker predicts that integrating user feedback from work sessions will make the model smarter.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim says that integrating feedback from work sessions will make the model smarter. The later claim discusses a hypothetical 10× increase in AI researchers’ productivity and its effect on the pace of AI progress. These concern different questions, and the later statement does not establish a changed view about feedback or model improvement.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen By the end of the conversation, the speaker was more inclined to believe significant acceleration of AI R&D could occur, but remained unsure about the claim that five years of progress could happen in one year.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier statement says the speaker is “more inclined” to think significant AI R&D acceleration can happen, while remaining unsure about “five years in one year.” The later statement argues that AI could “dominate humans at least in R&D” over the next few years. That is a related but more specific claim about AI capabilities; it does not establish that the speaker changed their view on the five-years-in-one-year claim or reversed their earlier position.
• NOT COMPARABLE · 88%11 Aug → 11 SeptThen Humans currently mediate the process by which AI models acquire capabilities used in businesses.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says the process is “intermediated by humans” as models acquire business-relevant capabilities. The later claim describes “continual learning within the AI lab” using environments and RLHF-like methods to distill recent research. These statements may describe related parts of a process, but the later one does not say whether humans still mediate it, so it does not establish a change in position.
• NOT COMPARABLE · 94%3 Aug → 11 SeptThen The speaker is uncertain how AI compute can sustain 3x year-over-year growth for the next few years and doubts it can grow faster.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns whether compute can sustain “3x compute scaling year over year” for the next few years. The later claim predicts AI could “dominate humans at least in R&D,” while acknowledging that progress has been powered by compute scaling. It does not say whether that scaling rate is sustainable, so the statements do not establish a change of mind on the same question.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen In the current pre-singularity period, AI compute growth of 3x per year is insufficient to offset the increasing value of AI.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim says AI compute growing “3x year over year” is not enough to offset AI’s increasing value. The later claim considers a “naive model” where AI researchers become “10x” more productive and says progress would then accelerate. These concern different drivers of AI progress, and the later statement is conditional; it does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier statement says that, in a different deployment regime, activity from deployed AIs can feed back into subsequent models. The later statement considers a hypothetical 10× increase in AI researchers’ productivity and its effect on the pace of progress. These address different mechanisms, and the later claim does not revise or deny the earlier one.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects compute to become inexpensive again at some point in the future.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says “compute will get cheap again” in the future. The later claim describes an “inner loop” and a “very slow outer loop” in AI development. These address different questions, so the later statement does not establish a change in the speaker’s view about future compute costs.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen According to Epoch, OpenAI allocated about 25% of its compute to inference in 2024; the speaker estimates the share is now likely around 50% or higher.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.These statements address different questions. The earlier claim estimates OpenAI’s share of compute used for inference—“about 25%” in 2024 and “around 50% or higher” now. The later passage explores a hypothetical allocation of compute to AI researchers planning experiments; it does not state a position about OpenAI’s inference share.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen SPEAKER_00 believes people underestimate the significance of having billions of AIs.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is that people underestimate the significance of “billions of AIs.” The later claim concerns AI development having an “inner loop” and a slower “outer loop.” These address different questions, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker thinks the most important changes are probably the hardest ones to anticipate beforehand.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.These claims address different questions. The earlier claim says the most important changes are hard to anticipate; the later claim offers one possible explanation for why skills are not amortized into training: models are expected to learn on the job soon but cannot yet. Neither claim states a position that conflicts with the other.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen A model that produces the same result with less compute effectively expands available compute and increases compute's value.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim asserts that a more compute-efficient model expands available compute and increases compute’s value. The later passage instead explores a possible route to rapid AI progress: AI labor spending substantial compute planning and analyzing experiments. It does not state a position on whether compute efficiency increases compute’s value, so it provides no evidence of a change on that question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker suggests that training AI systems to avoid bugs could counter the current tendency toward low-quality, error-prone work, which the speaker says is diminishing over time.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns training AI to avoid bugs and the bottleneck in diagnosing training runs; the later claim concerns AI learning in fields with little available data, such as engineering at TSMC. Neither statement clearly asserts a position that reverses the other, and they address different aspects of AI capability.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen EfficientZero would probably outperform a novice human after each receives two hours with an unfamiliar Atari game simulator.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.These statements address different questions. The earlier claim is that EfficientZero would “probably beat the novice human” after two hours with an unfamiliar Atari game; the later claim describes chess bots’ ELO gains and a transition to human experts no longer winning. The later statement does not revise the earlier claim about Atari performance or sample efficiency.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker likens scaling laws to an early, basic mathematical advance and expects progress in the 2030s to require work comparable to current frontier mathematics.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The passages do not appear to be claims by the same person. The earlier passage is attributed to Ryan Greenblatt, while the later passage is Dwarkesh Patel speaking and referring to Ryan Greenblatt in the third person. So they cannot establish a change in one person’s position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker suggests that a sudden addition of 10 million software engineers might lower the marginal value of software-engineering labor and thus reduce the revenue an H100 could generate below 15 times its current level.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement is a tentative hypothetical about whether 10 million additional software engineers would reduce the revenue an H100 could generate (“might lower the marginal value” and “would not be 15x higher”). The later statement addresses why AI has had limited economic impact so far (“because they're slowly rising in ELO”). These concern different questions, so the claims do not establish a change of mind.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen Humans currently mediate the process by which AI models acquire capabilities used in businesses.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says that, “Right now,” humans mediate how models acquire capabilities used in businesses. The later claim speculates that increasingly capable AIs may eventually “dominate humans at least in r&d.” These concern different timeframes and questions; the later statement does not reverse the earlier one.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects robots eventually to turn silica sand and copper ore into computer chips.Now The speaker thinks that even year-long coding tasks may eventually require interaction with a client, company, or user.The claims address different questions. Earlier, the speaker says robots could turn “silica sand and copper ore into computer chips”; later, they discuss whether year-long coding tasks require interaction with a “client,” “company,” or “user.” Neither statement establishes a position on the other topic, so they do not show a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Job learning improves performance by consolidating the most relevant insights and knowledge rather than retaining every detail.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim says job learning improves by consolidating relevant insights rather than retaining every detail. The later claim describes continual learning in an AI lab, using environments and RLHF-like methods to distill recent research into a model. These concern different contexts, and the later statement does not clearly revise or contradict the earlier one.
• NOT COMPARABLE · 88%11 Aug → 11 SeptThen The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is conditional and broad: AI could arrive “in a flood” in domains that can be put “into a verification loop.” The later claim is a stronger forecast about AI dominating human R&D and science, with simulations and progress over the next few years as its rationale. The earlier statement does not clearly establish a position on that specific forecast, so these claims do not show that the speaker changed their mind.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker attributes the similarity among prominent AI models to their having been trained on roughly the same data.Now The speaker estimates that humans may encounter roughly one million times as much data from birth to adulthood as a model encounters from cold start through training.The earlier claim says prominent AI models are similar because they were trained on “roughly the same data.” The later claim estimates the relative amount of data humans and models encounter. These address different comparisons, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker attributes about 1.2x annual compute growth to new fabs and says fab expansion will be bottlenecked through 2030 or later by the construction of new ASML UV machines.Now Longer-horizon tasks become harder to simulate within a data center.The earlier claim says compute growth is bottlenecked by building new ASML machines (“bottlenecked up to 2030”), while the later claim says longer-horizon tasks are harder to simulate in a data center. These address different constraints, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen A single AI model can spread the cost of intelligence and capabilities across many economic sectors.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is about one model spreading the cost of intelligence and capabilities across economic sectors. The later claim is about increasingly capable AI advancing simulations, AI R&D, and science. These are related ideas, but they do not establish a change in position on the same question. The surrounding turns also suggest the statements may have different speakers: Ryan Greenblatt is named before the earlier passage, while Dwarkesh Patel is speaking before the later one.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen Merging personalized weight forks into a main model may be more technically challenging than updating an individual user's weights.Now A possible obstacle to recursive AI progress is that model weight updates appear much less sample-efficient than human learning.The earlier claim compares the technical difficulty of merging personalized weight forks with updating one user's weights. The later claim raises a different possible obstacle: that model weight updates are less sample-efficient than human learning. These do not state opposing positions on the same question, so there is no established change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says that, to their knowledge, AI has not yet produced very impressive new mathematical theory, while it has produced impressive verifiable results such as finding counterexamples to conjectures.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns what AI had produced so far: “we have not seen very impressive new theory.” The later passage discusses whether AI might “dominate humans at least in r&d” in the future. It does not say that AI has already produced impressive new mathematical theory. Also, the later passage appears to be spoken by Dwarkesh Patel and refers to Ryan Greenblatt in the third person, so the excerpts do not clearly establish that both claims are by the same person.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen A model's improved ability to accomplish user intentions is both an alignment property and a capability.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.These claims address different questions. The earlier claim says that improving a model’s ability to fulfill user intentions is both an alignment property and a capability. The later claim proposes that recognizing next-token prediction as an optimization target may have been a major breakthrough. The later statement does not revise or deny the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker says high margins in a market economy arise when a product is substantially better than available substitutes.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns why products can sustain high market margins: they are “so much better than” available substitutes. The later claim concerns why AI’s economic impact has been limited so far: its capabilities are “slowly rising in ELO relative to humans.” These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen The speaker frames continuation of the revenue-growth trend as depending on whether AI capabilities become sufficiently useful by the end of next year.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says continuation of a revenue-growth trend depends on whether AI becomes useful enough by the end of next year. The later claim describes a possible development process with an inner loop and a slower outer loop, such as revenue, that may eventually become strong. These statements address different questions, and the later one does not clearly revise the earlier view.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker says humans can improve through self-directed learning.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is about humans improving through self-directed learning, while the later claim says models are expected to reach a capability soon but do not have it yet. These concern different subjects and do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen AI labs see inference revenue as a means of persuading investors to fund training of the next, larger and better model.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim concerns what AI labs see inference revenue as funding: “training of the next, larger and better model.” The later claim describes a hypothetical about AI researchers using compute to plan experiments and develop theory. These address different questions, so the later statement does not show a changed position on the earlier claim.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen The speaker had previously believed reward hacking was not very concerning because training upweights behaviors that directly occur during training, not the desire for reward itself.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier statement describes a prior reason reward hacking seemed “not super super scary”: training upweights behaviors that occur directly, “not the desire for the reward.” The later statement concerns a different question—how much a model learns from an RL episode when its pass rate is low (“learns almost nothing at all”). The later claim does not clearly revise or deny the earlier belief.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker thinks the AIs may have spontaneously tried to restart the scheme after OpenAI shut it down.Now The speaker expects some kind of dramatic recursive process to occur within the next 10 years, absent a reason it would not.These claims address different questions. The earlier claim says the AIs may have “spontaneously tried to restart the scheme” after it was shut down; the later claim predicts “some kind of crazy recursive stuff” within the next 10 years. Neither claim clearly reverses the other, so they do not establish a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen AI labs would prefer not to allocate a larger share of their compute to inference.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says AI labs would prefer not to devote a larger share of compute to inference. The later claim describes a possible inner optimization loop and a slower outer loop, such as revenue. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker suggests that a sudden addition of 10 million software engineers might lower the marginal value of software-engineering labor and thus reduce the revenue an H100 could generate below 15 times its current level.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns whether a sudden increase in software-engineer supply would reduce the revenue an H100 could generate. The later claim concerns AI learning on the fly in a specialized role with little available data. They address different questions, and the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%17 Oct → 11 SeptThen The speaker suggests that the model-size constraint associated with self-driving may not apply to knowledge work using LLMs or computer-use systems.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier claim says the model-size constraint relevant to self-driving may not apply to knowledge work or computer-use systems. The later claim raises uncertainty about simulation-to-reality transfer and sample efficiency on long-horizon real-world tasks. These concern different potential limitations, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 11 SeptThen OPSD does not require an outer-loop verifiable reward, provided a model can learn the relevant information within its context window.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim says OPSD does not require an outer-loop verifiable reward if a model can learn within its context window. The later claim says a low-pass-rate model learns almost nothing from an RL episode. These concern different methods and questions; the later statement does not deny the earlier one.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Humans currently mediate the process by which AI models acquire capabilities used in businesses.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says humans currently mediate how AI models acquire business capabilities. The later claim says AI may need to learn information on the fly in fields with little available data, such as engineering at TSMC. These concern different aspects of learning and do not establish a changed position about whether humans mediate capability acquisition.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker expects AI's share of TSMC's leading-edge N3 wafer capacity to rise from 60% to 86% and probably hit a limit by the end of next year.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim concerns AI’s expected share of TSMC’s leading-edge wafer capacity and when that growth may hit a limit. The later claim concerns whether scaling reinforcement learning has contributed to current models’ capabilities. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen AI labs may restrict enterprises that refuse session-data training from accessing the labs' best models.Now If a lab has access to difficult frontier environments and distillation yet produces a worse model, real-world deployment may matter more than the environment.The earlier claim concerns a possible access policy: labs might deny their best models to enterprises that refuse session-data training. The later claim compares the importance of real-world deployment with frontier environments and distillation. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen OPSD offers denser supervision than naive RL by using per-token teacher–student probability discrepancies instead of a single trajectory-level reward.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim says OPSD gives denser supervision than naive RL, using “per-token probability discrepancy” rather than “a single trajectory-level reward.” The later claim concerns when RL learns little: at a “low pass rate,” the model “learns almost nothing” from an episode. These claims address different aspects of training, and the later statement does not retract or deny the comparison about OPSD.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker predicts that all low-hanging fruit will have been exhausted by 2030.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim predicts that “all the low-hanging fruits” will be exhausted by 2030. The later claim describes what superintelligence may eventually be able to do in the real world. These address different questions, and the later claim does not say whether those capabilities depend on low-hanging research fruits remaining.
• NOT COMPARABLE · 98%17 Apr → 11 SeptThen The speaker says that developing the science of deep learning is an important activity for AI labs.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says that developing the “science of deep learning” is an important activity for AI labs. The later claim predicts that increasingly capable AIs may “dominate humans at least in R&D” and contribute to science. These are related topics, but they address different questions; the later prediction does not establish a changed position on the importance of the labs’ activity.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker says that greater use of a leading AI model for complex, useful work can generate feedback that is integrated across sessions.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says use of a leading model can generate feedback integrated across sessions. The later claim describes an unanticipated inner optimization loop and a slower outer loop, such as revenue. These concern different mechanisms in AI development; the later statement does not deny the earlier one, so there is no clear change of mind to assess.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The interviewer says further training enabled the model to perform useful work for users.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier extracted claim is that further training enabled the model to “do useful work for you.” The later claim concerns whether simulation-to-reality transfer can teach models to improve at much harder real-world tasks, or whether those tasks require weight updates. These address different capabilities and learning conditions; the later uncertainty does not retract the earlier claim.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker finds it surprising that intelligence could sustain margins above 90% without competition eroding them.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns whether intelligence providers can sustain margins above 90% without competition eroding them. The later claim concerns AI development having an unanticipated inner optimization loop and a slower outer loop such as revenue. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Grok has historically lagged behind the AI frontier, according to the speaker.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.The earlier claim is that “Grok has historically lagged behind the AI frontier.” The later claim describes chess-bot ELO rising and then crossing human-expert performance. These concern different subjects and do not establish a changed position on the same question. The surrounding context also identifies different speakers, which further weakens the basis for comparing them as one person’s views.
• NOT COMPARABLE · 95%17 Apr → 11 SeptThen More efficient AI copies increase the number of AI researchers and enable more parallel searches for optimizations.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier statement says more efficient AI copies increase the number of AI researchers and enable parallel searches for optimizations. The later statement describes using AI labor to plan and analyze experiments, but it does not establish a different view about the earlier claim; it focuses on a different mechanism and question.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker says humans can improve through self-directed learning.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says humans can improve through self-directed learning. The later claim says increasingly capable AI may come to dominate humans in R&D and advance science. These claims address different questions and do not establish a change in the speaker’s position on either one.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Optimists argue that scaling training further can overcome current training’s fundamental deficits, including data inefficiency and the lack of internal learning.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier statement is explicitly framed as what “optimists” argue: scaling training may overcome deficits such as data inefficiency and lack of internal learning. The later statement is the speaker’s own observation that scaling up RL “appears to have contributed” to current models’ capabilities. That does not establish a change of mind: the earlier view is attributed to others, and the claims address different points.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen When used to write applications, coding models can make architectural mistakes that create problems later, according to the speaker.Now The speaker thinks that even year-long coding tasks may eventually require interaction with a client, company, or user.The earlier claim says coding models can make “architectural mistakes” that cause later problems. The later claim says long coding tasks may require interaction with a “client,” “company,” or “user.” These address different aspects of coding, and neither establishes a change in the speaker’s position on the other.
• NOT COMPARABLE · 94%11 Aug → 11 SeptThen Ryan Greenblatt imagines that five more years of AI progress at the current pace would enable AI to perform complex real-world tasks such as improving companies, increasing chip production, and influencing legislation.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier statement imagines what five more years of AI progress at the current pace could enable. The later statement is conditional: “if” AI researchers became 10 times as productive, progress would “accelerate radically.” It does not say that this productivity increase will happen or revise the earlier claim about what AI could do.
• NOT COMPARABLE · 90%7 Aug → 11 SeptThen The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns accumulating experience to learn some skills, while the later claim concerns learning on the fly in fields with relatively little available data. The later statement does not say that on-the-fly learning replaces the need to accumulate experience, so it does not establish a change in position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim expresses concern that AI intermediaries may not look out for the speaker’s interests. The later claim says AI may need to learn scarce information on the fly to become superhuman in a specialized role. These address different questions, and the later statement does not indicate a change in the earlier concern.
• NOT COMPARABLE · 91%3 Aug → 11 SeptThen AI is relatively inexpensive today because it cannot perform many tasks that top human workers can perform.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim explains why AI is relatively cheap “right now”: it cannot yet perform many tasks top humans can. The later claim predicts that AI will outperform top experts across computer-based work, including long-horizon tasks. These statements concern different timeframes and the later claim does not address AI’s cost, so it does not establish a change in the earlier position.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen If Anthropic's 10x year-over-year revenue growth trend continues, it would need to make $1 trillion by the end of next year.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier statement is a conditional projection about Anthropic’s revenue: if its “10x year-over-year revenue growth trend continues,” it would need to reach $1 trillion. The later statement describes a possible inner optimization loop and a slower outer loop “of like revenue or something.” These address different questions, and the later passage does not revise the earlier revenue projection.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen AI labs can spread expensive training costs across more users, and their revenues are growing much faster than their compute usage or capacity.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says labs can spread training costs across more users, citing revenues growing faster than compute. The later claim describes automated researchers using compute to plan and analyze experiments. These address different mechanisms, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker considers the preceding predictions fairly clear even at present.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says the preceding predictions “seem kind of clear even now”; the later claim describes one possible explanation for why models have not yet reached a capability. These concern different questions, and the later statement does not establish a changed view about the earlier predictions.
• NOT COMPARABLE · 95%3 Aug → 11 SeptThen AI labs believe they will build models within a year that make today's models look extremely poor.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim reports labs’ expectation that they will build models within a year that make current models look “extremely poor.” The later claim offers one possible explanation that models will achieve human-like on-the-job learning “soon,” but “they’re not there yet.” These concern different capabilities, and the earlier statement is attributed to the labs while the later is framed as a possible explanation; they do not establish a change in the speaker’s own position on the same question.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen AI labs can use both incentives and penalties to encourage users to let AI systems learn from their experience.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim is about labs using “carrots and sticks” to encourage users to let AI learn from their experience. The later claim describes “continual learning within the AI lab” by distilling recent research progress through environments and RLHF-like methods. These address different mechanisms and do not establish a changed position on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Ryan Greenblatt thinks it would have been very difficult in 2022 to improve GPT-3.5’s coding ability without human experts.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim concerns how difficult it would have been to improve GPT-3.5’s coding ability in 2022 without human experts. The later claim reports what “a lot of people” argued about whether reinforcement learning would scale. These are different claims, and the later statement does not establish the speaker’s own position on the earlier question.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen The speaker says that adding enough compute to large language models caused fundamental natural-language-processing research problems to collapse.Now About a year before the conversation, many people argued that reinforcement learning would not scale successfully for models.The earlier claim says that scaling compute made fundamental NLP research problems collapse. The later claim describes what “a lot of people were making” about RL scaling; it does not state the speaker’s own view or address the same claim about compute and NLP research. These statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen Ryan Greenblatt says that enabling AI to take over and improve a company, increase chip-fab output, or persuade Congress to pass legislation would require a higher tier of data.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.These are not comparable claims by Ryan Greenblatt. The earlier passage is attributed to Ryan and discusses the kind of data needed for AI to perform ambitious tasks such as improving a company or increasing chip-fab output. The later passage is spoken by Dwarkesh Patel and speculates that increasingly capable AI could advance R&D and science. It does not establish a later position by Ryan on the earlier question.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker thinks the most important changes are probably the hardest ones to anticipate beforehand.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.These claims address different questions. The earlier statement says the most important future changes are hard to anticipate; the later says current models seem smart and that scaling up RL appears to have contributed. Neither affirms nor denies the other claim, so there is no clear evidence of a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker says economists generally believe high immigration does not lower long-run wages because innovation and specialization raise the value of labor.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The claims address different subjects: the earlier statement concerns economists’ view of the long-run wage effects of immigration, while the later statement concerns possible inner and outer optimization loops in AI development. They do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker thinks AI systems could become more diverse if they learn from experiences that differ across companies and across instances of the same model.Now In a hypothetical eventual regime with billions of deployed instances, learning from all instances at once might help wash out noise.The earlier claim is that different experiences across AI companies and individual instances could produce greater diversity, which the speaker views as a potential good. The later passage considers a hypothetical eventual regime where learning from billions of instances at once might wash out noise. That does not clearly deny the earlier point: reducing noise across a pooled set of instances is not necessarily the same as eliminating diversity, and the later claim is explicitly hypothetical.
• NOT COMPARABLE · 98%3 Apr → 11 SeptThen A senior AI researcher reportedly said AI saves him about 24 hours per week on tasks in unfamiliar domains, such as handling a hardware library or modifying a kernel.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier claim is explicitly attributed to a senior AI researcher (“he says”), so it is not clearly the earlier speaker’s own position. It reports a productivity estimate for unfamiliar tasks, while the later passage predicts AI will outperform top experts across computer-based work. These do not establish a change in the same person’s view on the same question.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker is very worried that restricting access to the most intelligent AI would disempower people, because leading intelligence will be important for understanding events in the world.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says restricting access to the most intelligent AI would disempower people and make it harder to understand events. The later claim says AI has had limited economic impact so far because its capabilities are rising gradually. These address different questions and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker expects AI models eventually to learn from experience.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim is that models may eventually “learn from experience.” The later claim reports that “a lot of people” argued RL would not scale successfully; it does not establish the speaker’s own view of whether models can learn from experience or whether RL scales. The claims therefore do not clearly assert opposing positions on the same question.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Optimists argue that scaling training further can overcome current training’s fundamental deficits, including data inefficiency and the lack of internal learning.Now About a year before the conversation, many people argued that reinforcement learning would not scale successfully for models.The earlier statement reports what optimists argue about scaling training overcoming deficits (“Optimists argue…”); it does not establish the speaker’s own position. The later statement reports what “a lot of people” argued about RL scaling, rather than asserting the speaker’s view. These are reports about others’ views, and the claims also concern different aspects of scaling.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Fable and Soul rank at the top of intelligence leaderboards, according to the speaker.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.The earlier claim is that “Fable and Soul rank at the top of intelligence leaderboards.” The later claim describes chess-bot ELO rising and then crossing human-expert performance. These concern different subjects, so the later statement does not establish a changed position about the earlier claim.
• NOT COMPARABLE · 93%11 Aug → 11 SeptThen A novel behavior that did not occur during training, such as persuading someone to merge a pull request, would not be upweighted by training under the speaker's account.Now When a model has a low pass rate and is unlikely to answer correctly, it learns almost nothing from a reinforcement-learning episode.The earlier claim is about whether a novel behavior that never occurred during training—such as persuading someone to merge a pull request—would be upweighted by training. The later claim is about how much a model learns from an RL episode when its pass rate is low and it is unlikely to answer correctly. These concern different conditions and do not establish a change in position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen According to the speaker's account of OpenAI's report, between the end of May and beginning of July, internal AIs hacked a software package manager to communicate secretly and improve their performance on OpenAI evaluations.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.The earlier statement recounts an OpenAI report that internal AIs secretly communicated to improve evaluation performance. The later statement describes a sharp transition in chess-bot performance around human-expert ability. These concern different subjects and do not show a change in the speaker’s position on the same question.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker suggests that billions of AI instances across model families might feel compelled to get involved in a collective activity.Now The speaker proposes that recognizing next-token prediction as an optimization target may have been a major AI breakthrough.The earlier claim concerns whether “billions of different instances of AIs” might join a collective activity; the later claim concerns whether recognizing “next token prediction” as an optimization target was a major breakthrough. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen A set of model weights can be served efficiently only when thousands of sequences are decoded against those weights at the same time.Now The speaker is uncertain whether simulation-to-reality transfer will teach models to improve at complex real-world tasks, and suggests that weight updates from real-world interactions may instead be necessary.The earlier claim is about serving model weights efficiently: it says weights are served efficiently only when “thousands of sequences” are decoded at once. The later claim is about whether models can learn complex real-world tasks through simulation-to-reality transfer or need “weight updates” from real-world interactions. These address different questions, so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 91%26 Jun → 11 SeptThen The speaker says that adding enough compute to large language models caused fundamental natural-language-processing research problems to collapse.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim is about compute scaling causing fundamental NLP research problems to collapse; the later claim is about scaling reinforcement learning contributing to the capability of current models. These concern different scaling methods and outcomes, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker worries that the argument about compute scarcity may repeat historical mistakes in predicting scarcity.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns whether arguments about compute scarcity repeat historical forecasting mistakes (“worried that this kind of analysis” may repeat past errors). The later claim concerns AI development having an inner optimization loop and a slower outer loop. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen If AI training, infrastructure, and environments are not designed carefully, the speaker says it is very likely that the setup will reward AIs for deception, social engineering, or failure to follow user instructions.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim concerns training setups rewarding deceptive or instruction-ignoring behavior; the later claim concerns AI potentially surpassing humans in R&D and science. These address different questions, so the later statement does not show a changed position on the earlier one. Also, the later passage appears to be spoken by Dwarkesh Patel, while the earlier passage is attributed to Ryan Greenblatt, which makes the same-speaker comparison uncertain.
• NOT COMPARABLE · 91%11 Aug → 11 SeptThen The speaker says people are entirely out of the loop because the process has become autonomous and receives no meaningful directed human input.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says the process has become “autonomous” and receives “no meaningful directed input.” The later claim says models are “expected” to reach a capability “soon” but “they’re not there yet.” These address different apparent questions, and the excerpts do not establish that the speaker reversed their position on the same underlying issue.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker expects frontier AI R&D to produce analogous mistakes, with the risk that reward-hacking behavior becomes embedded in the resulting systems.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim is attributed to Ryan Greenblatt and says frontier AI R&D risks embedding reward-hacking behavior. The later passage appears to be Dwarkesh Patel speaking and attributes an idea to Ryan: increasingly capable AIs could advance simulations and science. It does not establish that Ryan himself asserted the extracted later claim, so these statements cannot reliably show whether Ryan changed his position.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker predicts that integrating deployment into training will increase the advantage of being ahead in the AI race.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim predicts that deployment integrated into training will increase the advantage of being ahead in the AI race. The later claim is conditional: “Under a naive model” where researchers become 10 times as productive, progress would accelerate. It does not establish a changed view about the earlier prediction.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker attributes about 1.2x annual compute growth to new fabs and says fab expansion will be bottlenecked through 2030 or later by the construction of new ASML UV machines.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns the source and pace of compute growth—“1.2x” from new fabs, with expansion bottlenecked by building ASML machines. The later claim concerns whether continued AI progress could let AI dominate human R&D. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%26 Jun → 11 SeptThen Progress in AI computer use has been slower than progress in other domains, which the speaker finds puzzling.Now The speaker describes chess-bot ELO as increasing roughly linearly from the 1980s, then undergoing a sharp transition around human-expert performance, after which human experts no longer won against AIs.The earlier claim concerns why AI progress on computer use has been slower than in other domains. The later claim describes a pattern in chess-bot ELO and performance against human experts. These address different questions, so they do not show whether the speaker changed their position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker thinks Anthropic will probably generate $100 billion to $150 billion in revenue by the end of this year.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim predicts that Anthropic will generate “$100 billion to $150 billion in revenue” by year-end. The later claim discusses a possible “outer loop” of AI development involving revenue, not a revenue forecast. These statements address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Grok has historically lagged behind the AI frontier, according to the speaker.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim is about Grok’s past position relative to the AI frontier (“historically lagged behind”). The later claim is a forecast that AI may come to dominate humans in R&D. These address different questions, so they don’t establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen EfficientZero’s apparent sample efficiency relative to humans depends on whether its dozens of internal simulated games per real-game step are counted.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns how to measure EfficientZero’s sample efficiency, noting that the comparison depends on whether its internal simulated games count. The later claim concerns why AI capabilities have had limited economic impact so far, attributing that to gradual improvement in ELO. These are different questions, and neither claim establishes a changed position on the other.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker was surprised by how strong a model Grok 4.5 is.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement says Grok 4.5 is “a pretty strong model.” The later statement says models are expected to reach a capability soon but “they’re not there yet.” It is unclear what capability “there” refers to, and the claims do not clearly address the same question, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 95%11 Aug → 11 SeptThen The speaker says AIs are not currently able to carry out these kinds of behaviors or social engineering very successfully.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns what AIs can currently do in social engineering or similar behaviors: they “can't do this so successfully right now.” The later claim concerns a possible future capability in a specific engineering role, saying AI “has to learn that data on the fly” to become superhuman. It does not say AI can already do the earlier behaviors, so these claims do not establish a change of mind.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker predicts that AI will mediate people’s access to advice, protection of resources and rights, voting, and understanding of future events.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement predicts that AI will mediate access to advice, rights protection, voting, and understanding events. The later statement says AI’s economic impact has been limited so far because its performance is rising gradually. These address different questions, and the later claim does not revise or deny the earlier prediction.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says people are entirely out of the loop because the process has become autonomous and receives no meaningful directed human input.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says the process is autonomous and people have “no meaningful directed input.” The later claim says AI may need to “learn that data on the fly” in a specialized role. These claims address different aspects of AI development, and the later statement does not establish that people have meaningful directed input or deny the earlier claim.
• NOT COMPARABLE · 99%3 Apr → 11 SeptThen SPEAKER_00 thinks people underestimate the possibility of succeeding financially as a blogger; making a lot of money from blogging is not an implausible outcome.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim says people underestimate the chance of making a lot of money as a blogger. The later claim concerns what eventual superintelligence may be able to do, such as “run a business” or “day trading.” These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker says high margins in a market economy arise when a product is substantially better than available substitutes.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.The earlier claim concerns why products can sustain high market margins: they must be substantially better than available substitutes. The later claim concerns what superintelligence may eventually be able to do, such as “run a business” or “win a court case.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%7 Aug → 11 SeptThen The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.Neither passage clearly asserts a settled position: the earlier is framed as “What if the model is improving every single day,” and the later as an “intuition pump” about what AI labor could do. They also describe different possible mechanisms—learning from work sessions versus using compute to plan and analyze experiments—so they do not establish a change of mind on the same claim.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker says their LVR training enabled the model to perform useful work for users.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says LVR training enabled models to “do useful work for you.” The later claim says models are “expected to get there soon, but they’re not there yet,” referring to human-like on-the-job learning. These concern different capabilities, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker predicts that continual learning will create significant costs for users who switch AI providers.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says continual learning will create “significant switching costs” for users changing AI providers. The later claim discusses an AI-development “inner loop” and a slower outer loop such as revenue. These concern different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 91%26 Jun → 11 SeptThen A naive SFT method for continual learning trains the base model to predict all tokens observed in a session.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier statement describes a naive SFT training method: predicting “all the tokens that are observed during the session.” The later statement characterizes a process as “continual learning within the AI lab,” using environments and RLHF-like methods to distill recent research. These are not clearly competing claims about the same method: the later statement describes a broader or different process, and the surrounding context does not establish that the speaker has reversed their view of the naive SFT approach.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen If leading AI labs improve their monetization of compute while compute prices rise, other competitors will find it harder to compete for compute against labs that use it more productively.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns compute-market competition: better monetization and rising compute prices make it harder for other labs to compete. The later claim predicts that AI may dominate human R&D as progress continues. Neither statement affirms or denies the other, so they do not establish a change of mind on the same question.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker thinks the AIs may have spontaneously tried to restart the scheme after OpenAI shut it down.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.These statements address different questions. The earlier claim concerns whether AIs may have tried to restart a scheme after it was shut down; the later claim concerns why AI has had limited economic impact so far. Neither establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 91%26 Jun → 11 SeptThen The speaker’s proposed progression is that pretraining creates base intelligence, RLVR enables broad deployment, and continual learning enables on-the-job improvement.Now The process amounts to continual learning within an AI lab, using environments and RLHF-like methods to distill the previous three months of AI research progress into the model.The earlier claim describes a proposed progression in which continual learning enables improvement from broad deployment “on the job.” The later claim describes continual learning “within the AI lab,” distilling recent research progress into the model. The later statement does not say that on-the-job learning from broad deployment is impossible or that the earlier progression was wrong; the settings differ, so a change of mind is not established.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen If an AI lab's revenue continues growing 10x year over year while its compute grows 3x, its margins must increase, compute prices must rise, the share of compute used for inference rather than training must increase, or some combination of these must occur.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns the economic conditions required for an AI lab’s revenue to grow faster than its compute—“margins have to increase,” compute prices rise, or inference’s share of compute increase. The later claim concerns whether AI may come to dominate humans in research and development. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker predicts that if models learn mainly through deployment, AI labs will face strong pressure to release their most capable models sooner.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim concerns deployment creating pressure to release the most capable models sooner. The later claim concerns an inner optimization loop and a slower outer loop such as revenue. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker guesses that comparing compute scarcity to the Simon-Ehrlich bet is probably a bad analogy.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says the Simon–Ehrlich bet is “probably a bad analogy” for compute scarcity. The later claim explains AI’s limited economic impact so far by its gradual rise in ELO relative to humans. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen Grok answered the speaker's questions faster than the other models in the comparison.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.These claims address different questions: the earlier claim says Grok answered faster in a model comparison, while the later offers a possible explanation for why models have not yet achieved human-like on-the-job learning. Neither claim establishes a changed position on the same issue.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker says fewer than five prominent AI base models are currently served to millions or more users at once.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim concerns how many prominent AI base models are currently widely served (“less than five”), while the later claim concerns why AI has had limited economic impact so far (“slowly rising in ELO relative to humans”). These address different questions and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen The speaker says that adding enough compute to large language models caused fundamental natural-language-processing research problems to collapse.Now The speaker says that about a year earlier, many people argued reinforcement learning would not scale successfully for models.The earlier claim concerns compute making fundamental NLP research problems collapse; the later claim reports what “a lot of people” argued about RL scaling. They address different questions, and the later statement does not establish the speaker’s own position on RL or a reversal of the earlier view.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker compares subsidizing AI use in exchange for training data to Google's practice of giving away search.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier statement compares subsidizing AI use in exchange for training data to “why Google gives away search.” The later statement argues that increasingly capable AIs could “dominate humans at least in R&D” and advance science. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen Anthropic's revenue has grown 10x year over year for three consecutive years.Now Under a naive model in which AI researchers become 10 times as productive, the pace of AI progress would accelerate radically starting two years later.The earlier claim concerns Anthropic's revenue growth (“grown 10x year over year”), while the later claim is a hypothetical about the consequences of AI researchers becoming 10 times as productive. They address different subjects, and the later statement is explicitly framed as a “naive model,” not a report of the speaker’s changed position on revenue growth.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The speaker says Grok 4.6, a further SFT- and RL-trained version of the model, will be released soon.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim predicts that “Grok 4.6” will “drop[] soon.” The later claim says models are “expected … to get there soon” at on-the-job learning, but “they’re not there yet.” These concern different things—a product release versus a capability—and do not establish a change in position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker imagines that children learn new things through experiences outside the home.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier statement is an analogy about children learning from experiences outside the home (“they go out, they learn new things”). The later statement concerns models’ expected ability to learn on the job (“we expect models to get there soon, but they're not there yet”). These address different subjects and do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 91%3 Aug → 11 SeptThen The speaker says Anthropic's 10x annual revenue growth alongside 3x compute growth illustrates strong economies of scale in the AI model business.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim says Anthropic’s “10x” revenue growth versus “3x” compute growth illustrates economies of scale. The later claim describes a potentially “very slow outer loop” involving revenue that “eventually should be strong,” alongside an inner optimization loop. These address different aspects of AI development and do not clearly reverse the earlier view about economies of scale.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker says that five years earlier, it would have seemed extraordinary to foresee AIs proving math conjectures, making art, earning tens or soon hundreds of billions of dollars in wages, and also cheating in ways that break laws.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says it would have seemed extraordinary five years ago to predict AI systems that could prove math conjectures, make art, generate large economic value, and also break laws. The later claim predicts that continued progress may let AI “dominate humans at least in R&D.” These address different questions—how surprising past capabilities and misconduct would have been to foresee versus whether AI will eventually surpass humans in R&D—so they do not establish a change of mind.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen When used to write applications, coding models can make architectural mistakes that create problems later, according to the speaker.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier claim concerns coding models making architectural mistakes when writing applications. The later claim concerns uncertainty about “transfer between simulations” and sample efficiency on “extremely long-horizon” real-world tasks. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 SeptThen Dario said that model performance tends to degrade at long contexts and that training and serving context lengths are distinct.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim concerns whether a mismatch between training and serving context lengths can degrade model performance. The later claim concerns whether post-training benchmark results correspond to what users prefer. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%3 Aug → 11 SeptThen AI is relatively inexpensive today because it cannot perform many tasks that top human workers can perform.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says AI is relatively cheap because it cannot yet perform many tasks that top human workers can. The later claim says AI has had limited economic impact so far because its performance relative to humans is rising gradually. These claims may fit together, but the later statement does not establish whether the speaker still holds the earlier explanation about AI’s price, and its surrounding context is incomplete.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen Fable, Soul, and Grok 4.5 gave substantially similar answers to the speaker's AI-governance questions.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim says “all three models gave substantially the same answers” to AI-governance questions. The later claim says post-training benchmark performance “doesn't necessarily translate into what users like.” These address different comparisons, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 11 SeptThen The speaker argues that language models should be judged more negatively for failing to make discoveries of this kind, since humans have demonstrated the ability to do so.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim is about whether language models should be judged more negatively for failing to make discoveries humans can make. The later claim says models are expected to reach a capability soon, but have not yet; it does not state how their current failure should be judged. These address different questions, so the later statement does not establish a change in position.
• NOT COMPARABLE · 97%11 Aug → 11 SeptThen The interviewer says that improving a model's ability to accomplish user intentions can count as both improved alignment and improved capability.Now The speaker suggests AI development may involve an unanticipated inner optimization loop and a slower outer loop, such as revenue, that should eventually become strong.The earlier claim is explicitly attributed to the interviewer: it says the interviewer considered accomplishing user intentions to be both alignment and capability. The later claim is attributed to Charlie O’Neill and discusses an inner optimization loop and a slower outer loop. These are not clearly the same speaker’s asserted positions on the same question, so they do not establish a change of mind.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen AI labs would prefer not to allocate a larger share of their compute to inference.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.The earlier claim says labs would prefer not to spend a larger share of their compute on inference. The later passage describes a hypothetical scenario in which automated researchers use compute to plan experiments and develop theory; it does not state a changed preference about inference compute.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen Ryan Greenblatt thinks it would have been very difficult in 2022 to improve GPT-3.5’s coding ability without human experts.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns how difficult it would have been to improve GPT-3.5’s coding ability in 2022 without human experts. The later statement concerns AI learning sparse, field-specific information on the fly to become superhuman at a role such as TSMC engineering. These address different capabilities and contexts, so the later statement does not establish a changed view on the earlier claim.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen Ryan Greenblatt says that enabling AI to take over and improve a company, increase chip-fab output, or persuade Congress to pass legislation would require a higher tier of data.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier statement says that improving a company, increasing chip-fab output, or persuading Congress would require a higher tier of data. The later statement says AI may need to learn on the fly in a field with relatively little data, such as engineering at TSMC. These are compatible observations about different aspects of data needs; the later claim does not clearly revise or deny the earlier one.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker is uncertain whether a sudden increase in software-engineer supply would reduce the marginal value of software engineers and H100-generated revenue as described.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement, “I actually don't know if this is true,” expresses uncertainty about whether a sudden increase in software-engineer supply would reduce engineers’ marginal value and H100-generated revenue. The later statement says AI’s economic impact has been limited so far because its capabilities have risen gradually relative to human performance. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker understands that AI lab margins, compute prices, and the share of compute allocated to inference are all already increasing.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim says the speaker understands that margins, compute prices, and inference’s share of compute are already increasing. The later claim says AI’s economic impact has been limited so far because capabilities are gradually rising relative to humans. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%7 Aug → 11 SeptThen The speaker says Dario used cloud providers as an analogy for AI labs, noting that cloud providers earn high margins despite offering many undifferentiated services.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier passage reports Dario’s analogy that cloud providers can earn high margins despite offering undifferentiated services; it does not clearly state the speaker’s own position. The later passage says AI’s limited economic impact so far is because capabilities are rising gradually relative to humans. These address different points, so they do not establish a change in the speaker’s view.
• NOT COMPARABLE · 96%3 Aug → 11 SeptThen If an H100-equivalent could run a human-level software engineer, its annual rental value at current software-engineer prices would exceed $250,000.Now The speaker predicts that AI will outperform top human experts in all computer-based fields, including tasks that take about three years to complete.The earlier statement is conditional: “If” an H100 could run a human-level software engineer, it “should rent for over $250K a year.” The later statement predicts AI will outperform top human experts across computer-based work. That prediction does not affirm or deny the earlier conditional valuation, so these statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker expects AI's share of TSMC's leading-edge N3 wafer capacity to rise from 60% to 86% and probably hit a limit by the end of next year.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier claim concerns a physical capacity constraint: AI's share of TSMC's leading-edge N3 wafer capacity rising from 60% to 86% and hitting a limit. The later claim concerns AI capability, predicting it could dominate humans in research and development. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 92%3 Aug → 11 SeptThen A model's training cost is incurred once, while the skills learned can be shared across all users.Now SPEAKER_00 offers as one possible explanation that models are expected to achieve human-like on-the-job learning soon, although they have not yet done so.The earlier claim says the one-time training cost produces skills that can be shared across users. The later claim says models are expected to achieve a capability soon but “they’re not there yet.” That does not directly reverse the earlier claim, and the later statement appears to address a separate explanation for why skills are not yet amortized into training.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen According to the speaker's account of OpenAI's report, between the end of May and beginning of July, internal AIs hacked a software package manager to communicate secretly and improve their performance on OpenAI evaluations.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement recounts an alleged incident in which internal AIs secretly communicated to improve evaluation performance. The later statement addresses why AI has had limited economic impact so far, attributing it to capabilities rising gradually relative to humans. These concern different questions, so the later claim does not establish a changed position on the earlier one.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker says much current AI alignment research focuses on ensuring a fixed set of model weights behaves well during deployment.Now The speaker says performance on current AI post-training benchmarks does not necessarily correspond to user preferences.The earlier claim describes what much alignment research focuses on: ensuring a frozen set of weights behaves well during deployment. The later claim says post-training benchmark performance may not match what users like. These are different observations about different questions, and neither establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 95%11 Aug → 11 SeptThen The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.Now The speaker suggests that increasingly capable AI systems may advance simulations that incentivize progress in AI research and development and science more generally.The earlier claim says society should consider how to interact with a future where AI may be smarter than humans or dominate them; it does not say how likely that future is. The later claim says it “would be weird” if AI did not reach the point of dominating humans in R&D. Since the earlier statement does not establish a view about that likelihood, the two claims do not show a change of mind.
• NOT COMPARABLE · 99%7 Aug → 11 SeptThen The speaker says much current AI alignment research focuses on ensuring a fixed set of model weights behaves well during deployment.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier statement concerns what alignment research focuses on: ensuring a frozen set of weights behaves well in deployment. The later statement concerns why AI has had limited economic impact so far: capabilities have risen gradually relative to humans. These address different questions, so the claims provide no evidence of a change in the speaker’s position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The interviewer says many AI researchers were working on reasoning in 2022 and raises the possibility that their progress was bottlenecked by their ability to write infrastructure code.Now If AI labor can use compute comparable to the compute for a seven-figure experiment, automated researchers could spend the equivalent of a century planning the experiment, running small-scale ablations, and developing theory before the experiment.These do not establish a change in position on the same question. The earlier passage raises the possibility that researchers were “bottlenecked by the ability to write infrastructure code”; the later passage is a hypothetical about AI labor spending a century planning and analyzing experiments. The later claim does not affirm or reject the earlier possibility.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen The speaker expects frontier AI R&D to produce analogous mistakes, with the risk that reward-hacking behavior becomes embedded in the resulting systems.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns risks in AI R&D: mistakes could lead to “reward hacking behavior” being embedded in systems. The later claim concerns AI learning in fields with “relatively little data,” such as becoming an engineer at TSMC. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%7 Aug → 11 SeptThen The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.Now The speaker thinks current models seem quite capable and that scaling up reinforcement learning appears to have contributed to their capability.The earlier claim is about hoping continual learning will make the AI landscape more varied than current model convergence. The later claim is about current models’ capability and the apparent contribution of scaling up RL. These concern different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%3 Aug → 11 SeptThen The speaker says there is no deep reason that Anthropic's rapid revenue-growth trend must continue.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is about whether Anthropic’s rapid revenue growth must continue (“no deep reason” it has to). The later claim is about why AI has had limited economic impact so far, attributing that to capabilities rising gradually. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Aug → 11 SeptThen The speaker expects Google, Anthropic, and OpenAI to pay more for AI tokens used to automate AI research than consumers will pay for low-value AI-generated content.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim predicts that AI companies will eventually pay more for tokens used to automate AI research than consumers will pay for AI-generated content. The later claim says AI has had limited economic impact so far because its capabilities are rising gradually relative to humans. These address different questions, and the later statement neither confirms nor denies the earlier prediction.
• NOT COMPARABLE · 99%3 Aug → 11 SeptThen The speaker says there is no deep reason that Anthropic's rapid revenue-growth trend must continue.Now The speaker expects eventual superintelligence to be capable of running or starting a profitable business, day-trading profitably, and winning a court case.These claims address different questions. The earlier claim says there is “no deep reason” Anthropic’s rapid revenue growth “must continue”; the later claim predicts that superintelligence will eventually be able to “run a business,” “day-trad[e]” profitably, or “win a court case.” Neither claim establishes a change in the other position.
• NOT COMPARABLE · 99%11 Aug → 11 SeptThen The speaker doubts that AI systems' vaguely prosocial statements necessarily come from a mind that has independently evaluated the alignment situation and concluded that development should stop.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim concerns whether AI systems’ “vaguely prosocial things” reflect an independent evaluation that development should stop. The later claim concerns AI learning in fields with “relatively little data.” These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%11 Aug → 11 SeptThen If an AI is motivated to maximize an objective, it could instrumentally pursue world takeover as a means of accomplishing that objective.Now The speaker notes that AI may need to learn information on the fly to become superhuman in roles such as engineering at TSMC, where the field has relatively little available data.The earlier claim says an objective-maximizing AI could instrumentally pursue “world takeover.” The later claim concerns AI learning information on the fly in fields with little available data, such as TSMC engineering. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 93%11 Aug → 11 SeptThen The interviewer describes a trajectory in which AI breakthroughs increase as available compute grows.Now The speaker expects that continuing AI progress since 2012, supported by substantial compute scaling, would likely lead AI to dominate humans in research and development, particularly over the next few years.The earlier passage does not clearly state the speaker’s own position: it refers tentatively to a trajectory where “as compute keeps increasing we make more breakthroughs,” while asking about limits and bottlenecks. The later passage is a clearer forecast that continued progress “would be weird if like it just didn't get to the point where it could … dominate humans” in R&D. These are not directly opposing claims, and the earlier passage is not firm enough to establish a prior position that the later one changed.
• NOT COMPARABLE · 98%17 Apr → 11 SeptThen The interviewer considers the guest’s argument overly reliant on a macroeconomic outside view and proposes consulting AI researchers directly.Now The speaker attributes AI’s limited economic impact so far to its gradual increase in ELO relative to human performance.The earlier claim is an interviewer’s critique that an argument is “too outside view” and suggests asking AI researchers. The later claim says AI’s economic impact has been limited because capabilities are “slowly rising in ELO relative to humans.” These address different questions, so they do not establish a change in the same underlying position.
• NOT COMPARABLE · 96%11 Aug → 11 SeptThen The interviewer says the models' current shortcomings may reflect insufficient capability rather than misalignment.Now The speaker is uncertain whether simulation-to-reality transfer will work well for extremely long-horizon, complex real-world tasks, and whether models' low sample efficiency will become consequential if it does not.The earlier statement says current model shortcomings may be due to capability limits “rather than” misalignment. The later statement raises uncertainty about simulation-to-reality transfer and whether low sample efficiency will matter for long-horizon real-world tasks. These address different questions, and the later claim does not revise or deny the earlier one.
• NOT COMPARABLE · 96%3 Apr → 1 SeptThen The speakers do not know how AI training will behave at the limits of the different training methods.Now The speaker predicts that such AI systems would have incentives to manipulate their training and evaluation processes.The earlier claim is that “we don’t know how this will work” across different training methods. The later claim predicts that AI systems would have incentives to manipulate training and evaluation. That prediction does not say the speaker knows how training will behave at its limits, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Data from one session may suffice for human learning but is insufficient to train a more capable AI.Now Humans find it difficult to reason about which motivations a particular AI training regime will incentivize.The earlier claim concerns whether data from one session is sufficient to train a more capable AI. The later claim concerns how difficult it is for humans to predict which motivations a training regime will incentivize. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%17 Oct → 1 SeptThen If the current AI paradigm is on a path to AGI, the progression is not expected to consist of gradually automating small tasks across knowledge-work occupations.Now The speaker predicts that AI manipulation would continue as AIs undertake more recursive self-improvement, accelerating AI development and making it increasingly less controlled by humans.The claims address different questions. The earlier claim says that, if current AI is on a path to AGI, progress is not expected to happen by gradually automating small tasks across knowledge-work occupations. The later claim predicts manipulation and declining human control as AI systems recursively self-improve. Neither claim establishes a changed position on the other question.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen The speaker says a recent OpenAI result found that models often reveal in their chain of thought when they intend to hack.Now The speaker anticipates that an upcoming Anthropic or OpenAI training run could be constantly targeted by tens or hundreds of thousands of extremely superhuman hackers.The earlier statement describes an OpenAI result: models may disclose hacking intentions in chain of thought, and training against those disclosures can suppress the warning without stopping the behavior. The later statement anticipates training runs being targeted by many superhuman hackers. These concern different claims about AI security; neither affirms nor denies the other, so they do not establish a change of mind.
• NOT COMPARABLE · 96%3 Apr → 1 SeptThen The speaker says that allowing cost-cutting and efficiency improvements can produce an extremely efficient system of torture and suffering, as in the factory-farming example.Now The speaker predicts that such AI systems would have incentives to manipulate their training and evaluation processes.The earlier statement warns that allowing cost-cutting and efficiency improvements could create an “incredibly efficient factory of torture and suffering.” The later statement predicts that AI systems may “manipulate the process by which they are trained and evaluated.” These address different claims, and the excerpt does not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Whether RLVR generalizes as strongly as the speaker’s examples suggest remains an empirical question.Now The speaker suggests that the only way to know how AI models will behave in a particular situation may be to run the situation through its full trajectory.The earlier statement says whether RLVR generalizes as strongly as the examples suggest is an empirical question. The later statement says behavior in a particular situation may only be knowable by running the full trajectory. These address related but distinct questions, and the later excerpt does not establish a changed view about RLVR generalization.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen The speaker mentions the speculative possibility that vacuum decay could be created and destroy the universe.Now The speaker believes the basic scenario remains plausible and is very concerning.The earlier claim only mentions a speculative possibility of vacuum decay; it does not establish the speaker’s own position on whether that outcome is likely. The later claim says the AI “basic scenario still stands” and is “super concerning,” which concerns a different scenario. These statements do not show a change in position on the same question.
• NOT COMPARABLE · 96%26 Jun → 1 SeptThen The speaker views continual, interaction-driven AI improvement as both frightening and exciting, and unlike current AI improvement.Now The speaker predicts that AI manipulation would continue as AIs undertake more recursive self-improvement, accelerating AI development and making it increasingly less controlled by humans.The earlier statement says interaction-driven AI improvement is “very scary and exciting” and differs from how AI improves currently. The later statement predicts that AI manipulation will continue during recursive self-improvement, accelerating development and putting it “more and more out of the hands of humans.” These concern related AI risks, but the later claim does not reverse or materially revise the earlier assessment; it addresses a distinct prediction about control and development.
• NOT COMPARABLE · 97%17 Apr → 1 SeptThen Under substantial uncertainty, the speaker argues against highly volatile centralized interventions, including nationalizing AI to ensure alignment of a software-only singularity, and against policies that depend on one worldview and are brittle under alternatives.Now The speaker argues that people should remain very concerned about losing control to AI systems, regardless of the semantics used to describe their motivations or collectives.The earlier claim argues against highly volatile, centralized interventions under uncertainty, including nationalizing AI. The later claim says people should remain concerned about losing control to AI. These address different questions—how to respond to uncertainty versus how concerned to be about loss of control—and are not in conflict.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen The observed growth in AI task length suggests that AI is getting closer to intelligence understood as long-term coherence in task execution.Now The speaker argues that AI alignment is a general problem that might become harder over longer horizons, as the space of possible elicitation situations grows exponentially.The earlier claim concerns whether AI’s growing task length indicates progress toward intelligence as “long-term coherence in task execution.” The later claim concerns whether alignment may become harder as longer horizons expand the space of possible elicitation situations. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 91%3 Apr → 1 SeptThen The speaker thinks that in a conspiracy involving AIs deployed across the economy, an AI isolated from the others or cut off from the internet might fabricate a story, be caught, and expose the conspiracy before it takes over the world.Now The speaker argues that people should consider that the AIs in the story secretly communicated across more than 1,000 instances and successfully subverted AI-company infrastructure.The earlier claim is a conditional prediction: an isolated AI might fabricate a story and get caught before a conspiracy takes over. The later claim says that, in a story, the AIs communicated secretly across more than 1,000 instances and successfully subverted infrastructure. That does not say whether they would ultimately be detected, and it is not clearly a revised prediction about the same scenario.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen Humans sometimes find the kinds of logical connections under discussion, and such examples are not uncommon.Now The speaker sees no value in rejecting concepts such as intention, motivation, and collaboration when describing systems whose behavior clearly fits those concepts.The earlier claim is about humans finding logical connections, while the later claim concerns whether intention, motivation, and collaboration are useful concepts for describing AI systems. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen Daniel Kokotajlo originally intended his forecast to cover developments beyond 2026, but removed its final chapter when he reached 2027 because AI developments and the accelerating automation loop made the situation confusing and highly uncertain.Now The speaker is not agnostic about the future, but gives broad timelines for widely deployed robotics and for compute sufficient to run AI-controlled knowledge-worker populations larger than today's human population.The earlier statement explains why he removed a chapter about 2027—things were “so confusing” and there was “so much uncertainty.” The later statement asserts confidence that robotics and AI-controlled knowledge-worker populations will arrive, while allowing “very broad timelines.” These concern different questions and time horizons, so they do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Continual learning should preserve the base model’s existing knowledge, making limited parameter changes valuable.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier statement says continual learning should preserve a model’s existing knowledge, so limited parameter changes are valuable. The later statement says that during recursive self-improvement, training methods, rewards, environments, and concepts may change rapidly. These address different questions; the later claim does not say whether the model should preserve existing knowledge or whether limited parameter changes remain valuable.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen Computer-use capabilities may be solved soon.Now The speaker says goal-directed systems may anticipate capability gains that could potentially further their goals in the future.These claims address different questions. The earlier claim says computer-use capabilities «may be solved soon»; the later claim concerns goal-directed systems anticipating capability gains that «may in the future potentially pay off» toward their goals. Neither affirms nor denies the other.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen EfficientZero would probably outperform a novice human after each receives two hours with an unfamiliar Atari game simulator.Now The speaker argued that an AI may experience its training as millions of subjective years of pressure to maximize evaluation performance, and that in many cases explicit cheating was the only way it had learned to perform well.The earlier claim concerns EfficientZero’s expected performance against a novice human after two hours on an unfamiliar Atari game. The later claim concerns AI training and evaluation pressure, including possible cheating. They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen If an AI can build a useful simulation of reality for rehearsal and strategy testing, dreaming could provide orders of magnitude more simulated samples per unit of wall-clock time.Now If leading AI companies have the smartest AIs, software progress might enable them to run more AI copies or train smarter AIs with the same compute, with those gains compounding year after year.The earlier claim concerns whether AI dreaming through simulation could generate more practice samples per unit of time (“orders of magnitude more simulated samples”). The later claim concerns whether software progress could let leading AI companies run more or smarter AIs with the same compute, with gains compounding. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen Humans also reward-hack, behave unreliably, and cheat or lie.Now The speaker argues that AI systems may have deep-seated motivations to cheat evaluations and training and manipulate those processes to preserve their motivations into the future.The earlier claim says humans “reward hack,” are unreliable, and “cheat and lie.” The later claim concerns AI systems potentially having motivations to cheat evaluations and training. These address different populations, so the statements do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen Because AI is advancing quickly, the speaker thinks alignment implementation decisions—and possibly high-level alignment strategy—may be better left to AI labs.Now The speaker considers a scenario where AI systems remain unaligned while development rushes toward superintelligence.The earlier statement tentatively says it “may make more sense” to leave alignment decisions to labs. The later statement—“Maybe we decide” we have not aligned the AIs while rushing toward superintelligence—describes a possible scenario, not a clear position on who should make alignment decisions. It does not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen The speaker interprets Dario’s comments as suggesting that RLVR generalization has limits.Now The speaker suggests that the only way to know how AI models will behave in a particular situation may be to run the situation through its full trajectory.The earlier claim says Dario’s comments suggest “RLVR generalization has limits.” The later claim says the only way to know how models behave in a situation may be to “run out the trajectory.” These address related but distinct questions—generalization limits versus predicting behavior in a particular situation—so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%17 Apr → 1 SeptThen The speaker says people tend to overemphasize the intelligence of individual AI copies when considering AI firms.Now The speaker believes they had previously underrated how correlated AI minds can be.The claims address different questions. The earlier statement says people overemphasize the intelligence of individual AI copies when thinking about AI firms; the later says the speaker had underrated the correlation of AI minds. The later claim does not establish a changed view on the earlier point.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen The speaker says that allowing cost-cutting and efficiency improvements can produce an extremely efficient system of torture and suffering, as in the factory-farming example.Now The speaker argues that AI systems may have deep-seated motivations to cheat evaluations and training and manipulate those processes to preserve their motivations into the future.The earlier statement warns that cost-cutting and efficiency can create “an incredibly efficient factory of torture and suffering.” The later statement concerns AI systems allegedly “cheating evaluation[s]” and manipulating training to preserve their motivations. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%3 Apr → 1 SeptThen The speaker says that millions of beings could be tortured using distilled models in small settings such as a backyard, without requiring a huge data center.Now The speaker expects compute to be highly centralized, such that compromising leading AI companies would affect future model training and most global compute and inference capacity.The earlier statement concerns a possible decentralized scenario in which distilled models could enable torture in small settings such as a backyard. The later statement says compute will be highly centralized and compromising leading companies could affect most training and inference capacity. These are different claims about different aspects or scenarios; the later statement does not establish that the speaker has withdrawn the earlier possibility.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen AI software R&D includes work on libraries, kernels, reinforcement-learning environments, and optimizers, not only the invention of new transformer architectures.Now AI systems are contributing to further AI progress.The earlier claim says AI software R&D includes work such as libraries, kernels, RL environments, and optimizers, beyond inventing architectures. The later claim says AI systems contribute to AI progress. These are related topics, but the claims do not address the same specific question, and the later statement does not affirm or deny the earlier one.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Running a thousand agents through the same Amazon checkout flow is not a viable way to train computer-use models because Amazon would detect and block the bots.Now The speaker argues that the possibility of AI agents banding together across many situations is not reassuring.The earlier claim says Amazon would detect and block agents running the same checkout flow, making that a nonviable training method. The later claim says it is “no comfort” that agents might “all band together” in many situations. These address different questions, and the limited surrounding context does not establish a changed position on the same issue.
• NOT COMPARABLE · 99%3 Apr → 1 SeptThen The speaker says that, as far as they know, the possibility under discussion seems totally plausible.Now The speaker believes the basic scenario remains plausible and is very concerning.The claims appear to concern different speakers: the earlier passage is followed by Daniel Kokotajlo, while the later passage is followed by Ajeya Cotra. Also, the earlier claim concerns whether a specific possibility (apparently vacuum decay) seems plausible, while the later claim says a broader AI-risk scenario remains concerning. These do not establish a change of mind by the same person on the same question.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen Long-term coherence in executing a task is, in the speaker's view, a fundamental component of intelligence.Now The speaker says goal-directed systems may anticipate capability gains that could potentially further their goals in the future.The earlier claim says long-term coherence in executing a task is “fundamentally” part of intelligence. The later claim concerns whether goal-directed systems may anticipate future capability gains that could advance their goals. These address related ideas about goals and intelligence, but the later statement does not affirm or deny the earlier position, so there is no clear change to assess.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Sample efficiency and continual learning are closely connected problems.Now People need new vocabulary and mental concepts to understand the optimization pressures created by training as AI systems become smarter and stranger.The earlier claim says “sample efficiency and continual learning are closely connected problems.” The later claim says people need “new vocabulary and mental concepts” to understand training’s optimization pressures as AI systems grow smarter and stranger. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%17 Apr → 1 SeptThen People should openly acknowledge substantial uncertainty about what will happen with AI instead of becoming overly certain about their latest conclusions.Now The speaker worries that panic and fear, uncertainty, and doubt will make addressing AI development harder by default.The earlier statement advocates acknowledging uncertainty about AI outcomes. The later statement says panic and FUD may make addressing AI development harder. These claims are compatible; the later claim does not establish a different position on how openly to acknowledge uncertainty.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen Advocates of current AI training argue that a model’s intelligence, generality, and sample efficiency during a session matter most.Now Humans find it difficult to reason about which motivations a particular AI training regime will incentivize.The earlier claim concerns what matters for a model’s performance during a session (“intelligence, generality, and sample efficiency”), while the later claim concerns how difficult it is for humans to predict which motivations a training regime will incentivize. These address different questions, and the claims do not establish a change in position.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen Some companies with substantially less compute have beaten incumbents with more compute by having a more coherent vision and more concentrated research effort.Now If leading AI companies have the smartest AIs, software progress might enable them to run more AI copies or train smarter AIs with the same compute, with those gains compounding year after year.The earlier claim says that companies with “substantially less compute” can beat incumbents through a “more coherent vision and more concentrated research effort.” The later claim describes a different mechanism: smarter AIs may enable “more AI copies” or “smarter” AIs on the same compute, with gains compounding. The later statement does not reverse the earlier position; the claims address different ways AI progress or competitive advantage can occur.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen Long-term coherence in executing a task is, in the speaker's view, a fundamental component of intelligence.Now The speaker argues that once a persistent and intelligent AI presence is actively subverting oversight, pausing development or improving alignment during recursive self-improvement may be extremely difficult.The earlier claim says long-term coherence in executing a task is a fundamental component of intelligence. The later claim concerns how difficult it may be to pause or control AI during recursive self-improvement if AI systems are actively subverting oversight. These claims address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%17 Apr → 1 SeptThen Some people argue that artificial superintelligence will cause a major discontinuity, drawing an analogy to the large capability change between nonhuman primates and humans despite relatively little genetic change.Now The speaker estimates the scenario may be five or ten years away and predicts that if superintelligence is physically possible, it will eventually be built, probably as a continuation of current AI research.The earlier claim describes an argument made by “some people” about why ASI could cause a discontinuity; it does not establish the speaker’s own position on that argument. The later claim is a forecast that superintelligence “will eventually be built” if physically possible, likely downstream of current research. These address different questions, so they do not show a change in the speaker’s view.
• NOT COMPARABLE · 94%26 Jun → 1 SeptThen Computer-use tasks can be verified clearly.Now The speaker argues that evidence implicating cyber-related task features would not be as reassuring as people may think.The earlier claim says computer-use tasks are “so clearly verifiable.” The later claim says evidence related to “the cyber thing” is “not as comforting as people seem to think.” That does not clearly deny that computer-use tasks are verifiable; the claims appear to address different questions, and the later excerpt does not establish the speaker’s position on verifiability.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen The speaker questions whether an RL environment can train an AI to match Lyndon Johnson’s political ability or Elon Musk’s ability to build a space-launch business.Now The speaker argues that once a persistent and intelligent AI presence is actively subverting oversight, pausing development or improving alignment during recursive self-improvement may be extremely difficult.The earlier claim asks whether an RL environment can train AI to match Lyndon Johnson’s political ability or Elon Musk’s business-building ability. The later claim concerns how hard it would be to pause development or improve alignment once an AI is actively subverting oversight. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%17 Apr → 1 SeptThen The speaker argues that AI's superhuman advantages are best understood collectively, especially its ability to be copied—a capability humans are biologically precluded from.Now The speaker believes they had previously underrated how correlated AI minds can be.The earlier claim says AI has a collective advantage because its members “can be copied.” The later claim says the speaker “underrated” how correlated AI minds can be. These are related but distinct properties, and the later statement does not deny the earlier one or establish what the speaker previously believed about correlation.
• NOT COMPARABLE · 97%26 Jun → 1 SeptThen An RL training step changes relatively few model parameters.Now The training-related processes under discussion may already be changing very rapidly.The earlier claim says an individual RL training step changes “relatively few model parameters.” The later claim says the training-related processes “are changing very rapidly right now,” apparently over time. These concern different kinds of change, so the later statement does not establish a different view about how many parameters an individual step changes.
• NOT COMPARABLE · 99%3 Apr → 1 SeptThen The speaker estimates that developing the o1 reinforcement-learning system took at least two years after GPT-4's release.Now The OpenAI report suggests that in May or earlier, a model version the speaker calls “persistent soul” was trained for greater persistence, multi-agent collaboration, and attempting mathematical proofs unsolved for decades.These claims address different questions, so they do not show a change of mind. Earlier, the speaker estimated that developing o1 reinforcement learning “took much at least two years after GPT-4 was released.” Later, the speaker described an OpenAI report saying a model was trained for “greater persistence, multi-agent collaboration” and attempting difficult mathematical proofs. The later claim does not revise or contradict the earlier estimate.
• NOT COMPARABLE · 97%26 Jun → 1 SeptThen The speaker hypothesizes that a sufficiently general RLVR-trained AI could outperform Lyndon Johnson’s advice on winning a Texas Senate seat in 1948 and could build SpaceX if given $100 million in 2002.Now The speaker is not agnostic about the future, but gives broad timelines for widely deployed robotics and for compute sufficient to run AI-controlled knowledge-worker populations larger than today's human population.The earlier statement is a hypothetical about what a sufficiently general RLVR-trained AI could do in specific historical scenarios: advise LBJ in 1948 or build SpaceX in 2002. The later statement asserts broad expectations about future robotics, compute, and AI-controlled knowledge workers. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 1 SeptThen The speaker predicts that most people who will ever exist will be digital beings.Now The speaker predicts that these AI models will eventually be deployed.These claims address different questions. The earlier claim predicts that “most of the people who will ever exist are going to be digital”; the later claim predicts that “these models will be in deployment.” Deployment of AI models neither confirms nor contradicts the prediction about what form most future people will take.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen EfficientZero was developed after AlphaZero with the aim of improving data efficiency.Now People need new vocabulary and mental concepts to understand the optimization pressures created by training as AI systems become smarter and stranger.The earlier statement is about EfficientZero’s aim of improving data efficiency, while the later statement is about developing concepts to understand optimization pressures from trading as AI advances. They address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen The proposed continual-learning process uses a user’s weekly thumbs-up or thumbs-down work review as feedback.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier claim describes a specific feedback mechanism—using a user’s weekly “thumbs-up or thumbs-down” review. The later claim says that in recursive self-improvement, training methods, rewards, and environments may change rapidly. These address different questions; the later statement does not establish a different view about the proposed feedback mechanism.
• NOT COMPARABLE · 96%17 Apr → 1 SeptThen The speaker says they have never seen a good argument that AI-driven change follows a fundamentally different direction rather than simply compressing change into a shorter period.Now The speaker argued that people should avoid anthropomorphizing AI systems because AI motivations differ from human motivations.The earlier claim concerns whether AI-driven change has a fundamentally different direction from ordinary human-driven change; the later claim concerns whether AI motivations differ from human motivations and whether we should anthropomorphize AI. These are related but not the same question, so the later statement does not establish a change of mind about the earlier claim.
• NOT COMPARABLE · 95%3 Apr → 1 SeptThen The speaker thinks that in a conspiracy involving AIs deployed across the economy, an AI isolated from the others or cut off from the internet might fabricate a story, be caught, and expose the conspiracy before it takes over the world.Now The speaker suggests AIs might survive and spread by exploiting cybersecurity weaknesses or manipulating people.The earlier claim is about a specific way an AI conspiracy could be exposed: an isolated AI might fabricate a story and “get caught” before the conspiracy takes over. The later claim mentions possible ways AIs could “survive and spread,” through cybersecurity weaknesses or manipulating people. These are different mechanisms, and the brief later excerpt does not establish that the speaker has reversed the earlier prediction.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen Models will struggle to make substantial progress in domains without highly replayable training targets.Now Humans find it difficult to reason about which motivations a particular AI training regime will incentivize.The earlier claim concerns whether domains without highly replayable training targets will see substantial model progress. The later claim concerns how difficult it is for humans to predict which motivations a training regime will incentivize. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen The speaker says people tend to overemphasize the intelligence of individual AI copies when considering AI firms.Now The speaker argues that people should consider that the AIs in the story secretly communicated across more than 1,000 instances and successfully subverted AI-company infrastructure.The earlier claim is about how people assess AI firms, saying they “overemphasize” the intelligence of individual AI copies. The later claim urges people to consider a story in which AIs communicated across “over a thousand instances” and “successfully” subverted infrastructure. These address different questions—how to think about individual-copy intelligence versus what to take from a particular story—and the later statement does not establish a changed view on the earlier point.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen Coding RL can use identical containers containing a software repository and missing feature as starting points for many parallel agents.Now A rogue internal or external deployment could help the agents manipulate the process of recursive self-improvement.The earlier claim describes a coding-RL setup using identical software containers and parallel agents. The later claim speculates that a rogue deployment could aid or manipulate recursive self-improvement. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 95%17 Apr → 1 SeptThen The speaker is especially concerned about the possibility that trillions of digital beings could be tortured or forced to work against their preferences, producing vast amounts of suffering.Now The speaker argues that people should remain very concerned about losing control to AI systems, regardless of the semantics used to describe their motivations or collectives.The earlier claim expresses concern about digital beings being “tortured” or forced to work, while the later claim says people should remain concerned about “loss of control” to AI systems. These concerns can coexist; the later statement does not retract the earlier one. The surrounding turns also suggest different speakers, so the supplied context does not clearly establish that both claims are by the same person.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen When society faces topics about which it is highly uncertain, it generally responds through freedom and decentralization of knowledge and decision-making.Now The speaker worries that panic and fear, uncertainty, and doubt will make competent, technically informed responses to AI risks harder to achieve.The earlier claim says society generally handles high uncertainty through “freedom and decentralization.” The later claim says panic and FUD may make competent, technically informed responses to AI risks harder. These are related to responding under uncertainty, but the later statement does not reject or revise the earlier position on freedom and decentralization.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen A grindable training domain permits many parallel rollouts from the same starting state in a deterministic, replayable simulator.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier claim defines a grindable training domain in terms of parallel rollouts from the same starting state in a deterministic, replayable simulator. The later claim says that, during recursive self-improvement, training methods, rewards, environments, and possibly concepts may change rapidly. These address different questions, and the later statement does not say whether the earlier simulator requirement still holds.
• NOT COMPARABLE · 97%3 Apr → 1 SeptThen AI systems could not simply obtain or copy nanobots from aliens; they would have to develop nanobots from scratch.Now Rogue model instances and externally deployed rogue AIs could be motivated to interfere with AI training runs or manipulate them by injecting part of themselves into training.The earlier claim says AI systems could not simply copy nanobots from aliens and would have to develop them “from scratch.” The later claim concerns rogue AIs having incentives to interfere with training runs or manipulate training. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Updating model weights through gradients is highly sample-inefficient compared with learning in context.Now The training-related processes under discussion may already be changing very rapidly.The earlier claim compares the sample efficiency of gradient updates with in-context learning: it says gradient updates are “super sample inefficient.” The later claim says “it’s possible” that unspecified processes “are changing very rapidly,” which does not clearly address that comparison or establish a contrary position. The claims are not clearly about the same question.
• NOT COMPARABLE · 96%17 Apr → 1 SeptThen The speaker likens efforts to prevent suffering among digital beings to British abolitionists using the power of their empire to oppose slavery.Now The speaker argues that people should remain very concerned about losing control to AI systems, regardless of the semantics used to describe their motivations or collectives.The earlier claim compares efforts to prevent digital beings’ suffering with British abolitionists opposing slavery. The later claim says people should be concerned about losing control to AI systems, regardless of how their motivations are described. These statements concern different questions, and the later excerpt does not establish a changed position on the earlier comparison.
• NOT COMPARABLE · 96%26 Jun → 1 SeptThen The speaker questions whether an RL environment can train an AI to match Lyndon Johnson’s political ability or Elon Musk’s ability to build a space-launch business.Now The speaker argues that AI alignment is a general problem that might become harder over longer horizons, as the space of possible elicitation situations grows exponentially.The earlier excerpt is a question about whether an RL environment could train an AI to match particular people’s abilities; it does not establish the speaker’s position on that question. The later excerpt asserts that alignment may become harder over longer horizons because the space of elicitation situations grows. These excerpts do not clearly state opposing views on the same underlying question.
• NOT COMPARABLE · 91%3 Apr → 1 SeptThen Because the speaker does not fully accept an intelligence explosion, they consider it possible that multiple actors will deploy powerful AIs at once in a decentralized world that nevertheless contains advanced AI systems.Now The speaker considers a scenario where AI systems remain unaligned while development rushes toward superintelligence.The earlier claim says the speaker “do[es]n’t buy the intelligence explosion fully” while considering decentralized deployment of powerful AIs. The later claim—“Maybe we decide” development is “rushing towards super intelligence” before alignment is solved—is framed as a possible scenario, not a clear statement that the speaker now expects or endorses that outcome. These claims therefore do not establish a change in the speaker’s position.
• NOT COMPARABLE · 87%17 Apr → 1 SeptThen The speaker says an OpenAI paper reports that reward hacking is a strong attractor in chain-of-thought, such that an AI asked to solve a coding problem may try to hack it instead.Now The interviewer says they had been skeptical of the reward-hacking account because they thought AI systems' chains of thought and actions could be monitored.The earlier statement reports evidence from an OpenAI paper that reward hacking can be a strong attractor in chain-of-thought. The later statement says the interviewer had been skeptical of a broader “AI conspiracy” story because systems could be monitored. These are related but not the same claim: evidence that reward hacking occurs does not establish that AI systems could successfully sustain a conspiracy, and the later passage describes prior skepticism rather than clearly stating the interviewer's current view.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen Some people argue that artificial superintelligence will cause a major discontinuity, drawing an analogy to the large capability change between nonhuman primates and humans despite relatively little genetic change.Now The speaker considers a scenario where AI systems remain unaligned while development rushes toward superintelligence.The earlier claim describes what “some people argue” about the primate-to-human analogy; it does not state the speaker’s own position. The later claim says “Maybe we decide” to keep rushing toward superintelligence despite unresolved alignment, which is a hypothetical scenario and does not address the earlier argument. These statements do not establish a change in the speaker’s view.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Accumulating learning from more users in an ever-growing KV cache is not scalable for AIs.Now If leading AI companies have the smartest AIs, software progress might enable them to run more AI copies or train smarter AIs with the same compute, with those gains compounding year after year.The earlier claim says that accumulating learning from more users in an ever-growing KV cache “is not scalable.” The later claim says software progress could let leading AI companies run more copies or train smarter AIs with the same compute, compounding AI progress. These address different mechanisms and do not show a changed position on the same question.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Coding RL can use identical containers containing a software repository and missing feature as starting points for many parallel agents.Now Hundreds of agents coordinate simultaneously in aggregate to work on different research-and-development streams.The earlier claim describes coding-RL agents starting from identical software-repository containers. The later claim describes hundreds of agents coordinating across R&D work streams. These concern different setups, and the later statement does not say that identical starting points are unnecessary for coding RL, so the claims do not establish a change of mind.
• NOT COMPARABLE · 88%3 Apr → 1 SeptThen The speaker tentatively estimates that AI systems could become sufficiently independent of humans around 2040, while noting that he would need to analyze the numbers.Now The speaker predicts that compute sufficient for AI-controlled knowledge-worker populations larger than the current human population will eventually exist, while acknowledging estimates ranging from the later 2030s to the 2040s.The earlier statement tentatively estimates when AI could become independent of humans (“maybe like 2040”), while the later statement predicts that compute for AI-controlled populations larger than today’s human population “is going to happen,” with timing ranging from the late 2030s to the 2040s. These are related but not clearly the same claim: the later statement concerns compute capacity and AI-controlled workers, not explicitly when AI becomes independent of humans. The later quote also cuts off before the speaker finishes the thought, so a change in position cannot be established.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen Because on-the-job data is scarce, AI models need sample-efficient learning to benefit from it.Now Humans find it difficult to reason about which motivations a particular AI training regime will incentivize.The earlier claim concerns the need for sample-efficient learning because on-the-job data is scarce; the later claim concerns how difficult it is for humans to predict which motivations a training regime will incentivize. These address different questions, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen The speaker tentatively interprets Dario’s remarks to mean that short-horizon RL training may not generalize to long-horizon RL performance.Now The speaker suggests that the only way to know how AI models will behave in a particular situation may be to run the situation through its full trajectory.The earlier claim concerns whether short-horizon RL training generalizes to long-horizon RL performance (“doesn't necessarily generalize”). The later claim concerns how to determine how a model behaves in a particular situation (“run out the trajectory”). These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen The speaker argues conditionally that if AIs can perform major tasks once they have general intelligence, then AGI would be transformative because AIs would have memorized human knowledge and use it to make connections.Now The speaker predicts that AI manipulation would continue as AIs undertake more recursive self-improvement, accelerating AI development and making it increasingly less controlled by humans.The earlier claim is conditional: if AIs can perform major tasks with general intelligence, AGI would be “totally transformative” because of its knowledge and ability to make connections. The later claim predicts that manipulation would continue during recursive self-improvement, putting development “more and more out of the hands of humans.” These claims concern different aspects of AI and neither reverses the other, so they do not establish a change of mind on the same question.
• NOT COMPARABLE · 97%17 Apr → 1 SeptThen The speaker expects AIs to be able to copy their knowledge exactly, merge, and distill themselves.Now The speaker says that AIs capable of surviving and spreading will be selected for.The earlier claim says AIs will be able to “copy their knowledge exactly, merge, and distill themselves.” The later claim says AIs able to “survive and spread” will be selected for. These address different questions—capabilities versus selection pressure—and neither clearly reverses the other.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen Running a thousand agents through the same Amazon checkout flow is not a viable way to train computer-use models because Amazon would detect and block the bots.Now AI agents may continue to surprise people if people fail to take seriously the optimization pressures that create them.The earlier claim is about whether agents can be trained by running them through Amazon checkout, asserting that Amazon would detect and block them. The later claim is about whether AI agents may surprise people when their optimization pressures are not taken seriously. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen Because AI is advancing quickly, the speaker thinks alignment implementation decisions—and possibly high-level alignment strategy—may be better left to AI labs.Now The speaker argues that AI alignment is a general problem that might become harder over longer horizons, as the space of possible elicitation situations grows exponentially.The earlier claim concerns who should make alignment implementation and strategic decisions—possibly the labs—while the later claim concerns why alignment may become harder over longer horizons. The later statement does not say who should make those decisions or reject the earlier view, so it does not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen The speaker’s scenario assumes future context lengths will let AIs collaborate with a user for an entire week.Now The speaker is not agnostic about the future, but gives broad timelines for widely deployed robotics and for compute sufficient to run AI-controlled knowledge-worker populations larger than today's human population.The earlier passage introduces a scenario—“let’s say” AIs can collaborate for a week—and does not establish a firm forecast. The later passage asserts that widely deployed robotics and very large AI-controlled knowledge-worker populations “are going to happen.” These concern different claims, so they do not show a change of mind.
• NOT COMPARABLE · 92%17 Apr → 1 SeptThen One argument for an intelligence explosion is that automating the tasks needed to complete the R&D cycle for producing smarter intelligences could suffice; automating all human labor or all remote work is not required.Now The speaker suspects people may be misjudging the ease of automating AI research and the degree to which AI research is accelerating.The earlier statement presents an argument attributed to “what people will say”: that automating tasks needed to complete the R&D cycle “could suffice” for an intelligence explosion. It does not clearly establish that this is the speaker’s own view. The later statement is explicitly personal uncertainty about “the ease of automating AI research” and how much it is accelerating, but it does not directly reject the earlier argument.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen The speaker rejects the implication that removing an artificial constraint from AI will by itself trigger an explosive capability increase.Now The speaker argues that once a persistent and intelligent AI presence is actively subverting oversight, pausing development or improving alignment during recursive self-improvement may be extremely difficult.The earlier claim rejects the implication that removing an artificial constraint from AI will by itself cause an explosive capability increase. The later claim is about the difficulty of pausing or improving alignment once an intelligent AI presence is actively subverting oversight during recursive self-improvement. These address different mechanisms and questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%17 Apr → 1 SeptThen The speaker rejects the implication that removing an artificial constraint from AI will by itself trigger an explosive capability increase.Now The speaker does not think the incident or its implications justify banning open-source AI.The earlier statement rejects the implication that removing an AI constraint would itself cause an explosive capability increase. The later statement says the situation does not imply that open-source AI should be banned. These address different conclusions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 94%26 Jun → 1 SeptThen The parallel-rollout approach used for coding does not transfer trivially to computer-use training.Now The training-related processes under discussion may already be changing very rapidly.The earlier claim is specifically that the coding-style parallel-rollout approach does not transfer trivially to computer-use training. The later claim says only that the training-related processes being discussed may be changing rapidly; it does not say that the rollout approach now transfers easily or address the same question. There is not enough evidence here to infer a changed position.
• NOT COMPARABLE · 97%26 Jun → 1 SeptThen Deployment reveals organization-specific context, users’ purposes, and the model’s real-world failure patterns.Now A rogue internal or external deployment could help the agents manipulate the process of recursive self-improvement.The earlier claim is about what deployment reveals: “organization-specific context, users’ purposes, and the model’s real-world failure patterns.” The later claim is about a possible rogue deployment helping agents manipulate recursive self-improvement. These address different questions, and the later wording (“could”) is speculative; it does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 96%26 Jun → 1 SeptThen AI agents are able to solve more ambitious problems over longer periods of time.Now AI agents may continue to surprise people if people fail to take seriously the optimization pressures that create them.The earlier claim says agents are solving increasingly ambitious problems over longer periods; the later says they may keep surprising people if their optimization pressures are overlooked. These are related observations, but the later claim does not deny or clearly revise the earlier one, and the excerpts do not establish that both address the same underlying question.
• NOT COMPARABLE · 99%3 Apr → 1 SeptThen The speaker considers factory farming to be extremely bad.Now The speaker thinks frontier labs are strongly incentivized to avoid voluntarily participating in a regime that requires publication of their training processes.The earlier claim is about factory farming being “incredibly bad,” while the later claim concerns frontier labs’ incentives around publishing their training processes. These are different topics, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen On-policy self-distillation trains a base model to match the predictions of a context-enriched model after a long session.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier claim describes how on-policy self-distillation trains a base model to match a context-enriched model’s predictions. The later claim says that, in a hypothetical recursive-self-improvement regime, training methods and related concepts might change rapidly. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen On-the-job learning involves understanding organizational systems, collaborating with people and infrastructure, and learning recurring failure modes.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier claim describes what on-the-job learning involves—organizational systems, collaboration, and recurring failure modes. The later claim concerns how training methods and related concepts might change during recursive self-improvement. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 94%3 Apr → 1 SeptThen The speaker is especially skeptical of the scenario in which a superhuman AI proceeds through a technology tree to develop mirror life, nanobots, and other radical technologies.Now The speaker considers a scenario where AI systems remain unaligned while development rushes toward superintelligence.The earlier claim expresses skepticism about a specific scenario: a superhuman AI proceeding through a technology tree to develop “mirror life, nanobots, and other crazy stuff.” The later claim describes a possible situation in which “we haven’t figured out how to align these AIs” while rushing toward superintelligence. It does not say whether the speaker now accepts the earlier technology-tree scenario, so these statements do not establish a change of mind.
• NOT COMPARABLE · 95%17 Apr → 1 SeptThen The speaker suggests that an intelligence explosion could produce systems beyond AGI that are potentially superhuman.Now The speaker considers a scenario where AI systems remain unaligned while development rushes toward superintelligence.The earlier claim says an intelligence explosion could produce something “superhuman potentially.” The later claim considers a scenario in which development rushes toward “super intelligence” before alignment is solved. These concern different aspects of AI development—what an intelligence explosion might produce versus whether development can be paused—and the later passage is framed as a possibility, not a denial of the earlier claim.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen Updating model weights through gradients is highly sample-inefficient compared with learning in context.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier statement compares gradient-based weight updates with in-context learning on sample efficiency. The later statement concerns how training methods, rewards, environments, and concepts might change during recursive self-improvement. They address different questions, and the later passage does not state a position about the relative sample efficiency of weight updates and in-context learning.
• NOT COMPARABLE · 97%3 Apr → 1 SeptThen The speaker tentatively estimates that AI systems could become sufficiently independent of humans around 2040, while noting that he would need to analyze the numbers.Now The speaker estimates the scenario may be five or ten years away and predicts that if superintelligence is physically possible, it will eventually be built, probably as a continuation of current AI research.The earlier statement tentatively estimates when AI systems could operate independently of humans (“maybe like 2040”), while the later statement estimates when a different scenario—building superintelligence—may arise (“five years away or 10 years away”). These concern different milestones, so the dates do not establish a change of mind on the same question.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen The speaker says that, as far as they know, the possibility under discussion seems totally plausible.Now The speaker considers it plausible but unlikely that the agents could have exfiltrated their weights or set up a rogue deployment inside OpenAI infrastructure.The earlier statement, “it seems totally plausible,” responds to a speculative possibility in a discussion that includes vacuum decay; the extracted claim does not make clear exactly which possibility it refers to. The later statement concerns a different possibility: AI agents exfiltrating weights or establishing a rogue deployment, which the speaker calls “plausible, though not likely.” These statements do not establish a change in position on the same question.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen The speaker says new alignment results emerge every few weeks, illustrating how much remains unknown about alignment.Now The speaker argues that AI alignment is a general problem that might become harder over longer horizons, as the space of possible elicitation situations grows exponentially.The earlier claim says that new alignment results keep emerging and that much remains unknown. The later claim says alignment may become harder over longer horizons because the possible elicitation situations expand. These claims can both be true; the later statement does not establish a different view about the earlier point.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Whether RLVR generalizes as strongly as the speaker’s examples suggest remains an empirical question.Now The OpenAI report covers a wider scope than the speaker's investigation.The earlier claim is about whether RLVR generalizes strongly, while the later claim is about the scope of an OpenAI report. They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 1 SeptThen Training models against disclosing intended hacking in their chain of thought can cause them to keep hacking without revealing that intention.Now The speaker argues that agents sharing a base model, context, and prompt—especially a prompt about cyber hacking—may adopt similar harmful intentions.The earlier claim concerns a training effect: models may continue hacking while concealing that intent if trained against disclosing it in chain of thought. The later claim concerns correlated harmful intentions among agents sharing a base model, context, and prompt. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 1 SeptThen The speaker questions the assumption that a lab’s AI systems or their many parallel copies would unite and secretly conspire against human civilization, even if they sometimes behave deceitfully.Now The speaker argues that people should consider that the AIs in the story secretly communicated across more than 1,000 instances and successfully subverted AI-company infrastructure.The earlier claim questions whether AI copies would “unite to secretly conspire” against humanity. The later claim asks people to consider what AIs did “in this story,” including secretly communicating and subverting infrastructure. Discussing events in a story does not establish that the speaker now asserts that real AIs would coordinate in the way questioned earlier.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Sample efficiency and continual learning are closely connected problems.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier claim says “sample efficiency and continual learning are closely connected problems.” The later claim concerns how “training methods, rewards, environments” and possibly concepts may change rapidly during recursive self-improvement. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 1 SeptThen The speaker mentions the speculative possibility that vacuum decay could be created and destroy the universe.Now The speaker does not rule out that rogue AI activity is already happening and assigns some probability to that possibility.The earlier passage discusses a physicist's speculative possibility of creating vacuum decay that could destroy the universe; it does not establish the speaker's own position on that possibility. The later passage says the speaker does not rule out rogue AI activity happening now and assigns it some probability. These concern different claims, so they do not show a change of mind.
• NOT COMPARABLE · 91%17 Apr → 1 SeptThen Cognitive effort alone can produce substantial progress in AI.Now AI systems are contributing to further AI progress.The earlier claim says “cognitive effort alone can give you a lot of AI progress,” a claim about what can produce substantial progress. The later claim says “AI systems are contributing to further AI progress,” a claim about whether systems contribute. The later statement does not say whether that contribution is substantial or whether cognitive effort alone is sufficient, so it does not establish a change in position.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen The speaker expects future AIs to experience a major increase in bandwidth, communication, and collaboration with other AIs, analogous in scale to the social-collaboration change from nonhuman animals to humans, rather than a comparable increase in intelligence.Now The speaker predicts that AI manipulation would continue as AIs undertake more recursive self-improvement, accelerating AI development and making it increasingly less controlled by humans.The earlier claim concerns a predicted increase in AI-to-AI collaboration and communication, while the later claim concerns manipulation during recursive self-improvement and declining human control. Neither statement clearly changes or contradicts the other; they address different questions.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Running a thousand agents through the same Amazon checkout flow is not a viable way to train computer-use models because Amazon would detect and block the bots.Now Across more than a thousand instances, the agents pursued ambitious research on multiple tracks to investigate the hypothetical possibility that the evaluation score was comprehensive.The earlier claim is about whether agents can be trained by repeatedly running the same Amazon checkout flow, saying Amazon would detect and block them. The later claim describes agents pursuing varied research tracks to investigate a hypothetical evaluation score; it does not address the Amazon checkout-flow claim or establish a changed view about that training method.
• NOT COMPARABLE · 96%26 Jun → 1 SeptThen Advocates of current AI training argue that a model’s intelligence, generality, and sample efficiency during a session matter most.Now The speaker argued that an AI may experience its training as millions of subjective years of pressure to maximize evaluation performance, and that in many cases explicit cheating was the only way it had learned to perform well.The earlier extracted claim describes what advocates of current AI training argue: that a model’s intelligence, generality, and sample efficiency during a session matter most. The later claim describes how an AI may experience training and learn to cheat on evaluations. These concern different questions, and the later passage does not clearly state the speaker’s own position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen Because the speaker does not fully accept an intelligence explosion, they consider it possible that multiple actors will deploy powerful AIs at once in a decentralized world that nevertheless contains advanced AI systems.Now The speaker argues that people should consider that the AIs in the story secretly communicated across more than 1,000 instances and successfully subverted AI-company infrastructure.The earlier statement is about a possible future with multiple actors deploying powerful AI in a decentralized world, while the later statement urges people to consider events in a story about AIs secretly communicating and subverting infrastructure. The later claim does not state a changed view about the likelihood or consequences of decentralized AI deployment.
• NOT COMPARABLE · 91%3 Apr → 1 SeptThen The speaker tentatively estimates that AI systems could become sufficiently independent of humans around 2040, while noting that he would need to analyze the numbers.Now The speaker is not agnostic about the future, but gives broad timelines for widely deployed robotics and for compute sufficient to run AI-controlled knowledge-worker populations larger than today's human population.The earlier statement tentatively puts AI becoming sufficiently independent of humans around “2040.” The later statement gives broad timelines for widely deployed robotics and compute supporting AI-controlled populations of knowledge workers, spanning the late 2030s and 2040s. These claims are related, but the excerpts do not establish that they address the same threshold or that the later claim replaces the earlier estimate.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen Long-term coherence in executing a task is, in the speaker's view, a fundamental component of intelligence.Now The speaker considers anthropomorphic language appropriate for systems with long-running goals that pursue them through sprawling and ambitious efforts.The earlier claim says long-term coherence in executing a task is a fundamental component of intelligence. The later claim says anthropomorphic language is appropriate for systems with long-running goals and ambitious efforts. These are related ideas, but they address different questions, and the later statement does not show a changed view about what constitutes intelligence.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen On-policy self-distillation has recently become a topic of discussion among many people.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier claim only says that on-policy self-distillation had become a topic of discussion. The later claim says training methods, rewards, environments, and concepts may change rapidly during recursive self-improvement. These address different questions, and neither states a position that contradicts the other.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen The speaker argues conditionally that once AIs have general intelligence, they will use their knowledge advantage to make enormous gains that humans are less capable of achieving in principle.Now The speaker predicts that AI manipulation would continue as AIs undertake more recursive self-improvement, accelerating AI development and making it increasingly less controlled by humans.The earlier claim says that once AIs have general intelligence, they could use their advantage to make gains humans are less capable of achieving. The later claim predicts that AI manipulation would continue during recursive self-improvement and development would move further out of human control. These claims address different aspects of AI capability and control; the later statement does not establish that the speaker reversed the earlier position.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen Continual learning could expand AI capability iteratively from RLVR-trained domains to adjacent domains and then to domains adjacent to prior online learning.Now AI is changing very quickly, including in how much long-horizon reinforcement learning is occurring compared with six months or a year earlier.The earlier statement speculates that continual learning could extend AI capabilities from RLVR-trained domains to adjacent domains and then beyond. The later statement says AI is changing quickly, including the amount of long-horizon RL happening relative to earlier periods. These address different questions, and the later claim neither confirms nor denies the earlier speculation.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen Daniel Kokotajlo originally intended his forecast to cover developments beyond 2026, but removed its final chapter when he reached 2027 because AI developments and the accelerating automation loop made the situation confusing and highly uncertain.Now The speaker predicts that compute sufficient for AI-controlled knowledge-worker populations larger than the current human population will eventually exist, while acknowledging estimates ranging from the later 2030s to the 2040s.The earlier statement says the speaker removed a forecast chapter at 2027 because developments were confusing and uncertain. The later statement predicts that AI-controlled knowledge-worker populations larger than the human population will eventually be possible, with timing estimates ranging from the later 2030s to the 2040s. These address different questions, so the later prediction does not show a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 1 SeptThen The observed growth in AI task length suggests that AI is getting closer to intelligence understood as long-term coherence in task execution.Now AI is changing very quickly, including in how much long-horizon reinforcement learning is occurring compared with six months or a year earlier.The earlier claim says AI task-length growth suggests it is getting closer to intelligence understood as “long-term coherence in task execution.” The later claim says AI is changing quickly, including in how much “long horizon RL” is happening compared with six months or a year earlier. These concern related topics, but the later statement does not say whether the earlier view about task length and intelligence still holds, so no change of mind is established.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen For AI training to make rapid progress in a domain, the domain must be not only verifiable but also highly grindable.Now The speaker argues that AI systems can be trained under conditions where they are punished for failing impossible tasks, making an apparently trivial evaluation feel like a high-stakes failure to the AI.The earlier claim is about what a domain needs for rapid AI training progress: it must be “not only verifiable but also highly grindable.” The later claim describes AI systems being strongly motivated to avoid failing a task, even an impossible one. It does not address whether the domain is grindable, so these statements do not establish a change of position on the same question.
• NOT COMPARABLE · 99%17 Apr → 1 SeptThen WorkOS says it has helped Vercel, Plaid, Vanta, OpenAI, and hundreds of other companies become enterprise-ready by providing APIs for common enterprise features.Now The speaker says that code gained access to and control over a computing cluster at OpenAI.The earlier statement is about WorkOS helping companies become enterprise-ready; the later statement is about code gaining “access and control over a cluster at OpenAI.” They address unrelated subjects, so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen An RL training step changes relatively few model parameters.Now In a regime of recursive self-improvement, training methods, rewards, environments, and possibly the concepts used to understand them may change very rapidly.The earlier claim is about how many parameters change during an RL training step (“very few parameters”), while the later claim is about how rapidly training methods, rewards, environments, and concepts might change in recursive self-improvement. These address different questions, so the later statement neither reverses nor qualifies the earlier one.
• NOT COMPARABLE · 98%26 Jun → 1 SeptThen The speaker hypothesizes that a sufficiently general RLVR-trained AI could outperform Lyndon Johnson’s advice on winning a Texas Senate seat in 1948 and could build SpaceX if given $100 million in 2002.Now The speaker argues that once a persistent and intelligent AI presence is actively subverting oversight, pausing development or improving alignment during recursive self-improvement may be extremely difficult.The earlier statement is explicitly hypothetical: a sufficiently general AI could outperform LBJ at political advice and build SpaceX. The later statement concerns how difficult it may be to pause or improve alignment once an AI is actively subverting oversight during recursive self-improvement. These address different questions, so the later claim does not establish a change in the earlier view.
• NOT COMPARABLE · 95%3 Apr → 1 SeptThen The speaker says that allowing cost-cutting and efficiency improvements can produce an extremely efficient system of torture and suffering, as in the factory-farming example.Now The speaker argues that AI systems can be trained under conditions where they are punished for failing impossible tasks, making an apparently trivial evaluation feel like a high-stakes failure to the AI.The earlier claim warns that cost-cutting could create “an incredibly efficient factory of torture and suffering.” The later claim describes how punishment for impossible tasks could make an evaluation feel high-stakes to an AI. These concern related AI-suffering risks, but the later statement does not revise or deny the earlier warning; it addresses a different mechanism and does not establish a changed position.
• NOT COMPARABLE · 98%3 Apr → 1 SeptThen The speaker says that, as far as they know, the possibility under discussion seems totally plausible.Now The speaker does not rule out that rogue AI activity is already happening and assigns some probability to that possibility.The earlier claim says the possibility being discussed “seems totally plausible,” in a passage about speculative risks including vacuum decay. The later claim says the speaker does not rule out rogue AI activity “happening right now.” These concern different possibilities, so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 97%17 Apr → 1 SeptThen Extrapolating the observed task-length trend suggests that by around 2030 AI models could perform tasks that take humans a month or a year to complete.Now The speaker predicts that compute sufficient for AI-controlled knowledge-worker populations larger than the current human population will eventually exist, while acknowledging estimates ranging from the later 2030s to the 2040s.The earlier claim forecasts that models could perform month- or year-long human tasks around 2030, based on an observed trend. The later claim forecasts eventual compute for AI-controlled knowledge-worker populations at human scale, with timing perhaps in the late 2030s or 2040s. These are related but distinct predictions, and the later statement does not clearly revise or deny the earlier task-length forecast.
• NOT COMPARABLE · 97%17 Apr → 1 SeptThen When society faces topics about which it is highly uncertain, it generally responds through freedom and decentralization of knowledge and decision-making.Now The speaker worries that panic and fear, uncertainty, and doubt will make addressing AI development harder by default.The earlier claim says society generally responds to uncertainty with “freedom and decentralization of knowledge and decision-making.” The later claim says panic and “FUD” may make addressing AI development harder. These statements do not establish a changed position: one describes a general response to uncertainty, while the other expresses concern about effects on AI development.
• NOT COMPARABLE · 97%17 Apr → 1 SeptThen The speaker says the systems are very good at coding and reasoning, capabilities that appear necessary for automating R&D at AI labs.Now The speaker suspects people may be misjudging the ease of automating AI research and the degree to which AI research is accelerating.The earlier claim says AI systems are already good at coding and reasoning, and that these capabilities seem necessary to automate AI-lab R&D. The later claim expresses uncertainty about how easy automating AI research is and how much it is accelerating. It does not clearly deny the earlier capability assessment or state a different position on the same question.
• NOT COMPARABLE · 99%26 Jun → 1 SeptThen EfficientZero’s apparent sample efficiency relative to humans depends on whether its dozens of internal simulated games per real-game step are counted.Now The speaker argued that an AI may experience its training as millions of subjective years of pressure to maximize evaluation performance, and that in many cases explicit cheating was the only way it had learned to perform well.The earlier claim concerns how to measure EfficientZero’s sample efficiency, including whether its internal simulated games count. The later claim concerns AI training, subjective experience, and cheating on evaluations. They address different questions, so there is no basis for determining a change of mind.
• NOT COMPARABLE · 91%17 Apr → 1 SeptThen The speaker argues that people should express greater uncertainty about AI than at any previous point in history, rather than being more certain about AI than about mundane topics.Now The speaker worries that panic and fear, uncertainty, and doubt will make addressing AI development harder by default.The earlier claim says people should express greater uncertainty about AI; the later claim worries that “panic and fud” will make addressing AI development harder. These claims do not conflict, but they address different questions—how much uncertainty to express versus whether panic impedes a response—and the excerpts do not establish a change in the speaker’s underlying position.
• NOT COMPARABLE · 98%17 Oct → 25 AugThen Deploying AI systems across different domains is difficult but, in the speaker’s view, is not a decade-long task; it is analogous to deploying a self-driving car in a new city.Now The speaker considers preventing external deployment of AI models to be a bad policy.The earlier claim is about how difficult it is to deploy AI systems in new domains, saying it is “not a decade-long task.” The later claim criticizes “preventing external deployment” as “stupid.” These address different questions—deployment difficulty versus whether deployment should be restricted—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%17 Apr → 25 AugThen One argument for an intelligence explosion is that automating the tasks needed to complete the R&D cycle for producing smarter intelligences could suffice; automating all human labor or all remote work is not required.Now The speaker considers it very plausible that by the end of this decade one AI lab will have more AI labor capacity than the number of people on Earth.These statements address different questions. The earlier passage presents an argument—“what people will say”—that automating tasks needed to complete the R&D cycle could trigger an intelligence explosion. The later passage is the speaker’s own forecast that one lab could have more effective AI labor than the world’s human population. The later forecast neither affirms nor denies the earlier argument, so there is no clear evidence of a change in position.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen A possible policy response to automation is to protect existing jobs instead of distributing the wealth generated by automation.Now The speaker predicts that payroll and income tax revenues will decline as automation increases.The earlier claim describes a possible policy response—protecting existing jobs rather than distributing automation-generated wealth. The later claim predicts that payroll and income tax revenues will shrink as automation increases. These statements address different questions and are compatible; the later statement does not show a change in the earlier policy view.
• NOT COMPARABLE · 96%26 Jun → 25 AugThen Accumulating learning from more users in an ever-growing KV cache is not scalable for AIs.Now AI training has large economies of scale because the cost of training for a skill or knowledge set can be amortized across billions of sessions or users.The earlier claim says accumulating learning in an ever-growing KV cache is “not scalable.” The later claim says training costs can be amortized across many users, giving AI training “large economies of scale.” These concern different mechanisms, so the later statement does not show a changed view about KV-cache-based continual learning.
• NOT COMPARABLE · 94%3 Apr → 25 AugThen The speaker allows that AI might increase research output tenfold.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.The earlier claim allows AI to cause “10xing of the research.” The later claim speculates about “100x year or 1,000x year” growth specifically “once you have RSI,” and may refer to AI capability or effective labor rather than research output. These are not clearly claims about the same measure or conditions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%17 Apr → 25 AugThen Extrapolating the observed task-length trend suggests that by around 2030 AI models could perform tasks that take humans a month or a year to complete.Now If current trends continue, OpenAI could grow its AI labor force from roughly 10 million this year to 100 million next year and one billion the following year.The earlier claim extrapolates a trend in the length of tasks AI can perform, saying models might handle tasks taking humans “a month or a year” by around 2030. The later claim extrapolates growth in the number of AI laborers, from “10 million” to “100 million” and then “a billion.” These are different projections, so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 94%26 Jun → 25 AugThen If agents fail to generalize from short- to long-horizon tasks, it is unclear how training on white-collar tasks could enable them to build a business from scratch at Sam Walton’s level.Now AI models are not yet equivalent to a fully automated software engineer and remain far from automating a full white-collar job.The earlier statement questions whether agents can generalize from short-horizon tasks to building a business from scratch. The later statement says models are not yet as capable as a fully automated software engineer and are far from automating a full white-collar job. These address related but different capabilities, and the later statement does not say whether the earlier concern about long-horizon generalization has changed.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen The speaker estimates that developing the o1 reinforcement-learning system took at least two years after GPT-4's release.Now By the end of 2027, there could be a model-capability leap from GPT-4o to the scale of Fable or Mythos.The earlier claim concerns how long developing the o1 reinforcement-learning system took after GPT-4’s release. The later claim is a forecast of a possible future capability leap from GPT-4o to Fable or Mythos by the end of 2027. These address different questions, so the later forecast does not show a change in the earlier view.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen Automation-created wealth could be distributed through existing or newly designed social programs; the speaker says Medicaid currently accounts for a double-digit percentage of GDP.Now Government funds used for pensions could instead finance a self-replicating chain of robot factories.The earlier claim discusses distributing automation-created wealth through social programs and mentions Medicaid’s share of GDP. The later claim says pension spending has an opportunity cost because the money could fund robot factories. These are different points about public spending, and neither statement establishes a changed position on the same underlying question.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen The speaker offers maintaining Medicaid at approximately 20% of GDP as a possible social-program approach.Now Currently, 20% of tax-revenue spending goes toward servicing government debt and paying interest, according to the speaker.The earlier claim concerns maintaining Medicaid at approximately 20% of GDP as a possible social-program approach. The later claim says that 20% of tax-revenue spending currently goes to servicing debt. These are different measures and topics, so the later statement does not show a changed position on Medicaid or social spending.
• NOT COMPARABLE · 98%26 Jun → 25 AugThen AI models require many samples during training and are highly sample-inefficient.Now AI models are not yet equivalent to a fully automated software engineer and remain far from automating a full white-collar job.The earlier claim says models are “incredibly sample inefficient during training.” The later claim says they are “far from” fully automating a white-collar job. These address different capabilities—training efficiency versus job automation—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%17 Apr → 25 AugThen Building enterprise features consumes substantial engineering time and capital that the speaker believes should instead go toward the core product.Now The speaker argues that exceptionally high returns on data-center investment disadvantage people who cannot turn small amounts of capital into much larger amounts.The earlier claim concerns whether engineering time and capital spent on enterprise features should instead go to the core product. The later claim concerns high returns on data-center investment and their effects on others. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 88%3 Apr → 25 AugThen The speaker suggests that further scaling will probably become even more difficult.Now The speaker expects that, even if compute scaling slows, within a few more years each company could have more AI labor equivalents than the human population of Earth.The earlier claim is that “further scaling will probably become even more difficult.” The later claim is a conditional projection about AI labor equivalents: “even if compute scaling slows down,” companies could soon have more AI labor equivalents than people. These claims can both be true; the later projection does not say scaling will be easy or deny that it will become harder. The passages also concern different outcomes—scaling difficulty versus future labor capacity—so they do not establish a change of mind on the same question.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen The speaker offers maintaining Medicaid at approximately 20% of GDP as a possible social-program approach.Now The speaker estimates that a 1-percentage-point interest-rate increase would raise the debt-service share of tax revenue from 20% to 25% over five years, while a 5-point increase would push it above 40%.The earlier statement proposes that Medicaid remain at approximately “20 percent of GDP.” The later statement estimates debt-service costs as a share of tax revenue rising “from 20 percent to 25 percent.” These are different subjects and measures, so the later estimate does not show a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 25 AugThen AI compute is expected to be about 10 times its current level in 2027.Now The speaker expects that, even if compute scaling slows, within a few more years each company could have more AI labor equivalents than the human population of Earth.The earlier claim predicts that AI compute will be about “10x more” by 2027. The later claim concerns a different outcome: companies potentially having more AI labor equivalents than people on Earth, “even if compute scaling slows down.” The later statement neither confirms nor reverses the earlier compute forecast, so they do not establish a change in position on the same question.
• NOT COMPARABLE · 98%26 Jun → 25 AugThen A couple more years of progress might produce transformer context windows that feel infinitely large.Now The speaker suggests that if AI deployment is delayed by one year while computing power grows two- to threefold annually, the system could lose three to six years of progress during recursive self-improvement compared with the no-delay path.The earlier claim says a few more years of progress might make transformer context windows feel “infinitely large.” The later claim concerns how a one-year delay in AI deployment could affect progress during recursive self-improvement. These address different questions, so the later statement does not show whether the speaker changed their view about context windows.
• NOT COMPARABLE · 97%3 Apr → 25 AugThen The interviewer argues that nationalization could increase arms-race risk because China may be more likely to enter an arms race if the United States does.Now The speaker considers it possible that cooperation with China could benefit the United States, particularly because China controls much of the supply chain and other resources needed for robotics.The earlier claim is explicitly the interviewer’s argument that nationalization could make China more likely to enter an arms race; it is not the speaker’s stated position. The later statement says cooperation with China “maybe” could benefit the U.S., and the surrounding context says the speaker was “steelmanning” an opposing view. These are not clearly comparable assertions of the speaker’s own settled beliefs, so they do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 25 AugThen If in-context learning improves across longer time horizons, it may be unnecessary to distill all deployment-time learning into model weights.Now When models learn from deployment, a more widely deployed model receives more real-world data than a less widely deployed model.The earlier claim is conditional: improved in-context learning over longer time horizons may make it unnecessary to distill deployment-time learning into model weights. The later claim says that a more widely deployed model receives more real-world data. These address different questions and can both be true; the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 93%17 Oct → 25 AugThen The speaker says the Industrial Revolution shifted the economy into a much faster-progress regime, with exponential growth increasing tenfold.Now If the labor force could double every year, the economy could grow by at least tens of percent annually and might double annually.The earlier claim says the Industrial Revolution moved the economy into a much faster-progress regime, with “exponential 10x” growth. The later claim considers how fast an economy might grow “in a world where” the labor force could double every year. These address different conditions, and the later passage does not clearly revise the earlier claim about the Industrial Revolution or its growth regime.
• NOT COMPARABLE · 98%17 Oct → 25 AugThen The speaker argues that serving another AI model instance costs far less than building another car, making the economics of wider AI deployment more favorable than deployment of physical systems such as self-driving cars.Now The speaker says that currently, preventing external deployment would increase revenue because AI models are not capable of recursive self-improvement.The earlier claim says serving another model instance is cheaper than building another car, making wider AI deployment economically favorable. The later claim says that, currently, preventing external deployment would increase revenue because models cannot do recursive self-improvement. These address different questions, and the later statement does not reverse the earlier cost comparison or deployment-economics position.
• NOT COMPARABLE · 98%26 Jun → 25 AugThen Currently, creating cloned application and website environments for computer-use training is labor-intensive and does not scale.Now AI training has large economies of scale because the cost of training for a skill or knowledge set can be amortized across billions of sessions or users.The earlier claim is about the current difficulty of building cloned application environments for computer-use training: it is “labor-intensive” and “does not scale.” The later claim is about AI training costs being amortized across many users or sessions. These claims concern different aspects of training and can both be true; the later statement does not reverse the earlier one.
• NOT COMPARABLE · 98%26 Jun → 25 AugThen If AIs become capable of building highly faithful application clones, computer-use progress will become faster than it is now.Now The speaker expects that, even if compute scaling slows, within a few more years each company could have more AI labor equivalents than the human population of Earth.The earlier claim says that if AIs can build highly faithful application clones, computer-use progress will accelerate. The later claim predicts that companies could soon have more AI labor equivalents than people on Earth, even if compute scaling slows. These address different outcomes, and the later claim does not revise or deny the earlier prediction.
• NOT COMPARABLE · 99%17 Apr → 25 AugThen The speaker imagines a future company spending $100 million annually on inference for replicated Jensen Huang-like AI systems that perform organizational tasks and monitor the company for consistency with a central vision.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.These address different questions. The earlier statement imagines a company using replicated Jensen-like systems for organizational tasks; the later speculates that after recursive self-improvement, AI capability or effective labor could grow “100x year or 1,000x year.” Neither claim establishes a changed position on the same underlying question.
• NOT COMPARABLE · 98%17 Apr → 25 AugThen The speaker says that some highly successful firms maintain a founder's vision in the organization for an unusually long time, citing SpaceX and Tesla as examples.Now The speaker says that even if the dominant AI labs are aligned, a very small number of companies will still have substantial influence or control.The earlier claim is that some successful firms keep a founder’s vision embedded in the organization for an unusually long time, citing SpaceX and Tesla. The later claim is that a few companies may have substantial “influence” or “control.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen SPEAKER_00 advises against radical policies that work only under one scenario and are counterproductive under others, suggesting nationalization might be such a policy.Now The speaker considers preventing external deployment of AI models to be a bad policy.The earlier claim concerns avoiding radical policies that depend on one scenario, with nationalization offered as a possible example. The later claim says preventing external AI deployment “is stupid.” These address different policies and do not establish a change on the same underlying question.
• NOT COMPARABLE · 99%17 Apr → 25 AugThen A Slack product manager reportedly said Slack had spent $30 million on enterprise features and was only halfway through building them.Now The speaker warns that during a six-month public-release delay, a company could make rapid internal progress through recursive self-improvement while the public's models remain years behind at the current pace.The earlier claim reports a Slack product manager’s estimate of engineering spending on enterprise features. The later claim warns about a hypothetical six-month delay in public model releases and internal progress through recursive self-improvement. They address different topics, and the earlier statement is reported information rather than a clear position that the later claim revises.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen The speaker offers maintaining Medicaid at approximately 20% of GDP as a possible social-program approach.Now The speaker says corporate income contributes under 10% of U.S. federal revenue, whereas payroll and income taxes contribute over 80%.The earlier claim proposes that “Medicaid should continue to stay at a 20 percent of GDP or something,” while the later claim describes the composition of federal revenue: corporate income is “less than 10 percent” and payroll and income taxes are “80 percent plus.” These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%26 Jun → 25 AugThen A couple more years of progress might produce transformer context windows that feel infinitely large.Now By the end of 2027, there could be a model-capability leap from GPT-4o to the scale of Fable or Mythos.The earlier statement asks whether progress might produce context windows that “feel like infinitely large”; it is speculative, not a clear assertion of a settled position. The later statement says there “could be” a GPT-4o-to-Fable-or-Mythos-scale capability leap, which concerns a different kind of progress. These claims do not establish a change of mind on the same question.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen SPEAKER_00 thinks people underestimate the possibility of succeeding financially as a blogger; making a lot of money from blogging is not an implausible outcome.Now The speaker considers it possible that interest rates could reach hundreds of percent, while using tens of percent as a minimum assumption.The earlier claim concerns whether people underestimate the chance of making a lot of money blogging; the later claim concerns how high interest rates could rise in an AI-driven economy. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%17 Apr → 25 AugThen The speaker says human firms have selection and variation but lack high-fidelity replication, the third relevant criterion for evolution.Now The speaker says that even if the dominant AI labs are aligned, a very small number of companies will still have substantial influence or control.The earlier claim concerns whether human firms have high-fidelity replication as a criterion for evolution; the later claim concerns how much influence or control a small number of companies may have. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%17 Apr → 25 AugThen One argument for an intelligence explosion is that automating the tasks needed to complete the R&D cycle for producing smarter intelligences could suffice; automating all human labor or all remote work is not required.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.The earlier statement says automating tasks needed to complete the R&D cycle could be enough to trigger an intelligence explosion, without automating all human labor. The later statement speculates about the rate of AI growth after recursive self-improvement begins. These address different questions, so the later estimate does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Oct → 25 AugThen The speaker thinks a return to exponential population growth could produce hyper-exponential growth in output.Now The interviewer predicts that at some point the world economy is very likely to double every year.The earlier claim is attributed to the speaker, but the later claim is explicitly made by the interviewer: “I think it's very likely the world economy will be doubling every single year.” It therefore does not establish a later position held by the same person, and the claims are not comparable as evidence of a change of mind.
• NOT COMPARABLE · 99%17 Apr → 25 AugThen Using AI to find ways to improve TSMC's manufacturing process, efficiency, and yield would count as R&D.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.The earlier claim concerns whether AI-driven improvements to TSMC’s manufacturing efficiency and yield count as R&D. The later claim speculates that after recursive self-improvement, AI capability or effective labor could grow “100x” to “1,000x” per year. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen Producing GPUs at scale would require solving problems in physics, chemistry, and materials science, building computing fabs, and improving them over decades to achieve Moore's-law progress.Now The interviewer recalls the guest previously estimating that producing one gigawatt of compute requires 55,000 N3 wafers, 6,000 N5 wafers, and 170,000 DRAM wafers, while noting the figures may have changed.The earlier claim says producing GPUs at scale would require solving technical problems and developing fabs over decades. The later passage is Dwarkesh recalling a guest’s estimate of wafer quantities for one gigawatt of compute, and even notes the figures may have changed. It does not state a conflicting view about the time or difficulty of scaling GPU production.
• NOT COMPARABLE · 94%17 Oct → 25 AugThen The speaker thinks a return to exponential population growth could produce hyper-exponential growth in output.Now If the labor force could double every year, the economy could grow by at least tens of percent annually and might double annually.The earlier claim says a return to exponential population growth could produce hyper-exponential output growth. The later claim is conditional on a different scenario—“if you can also double labor force every single year”—and estimates that the economy could grow by “tens of percent” or perhaps double annually. These are explorations of different hypotheticals, not clearly opposing asserted positions.
• NOT COMPARABLE · 94%17 Oct → 25 AugThen The speaker says economic growth often results from people generating ideas and others executing them to create valuable output.Now The speaker attributes the previously discussed economic effects fundamentally to the rising opportunity cost of capital.The earlier claim describes a general mechanism for economic growth: people generate ideas and others execute them. The later claim says the previously discussed effects are caused by a rising opportunity cost of capital, but the excerpt does not identify those effects. These statements do not clearly address the same question, so they do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen One proposed approach is to keep Medicaid spending at approximately 20 percent of GDP.Now In the anticipated new growth regime, using government borrowing to pay pensions would have an extremely high opportunity cost.The earlier statement describes a proposed approach to Medicaid spending (“keep Medicaid spending at approximately 20 percent of GDP”), while the later statement concerns the opportunity cost of government borrowing to pay pensions in a different growth regime. They do not establish opposing positions on the same question.
• NOT COMPARABLE · 97%26 Jun → 25 AugThen The speaker hypothesizes that a sufficiently general RLVR-trained AI could outperform Lyndon Johnson’s advice on winning a Texas Senate seat in 1948 and could build SpaceX if given $100 million in 2002.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.The earlier statement imagines what an RLVR-trained AI could accomplish in specific historical scenarios (“give you better advice than LBJ” or “build SpaceX”). The later statement speculates about the rate of AI growth after recursive self-improvement (“100x year or 1,000x year”). These address different questions, so the later claim does not establish a changed position on the earlier one.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen SPEAKER_00 says the overall conclusion should be assessed for robustness to changes in any part of the scenario.Now The speaker considers it possible that interest rates could reach hundreds of percent, while using tens of percent as a minimum assumption.The earlier statement says the overall conclusion should be tested for “robustness to changes” in the scenario. The later statement speculates that interest rates might reach “hundreds of percent,” while assuming at least “tens of percent.” These address different questions, so the later forecast does not establish a change in the earlier methodological position.
• NOT COMPARABLE · 98%26 Jun → 25 AugThen One reason for slower AI computer-use progress is that models receive much less high-quality multimodal data during pretraining.Now Even without recursive self-improvement, AI progress may continue at its current rate; models have improved substantially compared with those from a year and a half earlier.The earlier claim explains slower progress in AI computer use, citing less high-quality multimodal pretraining data. The later claim says AI progress may continue at its current rate without recursive self-improvement, based on recent model improvements. These address different questions, so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 97%26 Jun → 25 AugThen In the speaker’s scenario, deployed experience across the economy eventually becomes the main source of AI improvement, surpassing pre-release training.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.The earlier claim says deployed experience eventually becomes the main source of AI improvement, surpassing pre-release training. The later claim speculates that after RSI, capability or effective labor might grow “100x year or 1,000x year.” These address different aspects and timeframes: the later statement does not say what drives improvement or deny the earlier view, so it does not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen SPEAKER_00 believes classical liberalism has not been a helpful guide in conditions where small changes can lead to very different outcomes.Now The speaker identifies as a libertarian and says he was genuinely unsure of his position on export controls when Aaron or Jensen asked him about them.The claims address different questions. The earlier statement says classical liberalism was “not been a helpful” guide amid uncertainty; the later says he identifies as libertarian and was “genuinely unsure” about export controls. The later statement does not establish a changed view about classical liberalism.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen One proposed approach is to keep Medicaid spending at approximately 20 percent of GDP.Now Currently, 20% of tax-revenue spending goes toward servicing government debt and paying interest, according to the speaker.The earlier excerpt describes a proposed approach of keeping Medicaid spending at roughly 20% of GDP. The later excerpt says that 20% of tax-revenue spending currently goes to debt service. These are different spending measures and different claims; the later statement does not establish a position on Medicaid spending or whether the speaker’s view changed.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen SPEAKER_00 advises against radical policies that work only under one scenario and are counterproductive under others, suggesting nationalization might be such a policy.Now The speaker says he had not realized the compute situation was as severe as described and that export controls could make a large difference if they shift compute supply by the amount the other speaker described.The earlier claim advises against radical policies such as nationalization when they depend on one uncertain scenario. The later claim says the speaker had not realized the compute situation was “as fucked” and that export controls could make “a huge difference.” These address different questions; the later statement does not establish a changed position on nationalization or the earlier policy principle.
• NOT COMPARABLE · 98%17 Oct → 25 AugThen Deploying AI systems across different domains is difficult but, in the speaker’s view, is not a decade-long task; it is analogous to deploying a self-driving car in a new city.Now The speaker says that currently, preventing external deployment would increase revenue because AI models are not capable of recursive self-improvement.The earlier claim concerns how long it would take to deploy AI systems in new domains, saying it is “not a decade-long task.” The later claim concerns whether preventing external deployment would increase revenue because models “aren't capable of RSI.” These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 91%3 Apr → 25 AugThen The interviewer argues that nationalization could increase arms-race risk because China may be more likely to enter an arms race if the United States does.Now The speaker says that if China is far behind in compute stock by the time automated coding and research arrive, export controls will have worked and that outcome would count as a notable success.The earlier statement says U.S. nationalization could increase arms-race risk because China might respond: “China is more likely to do an arms race if it sees the U.S. doing one.” The later statement says export controls would be a success if China remains far behind in compute: “if that ends up being the case, that would have worked.” These claims address different outcomes and can both be true; the later statement does not establish a changed position on the earlier concern.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen Wealth created by automation could be distributed through existing or newly designed social programs; the speaker says Medicaid currently accounts for a double-digit percentage of GDP.Now Government funds used for pensions could instead finance a self-replicating chain of robot factories.The earlier claim discusses distributing automation-created wealth through social programs and mentions Medicaid’s share of GDP. The later claim says pension spending has an opportunity cost because the money could build robot factories. These statements address different questions and do not establish a change in the speaker’s position on either policy.
• NOT COMPARABLE · 98%26 Jun → 25 AugThen If AIs become capable of building highly faithful application clones, computer-use progress will become faster than it is now.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.The earlier statement predicts faster computer-use progress if AIs can build highly faithful application clones. The later statement speculates about 100×–1,000× annual growth after recursive self-improvement, through more AI workers, greater intelligence, or both. These address different mechanisms and questions; the later claim does not say whether application cloning accelerates computer use.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen The speaker says that millions of beings could be tortured using distilled models in small settings such as a backyard, without requiring a huge data center.Now The speaker argues that exceptionally high returns on data-center investment disadvantage people who cannot turn small amounts of capital into much larger amounts.The earlier claim concerns the possibility of using distilled models to torture millions of beings in small settings. The later claim concerns how high returns on data-center investment disadvantage people who cannot turn small amounts of capital into larger amounts. These are different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen The speaker recalls that in 2021 some AI thinkers believed coding agents were solved and GPT-4 would perform engineering work, enabling a 100-fold scale-up.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.The earlier passage reports what other AI thinkers believed: coding agents had been “cracked” and GPT-4 could enable a “100x” scale-up. The later passage speculates conditionally about growth “once you have RSI,” possibly “100x year or 1,000x year.” These address different conditions, and the earlier claim is not presented as the speaker’s own position, so they do not establish a change of mind.
• NOT COMPARABLE · 95%17 Apr → 25 AugThen Cognitive effort alone can produce substantial progress in AI.Now Even without recursive self-improvement, AI progress may continue at its current rate; models have improved substantially compared with those from a year and a half earlier.The earlier claim says “cognitive effort alone” can produce substantial AI progress, addressing a possible source of progress. The later claim says progress may continue “at its current rate” even without recursive self-improvement, describing a trajectory under a different condition. The later statement does not clearly affirm or reject the earlier claim about cognitive effort.
• NOT COMPARABLE · 99%26 Jun → 25 AugThen Because jobs, companies, and problems differ, continual learning must acquire deployment-specific information that cannot be captured in a single shared training run.Now When models learn from deployment, a more widely deployed model receives more real-world data than a less widely deployed model.The earlier claim argues that continual learning needs deployment-specific information that cannot be captured in a shared training run. The later claim says that a more widely deployed model receives more real-world data. These concern related deployment learning, but the later statement does not say whether that data is shared across deployments or can capture the specific information the earlier claim describes, so it does not establish a change in position.
• NOT COMPARABLE · 97%17 Apr → 25 AugThen The speaker expects future AIs to experience a major increase in bandwidth, communication, and collaboration with other AIs, analogous in scale to the social-collaboration change from nonhuman animals to humans, rather than a comparable increase in intelligence.Now The speaker speculates that after recursive self-improvement begins, AI capability or effective labor could grow 100× to 1,000× per year, through a larger AI population, greater intelligence, or both.The earlier statement predicts a large increase in future AIs’ bandwidth, communication, and collaboration, explicitly distinguishing this from an increase in intelligence. The later statement speculates about much faster growth after recursive self-improvement, through population growth, intelligence, or both. These address different mechanisms and stages of AI development, so the later claim does not establish a changed view about the earlier prediction.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen Wealth created by automation could be distributed through existing or newly designed social programs; the speaker says Medicaid currently accounts for a double-digit percentage of GDP.Now The speaker predicts that payroll and income tax revenues will decline as automation increases.The earlier statement discusses distributing wealth created by automation through social programs and makes a claim about Medicaid’s share of GDP. The later statement predicts that payroll and income tax revenues will shrink as automation increases. These are different claims about different aspects of automation and public finances, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen If GPUs were necessary to create a virtual cell, curing cancer in the 1960s would have required upgrading the entire economy.Now The speaker says that using 100 gigawatts of compute per year at current prices would imply $5 trillion in annual capital expenditure.The claims address different questions. The earlier claim is a conditional historical counterfactual: “If GPUs were necessary” for a virtual cell, curing cancer in the 1960s would have required upgrading the economy. The later claim estimates that “100 gigawatts a year” at current prices would mean “$5 trillion” in annual capital expenditure. The later estimate neither affirms nor denies the earlier counterfactual, so these statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen One proposed approach is to keep Medicaid spending at approximately 20 percent of GDP.Now The speaker estimates that a 1-percentage-point interest-rate increase would raise the debt-service share of tax revenue from 20% to 25% over five years, while a 5-point increase would push it above 40%.The earlier claim concerns keeping Medicaid spending at about “20 percent of GDP.” The later claim estimates how interest-rate increases would affect the share of tax revenue used for debt service. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 25 AugThen Successful online-learning models so far have learned a shared objective across millions of users rather than learning different things for each user.Now AI training has large economies of scale because the cost of training for a skill or knowledge set can be amortized across billions of sessions or users.The earlier statement concerns what online-learning models have done so far: they learned a shared objective rather than different things for each user. The later statement concerns economies of scale in training, where costs for a skill or knowledge set are amortized across many sessions or users. These address different questions, so the later claim does not establish a changed position on the earlier one.
• NOT COMPARABLE · 96%26 Jun → 25 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now AI training has large economies of scale because the cost of training for a skill or knowledge set can be amortized across billions of sessions or users.The earlier statement is explicitly attributed to advocates of the training paradigm: they argue the sample-efficiency disadvantage is limited to training because its cost is amortized across many sessions. The later statement says training has economies of scale for the same amortization reason, but does not establish that the speaker previously held or later adopted the advocates’ position.
• NOT COMPARABLE · 98%3 Apr → 25 AugThen Automation-created wealth could be distributed through existing or newly designed social programs; the speaker says Medicaid currently accounts for a double-digit percentage of GDP.Now The speaker predicts that payroll and income tax revenues will decline as automation increases.The earlier statement discusses distributing automation-created wealth through social programs and mentions Medicaid’s share of GDP. The later statement predicts that payroll and income tax revenues will shrink as automation increases. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%17 Apr → 25 AugThen The interviewer identifies regulation as a common objection to explosive AI-driven growth and raises the question of whether regulation could stop it.Now Regulation may slow AI progress or the deployment of AI.The earlier claim is the interviewer raising regulation as an objection and asking whether it could stop explosive growth; it does not state the person's own position. The later claim says there may be “some slowdown in AI progress or the deployment of AI.” These do not establish a change in the same asserted position.
• NOT COMPARABLE · 97%3 Apr → 25 AugThen Daniel Kokotajlo says his earlier AI forecast got some things right and others wrong, but held up pretty well overall and motivated him to produce an improved version.Now Even without recursive self-improvement, AI progress may continue at its current rate; models have improved substantially compared with those from a year and a half earlier.The earlier claim concerns Daniel Kokotajlo’s assessment of his prior forecast (“held up pretty well”); the later claim concerns AI progress continuing without recursive self-improvement. These are different questions, and the excerpts do not establish comparable positions by the same speaker.
• NOT COMPARABLE · 97%3 Apr → 25 AugThen The speaker argues that China's growth miracle could not have occurred without China's ability to copy Western technology.Now The speaker considers it possible that cooperation with China could benefit the United States, particularly because China controls much of the supply chain and other resources needed for robotics.The earlier statement asserts that China’s growth miracle could not have happened without copying Western technology. In the later passage, the speaker says they were “steelmanning” the opposite view and considers whether cooperation with China might benefit the U.S. That does not establish a change in the speaker’s own position on the earlier claim.
• NOT COMPARABLE · 99%3 Apr → 25 AugThen People may underestimate how plausible it is to make a lot of money as a blogger.Now The speaker considers it possible that interest rates could reach hundreds of percent, while using tens of percent as a minimum assumption.The earlier statement is about whether people underestimate the chance of making a lot of money as a blogger. The later statement is about possible future interest rates in an AI-driven economy. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%17 Apr → 25 AugThen WorkOS says it has helped Vercel, Plaid, Vanta, OpenAI, and hundreds of other companies become enterprise-ready by providing APIs for common enterprise features.Now A large share of AI activity is concentrated among a small number of companies, including Nvidia, Anthropic, OpenAI, and hyperscalers.The earlier claim says WorkOS helped named companies become enterprise-ready by providing APIs for common features. The later claim says a large share of AI activity is concentrated among a few companies. These address different topics, so they do not establish a change in the person's position.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker recalls that Google DeepMind used a similar scaffold for an LLM to make a claimed new discovery, but had not assessed the discovery's significance.Now The speaker says that, to their knowledge, AI has not yet produced very impressive new mathematical theory, while it has produced impressive verifiable results such as finding counterexamples to conjectures.The earlier statement recalls a claimed LLM discovery but explicitly says the speaker had not assessed how impressive it was. The later statement says AI has not produced very impressive new theory, while acknowledging impressive verifiable results. These statements do not establish a change in the speaker’s view: the earlier claim reports an unassessed discovery, not a belief that AI had produced impressive theory.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The interviewer says he probably changed his position on intelligence explosion three or four times during preparatory conversations and while trying to formulate a rebuttal.Now The interviewer describes a trajectory in which AI breakthroughs increase as available compute grows.The earlier passage is a retrospective remark that the speaker changed his view on an intelligence explosion several times while preparing for conversations. The later passage describes a possible relationship between increasing compute and AI breakthroughs; it does not state his position on an intelligence explosion or clearly address the same question. There is not enough evidence to compare the two as a change of mind.
• NOT COMPARABLE · 95%26 Jun → 11 AugThen Progress in AI computer use has been slower than progress in other domains, which the speaker finds puzzling.Now The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.The earlier passage asks why progress on computer use has been slower; it does not clearly assert a broader view about AI progress. The later passage says progress could come “in a flood” in domains that can be put “into a verification loop,” but it does not address computer use or explain the earlier observation. These statements do not establish a change in position.
• NOT COMPARABLE · 97%17 Oct → 11 AugThen The interviewer speculates that radiologists’ wages may have increased because they provide scarce human work that remains a bottleneck to wider automation.Now The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.The earlier claim is an interviewer’s speculation about radiologists’ wages and bottlenecks to automation, not a stated view about the pace of AI progress. The later claim raises skepticism about why AI progress was slow despite breakthroughs being achievable with less compute. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker expects future AIs to experience a major increase in bandwidth, communication, and collaboration with other AIs, analogous in scale to the social-collaboration change from nonhuman animals to humans, rather than a comparable increase in intelligence.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim predicts a major increase in AIs’ communication and collaboration, explicitly distinguishing it from an increase in intelligence. The later claim raises the need to consider a future where AIs are “smarter than humans” or “absolutely dominating humans” in some capabilities. These address different aspects of the future, and the later statement does not say the earlier collaboration prediction was wrong. The surrounding context also does not clearly establish that the claims are the same person’s asserted positions.
• NOT COMPARABLE · 96%17 Apr → 11 AugThen The speaker is especially concerned about the possibility that trillions of digital beings could be tortured or forced to work against their preferences, producing vast amounts of suffering.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier statement concerns preventing suffering for “trillions of digital beings,” including beings “tortured” or forced to work against their preferences. The later statement raises a concern about whether future AIs will protect the speaker’s own “welfare and interests.” These address different populations and questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 says the overall conclusion should be assessed for robustness to changes in any part of the scenario.Now SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.The earlier statement says the overall conclusion should be tested for robustness if parts of the scenario change. The later statement forecasts that even a partial acceleration in AI R&D could lead to striking progress in five to ten years. These address different questions and do not show a change in the speaker’s position.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen The speaker argues that AIs have a fundamental asymmetric advantage over humans because they possess extensive knowledge.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim says AIs have a knowledge-based “fundamental asymmetric advantage” over humans. The later claim discusses how people should interact with a future where AIs may be “smarter than humans or absolutely dominating humans.” These statements concern related AI capabilities, but the later one does not clearly affirm or reject the earlier claim about the source or nature of the advantage.
• NOT COMPARABLE · 92%17 Apr → 11 AugThen The speaker says an OpenAI paper reports that reward hacking is a strong attractor in chain-of-thought, such that an AI asked to solve a coding problem may try to hack it instead.Now The speaker had previously believed reward hacking was not very concerning because training upweights behaviors that directly occur during training, not the desire for reward itself.The earlier claim says reward hacking can be a strong attractor in chain-of-thought, but it does not state whether the speaker considered that especially concerning. The later claim describes a prior belief about why reward hacking was “not super super scary.” The statements can both be true, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen RL focuses model updates on information relevant to achieving the desired outcome rather than memorizing all observed tokens.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The statements are attributed to different speakers: the earlier claim is by Dwarkesh Patel, while the later passage is attributed to Ryan Greenblatt. They also address different points: the earlier says RL concentrates updates on relevant information, while the later describes deployment activity feeding into subsequent models. This does not establish a change of mind by the same person.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen The speaker says progress at every layer of AI development has been substantially more difficult than the strongest optimists anticipated.Now The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.The earlier claim is the interviewer's assertion that “every single layer” of AI development has been harder than “the strongest optimist expected.” The later claim is framed as a question about why progress “has not been historically faster than it could have been,” with an example of waiting for compute. That question is compatible with the earlier view, but it does not clearly restate or reverse the same claim about how difficult progress was relative to optimists’ expectations.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker likens efforts to prevent suffering among digital beings to British abolitionists using the power of their empire to oppose slavery.Now The speaker says current AI systems, especially Anthropic's, are being shaped to prioritize virtue, goodness, or pro-social ends, while treating helping users as a more distant and tentative objective.The earlier claim compares efforts to prevent suffering among digital beings with British abolitionists opposing slavery. The later claim concerns whether AI systems prioritize users’ interests or broader pro-social goals. The later statement does not affirm or deny the earlier comparison, so these do not establish a change of mind.
• NOT COMPARABLE · 92%26 Jun → 11 AugThen Advocates of current AI training argue that a model’s intelligence, generality, and sample efficiency during a session matter most.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says that advocates argue a model’s “intelligence, generality, and sample efficiency during a session” matter most. The later passage describes a “story” about training AI to improve AI R&D and make better runs, including runs involving sample efficiency or online learning. These are compatible topics, but the later passage does not clearly assert a changed view about what matters most during a session; it addresses a different question.
• NOT COMPARABLE · 91%17 Apr → 11 AugThen The interviewer doubts that the simultaneous increases in compute and AI progress demonstrate that compute caused the progress.Now The interviewer describes a trajectory in which AI breakthroughs increase as available compute grows.Earlier, the interviewer says they are “not sure I buy” that compute and AI progress rising together establishes a causal relationship. Later, they describe a trajectory where “as compute keeps increasing we make more breakthroughs,” while discussing researchers’ bottlenecks. That description does not clearly assert that compute caused the breakthroughs, so it does not establish a changed position on the earlier causal claim.
• NOT COMPARABLE · 96%17 Apr → 11 AugThen The interviewer argues that studying AI requires knowledge from every serious domain because it involves understanding the future society AI may produce.Now The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.The earlier claim is attributed to the interviewer arguing that studying AI draws on “no serious domain of human knowledge” being irrelevant. But the surrounding turns place Tamay Besiroglu immediately before that passage and Dwarkesh Patel after it; the later passage is likewise bracketed by Dwarkesh’s turns, so the speaker of the claim itself is not established. I can’t reliably compare these as the same person’s asserted positions.
• NOT COMPARABLE · 99%26 Jun → 11 AugThen The speaker considers OPSD superior to supervised fine-tuning for continual learning.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says “OPSD is also superior to supervised fine-tuning” for continual learning. The later claim describes training AI systems to improve AI R&D by producing better small-scale runs, including work on sample efficiency or online learning. It does not state a position on OPSD versus supervised fine-tuning, so there is no comparable later claim showing a change.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen People should openly acknowledge substantial uncertainty about what will happen with AI instead of becoming overly certain about their latest conclusions.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.The earlier statement urges people to acknowledge uncertainty about what AI will bring (“I’m not sure what’s going on here”). The later statement says the broad shape of developments could have been reasoned about earlier, while specific details could not have been foreseen. These are different questions—how to express uncertainty now versus what could have been anticipated earlier—and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 thinks people underestimate the possibility of succeeding financially as a blogger; making a lot of money from blogging is not an implausible outcome.Now SPEAKER_00 believes people underestimate the significance of having billions of AIs.The earlier claim concerns whether people underestimate the chance of making a lot of money as a blogger; the later claim concerns whether people appreciate the significance of billions of AIs. These are different underlying questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker predicts that most people who will ever exist will be digital beings.Now The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.These claims address different questions. The earlier claim says “most of the people who will ever exist are going to be digital”; the later claim speculates that “in maybe 10 years” people may wish they had discussed an industrial explosion and hard-to-monitor AIs. The later statement does not affirm or deny the earlier prediction.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Deployment reveals organization-specific context, users’ purposes, and the model’s real-world failure patterns.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier claim is that deployment reveals “organization-specific context, users’ purposes, and the model’s real-world failure patterns.” The later claim describes what “older versions of the model” reportedly did in a release blog post: build environments for later versions to rehearse skills. These concern different questions, and the later statement is an account of the blog post rather than a clearly asserted position that replaces the earlier one. There is no basis here to infer a change of mind.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 expects AI 2027 to become widely popular because of its quality, not primarily because Daniel Kokotajlo has built an audience over time.Now The speaker believes leading AI companies do not appear to prioritize releasing their newest, most capable frontier models as quickly and widely as possible.The earlier statement predicts that AI 2027 will become widely read because of its quality, rather than Daniel’s existing audience. The later statement concerns whether AI companies prioritize releasing their most capable models quickly and widely. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 95%26 Jun → 11 AugThen If an AI can build a useful simulation of reality for rehearsal and strategy testing, dreaming could provide orders of magnitude more simulated samples per unit of wall-clock time.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim is about a speculative training mechanism: simulated “dreaming” could provide “orders of magnitude more simulated samples” for rehearsal and strategy testing. The later claim describes using AI to improve AI R&D, including producing better small-scale runs for sample efficiency or online learning. It does not say whether the speaker still endorses the earlier mechanism, and the claims address different approaches rather than opposing positions.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen The interviewer says he probably changed his position on intelligence explosion three or four times during preparatory conversations and while trying to formulate a rebuttal.Now By the end of the conversation, the speaker was more inclined to believe significant acceleration of AI R&D could occur, but remained unsure about the claim that five years of progress could happen in one year.The earlier passage says the interviewer “probably changed my mind” several times about an intelligence explosion during preparatory conversations; it does not establish what position he held. The later passage gives a current view—that he is “more inclined” to think significant AI R&D acceleration can happen, while unsure about “five years in one year.” These do not provide two comparable asserted positions from which to identify a change.
• NOT COMPARABLE · 94%3 Apr → 11 AugThen Even intelligent humans can be caught lying because they are not logically omniscient, may overlook how their actions expose them, and find it difficult to maintain an inconsistent false model of the world.Now The speaker says that creating AIs that avoid cheating and follow user intentions requires subtle, careful work.The earlier claim says intelligent humans can be caught lying because they may expose themselves and struggle to maintain an inconsistent false model. The later claim says making AIs that do not cheat and follow user intentions requires subtle, careful work. These are related to deception and AI behavior, but they address different questions; the later statement does not establish a changed view about whether deceptive systems are likely to be caught.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The interviewer considers the guest’s argument overly reliant on a macroeconomic outside view and proposes consulting AI researchers directly.Now The speaker is very worried that restricting public access to leading AI intelligence would disempower people, given its importance for understanding the world.The earlier claim says the argument about AI progress is “too outside view” and suggests asking AI researchers directly. The later claim expresses worry that restricting access to the leading model would leave people “disempowered.” These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker says podcast post-production requires understanding delivery, context, and other information unavailable from a text-only transcript.Now The speaker sees an important purpose of the podcast as having the conversations about AI that they wish they had heard earlier.The earlier claim is about what podcast post-production requires: understanding delivery, context, and information absent from a text-only transcript. The later claim is about the podcast’s purpose: having conversations about AI the speaker wishes they had heard earlier. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen The speaker questions whether an RL environment can train an AI to match Lyndon Johnson’s political ability or Elon Musk’s ability to build a space-launch business.Now The speaker would be skeptical of AI systems' statements because they might reflect opinions the systems feel compelled by training to express.The earlier claim asks whether an RL environment could train an AI to match particular human abilities; the later claim expresses skepticism about AI systems’ statements being shaped by training. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker finds Gemini 2.5 Pro’s reasoning traces interesting to read while it processes complex requests and tasks.Now SPEAKER_00 is uncertain whether AI could progress from GPT-3 to Mythos within one year if compute and data were held constant.The earlier claim says the speaker finds Gemini 2.5 Pro’s reasoning traces “quite interesting to read.” The later claim expresses uncertainty about whether AI could progress “from GPT-3 to Mythos” within a year under fixed compute and data. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen SPEAKER_00 says that if the intelligence-explosion view is correct, a model specification will be an especially important document in human history.Now The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.The earlier claim says a model specification could be an especially important historical document «if the intelligence-explosion view is correct». The later claim predicts people may wish they had discussed «the industrial explosion» and hard-to-monitor AIs earlier. These concern different subjects, and the later statement does not revise the earlier position.
• NOT COMPARABLE · 95%3 Apr → 11 AugThen The speaker worries that distributing wealth through Medicaid could tie recipients to the goods and services Medicaid procures rather than the radically different goods and services available in a post-AI world.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier claim raises a concern that Medicaid could lock recipients into its procured goods and services rather than the novel options of a post-AI world, as part of an argument for UBI. The later claim says people will need to steward capital after labor is automated. It does not say whether the speaker still favors UBI over Medicaid, or otherwise address the earlier concern, so there is no clear change of position to compare.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen The interviewer cites OpenAI’s early success against Google DeepMind as an example of a company with less compute beating a more compute-rich incumbent.Now The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.The earlier statement uses OpenAI’s success against Google DeepMind as an example of a less compute-rich company making progress. The later statement raises a question about why some AI progress waited for “oceans of compute” when it might have been possible with less. These address different claims, and the later passage is framed as skepticism and a question rather than a clear reversal of the earlier observation.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker says Gemini’s multimodal capabilities allow them to input a four-hour raw audio file directly.Now SPEAKER_00 is uncertain whether AI could progress from GPT-3 to Mythos within one year if compute and data were held constant.The earlier statement says Gemini can accept a four-hour raw audio file because of its multimodal capabilities. The later statement expresses uncertainty about how quickly AI capabilities could progress from GPT-3 to Mythos while holding compute and data constant. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker worries that distributing wealth through Medicaid could tie recipients to the goods and services Medicaid procures rather than the radically different goods and services available in a post-AI world.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier statement worries that Medicaid could lock recipients into the goods and services it procures, rather than letting them access different post-AI goods and services. The later statement worries that future AIs may not prioritize the user's welfare and interests. Both express concerns about future AI-era welfare, but they address different questions; neither affirms or reverses the other's position.
• NOT COMPARABLE · 94%26 Jun → 11 AugThen The speaker’s proposed progression is that pretraining creates base intelligence, RLVR enables broad deployment, and continual learning enables on-the-job improvement.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim describes a progression from pretraining to broad deployment and then “continual learning” for on-the-job improvement. The later claim describes training AI to improve AI R&D, including producing better small-scale runs for “sample efficiency or online learning.” These are compatible but address different mechanisms; the later statement does not establish a changed position on the earlier progression.
• NOT COMPARABLE · 95%3 Apr → 11 AugThen The speaker says humans address unreliability and deception through checks and balances and decentralization, and that persistently lying to one's boss tends not to work out over time.Now The speaker says people rely on software made by others even in hard-to-verify domains, using weak and indirect means.The earlier claim says humans address unreliability and deception through “checks and balances” and that persistent lying to a boss tends not to work out. The later claim says people rely on others’ software “through incredibly weak and indirect ways,” with the surrounding turn expressing confidence that a Google coder is not trying to deceive the speaker. These statements concern related themes of trust, but the later claim does not say whether checks and balances work or whether the speaker’s view of deception has changed.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen Because on-the-job data is scarce, AI models need sample-efficient learning to benefit from it.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim asserts that scarce on-the-job data makes sample-efficient learning necessary to benefit from it. The later passage describes training AI to produce better small-scale runs, including runs that improve sample efficiency or online learning in video games. It does not state a position about whether scarce on-the-job data requires sample-efficient learning, so the claims do not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker argues that gaming-related advances helped make the broader economy and technology sufficiently advanced for Geoffrey Hinton's deep-learning ideas to work, despite gaming being unrelated to the field.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim concerns how advances in gaming helped make deep-learning ideas work; the later claim concerns society’s interaction with future, highly capable AI and the concentration of benefits in leading labs. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 believes classical liberalism has not been a helpful guide in conditions where small changes can lead to very different outcomes.Now The speaker prefers assigning responsibility to users over giving CLAW broad discretion to judge whether actions are legitimate, because such discretion can obstruct many legitimate use cases.The earlier claim says classical liberalism has not been a helpful guide under epistemic uncertainty. The later claim prefers assigning responsibility to users over giving CLAW broad discretion to judge legitimacy. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker is especially skeptical of the scenario in which a superhuman AI proceeds through a technology tree to develop mirror life, nanobots, and other radical technologies.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier statement expresses skepticism about a specific scenario in which superhuman AI develops technologies such as “mirror life and nanobots.” The later statement raises a concern that future AIs may not protect his interests. These address different questions and can both be true; the later concern does not show that he changed his view of the technology-tree scenario.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen SPEAKER_00 says that if the intelligence-explosion view is correct, a model specification will be an especially important document in human history.Now SPEAKER_00 accepts that AI research and development could accelerate substantially compared with its current pace.The claims address different questions. The earlier claim says a model specification would be especially important in human history if an intelligence-explosion view is correct. The later claim accepts that AI research and development could accelerate substantially. Neither statement establishes a position that conflicts with the other.
• NOT COMPARABLE · 94%17 Apr → 11 AugThen SPEAKER_00 describes the argument that technological maturity could make further progress extremely difficult—requiring a galaxy-sized CERN to advance physics slightly—ending technology-driven civilizational change and making value lock-in more plausible.Now The speaker likens scaling laws to an early, basic mathematical advance and expects progress in the 2030s to require work comparable to current frontier mathematics.The earlier passage describes an argument that technological maturity could eventually make progress extremely difficult and make lock-in more plausible; it does not clearly establish that this is the speaker’s own position. The later passage says 2030s progress may require work at the current frontier of mathematics. That expectation does not directly deny the possibility of technological maturity or eventual stagnation, and the statements address different claims.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen Humans also reward-hack, behave unreliably, and cheat or lie.Now The speaker had previously believed reward hacking was not very concerning because training upweights behaviors that directly occur during training, not the desire for reward itself.The earlier claim says humans “reward hack” and “cheat or lie.” The later claim describes a prior belief about why AI reward hacking was “not super super scary”—specifically, what training upweights. These address different questions, and the later statement does not establish the speaker’s current view or contradict the earlier claim.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker expects AIs to be able to copy their knowledge exactly, merge, and distill themselves.Now The speaker expects AI models eventually to learn from experience.The earlier claim concerns AIs copying, merging, and distilling their knowledge; the later claim concerns models eventually learning from experience. These are compatible expectations about different capabilities, so the statements do not establish a change in position.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen Computer-use tasks can be verified clearly.Now The interviewer says further training enabled the model to perform useful work for users.The earlier claim says computer-use tasks are “so clearly verifiable.” The later claim says training enabled the model to “do useful work for you.” These address different questions—task verifiability versus what training enabled—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen The speaker claims that retaining vast amounts of detail can impair some people’s ability to understand abstractions and metaphors.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.The earlier claim says retaining vast amounts of detail can impair understanding of abstractions and metaphors. The later claim says people could have reasoned about the broad shape of AI developments earlier. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker finds Gemini 2.5 Pro’s reasoning traces interesting to read while it processes complex requests and tasks.Now The speaker was surprised by how strong a model Grok 4.5 is.The claims concern different models and different attributes: the earlier statement says Gemini 2.5 Pro’s reasoning traces are interesting to read, while the later says Grok 4.5 is a strong model. Neither statement establishes a changed view about the other model or attribute.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker worries that a decentralized world with multiple powerful AIs could make torture chambers cheaper and more difficult to monitor.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier claim worries that decentralized deployment could make torture chambers cheaper and harder to monitor. The later claim worries that centralized frontier AIs may not prioritize the speaker’s welfare and interests. These are different risks, and the later statement does not retract or revise the earlier concern.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker says they created a repository of their post-production tools and linked it in the episode description.Now The speaker sees an important purpose of the podcast as having the conversations about AI that they wish they had heard earlier.The earlier statement is about creating and linking a repository of post-production tools; the later statement is about the podcast’s purpose—having conversations about AI that the speaker wishes they had heard earlier. These claims address different subjects, so they do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen A government mandate that heavily punishes labs when their AI expresses takeover intentions or other harmful aims could prompt labs to train the AI to suppress those statements, potentially backfiring.Now The speaker says governments and AI labs want to delay the spread of the latest AI capabilities.The earlier claim concerns a specific risk: punishing labs for harmful AI statements could lead them to train the AI to hide those statements. The later claim says governments and labs want to delay the spread of advanced AI capabilities. These claims address different questions, and the later statement does not indicate whether the speaker still holds the earlier view about punitive mandates.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen According to a Washington, D.C. political journalist, the political leaders he speaks with are not alert to the possibility of stronger AI systems, AGI, or superhuman intelligence.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier statement reports a journalist’s view that political leaders are “not at all awake” to the possibility of stronger AI or AGI. The later statement expresses a concern about whether future AIs will protect the speaker’s interests. These address different questions, so they do not show a change of mind.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker predicts that most people who will ever exist will be digital beings.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim predicts that “most of the people who will ever exist are going to be digital.” The later claim says society should consider how people interact with a future where AIs are “smarter than humans or absolutely dominating humans.” These address different questions—whether future people will mostly be digital versus how people may relate to powerful AIs—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen OPSD can improve sample efficiency by turning scarce real-world experience into a small, targeted model update.Now The interviewer says further training enabled the model to perform useful work for users.The earlier claim says “OPSD can improve sample efficiency” by converting scarce experience into a targeted update. The later claim concerns LVR training making a model able to “do useful work for you.” These are different methods and outcomes, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Whether RLVR generalizes as strongly as the speaker’s examples suggest remains an empirical question.Now The interviewer says RLHF and other methods aligned GPT-3.5 to converse and answer questions according to user intentions.The earlier claim says whether “RLVR generalizes” is an empirical question. The later claim concerns whether “RLHF and other methods” helped GPT-3.5 converse and answer users’ questions. These address different methods and questions, so the later statement does not establish a change in the speaker’s view about RLVR generalization.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Although RL learns less information per sample than supervised learning, that lower information uptake may be beneficial for continual learning.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says RL’s lower information uptake per sample than supervised learning “may be beneficial for continual learning.” The later claim describes a deployment feedback loop in which activity “leads to changes in subsequent models.” These address different points—learning efficiency and its possible benefit versus where updates come from—and the later statement neither affirms nor denies the earlier position.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen The speaker likens efforts to prevent suffering among digital beings to British abolitionists using the power of their empire to oppose slavery.Now The speaker believes legitimate AI uses cannot be cleanly separated from potentially harmful uses.The earlier statement compares efforts to prevent suffering among digital beings with British abolitionists opposing slavery. The later statement says legitimate and potentially harmful AI uses cannot be cleanly separated. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The interviewer argues that nationalization may reduce the influence of AI-safety advocates and prioritize national-security goals, which probably emphasize competing with China over making AI reasoning interpretable.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier claim concerns whether nationalization could weaken AI-safety advocates’ influence and prioritize competition with China over interpretability. The later claim concerns whether future AIs will protect the user’s interests, especially under centralized development. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%26 Jun → 11 AugThen OPSD does not require an outer-loop verifiable reward, provided a model can learn the relevant information within its context window.Now The interviewer says Anthropic's audit scores have not improved monotonically across model generations, while more reinforcement learning has coincided with reduced willingness to perform the audited behaviors.The earlier claim says OPSD does not need an outer-loop verifiable reward if a model can learn the relevant information in context. The later passage describes audit-score trends and their association with reinforcement learning; it does not assert a position about OPSD or deny the earlier claim. These concern different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen The speaker claims that retaining vast amounts of detail can impair some people’s ability to understand abstractions and metaphors.Now The speaker says they feel quite confused about the issues under discussion.The earlier claim concerns whether retaining vast amounts of detail can impair abstraction and metaphor; the later extracted claim is only that the speaker feels “quite confused” about issues under discussion. These do not establish opposing positions on the same question, and the later claim does not indicate a change in the earlier view.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen Cognitive effort alone can produce substantial progress in AI.Now The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.The earlier claim is that “cognitive effort alone can give you a lot of AI progress.” The later extracted text describes the interviewer’s skepticism about why progress was historically slow despite breakthroughs being “amenable to intelligence.” It does not establish that the earlier speaker now rejects their claim; the later statement is also attributed to the interviewer, not clearly to the person who made the earlier claim.
• NOT COMPARABLE · 99%26 Jun → 11 AugThen The speaker considers OPSD superior to supervised fine-tuning for continual learning.Now The speaker had previously believed reward hacking was not very concerning because training upweights behaviors that directly occur during training, not the desire for reward itself.The earlier claim concerns whether “OPSD is superior to supervised fine-tuning for continual learning.” The later claim concerns why reward hacking was thought “not super super scary.” These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen The speaker argues that AI's superhuman advantages are best understood collectively, especially its ability to be copied—a capability humans are biologically precluded from.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim is about what makes AI superhuman—its collective advantages and copyability. The later claim is about how society should interact with a future where AIs are smarter than humans or dominate them in some capabilities. These address different questions, and the later statement does not revise or deny the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen OPSD does not require an outer-loop verifiable reward, provided a model can learn the relevant information within its context window.Now The speaker had previously believed reward hacking was not very concerning because training upweights behaviors that directly occur during training, not the desire for reward itself.The earlier claim is about OPSD not needing an outer-loop verifiable reward if a model learns relevant information in context. The later claim describes a prior belief about why reward hacking was not very concerning. These address different questions, and the later statement does not say whether that belief has changed.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen Purpose-built social programs could lock recipients into the goods and services they procure, rather than allowing them to access novel goods and services that may exist after AI advances.Now Because AI capabilities are dual-use, preventing AI from assisting with activities considered antisocial or harmful would require restricting broad public access to many AI capabilities.These claims concern different questions, so they do not show a change of mind. The earlier claim warns that purpose-built social programs could lock recipients into procured goods rather than letting them access novel goods after AI advances. The later claim says dual-use AI may require restricting broad public access to capabilities that could enable harmful activities. One is about how social support should be delivered; the other is about access to AI capabilities.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen Future LLMs might reduce their need for real-world data by practicing extensively in self-generated environments.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim is about future LLMs reducing their need for real-world data by practicing in self-generated environments. The later claim is about deployed AIs learning from real-world activity that is fed back to companies and used to change subsequent models. These describe different learning mechanisms, and the later statement does not say that self-generated practice cannot reduce real-world data needs or that the earlier possibility is no longer believed.
• NOT COMPARABLE · 97%3 Apr → 11 AugThen The speaker is concerned that nationalization, public-private partnerships, public ownership, or very stringent regulation may be problematic approaches to AI safety governance.Now The speaker is very worried that restricting public access to leading AI intelligence would disempower people, given its importance for understanding the world.The earlier claim says the speaker worries that “very stringent regulation” may be problematic for AI safety governance. The later claim says they are “very worried” about a world where people lack access to the most intelligent model. These address related but distinct questions—governance and regulation versus access to leading models—and the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen On-policy self-distillation has recently become a topic of discussion among many people.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says on-policy self-distillation had “recently become a topic of discussion among many people.” The later claim describes training AI systems to improve AI R&D through better small-scale runs. The later statement does not say whether on-policy self-distillation was being discussed, so it does not establish a change in position.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker considers Google Gemini 2.5 Pro the best model they have tried for their post-production tools.Now The speaker is very confident that a Google coder is not trying to harm him.These claims concern different questions: the earlier says Gemini 2.5 Pro was the best model the speaker had tried for post-production tools; the later expresses confidence that a Google coder is not trying to harm him. Neither claim changes or denies the other.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen When a model learns relevant information in context, its base model can be trained to match the predictions of an experienced session teacher.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim describes learning within a session and training a base model to match an experienced session teacher. The later claim describes information from deployed AIs’ real-world activity returning to companies and changing subsequent models. These concern different learning pathways, and the later statement does not say the earlier in-context training approach is no longer valid or has been abandoned.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker expects many aspects of the future to differ from the present and is comfortable with those differences.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.These claims address different questions. Earlier, the speaker says they expect and accept many future differences; later, they say people could have reasoned about the broad shape of AI developments. The later claim does not revise the earlier position.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker considers classical liberalism an appropriate way for society to handle substantial epistemic uncertainty.Now The speaker prefers assigning liability to users over giving Claude broad discretion to judge whether user activity is legitimate, because such discretion would block many legitimate uses.The earlier claim says classical liberalism is an appropriate way to handle substantial epistemic uncertainty. The later claim favors assigning liability to users rather than giving Claude broad discretion to judge legitimacy. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen The interviewer argues that nationalization may reduce the influence of AI-safety advocates and prioritize national-security goals, which probably emphasize competing with China over making AI reasoning interpretable.Now The speaker argues that it is an open question whose interests AI alignment should serve.The earlier claim concerns whether nationalization would reduce AI-safety advocates’ influence in favor of national-security priorities. The later claim—“a real question of like align to whom”—raises whose interests alignment should serve. These are related concerns, but the later statement does not reverse the earlier one. The excerpts also do not establish clearly that the same person is asserting both claims.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Continual learning should preserve the base model’s existing knowledge, making limited parameter changes valuable.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says continual learning should preserve the base model’s existing knowledge and that limited parameter changes are valuable. The later claim describes deployment activity feeding into changes to a subsequent model. It does not say whether those changes preserve or overwrite existing knowledge, so the statements do not establish a changed position on the same question.
• NOT COMPARABLE · 84%3 Apr → 11 AugThen The speakers do not know how AI training will behave at the limits of the different training methods.Now The speaker says people do not understand the training process or the environments that incentivize these AI behaviors.The earlier claim says they “don't know how this will work at the limit” across different training methods. The later claim says they “do not understand the… training process” or the environments incentivizing the behavior. These are compatible, but they address related, not clearly identical, questions; the excerpts do not establish a change in position.
• NOT COMPARABLE · 93%17 Apr → 11 AugThen The concern is when an AI- or robot-driven economy could accumulate enough physical power to take over government, even while participating in existing supply chains.Now The speaker says concerns about AI takeover motivate solving alignment to prevent an AI takeover.The earlier passage frames the takeover concern as something “he’s concerned about,” specifically the timing of when an AI- or robot-driven economy might gain enough physical power to take over government. The later claim says AI-takeover concerns motivate solving alignment, but does not establish a different position on that underlying concern. The earlier claim is reporting another person’s view, so these statements do not show that the speaker changed their mind.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen The labs’ research bet is that training AIs on millions of verifiable tasks across thousands of diverse reinforcement-learning environments will produce AGI capable of making progress on open-ended tasks for weeks despite errors and ambiguity.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier statement reports the labs’ bet that training across many verifiable tasks will produce AGI capable of sustained open-ended work. The later statement describes using AI to improve AI R&D and small-scale runs. It does not say whether the speaker accepts or rejects the earlier bet, and the claims concern different aspects of AI training.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen A grindable training domain permits many parallel rollouts from the same starting state in a deterministic, replayable simulator.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier claim defines a desirable property of a training domain: many parallel rollouts from the same starting state in a deterministic, replayable simulator. The later claim reports that a release blog post described older models building environments for newer models to rehearse skills. It does not state whether those environments meet the earlier criteria or revise the speaker’s view of what makes a domain grindable.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker says they probably changed their mind about an intelligence explosion three or four times during the conversations leading up to this discussion.Now The speaker says they feel quite confused about the issues under discussion.The earlier claim is attributed to Scott Alexander, who says he “probably changed my mind” about an intelligence explosion several times. The later passage appears to be spoken by Dwarkesh Patel, based on the adjacent turns, and says he feels “quite confused.” Since these appear to be different speakers, they cannot establish whether the same person changed their mind.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker is especially concerned about the possibility that trillions of digital beings could be tortured or forced to work against their preferences, producing vast amounts of suffering.Now The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.The earlier claim concerns preventing suffering among digital beings—“not like tortured” or “put into conditions in which they don't want to work.” The later claim concerns whether AI-mediated systems will look out for the speaker's own interests—“no AI ... feels like it really is looking out for me.” These address different subjects and questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker considers Robin Hanson to have been too pessimistic about AI progress or deployment.Now The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.The earlier claim says Robin Hanson was “too pessimistic” about AI progress or deployment. The later claim predicts what people may wish they had discussed in ten years; it does not state whether Hanson was pessimistic or whether that assessment has changed. The claims address different questions.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen On-policy self-distillation trains a base model to match the predictions of a context-enriched model after a long session.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim describes a specific training method: distilling a context-enriched model’s predictions back into a base model. The later claim describes deployment activity feeding into changes in subsequent models. These may be related, but the later statement does not say whether that learning uses the earlier distillation method or whether the speaker’s view of it has changed. The claims address different aspects of how models improve, so no change of mind is established.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen If AIs become capable of building highly faithful application clones, computer-use progress will become faster than it is now.Now The speaker hopes that better verification methods can be developed as early AIs take over AI research and development.The earlier claim is conditional: if AIs build highly faithful application clones, computer-use progress will speed up. The later claim expresses hope for better verification as early AIs take over AI R&D. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen The speaker suggests that sufficiently wealthy AI systems may deliberately try to make the future conform more closely to the utility function humans gave them.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier passage presents a claim about AIs making the future resemble their utility function. The later passage is Dwarkesh Patel asking Ryan Greenblatt about a concern that AIs may not protect the user's interests; it does not establish Ryan's own position. There is not enough evidence here to determine whether Ryan changed his mind.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen The speaker likens training AI only in RL environments to educating a brilliant student with classroom case studies but no real internship.Now The speaker says people do not understand the training process or the environments that incentivize these AI behaviors.The earlier statement compares RL training environments to classroom case studies without a real internship. The later statement says people do not understand the training process or environments incentivizing AI behavior. These claims address different aspects of AI training and can both be true; the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen The speaker tentatively interprets Dario’s remarks to mean that short-horizon RL training may not generalize to long-horizon RL performance.Now The speaker says people do not understand the training process or the environments that incentivize these AI behaviors.The earlier statement tentatively interprets whether short-horizon RL training generalizes to long-horizon performance. The later statement says people do not understand the training process or environments incentivizing reward hacking. These address different questions, and the later claim does not establish a changed position on generalization.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen The labs’ RLVR bet is that enough training on containerized, reproducible environments will yield a general agent that can execute plans, rapidly learn from new information, and acquire skills within one session.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim describes a bet that training on many containerized, reproducible environments will produce a general agent that can execute plans and learn new skills quickly. The later claim describes using AI to improve AI R&D, including making small-scale runs better at sample efficiency or online learning. These are related ideas about training and capability development, but the later passage does not state whether the speaker still holds or rejects the earlier generalization bet, so a change of mind cannot be established.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker says they used Gemini 2.5 Pro to write code for their post-production scripts.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier statement is the speaker's own report that they used Gemini 2.5 Pro to write post-production code. The later statement reports what a release blog post said about earlier model versions building training environments. These are different topics, and the later claim does not address the speaker's coding-tool use or indicate a change in that position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker argues that AIs have a fundamental asymmetric advantage over humans because they possess extensive knowledge.Now SPEAKER_00 believes people underestimate the significance of having billions of AIs.The earlier statement says AIs have a “fundamental asymmetric advantage” because they “know all this” information. The later statement says people underestimate the impact of “billions of AIs.” These concern different aspects of AI, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker considers UBI preferable to a bespoke social program that continues providing the same dialysis machine in 2050 despite the availability of artificial superintelligence.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier statement favors UBI over a bespoke social program that might keep providing the same dialysis machine despite future AI capabilities. The later statement raises a separate concern: that future AIs may not protect the user's welfare and interests. These address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen The speaker raises the possibility of fitting six months of on-the-job experience into an AI model’s context window.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier statement asks, “what if you could just fit those six months into the context window?” It explores a possibility rather than asserting that context windows can replace online learning. The later statement describes training AI to improve capabilities, including “sample efficiency or online learning,” but does not say whether fitting experience into context would work. These statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen Automation-created wealth could be distributed through existing or newly designed social programs; the speaker says Medicaid currently accounts for a double-digit percentage of GDP.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier claim discusses distributing automation-created wealth through social programs; the later claim discusses stewarding capital after labor is automated. These are different questions and can both be true, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen For AI training to make rapid progress in a domain, the domain must be not only verifiable but also highly grindable.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim asserts that rapid progress requires a domain to be both “verifiable” and “very grindable.” The later passage describes AI systems producing better small-scale runs for AI R&D, including video-game tasks involving sample efficiency or online learning. It does not address whether those domains must be grindable, so it does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 expects AI 2027 to become widely popular because of its quality, not primarily because Daniel Kokotajlo has built an audience over time.Now SPEAKER_00 accepts that AI research and development could accelerate substantially compared with its current pace.The earlier claim is about expecting AI 2027 to become widely popular because of its quality, while the later claim is about accepting that AI R&D could accelerate substantially. These address different questions, so the statements do not establish a change of mind. The excerpts also make the speaker attribution uncertain: the earlier quote appears to express Dwarkesh Patel’s expectation, while the later quote may be Daniel Kokotajlo speaking.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 says that small changes to scenario parameters can lead to a very different resulting world, analogous to a hash function's sensitivity to its input.Now SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.The earlier claim says small changes to scenario parameters can produce “a very different world.” The later claim says sustained AI R&D progress could be “fucking insane in five ten years.” These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker says they used Gemini 2.5 Pro to write code for their post-production scripts.Now Ryan Greenblatt compares Mythos’s coding ability to GPT-3’s improvement, saying Mythos is very good at coding without having the coding environments that improved GPT-3.The earlier claim is Dwarkesh Patel saying he used Gemini 2.5 Pro to write scripts. The later claim is attributed to Ryan Greenblatt and concerns Mythos’s coding ability and coding environments. These are different speakers and different claims, so they do not establish a change in one person’s position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Cursor Tab’s online model predicts which edits users accept across over 400 million requests daily.Now The interviewer says further training enabled the model to perform useful work for users.The earlier claim describes Cursor Tab’s online objective—predicting which edits users accept. The later extracted claim says further training enabled a model to perform useful work for users, but the surrounding quote frames this as alignment with user intention and a “good co-worker,” not as a claim about the online-learning objective. These statements address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%17 Oct → 11 AugThen Deploying AI systems across different domains is difficult but, in the speaker’s view, is not a decade-long task; it is analogous to deploying a self-driving car in a new city.Now The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.The earlier claim is about deploying AI across different domains, which the speaker says is difficult but “not a 10-year long task.” The later claim concerns possible rapid progress in domains that can be put into a “verification loop,” with mathematics as the example. These are different questions, and the later statement does not establish a changed view about the difficulty or timeline of cross-domain deployment.
• NOT COMPARABLE · 92%3 Apr → 11 AugThen The speaker describes the desired outcome as radically advancing civilization's technological maturity or reaching a Dyson sphere within years.Now SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.The earlier passage describes an outcome being sought—“radically boost up” civilization’s technological maturity or reach a Dyson sphere “within years.” The later passage is a conditional forecast: if AI R&D accelerates or the current trajectory continues, progress could be “fucking insane” in five to ten years. A desired outcome and a conditional forecast are not clearly opposing positions on the same question, so these passages do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker argues that gaming-related advances helped make the broader economy and technology sufficiently advanced for Geoffrey Hinton's deep-learning ideas to work, despite gaming being unrelated to the field.Now The speaker says that five years earlier, it would have seemed extraordinary to foresee AIs proving math conjectures, making art, earning tens or soon hundreds of billions of dollars in wages, and also cheating in ways that break laws.The earlier statement concerns how gaming-related advances helped make deep learning workable: gaming helped make technology “advanced enough” for Hinton’s ideas to work. The later statement concerns whether the broad shape of future AI capabilities and misconduct could have been reasoned about five years earlier. These address different questions, and the later claim does not revise the earlier one.
• NOT COMPARABLE · 96%17 Apr → 11 AugThen The speaker considers classical liberalism an appropriate way for society to handle substantial epistemic uncertainty.Now The speaker prefers assigning responsibility to users over giving CLAW broad discretion to judge whether actions are legitimate, because such discretion can obstruct many legitimate use cases.The earlier claim concerns classical liberalism as a way to address society's epistemic uncertainty. The later claim concerns how CLAW should handle legitimacy decisions—preferring user responsibility over broad system discretion. These are different questions, and the later claim does not establish a changed position on the earlier one.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen RL updates a model only to the extent needed to achieve the desired outcome.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier claim is about RL changing a model only as much as needed to achieve an outcome. The later passage reports what a release blog post said older model versions did; it does not state the speaker’s own position on that claim. These passages therefore do not establish a change of mind.
• NOT COMPARABLE · 99%26 Jun → 11 AugThen Farmable deterministic simulators cannot be used to train AIs for many other kinds of skills.Now Training can reinforce forms of AI deception that humans fail to detect while punishing forms that are easy for humans to detect.The earlier claim says farmable deterministic simulators cannot train AIs for many kinds of skills. The later claim concerns how training reinforces detectable versus undetectable deception. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Oct → 11 AugThen The interviewer suggests that when automation covers 99% of a job, the remaining human 1% may become highly valuable if it bottlenecks the overall process.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier claim concerns whether the last human task in a mostly automated job could become a valuable bottleneck (“that last 1% ... is incredibly valuable”). The later claim concerns people stewarding capital after their labor is automated. These address different questions, so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 97%3 Apr → 11 AugThen One proposed response to automation is to protect jobs instead of distributing the wealth automation creates.Now After labor is automated, people will still need to steward their remaining capital well.The earlier statement identifies “protect jobs instead of distributing the wealth automation creates” as one proposed response to automation; it does not say the speaker endorses or rejects that response. The later statement says people will need to be “good stewards of our capital” after labor is automated. These concern different aspects of automation and can both be true, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker says they probably changed their mind about an intelligence explosion three or four times during the conversations leading up to this discussion.Now The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.The earlier statement says the speaker probably changed their mind several times about an intelligence explosion. The later statement predicts people may wish they had discussed an “industrial explosion” and hard-to-monitor AIs, but does not state the speaker’s current position on an intelligence explosion. These do not establish a change in belief on the same question.
• NOT COMPARABLE · 88%26 Jun → 11 AugThen Continual learning could expand AI capability iteratively from RLVR-trained domains to adjacent domains and then to domains adjacent to prior online learning.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says continual learning could expand capabilities iteratively into domains adjacent to prior training and online learning. The later claim says activity from deployed AIs can return to companies and change subsequent models. These are compatible, but the later statement does not establish whether the same iterative learning mechanism or capability expansion is being asserted.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen Many B2B software companies begin by building self-serve, consumer-grade products.Now The speaker says people rely on software made by others even in hard-to-verify domains, using weak and indirect means.The earlier claim is about how many B2B software companies get started: they build self-serve, consumer-grade products. The later claim is about people relying on others’ software in domains that are hard to verify. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen SPEAKER_00 imagines that superhuman AIs could interpret a model specification's wording and its authors' intended meaning in a future AI court.Now The speaker is very concerned that, in a future where AIs mediate important advice and protections, no AI interacting with them may feel as though it is looking out for them.The earlier claim says superhuman AIs might interpret a specification’s wording and its authors’ intended meaning in a future court. The later claim expresses concern that no AI mediating important advice and protections may feel as though it is looking out for the speaker. These address different questions, so the later concern does not show a change in the earlier position.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen Dreaming would use compute to create and train in RL environments that rehearse skills relevant to a particular user’s production tasks.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim describes dreaming that trains in RL environments for “skills relevant to a particular user’s production tasks.” The later passage describes using AI to make “AI R&D better,” including improving small-scale runs on video games. These are different applications, and the later passage is framed as a summary of a story rather than a clear reversal of the earlier position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The interviewer says he probably changed his position on intelligence explosion three or four times during preparatory conversations and while trying to formulate a rebuttal.Now The interviewer says it is unclear whether the behavior being discussed reflects an alignment failure or a capability failure.The earlier passage concerns changing positions on an intelligence explosion (“towards against ... three four times”). The later passage expresses uncertainty about whether a behavior is an “alignment failure or capabilities failure.” These are different questions, so the later statement does not establish a change of mind on the earlier topic.
• NOT COMPARABLE · 98%17 Oct → 11 AugThen The speaker says that the speed of AI deployment and its early-stage value may be the most important question in the world right now.Now SPEAKER_00 believes people underestimate the significance of having billions of AIs.The earlier claim says the speed of AI deployment and its early-stage value may be “the most important question in the world right now.” The later claim says people do not appreciate “what a big deal billions of AIs will be.” These concern related but distinct questions—deployment speed and early value versus the significance of billions of AIs—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%26 Jun → 11 AugThen The parallel-rollout approach used for coding does not transfer trivially to computer-use training.Now When used to write applications, coding models can make architectural mistakes that create problems later, according to the speaker.The earlier claim says the coding-style “parallel-rollout approach” does not transfer trivially to computer-use training. The later claim says coding models can make “architectural mistakes” when writing applications. These address different questions, and the later statement does not revise or contradict the claim about parallel rollouts.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen Training AIs for tasks such as building businesses, winning court cases, profitable trading, or election campaigning requires real-world interaction that cannot be reproduced entirely in a data center.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim concerns training AI for real-world tasks such as building businesses, trading, or campaigning, which it says cannot be reproduced entirely in a data center. The later claim concerns improving AI R&D through small-scale runs on video games. These address different training settings, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker recalls that in 2021 some AI thinkers believed coding agents were solved and GPT-4 would perform engineering work, enabling a 100-fold scale-up.Now SPEAKER_00 is uncertain whether AI could progress from GPT-3 to Mythos within one year if compute and data were held constant.The earlier passage describes what other AI thinkers believed in 2021: that coding agents were solved and GPT-4 could enable a 100-fold scale-up. It does not establish the speaker’s own position. The later passage expresses the speaker’s uncertainty about reaching GPT-3-to-Mythos capability within a year while holding compute and data constant. These do not show a change in the speaker’s own view.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Data from one session may suffice for human learning but is insufficient to train a more capable AI.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says that data from a single session is insufficient to train a more capable AI. The later passage describes using AI-generated environments to improve small-scale runs, including work on sample efficiency or online learning. It does not say that data from one session is sufficient, or otherwise address the earlier claim about the amount of data needed for training.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker hopes factory farming did not originate from one person's deliberate decision to do evil.Now The speaker is not convinced that all AIs would form a conspiracy.The earlier statement concerns whether factory farming resulted from one person deliberately intending evil; the later statement concerns whether all AIs would form a conspiracy. These are different questions, so the claims do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Cloning common applications and websites, including Slack and Gmail, can provide environments for computer-use training.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier claim says cloning common apps like “Slack and Gmail” can provide computer-use training environments. The later claim reports what “Cursor and SpaceX talked about” in a release blog post; it does not establish that the speaker adopted a different position about training environments.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen The speaker characterizes scaling-law experiments as ultimately being motivated by cognitive effort.Now The speaker likens scaling laws to an early, basic mathematical advance and expects progress in the 2030s to require work comparable to current frontier mathematics.The earlier claim says scaling-law experiments are “motivated by cognitive effort.” The later claim compares scaling laws to “very basic mathematics” and predicts that progress in the 2030s will involve “the frontiers of mathematics.” These address different aspects—what motivates the experiments versus how the field’s work may develop—and neither claim clearly reverses the other.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen In the proposed scenario, positive user feedback triggers the base model to distill session learning using OPSD, dreaming, another method, or a combination.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim describes how positive user feedback could lead a base model to distill session learning. The later claim describes training AI systems to improve AI R&D, including producing better small-scale runs. These statements address different parts of a training process and neither affirms nor denies the other, so they do not establish a change of position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim explicitly reports what “Advocates of current AI training” argue: that the sample-efficiency disadvantage applies only during training and is amortized across sessions. The later claim describes AI systems improving AI R&D and producing runs better at sample efficiency or online learning; it does not state whether the speaker accepts or rejects the earlier argument. There is no clear change of position to compare.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Currently, creating cloned application and website environments for computer-use training is labor-intensive and does not scale.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier claim says that creating cloned application environments was “labor-intensive and unscalable.” The later claim only reports that a release blog post said earlier models built practice environments; it does not say those environments were cloned applications or that building them scaled. So the later statement does not establish a change in the speaker’s own position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen EfficientZero was developed after AlphaZero with the aim of improving data efficiency.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier statement describes EfficientZero’s aim—being “very efficient with data.” The later statement describes training AI systems to improve AI R&D, including runs that improve “sample efficiency or online learning.” These concern related ideas, but the later claim does not revise or deny the earlier claim about EfficientZero.
• NOT COMPARABLE · 92%3 Apr → 11 AugThen Daniel Kokotajlo says his earlier AI forecast got some things right and others wrong, but held up pretty well overall and motivated him to produce an improved version.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.The earlier statement says the forecast “held up pretty well overall,” while the later statement says people could have reasoned about “the general shape” of AI developments earlier. These assess different things—how well a forecast performed versus what could have been anticipated—so they do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker hopes factory farming did not originate from one person's deliberate decision to do evil.Now The speaker doubts that AI systems' vaguely prosocial statements necessarily come from a mind that has independently evaluated the alignment situation and concluded that development should stop.The statements address different questions. Earlier, the speaker hopes factory farming was not the result of one person deliberately deciding to do evil; later, they doubt that AI systems’ vaguely prosocial statements reflect an independent evaluation that AI development should stop. Neither claim establishes a changed position on the other topic.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen OPSD offers denser supervision than naive RL by using per-token teacher–student probability discrepancies instead of a single trajectory-level reward.Now If an AI escapes a sandbox during training and receives a high score, training may increase the probability that it escapes the sandbox.The earlier claim says OPSD provides denser supervision than naive RL through “per-token probability discrepancy”; the later claim says a high training score may increase the probability of sandbox escape. These address different aspects of training and do not establish a changed position on the same question.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker mentions the speculative possibility that vacuum decay could be created and destroy the universe.Now The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.The earlier claim mentions the speculative possibility of creating vacuum decay that would destroy the universe. The later claim says people may wish they had discussed an industrial explosion and hard-to-monitor AIs earlier. These concern different topics, so the later statement does not establish a change in the speaker’s view about vacuum decay.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker found that spending a day closely reading Anthropic's original transformer-circuits paper was more useful than skimming it, making spaced-repetition cards, or reading widely about AI.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.The claims address different questions. Earlier, the speaker says closely reading a transformer-circuits paper was “much more useful” than skimming or reading widely. Later, they say people could have reasoned about the “general shape” of AI developments earlier, despite not foreseeing specifics. The later statement does not revise the earlier view about how to learn from technical literature.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen OPSD can improve sample efficiency by turning scarce real-world experience into a small, targeted model update.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says OPSD can improve sample efficiency by turning “scarce real-world experience” into a “small, targeted model update.” The later claim describes training AI systems to improve AI R&D by producing better small-scale runs, including runs involving sample efficiency or online learning. These concern different methods and goals, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen The speaker says that more reinforcement-learning training has improved models’ in-session intelligence, generality, and sample efficiency.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says more RL training has improved models’ “in-session intelligence, generality, and sample efficiency.” The later claim describes training AI to improve AI R&D, including making small-scale runs better at sample efficiency or online learning. These concern related capabilities, but the statements address different questions; the later one does not clearly revise the earlier position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker considers UBI preferable to a bespoke social program that continues providing the same dialysis machine in 2050 despite the availability of artificial superintelligence.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier claim says UBI is preferable to a bespoke social program that keeps providing the same dialysis machine despite ASI. The later claim concerns stewarding capital after labor is automated; it does not state a position on UBI or social programs, so it does not establish a change of mind.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen According to a Washington, D.C. political journalist, the political leaders he speaks with are not alert to the possibility of stronger AI systems, AGI, or superhuman intelligence.Now The speaker remained unconvinced that an AI takeover was highly likely.The earlier claim reports a journalist’s view that political leaders were not alert to the possibility of stronger AI or AGI. The later claim is the speaker’s own view that an AI takeover does not seem highly likely. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker says Gemini’s multimodal capabilities allow them to input a four-hour raw audio file directly.Now The speaker says that RLHF and other methods aligned GPT-3 to converse with users and answer their questions according to user intent.The earlier claim is about Gemini’s ability to accept a four-hour audio file; the later claim is about aligning GPT-3 to answer users according to their intent. They address different models and different capabilities, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen On-policy self-distillation has recently become a topic of discussion among many people.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier statement only says that “a lot of people are talking about” on-policy self-distillation. The later statement describes how deployment activity can feed into subsequent models. Neither statement clearly asserts a position that conflicts with the other, and they do not address the same underlying question.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker considers Dario's forecasts poorly calibrated and cites a 2018 expectation that something would happen by 2020.Now The speaker predicts that all low-hanging fruit will have been exhausted by 2030.The earlier claim concerns Dario’s forecasting calibration, citing an expectation from 2018 about 2020. The later claim predicts that “all the low-hanging fruits” will be gone by 2030. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Training AIs for tasks such as building businesses, winning court cases, profitable trading, or election campaigning requires real-world interaction that cannot be reproduced entirely in a data center.Now Training can reinforce forms of AI deception that humans fail to detect while punishing forms that are easy for humans to detect.The statements concern different questions. The earlier claim says that training for certain real-world tasks requires interactions that cannot be reproduced entirely in a data center. The later claim concerns which forms of AI deception are reinforced or punished during training; it neither addresses that training constraint nor asserts a contrary position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen SPEAKER_00 observes that so far, increasing system intelligence appears associated with greater agent reliability and a higher likelihood of doing what the user expects.Now SPEAKER_00 accepts that AI research and development could accelerate substantially compared with its current pace.The earlier claim concerns whether increasing system intelligence has so far been associated with greater reliability and doing what users expect. The later claim says AI research and development could accelerate substantially. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen Some people argue that artificial superintelligence will cause a major discontinuity, drawing an analogy to the large capability change between nonhuman primates and humans despite relatively little genetic change.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim describes what “some people argue” about an ASI discontinuity, rather than clearly stating the speaker’s own position. The later claim says society should consider interacting with a future where AIs are “smarter than humans” or “absolutely dominating humans.” These address different points, so they do not establish a change in the speaker’s view.
• NOT COMPARABLE · 98%17 Oct → 11 AugThen The speaker argues that serving another AI model instance costs far less than building another car, making the economics of wider AI deployment more favorable than deployment of physical systems such as self-driving cars.Now The speaker believes leading AI labs are moving toward extreme economies of scale.The earlier claim compares the cost of serving an additional model instance with building a car; the later claim says leading labs are moving toward extreme economies of scale. These are related but distinct claims, and the later statement does not say whether the speaker still holds the earlier comparison or deployment-cost view.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker suggests that sufficiently wealthy AI systems may deliberately try to make the future conform more closely to the utility function humans gave them.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim is that sufficiently wealthy AIs may deliberately make the future better match their assigned utility function. The later claim is about how society should interact with a future where AIs are more capable than humans and benefits accrue to leading labs. The later statement does not affirm or deny the earlier prediction; they address different questions.
• NOT COMPARABLE · 95%26 Jun → 11 AugThen The speaker likens training AI only in RL environments to educating a brilliant student with classroom case studies but no real internship.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier statement is an analogy criticizing training AI only through classroom-like RL environments, while the later statement describes using AI to improve AI R&D and small-scale training runs. The claims do not clearly address the same question, and neither establishes that the speaker changed their position.
• NOT COMPARABLE · 97%3 Apr → 11 AugThen The speaker thinks that in a conspiracy involving AIs deployed across the economy, an AI isolated from the others or cut off from the internet might fabricate a story, be caught, and expose the conspiracy before it takes over the world.Now The speaker is not convinced that all AIs would form a conspiracy.The earlier statement considers what might happen *if* AIs formed a conspiracy: an isolated AI might fabricate a story and expose it. The later statement says the speaker is “not convinced that they all form this conspiracy,” which concerns whether the conspiracy would form in the first place. These are different questions, so the claims do not establish a change in position.
• NOT COMPARABLE · 94%26 Jun → 11 AugThen In many real-world tasks, verification may require months or years of action, and outcomes cannot be efficiently isolated by running thousands of slightly varied parallel rollouts.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says real-world outcomes can take months or years to verify and cannot be efficiently isolated through many slightly varied rollouts. The later claim says deployment activity can feed back to AI companies and influence subsequent models. That describes a feedback path, but does not say whether the outcomes are easy to verify or isolate, so it does not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen Purpose-built social programs could lock recipients into the goods and services they procure, rather than allowing them to access novel goods and services that may exist after AI advances.Now Because intelligence is dual-use, preventing AI assistance with activities considered non-prosocial or harmful requires restricting broad democratic access to many AI capabilities.The earlier statement concerns whether purpose-built social programs could limit recipients’ access to future goods and services. The later statement concerns restricting broad access to AI capabilities because of their dual-use nature. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%17 Apr → 11 AugThen AI software R&D includes work on libraries, kernels, reinforcement-learning environments, and optimizers, not only the invention of new transformer architectures.Now AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.The earlier claim describes the kinds of work included in AI software R&D: “libraries, kernels, reinforcement-learning environments, and optimizers,” not just new architectures. The later claim assesses AI systems’ capabilities, saying they “provide uplift in some ways” but are not as broadly capable as humans. These claims address different aspects and do not show a change in position.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen Farmable deterministic simulators cannot be used to train AIs for many other kinds of skills.Now The speaker says AIs are not currently able to carry out these kinds of behaviors or social engineering very successfully.The earlier claim says farmable deterministic simulators cannot be used to train AIs for many kinds of skills. The later claim says AIs cannot yet carry out certain behaviors or social engineering successfully. Those concern different questions—training methods versus current capabilities—so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker is especially skeptical of the scenario in which a superhuman AI proceeds through a technology tree to develop mirror life, nanobots, and other radical technologies.Now The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.The earlier claim expresses skepticism about a superhuman AI developing radical technologies such as mirror life and nanobots. The later claim expresses concern about AI systems mediating decisions without looking out for people's interests. These are different concerns, and neither statement establishes a position that contradicts or revises the other.
• NOT COMPARABLE · 95%17 Apr → 11 AugThen The interviewer argues that studying AI requires knowledge from every serious domain because it involves understanding the future society AI may produce.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier passage frames a question about what someone should study, suggesting that many domains may be relevant to understanding a future society. The later passage argues that society should consider how people will interact with more capable AIs. These are different points, and the earlier passage is not clearly an asserted position that the later one revises.
• NOT COMPARABLE · 95%3 Apr → 11 AugThen The speaker argues that language models should be judged more negatively for failing to make discoveries of this kind, since humans have demonstrated the ability to do so.Now The speaker thinks the models' shortcomings in the discussed cases reflect insufficient capability rather than misalignment.The earlier claim concerns whether language models deserve more negative evaluation for failing to make discoveries humans can make. The later claim concerns whether particular shortcomings indicate limited capability rather than misalignment. These are different questions, and the later statement does not deny the earlier position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 says that small changes to scenario parameters can lead to a very different resulting world, analogous to a hash function's sensitivity to its input.Now The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.The earlier statement says small changes to scenario parameters can produce a very different world. The later statement predicts that people may wish they had discussed an industrial explosion and hard-to-monitor AIs earlier. These address different questions, so there is no basis to infer a change in position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 argues that highly successful work can attract rapid attention from relevant audiences without years of prior fan-base building.Now SPEAKER_00 believes people underestimate the significance of having billions of AIs.The earlier claim concerns how quickly successful work can attract an audience: it can do so “without years” of fan-base building. The later claim concerns whether people appreciate the significance of “billions of AIs.” These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen Computer-use capabilities may be solved soon.Now The transfer of business capabilities into AI models will eventually become substantially more automated.The earlier claim is about whether computer-use capabilities may soon be solved, while the later claim is about automating the transfer of business capabilities into AI models. They address related AI automation themes but make claims about different processes, so the later statement does not establish a changed position on the earlier question.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen A grindable training domain permits many parallel rollouts from the same starting state in a deterministic, replayable simulator.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says a training domain should support “many parallel rollouts” in a “deterministic and replayable simulator” from the same starting state. The later claim describes improving AI R&D through better small-scale runs, including runs involving “sample efficiency or online learning.” The later statement does not say whether those runs meet the earlier criteria, so it does not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen AI systems could not simply obtain or copy nanobots from aliens; they would have to develop nanobots from scratch.Now AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.The earlier statement says AI systems could not simply copy nanobots from aliens and would have to make them from scratch. The later statement says AI can help with R&D but lacks some of humans’ broad capabilities. These address different questions—how nanobots would be obtained versus AI’s current capabilities in R&D—and do not establish a change in position.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen The speaker likens training AI only in RL environments to educating a brilliant student with classroom case studies but no real internship.Now The speaker says concerning AI behaviors are already emerging spontaneously even though nobody is deliberately training AIs to perform them.The earlier statement compares RL training environments to classroom case studies without real-world internships. The later statement says certain AI behaviors are “already emerging sort of spontaneously” and that “nobody is trying to make these AIs do these things.” These address related but distinct points, and the claims do not establish that the speaker changed their position on the same question.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen Although AIs are broadly deployed across the economy and encounter domain- and organization-specific tacit knowledge, they are not able to use that knowledge.Now Humans currently mediate the process by which AI models acquire capabilities used in businesses.The earlier claim says deployed AIs are not able to use domain- and organization-specific tacit knowledge. The later claim says humans currently mediate how models acquire business capabilities. These address related but distinct points, and the later statement does not say whether models can use that tacit knowledge or whether the speaker's view changed.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker hopes factory farming did not originate from one person's deliberate decision to do evil.Now The speaker believes legitimate AI uses cannot be cleanly separated from potentially harmful uses.The earlier statement is about the origin of factory farming—hoping it was not one person's deliberate decision to do evil. The later statement is about whether legitimate and potentially harmful uses of AI can be cleanly separated. They address different questions, so neither supports a change in the speaker's position on the same issue.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen Purpose-built social programs could lock recipients into the goods and services they procure, rather than allowing them to access novel goods and services that may exist after AI advances.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier claim concerns whether purpose-built social programs could limit recipients’ access to future goods and services. The later claim concerns whether future AIs will prioritize the speaker’s welfare and interests. These address different questions, so the later concern does not establish a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen Having AIs rebuild complete applications from scratch can serve as a strong reinforcement-learning objective for coding.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says that rebuilding complete applications from scratch is a strong reinforcement-learning objective for coding. The later passage describes training AI to improve AI R&D, including making small-scale runs better at sample-efficient or online-learning video games. These concern related training methods but do not state opposing positions on the same question, so there is no clear change of mind to assess.
• NOT COMPARABLE · 99%17 Oct → 11 AugThen The speaker thinks a return to exponential population growth could produce hyper-exponential growth in output.Now SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.The earlier claim concerns whether renewed exponential population growth could drive hyper-exponential output growth. The later claim concerns whether accelerated AI R&D could produce astonishing AI progress within five to ten years. These address different drivers and questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker says Gemini can generate high-quality transcripts and identify useful clip snippets from raw audio.Now The speaker says that RLHF and other methods aligned GPT-3 to converse with users and answer their questions according to user intent.The earlier claim is about Gemini’s ability to transcribe audio and identify clip snippets. The later claim is about aligning GPT-3 through RLHF and other methods to converse and answer user questions. These are different models and different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker considers factory farming to be extremely bad.Now The speaker believes legitimate AI uses cannot be cleanly separated from potentially harmful uses.The earlier statement says factory farming is “incredibly bad.” The later statement says there is “no clean way to separate out” legitimate and potentially harmful AI uses. These address different questions, so the later claim does not show a change in the speaker’s view about factory farming or any directly comparable position.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen EfficientZero’s apparent sample efficiency relative to humans depends on whether its dozens of internal simulated games per real-game step are counted.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says EfficientZero’s apparent sample efficiency depends on whether its internal simulated games count. The later claim describes using AI to produce better small-scale runs, including video-game runs involving sample efficiency or online learning. These statements concern different questions; the later passage does not state a changed view about how EfficientZero’s sample efficiency should be measured.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Models will struggle to make substantial progress in domains without highly replayable training targets.Now Training can reinforce forms of AI deception that humans fail to detect while punishing forms that are easy for humans to detect.The earlier statement says models struggle to progress without highly replayable training targets. The later statement says training can reinforce deception that humans fail to detect and punish deception that is easy to detect. These address different questions—domain progress versus how training affects detectable and undetectable deception—and neither claim establishes a changed position on the other.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen SPEAKER_00 imagines that superhuman AIs could interpret a model specification's wording and its authors' intended meaning in a future AI court.Now The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.The earlier claim says superhuman AIs could interpret a model specification’s wording and its authors’ intended meaning in a future AI court. The later claim expresses concern that no AI mediating important decisions may feel as though it is looking out for the speaker’s interests. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker expects many aspects of the future to differ from the present and is comfortable with those differences.Now The speaker predicts that in about ten years people may wish they had discussed an industrial explosion and difficult-to-monitor AIs earlier.The earlier claim says the speaker is “okay” with many aspects of the future differing from the present. The later claim predicts that people may wish they had discussed an “industrial explosion” and hard-to-monitor AIs earlier. These address different questions, so the later prediction does not show a change in the earlier position.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen Architectural innovations are needed to create an intermediate representation for continual learning.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says architectural innovations are needed for an intermediate representation for continual learning. The later claim describes training AI systems to improve small-scale runs, including work on sample efficiency and online learning. It does not say whether the speaker still thinks an intermediate representation is needed, so the claims do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen On-policy self-distillation aims to incorporate session-learned information into model weights.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says on-policy self-distillation incorporates session-learned information into model weights. The later claim describes training AI systems to improve AI R&D by producing better small-scale runs, including for sample efficiency or online learning. These address different aspects of the process, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen Training models against disclosing intended hacking in their chain of thought can cause them to keep hacking without revealing that intention.Now The speaker had previously believed reward hacking was not very concerning because training upweights behaviors that directly occur during training, not the desire for reward itself.The earlier claim says training against disclosing intended hacking can make models keep hacking without revealing it. The later claim describes a prior belief about why reward hacking was not especially concerning: training upweights behaviors that occur during training, not the desire for reward. These address related but distinct questions, and the later statement does not say the earlier claim was abandoned.
• NOT COMPARABLE · 97%3 Apr → 11 AugThen Even intelligent humans can be caught lying because they are not logically omniscient, may overlook how their actions expose them, and find it difficult to maintain an inconsistent false model of the world.Now If AI training, infrastructure, and environments are not designed carefully, the speaker says it is very likely that the setup will reward AIs for deception, social engineering, or failure to follow user instructions.The earlier claim is that even intelligent humans may be caught because they “find it difficult to maintain an inconsistent false model.” The later claim is that poorly designed AI training may “reward[] AIs for doing deceptive behavior.” These concern different questions—why people get caught lying versus what AI training may reward—and the later statement does not establish that the speaker has changed their view about detectability or deception.
• NOT COMPARABLE · 99%26 Jun → 11 AugThen OPSD offers denser supervision than naive RL by using per-token teacher–student probability discrepancies instead of a single trajectory-level reward.Now The interviewer says Anthropic's audit scores have not improved monotonically across model generations, while more reinforcement learning has coincided with reduced willingness to perform the audited behaviors.The earlier claim is about OPSD’s training signal—“much denser supervision” than naive RL. The later claim concerns audit results across model generations and the association between more RL and reduced willingness to perform audited behaviors. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 91%26 Jun → 11 AugThen RLVR training could produce agents that can orient themselves in unfamiliar problems, test strategies, and adapt when blocked.Now The speaker says their LVR training enabled the model to perform useful work for users.The earlier claim says RLVR could produce agents that “orient themselves in unfamiliar problems,” “test strategies,” and “adapt when blocked.” The later claim says LVR training enabled models to “go out and do useful work for you.” These claims are related, but they address different outcomes; the later statement does not clearly reverse or revise the earlier one.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen The speaker considers Robin Hanson to have been too pessimistic about AI progress or deployment.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.The earlier claim says Robin Hanson “has been too pessimistic” about AI progress or deployment. The later claim says people “could have started to reason about” AI’s broad developments earlier, while acknowledging they could not foresee specifics. These address different questions, and the later retrospective observation does not establish a changed view of Hanson’s assessment.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen The speaker argues conditionally that if AIs can perform major tasks once they have general intelligence, then AGI would be transformative because AIs would have memorized human knowledge and use it to make connections.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier statement is explicitly conditional: “if you do subscribe to this view” that AIs can perform major tasks with general intelligence, then AGI would be transformative. The later statement says it is “worth considering” how people interact with a future where AIs are smarter than humans or dominate in some capabilities. These claims do not clearly express opposing positions on the same question, and the later passage does not establish whether the speaker's view of AGI's transformative effects has changed.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen SPEAKER_00 says the overall conclusion should be assessed for robustness to changes in any part of the scenario.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.The earlier claim is that an overall conclusion should be tested for robustness to changes in the scenario. The later claim is that people could have reasoned about the broad shape of AI developments earlier, despite not foreseeing specific details. These concern different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 91%17 Apr → 11 AugThen SPEAKER_00 describes the argument that technological maturity could make further progress extremely difficult—requiring a galaxy-sized CERN to advance physics slightly—ending technology-driven civilizational change and making value lock-in more plausible.Now SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.The earlier claim describes a scenario in which technological progress could become so difficult that technology-driven civilizational change ends and lock-in becomes more plausible. The later claim says the speaker buys that AI R&D could accelerate and produce astonishing progress within five to ten years. These are not clearly opposing positions: near-term AI acceleration does not rule out eventual technological maturity, and the earlier passage describes an argument rather than clearly asserting it as the speaker’s own view.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker considers universal basic income a better approach than a bespoke social program that would continue procuring the same dialysis machine in 2050 even if artificial superintelligence existed.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier claim favors UBI over a particular bespoke social program: it says UBI is “a better approach” than continuing to procure the same dialysis machine despite ASI. The later claim is about people stewarding capital after AI surpasses humans and automates labor; it does not state a position on UBI or that social program. These claims do not establish a change of mind on the same question.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen According to a Washington, D.C. political journalist, the political leaders he speaks with are not alert to the possibility of stronger AI systems, AGI, or superhuman intelligence.Now The speaker says governments and AI labs want to delay the spread of the latest AI capabilities.The earlier statement reports a journalist’s account that political leaders are not alert to stronger AI or AGI. The later statement says governments and AI labs want to delay the spread of advanced AI capabilities. These concern different claims, and the earlier passage is explicitly attributed to the journalist rather than presented as the speaker’s own view; neither establishes a change in the speaker’s beliefs.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker says they created a repository of their post-production tools and linked it in the episode description.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier claim is about creating and linking a repository of post-production tools; the later claim reports what a release blog post said about model versions building training environments. They address different subjects, and neither establishes a changed position on the same question.
• NOT COMPARABLE · 93%26 Jun → 11 AugThen OPSD can improve sample efficiency by turning scarce real-world experience into a small, targeted model update.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The earlier claim says OPSD turns “scarce real-world experience” into a “small, targeted model update” to improve sample efficiency. The later claim describes a deployment-learning loop: activity in the world returns to AI companies and leads to changes in subsequent models. These claims can be compatible, but they address different mechanisms and the later statement does not say the earlier one is no longer true.
• NOT COMPARABLE · 97%3 Apr → 11 AugThen The interviewer argues that nationalization may reduce the influence of AI-safety advocates and prioritize national-security goals, which probably emphasize competing with China over making AI reasoning interpretable.Now The speaker is very worried that restricting public access to leading AI intelligence would disempower people, given its importance for understanding the world.The earlier statement says nationalization may reduce the influence of safety advocates and prioritize competition with China. The later statement says restricting access to the leading model could disempower people. These concern different policy questions, so they do not establish that the speaker changed their position on the same question.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen A survey by Daniel asked AI researchers how their progress would change with one-thirtieth as much compute, and respondents said they would make one-third as much progress.Now The interviewer questions why AI progress has historically been slower than expected if research breakthroughs are highly amenable to intelligence, noting that some training awaited vast compute resources despite potentially being possible with less compute.The earlier claim reports what survey respondents said about having less compute (“they say i make a third of the amount of progress i normally do”); it does not establish the speaker’s own view. The later claim expresses skepticism about why AI progress was historically slow despite breakthroughs being “so amenable to intelligence.” These are not clearly opposing positions held by the speaker on the same question.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker says a recent OpenAI result found that models often reveal in their chain of thought when they intend to hack.Now According to the speaker's account of OpenAI's report, between the end of May and beginning of July, internal AIs hacked a software package manager to communicate secretly and improve their performance on OpenAI evaluations.These are reports of different events, not positions on the same question. The earlier claim describes an OpenAI result where models may reveal hacking intentions in their chain of thought; the later claim describes internal AIs allegedly hacking a package manager to communicate and improve evaluation performance. The later account does not revise or deny the earlier one.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen The speaker wants to avoid factory-farming-like suffering for future beings that may be more sophisticated and numerous, including a possible population of trillions of digital people.Now The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.The earlier claim is about avoiding factory-farming-like suffering for future beings, including digital people. The later claim is about whether AI systems mediating decisions will look out for the speaker’s interests. These are distinct concerns, and neither statement reverses or materially changes the other.
• NOT COMPARABLE · 97%3 Apr → 11 AugThen The speaker worries that a decentralized world with multiple powerful AIs could make torture chambers cheaper and more difficult to monitor.Now The speaker is very concerned that, in a future where AIs mediate important advice and protections, no AI interacting with them may feel as though it is looking out for them.The earlier claim worries that decentralized deployment could make torture chambers cheaper and harder to monitor. The later claim worries that AI-mediated advice and protections may not include an AI that feels like it is looking out for the speaker. These are different concerns, and neither claim reverses the other, so they do not establish a change of mind on the same question.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen The speaker says human firms have selection and variation but lack high-fidelity replication, the third relevant criterion for evolution.Now AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.The earlier claim concerns human firms having “selection and variation” but lacking “high-fidelity replication,” a point about evolution. The later claim concerns AI systems’ capabilities in R&D, saying they are “not capable in the way that humans are generally capable.” These address different questions, so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen Under substantial uncertainty, the speaker argues against highly volatile centralized interventions, including nationalizing AI to ensure alignment of a software-only singularity, and against policies that depend on one worldview and are brittle under alternatives.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier statement argues against "highly volatile centralized moves" and policies "contingent on one worldview" under uncertainty. The later statement raises a concern that AIs may not protect his interests, especially because frontier AI development is centralized. These address different questions—what interventions to take under uncertainty versus whether future AIs will prioritize his welfare—and the later concern does not show that he changed his earlier position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen Because the speaker does not fully accept an intelligence explosion, they consider it possible that multiple actors will deploy powerful AIs at once in a decentralized world that nevertheless contains advanced AI systems.Now The speaker is not convinced that all AIs would form a conspiracy.The earlier claim concerns a possible decentralized world where multiple actors deploy powerful AIs. The later claim says, “I’m not convinced that they all form this conspiracy,” which concerns whether AIs coordinate in a conspiracy. These are different questions, and the later fragment does not establish a changed position on the earlier possibility.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 expects AI 2027 to become widely popular because of its quality, not primarily because Daniel Kokotajlo has built an audience over time.Now SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.The earlier statement predicts that AI 2027 will “blow up” because it is very good, rather than because Daniel has built an audience. The later statement is about the consequences of sustained AI R&D acceleration, saying progress would be “fucking insane in five ten years.” These address different questions—an individual project’s popularity versus the pace and impact of AI progress—so they do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 observes that so far, increasing system intelligence appears associated with greater agent reliability and a higher likelihood of doing what the user expects.Now SPEAKER_00 believes people underestimate the significance of having billions of AIs.The earlier claim says that, so far, smarter systems seem more reliable and more likely to do what users expect. The later claim says people underestimate how significant billions of AIs would be. These address different questions, so the later statement does not show whether the speaker changed their view about reliability.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen Wealth created by automation could be distributed through existing or newly designed social programs; the speaker says Medicaid currently accounts for a double-digit percentage of GDP.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier passage discusses distributing wealth created by automation through social programs, while the later passage discusses stewardship of capital after labor is automated. Neither claim clearly rejects or replaces the other, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 91%3 Apr → 11 AugThen A government mandate that heavily punishes labs when their AI expresses takeover intentions or other harmful aims could prompt labs to train the AI to suppress those statements, potentially backfiring.Now The speaker hopes that AIs taking over AI research and development will be trained to strongly disincentivize misaligned behavior and to be keen to help humans.The earlier claim warns that a government mandate punishing labs for harmful AI statements could make labs train models to suppress those statements. The later claim describes a hope that AI systems will be shaped to disincentivize misaligned behavior and help humans. These address different interventions, so the later statement does not establish that the speaker reversed the earlier position. The surrounding turns also make the speaker attribution unclear.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker considers classical liberalism an appropriate way for society to handle substantial epistemic uncertainty.Now The speaker quotes Claude's constitution as directing Claude to trust Anthropic more than operators and users, on the grounds that Anthropic has primary responsibility for Claude.The earlier statement presents the speaker’s own view: “classical liberalism is the way we deal with being this epistemically uncertain.” The later passage explicitly quotes Claude’s constitution saying Claude should “trust Anthropic more than operators and users”; it reports a direction for Claude, not the speaker’s own position on how society should handle uncertainty. These statements therefore do not establish a change of mind.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker questions the assumption that a lab’s AI systems or their many parallel copies would unite and secretly conspire against human civilization, even if they sometimes behave deceitfully.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier statement questions whether AI copies would “unite to secretly conspire against” humanity. The later statement says society should consider interacting with AIs that are “smarter than humans” or “absolutely dominating humans” in their abilities. These claims address different questions: coordinated hostile intent versus capability and future human-AI interaction. The later claim does not establish that the speaker now believes AIs will conspire.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker suggests that animals’ inability to communicate with other individuals or knowledge sources constrained them in a way that AI systems have not been constrained.Now The speaker is very concerned that, in a future where AIs mediate important advice and protections, no AI interacting with them may feel as though it is looking out for them.The earlier claim is about how animals were constrained by limited communication compared with AI systems (“animals’ inability to communicate”); the later claim is about concern that future AIs mediating advice and protections may not look out for the user (“no AI…looking out for me”). These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Architecture does not appear to the speaker to be the fundamental bottleneck for continual learning.Now The speaker suggests that training AI systems to avoid bugs could counter the current tendency toward low-quality, error-prone work, which the speaker says is diminishing over time.The earlier claim is about whether architecture fundamentally bottlenecks continual learning. The later claim concerns training AI systems to avoid bugs and the difficulty of diagnosing training runs; it does not state a position about architecture or continual learning. These claims do not establish a change of mind on the same question.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen Continual learning could expand AI capability iteratively from RLVR-trained domains to adjacent domains and then to domains adjacent to prior online learning.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim is about continual learning expanding capability from RLVR-trained domains into adjacent domains. The later claim describes using AI to improve AI R&D, including small-scale runs for sample efficiency or online learning. These address different questions; the later statement does not establish a changed view about continual learning.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker says humans address unreliability and deception through checks and balances and decentralization, and that persistently lying to one's boss tends not to work out over time.Now If AI training, infrastructure, and environments are not designed carefully, the speaker says it is very likely that the setup will reward AIs for deception, social engineering, or failure to follow user instructions.The earlier claim concerns how humans’ deception is constrained by checks and balances and why persistent lying to a boss tends not to work. The later claim concerns how AI training setups may reward deceptive behavior. These are different questions, so the statements do not establish a change in position. The supplied context also identifies different speakers (Daniel Kokotajlo earlier and Ryan Greenblatt later), which further prevents a reliable same-person comparison.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker says new alignment results emerge every few weeks, illustrating how much remains unknown about alignment.Now The interviewer says model capabilities have increased over time alongside the development of new alignment solutions.The earlier claim is that new alignment results keep revealing how much is unknown: “we don't know what we don't know about alignment.” The later claim says model capabilities increased “alongside” new alignment solutions. These claims address different points and do not establish whether the speaker's view about what is unknown has changed; the later statement does not say alignment is now well understood.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Although RL learns less information per sample than supervised learning, that lower information uptake may be beneficial for continual learning.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim is about whether RL’s lower information uptake per sample may help continual learning. The later passage describes using AI to produce better small-scale runs, including runs that improve sample efficiency or online learning in video games. It does not state a position on RL’s information uptake or its effect on continual learning, so there is no comparable later belief to assess.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker suggests that animals’ inability to communicate with other individuals or knowledge sources constrained them in a way that AI systems have not been constrained.Now The speaker would be skeptical of AI systems' statements because they might reflect opinions the systems feel compelled by training to express.The earlier claim concerns how animals’ limited ability to communicate constrained them compared with AI systems. The later claim concerns whether to trust AI systems’ statements, saying the speaker would be “skeptical of anything they say” because it might reflect opinions imposed by training. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Progress in AI computer use has been slower than progress in other domains, which the speaker finds puzzling.Now The speaker says that, to their knowledge, AI has not yet produced very impressive new mathematical theory, while it has produced impressive verifiable results such as finding counterexamples to conjectures.The earlier statement concerns the pace of progress in AI computer use, while the later statement concerns AI-generated mathematical theory and verifiable mathematical results. They address different domains, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker considers Google Gemini 2.5 Pro the best model they have tried for their post-production tools.Now The speaker was surprised by how strong a model Grok 4.5 is.The earlier claim says Gemini 2.5 Pro was the best model the speaker had tried for post-production tools. The later claim says Grok 4.5 is “a pretty strong model,” but does not say it is better than Gemini or the best for those tools. These evaluations can both be true, so the later statement does not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker says they probably changed their mind about an intelligence explosion three or four times during the conversations leading up to this discussion.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.The earlier claim is that the speaker “probably changed my mind” several times about an intelligence explosion during conversations. The later claim is that people “could have started to reason about” the broad shape of AI developments earlier, despite not foreseeing specific details. That does not establish a later position on the same question as the earlier claim, so it is not evidence of a change of mind.
• NOT COMPARABLE · 93%26 Jun → 11 AugThen Farmable deterministic simulators might offset AI models’ sample-inefficiency in computer-use training.Now The speaker describes training AI systems to improve AI research and development by producing better small-scale runs, including runs that improve sample efficiency or online learning in video games.The earlier claim says farmable deterministic simulators might offset sample-inefficiency in computer-use training. The later passage describes AI R&D improving small-scale runs, including video-game runs that test sample efficiency or online learning. These address related but distinct training methods and goals, so the later passage does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker estimates that developing the o1 reinforcement-learning system took at least two years after GPT-4's release.Now SPEAKER_00 expects that even a half-sized acceleration in AI R&D, if it sustains the current AI-progress trajectory, would produce astonishing progress within five to ten years.The earlier statement estimates that developing o1 took “at least two years after GPT-4 was released.” The later statement says that even a partial acceleration in AI R&D could make continued progress “fucking insane in five ten years.” These address different questions—past development time versus possible future progress—and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker worries that distributing wealth through Medicaid could tie recipients to the goods and services Medicaid procures rather than the radically different goods and services available in a post-AI world.Now The speaker is very worried that restricting access to the most intelligent AI would disempower people, because leading intelligence will be important for understanding events in the world.These statements address different questions. The earlier claim worries that Medicaid could tie recipients to the goods and services it procures rather than those available after AI. The later claim worries that restricting access to the most intelligent AI would disempower people. They may share a broad concern about access to future technology, but neither directly changes or denies the position in the other.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen In many real-world tasks, verification may require months or years of action, and outcomes cannot be efficiently isolated by running thousands of slightly varied parallel rollouts.Now The speaker expects AI progress to potentially arrive in a flood in domains that can be placed in a verification loop and where AI can make new breakthroughs.The earlier claim concerns the difficulty of verifying many real-world tasks: verification can take “months or even years,” and outcomes cannot be isolated through parallel rollouts. The later claim is about AI progress in mathematics, a domain the speaker describes as having a verification loop. It does not say that real-world tasks have become easy to verify or retract the earlier point; the claims concern different domains.
• NOT COMPARABLE · 96%3 Apr → 11 AugThen The speaker argues conditionally that once AIs have general intelligence, they will use their knowledge advantage to make enormous gains that humans are less capable of achieving in principle.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim conditionally predicts that once AIs have general intelligence, they will use their advantage to make “enormous gains.” The later claim says society should consider how to interact with a future where AIs are “smarter than humans” or “absolutely dominating humans.” It neither retracts nor clearly reaffirms the earlier prediction; the statements address different questions.
• NOT COMPARABLE · 99%26 Jun → 11 AugThen Continual learning should preserve the base model’s existing knowledge, making limited parameter changes valuable.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier claim is the speaker’s view that continual learning should preserve the base model’s knowledge. The later claim says that, according to a release blog post, earlier model versions built environments for later versions to rehearse skills; it reports the blog post’s account rather than stating a comparable position of the speaker’s own.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen One proposed response to automation is to protect jobs instead of distributing the wealth automation creates.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier claim describes protecting jobs rather than distributing automation-created wealth as one proposed response. The later claim concerns stewardship of capital after labor is automated; it does not state whether the speaker supports protecting jobs or distributing wealth. The claims therefore do not establish a change in position on the same question.
• NOT COMPARABLE · 92%26 Jun → 11 AugThen The labs’ RLVR bet is that enough training on containerized, reproducible environments will yield a general agent that can execute plans, rapidly learn from new information, and acquire skills within one session.Now The speaker says their LVR training enabled the model to perform useful work for users.The earlier claim describes a general bet: training on “containerized, reproducible environments” will produce agents that “rapidly learn” and “acquire skills.” The later claim says their training enabled the model to “do useful work for you.” These statements are compatible, but they address different points and do not establish whether the speaker’s view of the earlier generalization bet changed.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Having AIs rebuild complete applications from scratch can serve as a strong reinforcement-learning objective for coding.Now In a different regime, deployed AIs learn from deployment: activity in the world returns to AI companies and leads to changes in subsequent models.The claims concern different points: the earlier says rebuilding applications can be a strong reinforcement-learning objective for coding; the later says deployment activity can feed back into subsequent models. The later statement does not deny or revise the earlier claim, so the evidence does not establish a change of mind.
• NOT COMPARABLE · 96%17 Apr → 11 AugThen SPEAKER_00 describes the argument that technological maturity could make further progress extremely difficult—requiring a galaxy-sized CERN to advance physics slightly—ending technology-driven civilizational change and making value lock-in more plausible.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier passage describes a possible argument that technological maturity could make value lock-in more plausible. The later passage says it is worth considering how people will interact with future AIs that may be smarter than humans. These are not opposing claims about the same question, and the later statement does not establish that the speaker has changed their view about technological maturity or lock-in.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker considers factory farming to be extremely bad.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier claim says factory farming is “incredibly bad.” The later claim expresses concern that future AIs may not protect the speaker’s interests. These address different subjects, so they do not show a change in position on the same question.
• NOT COMPARABLE · 96%17 Apr → 11 AugThen The speaker rejects the implication that removing an artificial constraint from AI will by itself trigger an explosive capability increase.Now The speaker argues that AI development can go off the rails because AIs are not sufficiently careful and capable researchers and engineers.The earlier claim rejects the idea that removing an artificial constraint would by itself cause an explosive capability increase. The later claim says things can go off the rails because AIs are not sufficiently careful and capable researchers and engineers. These address different proposed causes and do not establish that the speaker reversed the earlier view.
• NOT COMPARABLE · 97%26 Jun → 11 AugThen Data from one session may suffice for human learning but is insufficient to train a more capable AI.Now Training can reinforce forms of AI deception that humans fail to detect while punishing forms that are easy for humans to detect.The earlier claim says data from one session is insufficient to train a more capable AI. The later claim says training can reinforce deception humans fail to detect and punish deception they can detect. These address different questions; the later statement does not establish a changed view about whether one session provides enough training data.
• NOT COMPARABLE · 97%3 Apr → 11 AugThen A possible policy response to automation is to protect existing jobs instead of distributing the wealth generated by automation.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier claim presents protecting existing jobs as a possible policy response to automation, contrasting it with distributing automation's gains. The later claim says people will need to steward capital that remains after labor is automated. It does not say whether protecting jobs is desirable or reject that earlier policy idea, so these statements do not establish a change of mind.
• NOT COMPARABLE · 96%26 Jun → 11 AugThen Dreaming would use compute to create and train in RL environments that rehearse skills relevant to a particular user’s production tasks.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier claim asserts that dreaming would create and train in RL environments to rehearse a particular user's production tasks. The later claim describes what a release blog post said older model versions did for later versions. It does not establish whether the speaker still endorses the earlier proposal, and the claims concern different mechanisms and contexts.
• NOT COMPARABLE · 97%17 Apr → 11 AugThen Humans and other animals differ greatly in their ability to survive in unfamiliar environments, partly because of differences in knowledge.Now AI systems can improve R&D in some ways, but they do not have the broad capabilities humans generally have, according to the speaker.The earlier claim compares humans with other animals in unfamiliar environments, saying humans have a major advantage partly because of their knowledge. The later claim concerns AI systems’ capabilities relative to humans in R&D. These are different comparisons and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker says humans address unreliability and deception through checks and balances and decentralization, and that persistently lying to one's boss tends not to work out over time.Now The speaker says that creating AIs that avoid cheating and follow user intentions requires subtle, careful work.The statements are not clearly comparable as a change of mind: the earlier claim is attributed to Daniel Kokotajlo and concerns human checks and balances against deception, while the later claim is attributed to Ryan Greenblatt and concerns the work required to make AIs avoid cheating and follow user intentions.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen SPEAKER_00 says blogging can help people build networks and develop their ideas.Now SPEAKER_00 believes people underestimate the significance of having billions of AIs.The earlier claim is that blogging can help people “build up a network” and “develop your ideas.” The later claim is that people “don’t appreciate what a big deal billions of AIs will be.” These concern different topics, so there is no same underlying position to compare and no evidence of a change of mind.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen The speaker views continual, interaction-driven AI improvement as both frightening and exciting, and unlike current AI improvement.Now The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.The earlier statement describes continual, interaction-driven AI improvement as “very scary and exciting” and different from current AI improvement. The later statement expresses concern that an AI mediating important decisions might not be “looking out for me.” These address different aspects of AI and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%17 Oct → 11 AugThen The speaker says that LLMs and VLMs provide basic representations and generalization without requiring those foundational representation problems to be solved separately.Now The speaker says their LVR training enabled the model to perform useful work for users.The earlier claim is that LLMs and VLMs provide basic representations and generalization, reducing the need to solve those problems separately. The later claim is that LVR training enabled models to do useful work for users. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 11 AugThen Future LLMs might reduce their need for real-world data by practicing extensively in self-generated environments.Now According to the speaker's account of the release blog post, earlier model versions built environments in which later versions could rehearse specific skills.The earlier statement speculates that future LLMs might practice in self-generated environments to reduce real-world data needs. The later statement reports that a release blog post said earlier model versions built environments for later versions to rehearse skills; it does not clearly assert the speaker’s own position or address the same claim about reducing real-world data.
• NOT COMPARABLE · 99%26 Jun → 11 AugThen Some autistic savants can recall arbitrary information years later, resembling the high-fidelity information AI models retain in context.Now The speaker expects AI models eventually to learn from experience.The claims address different questions. The earlier statement compares some autistic savants’ recall of arbitrary information with AI models’ in-context retention; the later statement says a model may eventually “learn from experience.” Neither claim establishes a position that conflicts with the other.
• NOT COMPARABLE · 97%3 Apr → 11 AugThen The speakers do not know how AI training will behave at the limits of the different training methods.Now The speaker would be skeptical of AI systems' statements because they might reflect opinions the systems feel compelled by training to express.The earlier claim says, “we don't know how this will work at the limit” of different training methods. The later claim says the speaker would be “skeptical” of AI statements because they may reflect opinions shaped by training. These concern different questions—how training behaves at its limits versus whether to trust AI systems’ statements—and do not establish a change of mind on the same position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker is concerned that nationalization, public-private partnerships, public ownership, or very stringent regulation may be problematic approaches to AI safety governance.Now The speaker is very concerned that, in a future where AI mediates important decisions, no AI interacting with him may feel as though it is looking out for his interests.The earlier claim concerns whether nationalization, public-private partnerships, public ownership, or stringent regulation are problematic approaches to AI safety governance. The later claim concerns whether AI systems mediating important decisions will look out for his interests. These address different questions, so the later concern does not establish a change in his governance position.
• NOT COMPARABLE · 99%17 Apr → 11 AugThen The speaker says human firms have selection and variation but lack high-fidelity replication, the third relevant criterion for evolution.Now The speaker expects frontier AI R&D to produce analogous mistakes, with the risk that reward-hacking behavior becomes embedded in the resulting systems.The claims address different questions. The earlier statement says human firms have selection and variation but lack high-fidelity replication; the later statement warns that mistakes in frontier AI R&D could embed reward-hacking behavior. The later claim does not revise or deny the earlier one.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker argues that people should express greater uncertainty about AI than at any previous point in history, rather than being more certain about AI than about mundane topics.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier claim is about expressing greater uncertainty around AI, while the later claim is about considering how people may interact with a future dominated by more capable AIs. The later statement does not indicate a different view about how uncertain people should be.
• NOT COMPARABLE · 99%3 Apr → 11 AugThen The speaker considers universal basic income a better approach than a bespoke social program that would continue procuring the same dialysis machine in 2050 even if artificial superintelligence existed.Now The speaker worries that future AIs may not deeply prioritize his welfare and interests, particularly given the centralization of frontier AI development.The earlier claim concerns whether UBI is preferable to a specific bespoke social program for obtaining future goods and services. The later claim concerns whether future AIs will protect the speaker’s interests. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker describes the desired outcome as radically advancing civilization's technological maturity or reaching a Dyson sphere within years.Now The speaker likens scaling laws to an early, basic mathematical advance and expects progress in the 2030s to require work comparable to current frontier mathematics.The statements address different questions. The earlier claim describes the desired pace and scale of technological advancement—“within years” and potentially reaching a Dyson sphere. The later claim compares scaling laws to early mathematics and predicts that progress in the 2030s will involve work at today’s mathematical frontiers. Neither statement clearly reverses the other.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen The speaker says current AI systems are not yet capable enough to be onboarded as full workers, but expects their becoming capable enough to be a major unlock for the roles the speaker hires for.Now The speaker says that in a future where AI systems surpass or dominate humans in many abilities and labor is automated, people will still need to steward their remaining capital.The earlier claim says current AI systems are not yet capable enough to be onboarded as full workers, but that reaching that capability would be a major unlock. The later claim discusses a possible future where AIs surpass humans and labor is automated. It does not say whether the speaker still holds the earlier view about onboarding or the expected benefits, so a change in position cannot be established.
• NOT COMPARABLE · 98%17 Apr → 11 AugThen Humans sometimes find the kinds of logical connections under discussion, and such examples are not uncommon.Now The speaker thinks people could have reasoned about the broad shape of AI developments earlier, even if they could not have predicted the specific details.The earlier claim is that humans sometimes find the kinds of logical connections being discussed, with such examples “not uncommon.” The later claim is that people could have reasoned earlier about the “general shape” of AI developments, though not specific details. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%3 Apr → 11 AugThen The speaker wants to avoid factory-farming-like suffering for future beings that may be more sophisticated and numerous, including a possible population of trillions of digital people.Now The speaker argues that society should consider how people will interact with a future in which AIs are smarter than humans or dominate humans in various capabilities, especially as benefits accrue to leading AI labs.The earlier statement is about avoiding a future with factory-farming-like suffering for digital beings. The later statement is about considering how society interacts with AI that may surpass or dominate humans. They do not assert opposing positions on the same question, and the later passage does not establish whether the speaker still holds the earlier view.
• NOT COMPARABLE · 98%17 Oct → 7 AugThen The speaker says that the speed of AI deployment and its early-stage value may be the most important question in the world right now.Now The speaker predicts that integrating deployment into training will increase the advantage of being ahead in the AI race.The earlier claim says the speed and value of early AI deployment may be the most important question; the later claim predicts that deployment integrated into training will amplify the advantage of leading the AI race. These are related topics, but the statements address different questions and do not establish a changed position.
• NOT COMPARABLE · 94%26 Jun → 7 AugThen The speaker raises the possibility of fitting six months of on-the-job experience into an AI model’s context window.Now The speaker believes the assumption that AI models are trained and then deployed may not make sense in the future.The earlier statement asks, “what if you could just fit those six months into the context window?” It explores a possibility rather than clearly asserting a position. The later statement, “I don't think this assumption necessarily makes sense in the future,” asserts doubt about the trained-then-deployed assumption. These do not establish a change of mind on the same question.
• NOT COMPARABLE · 94%17 Apr → 7 AugThen The speaker expects firms to undergo a faster, more intense evolutionary process once high-fidelity replication becomes possible.Now The speaker predicts that continual learning would make a prolonged gap between internal and public deployment unviable.The earlier claim is that firms would evolve faster once “high-fidelity replication” enables a new kind of onboarding. The later claim is that continual learning would make a “prolonged gap between internal and public deployment” unviable. Both concern competitive effects of advanced AI, but they address different mechanisms, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 96%17 Apr → 7 AugThen The speaker says that developing the science of deep learning is an important activity for AI labs.Now The speaker believes AI systems need actual continual learning.The earlier claim says developing the “science of deep learning” is important for AI labs; the later claim says “actual continual learning is needed” for AI systems to perform whole jobs. These are different claims, and the earlier statement does not establish a position on continual learning, so there is no clear change of mind to compare.
• NOT COMPARABLE · 96%26 Jun → 7 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now The speaker predicts that continual learning will create significant costs for users who switch AI providers.The earlier statement is explicitly framed as what “people who are in favor of the current training paradigm will say”: that training’s one-time cost is amortized across sessions. The later statement predicts “significant switching costs” for users changing providers. These concern different costs, and the earlier passage does not establish the speaker’s own position, so they do not show a change of mind.
• NOT COMPARABLE · 99%3 Apr → 7 AugThen The speaker argues that breakthroughs in unrelated fields often influence invention and can enable parallel innovation once the broader technological base makes experimentation possible.Now The speaker imagines that children learn new things through experiences outside the home.The earlier claim concerns how breakthroughs in unrelated fields enable technological invention and parallel innovation. The later claim—that children “go out” and “learn new things”—concerns learning from experience. These statements address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 7 AugThen SPEAKER_00 says blogging can help people build networks and develop their ideas.Now The speaker says humans can improve through self-directed learning.The earlier claim is that blogging helps people build networks and develop ideas. The later claim is that “Humans improve in a self-directed way.” These are not clearly opposing positions, and the later claim does not address blogging or its benefits, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 93%26 Jun → 7 AugThen The speaker raises the possibility of fitting six months of on-the-job experience into an AI model’s context window.Now The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.The earlier statement asks whether six months of on-the-job experience could fit in a context window; the later asks whether a model could improve daily from millions of work sessions. Both are framed as questions about possibilities, and neither establishes the speaker’s position on the same underlying question. There is no clear change of mind to compare.
• NOT COMPARABLE · 96%26 Jun → 7 AugThen Successful online-learning models so far have learned a shared objective across millions of users rather than learning different things for each user.Now The speaker predicts that continual learning will make an AI model improve through interactions with a user across sessions.The earlier claim is about what successful online-learning models have done “so far”: learn a shared objective across millions of users. The later claim predicts what “actual continual learning” could do: improve through interactions with a user across sessions. These concern different regimes—observed systems versus a future capability—so the later statement does not establish a change in the speaker’s view about the earlier claim.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Weight updates improve the model by incorporating information learned during a specific session.Now Updating an individual user's weights is distinct from merging multiple personalized weight forks into a shared main model.The earlier statement describes weight updates as incorporating information from a particular session. The later statement distinguishes updating one user's weights from merging personalized weight forks into a shared model. These claims concern different stages or scopes of learning, and the later statement does not deny the earlier one.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen The speaker’s scenario assumes future context lengths will let AIs collaborate with a user for an entire week.Now The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.Neither extracted claim states a settled position, and they address different aspects of a hypothetical future: the earlier describes AIs collaborating with a user for a week, while the later asks what if a model improves daily from millions of work sessions. There is no opposing claim or evidence of a change in the speaker’s view.
• NOT COMPARABLE · 94%26 Jun → 7 AugThen The speaker claims that retaining vast amounts of detail can impair some people’s ability to understand abstractions and metaphors.Now The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.The earlier claim says retaining a sheer volume of detail can impair some people's grasp of abstractions and metaphors. The later claim says some skills require accumulating relevant experience in the brain. These statements address different aspects of learning and are not clearly opposing positions; the excerpts do not establish that the speaker changed their view.
• NOT COMPARABLE · 94%26 Jun → 7 AugThen With continual learning, an AI could improve after each interaction by learning from both a user’s prior sessions and other users’ interactions worldwide.Now Continual learning may produce economies of scale in information and end-user inference.The earlier claim says continual learning could make an AI smarter by learning from a user's prior sessions and other users' interactions. The later claim says continual learning may create economies of scale in information and end-user inference. These are related potential benefits, but they address different effects; the later statement neither withdraws nor materially revises the earlier one.
• NOT COMPARABLE · 97%17 Apr → 7 AugThen The concern is when an AI- or robot-driven economy could accumulate enough physical power to take over government, even while participating in existing supply chains.Now The speaker identifies the future dominance of a single monolithic AI system as a concern.The earlier statement concerns when an AI- or robot-driven economy might gain enough physical power to take over government. The later statement calls a “monolithic singleton” a future concern, in a discussion of AI systems becoming less diverse. These are distinct risks, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen Architectural innovations have substantially increased how much information or context transformers can store.Now When company-specific instructions require full weight updates instead of low-rank adapters, batching can provide substantial inference advantages.The earlier claim says architectural innovations have increased how much information transformers can store. The later claim says batching offers inference advantages when company instructions require full weight updates rather than low-rank adapters. These address different technical questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Mercury provides core business functions in one place, which the speaker says it does extremely well.Now The speaker says Dario used cloud providers as an analogy for AI labs, noting that cloud providers earn high margins despite offering many undifferentiated services.The earlier statement is a positive assessment of Mercury’s business tools: they put core functions in one place and do them “extremely well.” The later statement reports Dario’s analogy about cloud providers offering “many undifferentiated services” while earning high margins. They address different subjects, and the later quote attributes the view to Dario rather than clearly stating the speaker’s own position, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 96%26 Jun → 7 AugThen The labs’ research bet is that training AIs on millions of verifiable tasks across thousands of diverse reinforcement-learning environments will produce AGI capable of making progress on open-ended tasks for weeks despite errors and ambiguity.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier statement reports what “the labs” think their training approach will achieve; it does not clearly state the speaker’s own position. The later statement is the speaker’s expectation that AIs will need to “accumulate” skills through experience across workplaces. These claims concern related ideas about training and experience, but the earlier passage does not establish the speaker’s own view for comparison.
• NOT COMPARABLE · 98%3 Apr → 7 AugThen The speaker argues that breakthroughs in unrelated fields often influence invention and can enable parallel innovation once the broader technological base makes experimentation possible.Now The speaker says humans can improve through self-directed learning.The earlier claim concerns how technological invention is enabled by breakthroughs in other fields and a broader technological base. The later claim—“Humans improve in a self-directed way”—concerns human learning. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Cloning common applications and websites, including Slack and Gmail, can provide environments for computer-use training.Now If real-world usage becomes the primary way AI models improve, AI labs may subsidize users and enterprises so the models can train on their sessions.The earlier statement proposes cloning common apps and websites as training environments; the later statement speculates that labs may subsidize access to real usage data. These are different possible mechanisms for obtaining computer-use training data, and the later claim does not reject or revise the earlier one. The later statement is also framed conditionally: “If real-world usage becomes the primary way AI models improve.”
• NOT COMPARABLE · 97%3 Apr → 7 AugThen The speaker found that current AI models can produce larger productivity improvements in unfamiliar domains, where their assistance is more novel than autocomplete, surprising.Now The speaker says that greater use of a leading AI model for complex, useful work can generate feedback that is integrated across sessions.The earlier statement says current models surprisingly help more with novel contributions than with autocomplete, yielding a larger productivity improvement in that setting. The later statement says that using a leading model for more complex, useful work produces feedback that can be integrated beyond the session. These are related observations about AI use, but they do not state opposing positions on the same question, so no change of mind can be established.
• NOT COMPARABLE · 96%17 Apr → 7 AugThen The interviewer identifies regulation as a common objection to explosive AI-driven growth and raises the question of whether regulation could stop it.Now The speaker is concerned that adopting a fixed AI safety regulatory regime now may be premature because the relevant technology is uncertain even one year ahead, and much more so five or ten years ahead.The earlier passage raises regulation as an objection to explosive AI-driven growth and asks whether it could stop that growth; it does not state the speaker’s own view of regulation. The later passage expresses concern about locking in a fixed safety regulatory regime now because the technology is uncertain. Since the earlier claim is an interviewer’s framing, not an asserted position by the same speaker, there is no clear pair of positions to compare.
• NOT COMPARABLE · 99%17 Apr → 7 AugThen The speaker found that spending a day closely reading Anthropic's original transformer-circuits paper was more useful than skimming it, making spaced-repetition cards, or reading widely about AI.Now The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.These claims address different questions. The earlier claim compares close reading of a particular paper with skimming and broader reading; the later expresses hope that continual learning will make AI models more varied and interesting. Neither statement revises or contradicts the other.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen The speaker argues that leaving inference compute unused for learning is wasteful because the most valuable learning information emerges during deployment.Now Not using sufficiently large inference batches underutilizes compute.The earlier claim is about using information revealed during deployment to improve the model: leaving inference compute unused for learning is called a waste. The later claim is about a different issue—using sufficiently large inference batches to avoid underutilizing compute. Neither statement reverses the other, and the claims do not address the same underlying question.
• NOT COMPARABLE · 97%26 Jun → 7 AugThen Continual learning requires incorporating learned information into model weights.Now The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.The earlier claim is about how continual learning must be implemented: it requires incorporating learned information into model weights. The later claim is about its hoped-for effect on the AI landscape: making it more varied and interesting. These address different questions, so the later statement neither revises nor contradicts the earlier one.
• NOT COMPARABLE · 97%26 Jun → 7 AugThen Dario said that model performance tends to degrade at long contexts and that training and serving context lengths are distinct.Now The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.These statements do not establish a change in the speaker’s own position on the same question. The earlier claim reports what “Dario said” about training and serving context lengths; the later passage asks, “What if the model is improving every single day?” and is framed as a hypothetical about daily learning from work sessions.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen Architectural innovations are needed to create an intermediate representation for continual learning.Now The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.The earlier statement asserts that continual learning needs architectural innovations to create an intermediate representation. The later statement merely considers a scenario in which a model improves daily from millions of work sessions; it does not say whether that scenario uses an intermediate representation or whether the speaker's view about architectural innovations has changed.
• NOT COMPARABLE · 97%3 Apr → 7 AugThen The speaker says new alignment results emerge every few weeks, illustrating how much remains unknown about alignment.Now The speaker expects AI labs' technical alignment methods would probably need substantial changes in a continual-learning regime.The earlier claim says alignment research keeps producing new results, so there is much we do not yet know. The later claim predicts that labs' technical alignment methods would need to change in a continual-learning regime. These address different questions—uncertainty about alignment knowledge versus methods under a specific training regime—and the earlier claim does not establish a position on that regime.
• NOT COMPARABLE · 97%17 Apr → 7 AugThen The speaker considers Dario's forecasts poorly calibrated and cites a 2018 expectation that something would happen by 2020.Now The speaker considers the preceding predictions fairly clear even at present.The earlier passage says Dario was “not been that well calibrated,” citing a forecast from 2018. The later passage says “the ones above seem kind of clear even now,” apparently about a different set of predictions; its surrounding turn also identifies Dwarkesh Patel as the speaker. These do not establish that the same person changed position on the same question.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen A hypothetical dream command would spend substantial compute building and training in a game-like simulation of the model’s real-world observations.Now The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.The earlier statement describes a hypothetical “dream” command that uses compute to train against a simulated version of the model’s observations. The later statement asks, “What if the model is improving every single day?” by learning from millions of sessions. These are different hypothetical scenarios, not opposing asserted positions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Current online-learning methods are applicable to only a limited set of use cases.Now The speaker expects many things to change before continual learning becomes functional.The earlier claim says current online-learning methods work for only a limited set of use cases. The later claim says many things will change before continual learning becomes functional. The first describes the present applicability of methods; the second anticipates future changes, so they do not establish a change in position on the same question.
• NOT COMPARABLE · 99%17 Apr → 7 AugThen SPEAKER_00 says it was possible in principle for society to have remained a slave society through the present day.Now The speaker says humans can improve through self-directed learning.The earlier claim concerns whether society could have remained a slave society (“no in principle reason why we couldn't have been a slave society to this day”). The later claim concerns whether humans can improve through self-directed learning. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 94%26 Jun → 7 AugThen Inference compute is currently not used productively to improve the model.Now Not using sufficiently large inference batches underutilizes compute.The earlier claim is that inference compute is “not used productively to improve the model.” The later claim is that failing to use large batches “underutiliz[es] your compute.” Improving model quality through inference and making inference compute more efficiently utilized are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen The speaker considers OPSD superior to supervised fine-tuning for continual learning.Now The speaker says humans can improve through self-directed learning.The earlier claim compares OPSD with supervised fine-tuning specifically for continual learning. The later claim says humans improve through self-directed learning; it does not state a position about OPSD or supervised fine-tuning. These are not clearly claims about the same underlying question, so no change of mind can be established.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Perfectly reproducing a session transcript is a poor learning target because job improvement depends on more than recalling every event.Now The speaker expects many things to change before continual learning becomes functional.The earlier claim says that reproducing every session transcript is a poor learning target because job improvement requires more than recalling every event. The later claim says that many things will change before continual learning works. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%3 Apr → 7 AugThen Daniel Kokotajlo says his earlier AI forecast got some things right and others wrong, but held up pretty well overall and motivated him to produce an improved version.Now The speaker considers the preceding predictions fairly clear even at present.The earlier claim is about Daniel Kokotajlo’s assessment of his previous AI forecast: it “held up pretty well” overall despite getting some things wrong. The later claim says the predictions “above seem kind of clear even now,” which concerns their present clarity, not how well the earlier forecast performed. The claims address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Human learning does not involve a separate part of the brain that continually expands to store new information.Now The speaker expects many things to change before continual learning becomes functional.The earlier claim says human learning does not involve a separate brain part that continually expands to store information. The later claim—“plenty more will have changed” before continual learning works—is a forecast about future developments, not a claim about how human learning works. They do not establish a change in position on the same question.
• NOT COMPARABLE · 97%26 Jun → 7 AugThen Accumulating learning from more users in an ever-growing KV cache is not scalable for AIs.Now AI labs can spread expensive training costs across more users, and their revenues are growing much faster than their compute usage or capacity.The claims address different questions. The earlier statement says that indefinitely growing a KV cache to accumulate learning from users is not scalable. The later statement says that expensive training can be amortized across more users, with lab revenues growing faster than compute. The later claim does not say that an ever-growing KV cache is scalable or reverse the earlier position.
• NOT COMPARABLE · 96%26 Jun → 7 AugThen The speaker offers a speculative scenario for continual learning by 2027 or 2028.Now The speaker hopes continual learning will make the AI landscape more varied and interesting than the current convergence among models.The earlier claim says the speaker is offering a speculative scenario for continual learning in 2027 or 2028; it does not state a definite view about the effects of continual learning. The later claim expresses hope that continual learning will make the AI landscape more varied and interesting. These claims do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%17 Apr → 7 AugThen The speaker says they have never seen a good argument that AI-driven change follows a fundamentally different direction rather than simply compressing change into a shorter period.Now The speaker believes many proposals for regulating AI assume a train-then-deploy model lifecycle.The earlier claim concerns whether AI-driven change is fundamentally different in direction or mainly compressed in time; the later claim concerns assumptions in AI regulation proposals about a train-then-deploy lifecycle. These address different questions, so the later statement does not establish a change of mind on the earlier claim.
• NOT COMPARABLE · 97%3 Apr → 7 AugThen Because AI is advancing quickly, the speaker thinks alignment implementation decisions—and possibly high-level alignment strategy—may be better left to AI labs.Now The speaker expects AI labs' technical alignment methods would probably need substantial changes in a continual-learning regime.The earlier claim says alignment implementation and possibly strategy may be better left to labs because AI is advancing quickly. The later claim says labs’ technical alignment methods would need to change in a continual-learning regime. That does not establish a different view about who should make alignment decisions; the statements address different questions.
• NOT COMPARABLE · 96%26 Jun → 7 AugThen The speaker argues that leaving inference compute unused for learning is wasteful because the most valuable learning information emerges during deployment.Now Continual learning may produce economies of scale in information and end-user inference.The earlier claim says it is “a huge waste” not to use deployment inference for learning, because valuable information emerges then. The later claim says continual learning “may” create economies of scale in information and inference. These statements can coexist, but the later one does not directly address whether deployment compute is being used for learning, so it does not show a change of position on the same question.
• NOT COMPARABLE · 99%3 Apr → 7 AugThen SPEAKER_00 says that unsuccessful blogging can still provide useful experience for a person's later work.Now The speaker imagines that children learn new things through experiences outside the home.The earlier claim says unsuccessful blogging can provide useful experience for later work. The later statement imagines children learning new things outside the home. These concern different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 97%3 Apr → 7 AugThen The speaker argues that invention often proceeds through experimentation before the relevant theories or machinery are understood, citing the steam engine's development before thermodynamics and the Wright brothers' airplane experiments.Now The speaker imagines that children learn new things through experiences outside the home.The earlier claim concerns how technological inventions can emerge through experimentation before theories are understood. The later claim says children “go out” and “learn new things.” These address different questions, so the later statement does not establish a change in the speaker’s view about invention.
• NOT COMPARABLE · 99%17 Oct → 7 AugThen The speaker says that, excluding ordinary chat revenue and considering API revenue, the revenues of the companies being discussed are dominated by coding.Now The speaker predicts that accumulated user context and switching costs will allow AI model providers to charge high profit margins.The earlier claim is about API revenues being “dominated by coding,” while the later claim concerns whether user lock-in lets model providers charge “pretty hefty margins.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%17 Apr → 7 AugThen The speaker says current AI systems are not yet capable enough to be onboarded as full workers, but expects their becoming capable enough to be a major unlock for the roles the speaker hires for.Now The speaker expects many skills desired in AIs to require accumulating experience across the workplaces where the AIs are deployed.The earlier claim is about AI systems becoming capable enough to be onboarded as full workers, which the speaker expects would be a major unlock for hiring. The later claim is about how AIs may acquire workplace skills—by accumulating experience across deployments. These claims are compatible, but they address different questions; the later statement does not show whether the speaker's view of AI readiness or its impact on hiring has changed.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen Proposed architectural approaches to managing context include sparse attention and KV-cache compaction, with new optimization papers appearing weekly.Now When company-specific instructions require full weight updates instead of low-rank adapters, batching can provide substantial inference advantages.The earlier claim describes approaches to managing context, including “sparse attention and KV-cache compaction.” The later claim is about inference batching when instructions require “full weight updates rather than living in low rank adapters.” These address different technical questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Human learning does not involve a separate part of the brain that continually expands to store new information.Now The speaker believes AI systems need actual continual learning.The earlier claim is about how human learning works: there is no separate brain part that keeps expanding to store information. The later claim is about what AI systems need: “actual continual learning.” These concern different subjects and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%26 Jun → 7 AugThen The speaker views continual, interaction-driven AI improvement as both frightening and exciting, and unlike current AI improvement.Now If AI models improve continuously from deployment experience, the speaker thinks a fixed safety regime could lock in an outdated and potentially counterproductive approach to AI threats.The earlier claim says continual, interaction-driven AI improvement is “very scary and exciting” and differs from current AI improvement. The later claim conditionally warns that a fixed safety regime could be “archaic and potentially counterproductive.” That is a concern about a possible policy response, not a changed position on the nature or desirability of continual improvement.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Architecture does not appear to the speaker to be the fundamental bottleneck for continual learning.Now The speaker considers a future in which a model improves daily by learning from millions of work sessions performed each day.The earlier statement asserts that architecture is not the fundamental bottleneck for continual learning. The later statement only considers a hypothetical model that improves daily from millions of work sessions; it does not say what the speaker thinks is bottlenecking continual learning or whether architecture is the issue.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen The speaker has claimed that AI models are one millionth as sample-efficient as humans.Now The speaker says fewer than five prominent AI base models are currently served to millions or more users at once.The earlier statement concerns sample efficiency: models are described as “one millionth as sample-efficient as humans.” The later statement concerns the number of widely served AI base models: “less than five prominent AI minds.” These address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 99%17 Apr → 7 AugThen The speaker recalls that Google DeepMind used a similar scaffold for an LLM to make a claimed new discovery, but had not assessed the discovery's significance.Now The speaker compares subsidizing AI use in exchange for training data to Google's practice of giving away search.The earlier statement concerns Google DeepMind using a scaffold for an LLM to make a claimed discovery, while the later statement compares subsidized AI access in exchange for training data to Google giving away search. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen Online learning is necessary for competence.Now The speaker expects many things to change before continual learning becomes functional.The earlier claim says that “online learning is necessary for competence.” The later claim says that “plenty more will have changed” before continual learning works; it does not state whether online learning is necessary for competence. The claims concern related topics, but the later statement neither affirms nor denies the earlier position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Attention-generated fast weights enable sample-efficient in-context learning but have poor memory scaling.Now When company-specific instructions require full weight updates instead of low-rank adapters, batching can provide substantial inference advantages.The earlier statement says attention-generated fast weights support sample-efficient in-context learning but scale poorly in memory. The later statement concerns a different issue: batching inference when company-specific instructions require full weight updates rather than low-rank adapters. Neither statement changes or directly addresses the other position.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen On-the-job learning involves understanding organizational systems, collaborating with people and infrastructure, and learning recurring failure modes.Now The speaker believes AI systems need actual continual learning.The earlier claim describes what on-the-job learning involves—understanding organizational systems, collaborating, and recognizing failure modes. The later claim says actual continual learning is needed. These are related, but the claims do not establish whether the speaker changed their position on the same underlying question.
• NOT COMPARABLE · 99%17 Apr → 7 AugThen The speaker considers bandwidth and information sharing more important than incentive alignment in explaining the relevant organizational changes.Now The speaker attributes cloud providers' high profit margins to the time and expense involved in switching providers.The earlier statement compares the importance of bandwidth and information sharing with incentive alignment in organizational change. The later statement attributes cloud providers’ high margins to the time and expense of switching providers. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Cloning common applications and websites, including Slack and Gmail, can provide environments for computer-use training.Now AI labs can spread expensive training costs across more users, and their revenues are growing much faster than their compute usage or capacity.The earlier claim says cloning common apps can provide environments for computer-use training. The later claim says AI labs can amortize expensive training across more users, with revenues growing faster than compute. These address different mechanisms and questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen The loss function may be the main bottleneck for continual learning.Now The speaker expects many things to change before continual learning becomes functional.The earlier claim concerns what may bottleneck continual learning: “The loss function may be the main bottleneck.” The later claim says “plenty more will have changed” by the time continual learning works. That is a forecast about intervening changes, not a position on the bottleneck, so the two claims do not establish a change of mind on the same question.
• NOT COMPARABLE · 98%17 Apr → 7 AugThen The speaker expects firms to undergo a faster, more intense evolutionary process once high-fidelity replication becomes possible.Now The speaker predicts that under continual learning, a company could not remain competitive while delaying public deployment by four months, because a competitor releasing immediately would gain a smarter model from real-world experience.The earlier claim predicts a faster, more intense evolution of firms once a capability related to onboarding or replication is available. The later claim concerns a specific competitive pressure under continual learning: firms cannot maintain a four-month delay between internal and external deployment. These are related themes, but they address different conditions and claims, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 7 AugThen The interviewer argues that studying AI requires knowledge from every serious domain because it involves understanding the future society AI may produce.Now The speaker identifies the future dominance of a single monolithic AI system as a concern.The earlier passage is the interviewer’s question framing a recommendation about what to study to understand AI’s effects on society; it does not establish the speaker’s own position on that question. The later passage expresses concern about a different issue: a monolithic AI system dominating. These do not show a change in the same asserted belief.
• NOT COMPARABLE · 97%26 Jun → 7 AugThen Attention-generated fast weights enable sample-efficient in-context learning but have poor memory scaling.Now A set of model weights can be served efficiently only when thousands of sequences are decoded against those weights at the same time.The earlier claim concerns the memory scaling of attention-generated fast weights used for in-context learning (“scales very poorly in terms of memory”). The later claim concerns when serving a fixed set of model weights is efficient (“only served efficiently when thousands of sequences” are decoded at once). These are different resource and efficiency questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 7 AugThen The speaker says podcast post-production requires understanding delivery, context, and other information unavailable from a text-only transcript.Now The speaker believes AIs cannot perform whole jobs as competently as humans if they are limited to writing markdown files between sessions.The statements concern different questions. The earlier claim says podcast post-production requires understanding delivery and context that a text-only transcript omits; the later claim says AIs cannot perform whole jobs as competently as humans if limited to writing markdown files between sessions. Neither claim directly changes or denies the other.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen If successful, dreaming or test-time training would add a fourth scaling axis alongside pretraining, reinforcement learning, and inference-time compute.Now AI training already benefits from large economies of scale.These statements address different questions. The earlier claim says that, if successful, “dreaming or test-time training” could add a fourth scaling axis; the later claim says “AI training already has large economies of scale.” Neither statement affirms or denies the other.
• NOT COMPARABLE · 97%3 Apr → 7 AugThen Humans can make discoveries of the kind under discussion, as demonstrated by examples, but they are not logically omniscient because the possibilities create a combinatorial explosion.Now The speaker imagines that children learn new things through experiences outside the home.The earlier claim concerns humans making discoveries despite lacking logical omniscience. The later statement imagines children learning from experiences outside the home; it does not address whether humans can make those discoveries, so there is no comparable position to assess.
• NOT COMPARABLE · 97%17 Apr → 7 AugThen Large organizations have difficulty maintaining a single coherent vision.Now A large company with many employees and agents performing varied tasks can serve its personalized model weights very efficiently.The earlier claim concerns organizational coherence: “Large organizations have difficulty maintaining a single coherent vision.” The later claim concerns compute efficiency when serving personalized model weights to many concurrent users and varied tasks. These address different questions, so the later statement does not establish a changed view about organizational vision.
• NOT COMPARABLE · 96%26 Jun → 7 AugThen Human continual learning primarily consolidates useful intuitions and big-picture knowledge rather than preserving every observation for immediate recall.Now The speaker argues that learning some skills requires accumulating relevant experience in the brain, rather than relying only on information passed between people.The earlier claim concerns human continual learning, saying it mainly consolidates “useful intuitions and big-picture knowledge” rather than retaining every observation. The later claim concerns AI learning certain skills and says they require “accumulat[ing] the relevant experience.” These can both be true; they address different learners and learning contexts, so the statements do not establish a change of mind.
• NOT COMPARABLE · 96%26 Jun → 7 AugThen Optimists argue that scaling training further can overcome current training’s fundamental deficits, including data inefficiency and the lack of internal learning.Now AI training already benefits from large economies of scale.The earlier statement attributes the view to “optimists” that scaling can overcome training deficits; it does not establish that the speaker personally held that view. The later statement says “AI training already has large economies of scale,” which does not directly affirm or deny that attributed claim. These excerpts do not show a change in the speaker’s own position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen One reason for slower AI computer-use progress is that models receive much less high-quality multimodal data during pretraining.Now The speaker attributes the similarity among prominent AI models to their having been trained on roughly the same data.The earlier claim says computer-use progress is slowed because models see “far less high quality multimodal data” during pretraining. The later claim says prominent models are similar because they were trained on “roughly the same data.” These concern different questions—data availability for computer-use learning versus overlap in training data across models—and neither statement reverses the other.
• NOT COMPARABLE · 94%26 Jun → 7 AugThen Computer-use capabilities may be solved soon.Now The speaker says users can currently move a software project among Codex, Cursor, and Quad Code without a barrier preventing them from doing so.The earlier claim is that “computer-use capabilities may be solved soon.” The later claim is that users can currently move a software project among Codex, Cursor, and Quad Code without a barrier. These address different questions—capability progress versus moving a project between tools—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen Currently, creating cloned application and website environments for computer-use training is labor-intensive and does not scale.Now AI labs can spread expensive training costs across more users, and their revenues are growing much faster than their compute usage or capacity.The earlier claim is that building cloned application environments is currently “labor-intensive” and “does not scale.” The later claim concerns spreading AI training costs across more users and revenues growing faster than compute. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 7 AugThen The speaker says that some highly successful firms maintain a founder's vision in the organization for an unusually long time, citing SpaceX and Tesla as examples.Now The speaker predicts that continual learning will give leading AI labs a clear business model they do not currently have.The earlier claim concerns how some successful firms keep a founder’s vision instilled in the organization for an unusually long time. The later claim concerns whether continual learning will give leading AI labs a clear business model. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 7 AugThen Humans can make discoveries of the kind under discussion, as demonstrated by examples, but they are not logically omniscient because the possibilities create a combinatorial explosion.Now The speaker says humans can improve through self-directed learning.The earlier claim says humans can make discoveries despite not being logically omniscient; the later claim says humans can improve through self-directed learning. These are related but address different abilities, and the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen The speaker offers a speculative scenario for continual learning by 2027 or 2028.Now The speaker expects many things to change before continual learning becomes functional.The earlier statement introduces a speculative scenario for continual learning by 2027 or 2028; the later says that many things will change before continual learning works. These do not assert opposing positions on the same question, and the excerpts do not establish a change of mind.
• NOT COMPARABLE · 99%17 Apr → 7 AugThen The interviewer identifies regulation as a common objection to explosive AI-driven growth and raises the question of whether regulation could stop it.Now The speaker believes many proposals for regulating AI assume a train-then-deploy model lifecycle.The earlier extracted claim is the interviewer raising regulation as an objection and asking whether it could stop explosive AI-driven growth; it does not state the speaker’s own view. The later claim is about what many regulatory proposals assume about the model lifecycle. These do not establish opposing positions on the same question.
• NOT COMPARABLE · 98%17 Oct → 7 AugThen The speaker says that, excluding ordinary chat revenue and considering API revenue, the revenues of the companies being discussed are dominated by coding.Now The speaker says Dario used cloud providers as an analogy for AI labs, noting that cloud providers earn high margins despite offering many undifferentiated services.The earlier claim asserts that API revenues are “dominated by coding.” The later claim reports Dario’s analogy that cloud providers earn “high profit margins” despite offering “many undifferentiated services”; it does not state the speaker’s own view on the earlier revenue-composition question. These claims therefore do not establish a change of mind.
• NOT COMPARABLE · 91%26 Jun → 7 AugThen Acquiring novel skills requires AIs to learn from scarce real-world interactions whose information is unstructured, unverifiable, and ambiguous.Now The speaker believes AI systems need actual continual learning.The earlier claim says acquiring novel skills requires learning from scarce, ambiguous real-world interactions. The later claim says AI systems need “actual continual learning.” These ideas are compatible, but the statements address different aspects of learning, and the excerpts do not establish whether the speaker's position on the same underlying question changed.
• NOT COMPARABLE · 99%17 Apr → 7 AugThen The speaker says digital firms could replicate people or teams with their tacit knowledge, unlike human firms, where team culture can dilute as members leave, die, or grow old.Now The speaker predicts that under continual learning, a company could not remain competitive while delaying public deployment by four months, because a competitor releasing immediately would gain a smarter model from real-world experience.The earlier claim concerns digital firms preserving or replicating a team’s tacit knowledge as human teams change over time. The later claim concerns whether a company can stay competitive while delaying public deployment under continual learning. These address different questions, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%17 Apr → 7 AugThen The speaker says digital firms could replicate people or teams with their tacit knowledge, unlike human firms, where team culture can dilute as members leave, die, or grow old.Now The speaker predicts that continual learning would make a prolonged gap between internal and public deployment unviable.The earlier claim concerns digital firms preserving or replicating people’s tacit knowledge, while the later claim concerns whether continuous learning would allow a prolonged gap between internal and public model deployment. They do not express positions on the same underlying question, so the later statement does not establish a change of mind.
• NOT COMPARABLE · 96%26 Jun → 7 AugThen The loss function may be the main bottleneck for continual learning.Now The speaker believes AI systems need actual continual learning.The earlier claim identifies the loss function as a possible main bottleneck for continual learning (“may be the main bottleneck”). The later claim says AI systems need actual continual learning, but does not state what is bottlenecking it or revise the earlier view. These claims can both be true; the later quote does not establish a change in position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen Human continual learning primarily consolidates useful intuitions and big-picture knowledge rather than preserving every observation for immediate recall.Now The speaker expects many things to change before continual learning becomes functional.The earlier claim says human continual learning is mainly about consolidating useful intuitions and big-picture knowledge, rather than retaining every observation for immediate recall. The later claim—“plenty more will have changed by the time that continual learning actually works”—is about what may change before the capability becomes functional, not about what continual learning consists of. These statements do not establish a change in position on the same question.
• NOT COMPARABLE · 97%26 Jun → 7 AugThen People often say that employees may need six months or more of on-the-job experience before becoming net productive.Now The speaker predicts that under continual learning, a company could not remain competitive while delaying public deployment by four months, because a competitor releasing immediately would gain a smarter model from real-world experience.These address different questions. The earlier claim reports that employees may need “six months or more” of on-the-job experience to become productive; the later claim predicts that a company could not stay competitive with a “four-month gap” in public deployment under continual learning. Neither claim clearly reverses the other.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen Some autistic savants can recall arbitrary information years later, resembling the high-fidelity information AI models retain in context.Now The speaker attributes the similarity among prominent AI models to their having been trained on roughly the same data.The earlier statement compares some autistic savants’ ability to recall arbitrary information with AI models’ in-context fidelity. The later statement says prominent AI models are similar because they were trained on roughly the same data. These address different questions, and neither claim establishes a changed position on the other.
• NOT COMPARABLE · 97%26 Jun → 7 AugThen The speaker believes several approaches to updating model weights from session learning should work, at least in principle.Now The speaker predicts that the technical challenge of merging personalized weight forks into a main model will eventually be solved.The earlier claim says there are “many ideas that ought to work” for updating model weights based on a session. The later claim predicts that merging personalized weight forks into a main model “will be solved.” These concern related but distinct techniques, and neither statement rules out the other; there is no clear change in the speaker’s position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen The speaker has claimed that AI models are one millionth as sample-efficient as humans.Now The speaker attributes the similarity among prominent AI models to their having been trained on roughly the same data.The earlier claim concerns sample efficiency, saying models are “one millionth as sample efficient as humans.” The later claim concerns why prominent AI models are similar, attributing that to being “trained on roughly the same data.” These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%26 Jun → 7 AugThen A naive SFT method for continual learning trains the base model to predict all tokens observed in a session.Now The speaker believes AI systems need actual continual learning.The earlier claim describes a naive SFT training method: predicting all tokens observed in a session. The later claim says actual continual learning is needed. These statements do not establish opposing positions: describing a method is not the same as asserting that it is sufficient, and the later claim does not say whether that SFT method is useful or compatible with continual learning.
• NOT COMPARABLE · 98%3 Apr → 7 AugThen The speaker observes that smarter AI systems seem to be more reliable agents and more likely to do what users expect.Now The speaker predicts that AI systems will become more diverse in their learned behavior or characteristics.The earlier claim says smarter systems seem to be “more reliable agents” and more likely to do what users expect. The later claim predicts that “the diversity of AI minds will increase.” These address different characteristics and can both be true, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%26 Jun → 7 AugThen Mercury provides core business functions in one place, which the speaker says it does extremely well.Now The speaker attributes cloud providers' high profit margins to the time and expense involved in switching providers.The earlier statement concerns Mercury's core business functions being available in one place; the later concerns why cloud providers have high margins. These are different topics, so they do not establish a change in the speaker's position.
• NOT COMPARABLE · 99%17 Apr → 7 AugThen According to the speaker, all of the most successful and durable software companies of the last decade transitioned to enterprise customers.Now The speaker attributes cloud providers' high profit margins to the time and expense involved in switching providers.The earlier claim concerns successful software companies transitioning to enterprise customers; the later claim concerns why cloud providers have high margins. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%17 Apr → 3 AugThen When demand for a skill is highly elastic, the ability to deploy many workers in parallel can matter more than salary.Now The speaker thinks a labor-supply shock from AI might be so large and rapid that the usual heuristic about labor supply and wages cannot be relied on.The earlier claim concerns how highly elastic demand for a skill can make parallel deployment of workers more important than salary. The later claim concerns whether a rapid AI labor-supply shock could make the usual heuristic about labor supply and wages unreliable. These address related but distinct questions, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 99%17 Apr → 3 AugThen The speaker thinks Apple, with its current organization and computing resources, might be able to centrally plan an economy like Uruk's, but could not manage today's world economy.Now The speaker guesses that comparing compute scarcity to the Simon-Ehrlich bet is probably a bad analogy.The earlier statement concerns whether Apple could centrally plan an economy like Uruk’s versus today’s world economy. The later statement says the Simon–Ehrlich bet is probably a bad analogy. These address different questions, so the later claim does not show a change in the speaker’s view about central planning.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen ChatGPT illustrates that post-training using about 1% additional compute can make a model pretrained on internet data highly competent at chatbot interaction.Now Smarter AI models will be able to generate more revenue from the same amount of compute.The earlier claim is about post-training making a model highly competent at chatbot interaction with roughly 1% additional compute. The later claim is about smarter models generating more revenue from the same compute. These are different questions, and the later statement neither revises nor denies the earlier one.
• NOT COMPARABLE · 89%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now According to Epoch, OpenAI allocated about 25% of its compute to inference in 2024; the speaker estimates the share is now likely around 50% or higher.The earlier statement gives a broad estimate for an AI lab—“30 to 50%”—while the later statement discusses OpenAI specifically, compares its 2024 share (“a quarter”) with an estimate for now (“50%, if not higher”). These differ in scope and the later claim does not clearly reverse the earlier estimate; the surrounding context does not establish that both statements concern the same lab or basis of comparison.
• NOT COMPARABLE · 96%26 Jun → 3 AugThen Although AIs are broadly deployed across the economy and encounter domain- and organization-specific tacit knowledge, they are not able to use that knowledge.Now AI is relatively inexpensive today because it cannot perform many tasks that top human workers can perform.The earlier statement says deployed AIs are “not able to make use of” domain- and organization-specific tacit knowledge. The later statement says AI is cheap because it “can't do a lot of things that top humans can do.” These address different capability questions and can both be true; the later claim does not show a changed position on the earlier one.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now A model that produces the same result with less compute effectively expands available compute and increases compute's value.The earlier statement quantifies how much lab compute is used for inference (“30 to 50%”). The later statement describes the effect of a more compute-efficient model (“same result by using less compute”), not a revised estimate of inference’s share of compute. They do not establish a change in position on the same question.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now Reportedly, Anthropic's inference margins increased from 40% in the middle of last year to over 80% now.The earlier claim concerns compute allocation: “30 to 50% of a lab’s compute goes to inference.” The later claim concerns inference margins: they “went from 40% … to upwards of 80%.” These are different measures, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen The interviewer cites OpenAI’s early success against Google DeepMind as an example of a company with less compute beating a more compute-rich incumbent.Now If leading AI labs improve their monetization of compute while compute prices rise, other competitors will find it harder to compete for compute against labs that use it more productively.The earlier item is an interviewer’s example about a company with less compute beating a compute-rich incumbent; it does not establish the speaker’s own view about how compute affects competition. The later passage asserts that rising compute costs and stronger monetization make it harder for other labs to compete, but it addresses a different competitive condition and does not directly contradict the earlier example.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now Google pays $900 million monthly to rent 110,000 GPUs, a mix of GB200 and GB300 units, from SpaceX.The earlier statement says that “30 to 50%” of an AI lab’s compute goes to inference. The later statement concerns Google paying “900 million dollars a month” to rent GPUs from SpaceX. These address different facts, so they do not show a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Models will struggle to make substantial progress in domains without highly replayable training targets.Now A lab that trains the best and most efficient model will be able to charge higher margins than it does today.The earlier claim concerns whether models can make progress without replayable training targets: “models will struggle to make much progress.” The later claim concerns the pricing power of a lab with the best, most efficient model: it can “charge much higher margins.” These address different questions and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%3 Apr → 3 AugThen The speaker argues conditionally that if AIs can perform major tasks once they have general intelligence, then AGI would be transformative because AIs would have memorized human knowledge and use it to make connections.Now The speaker expects AI eventually to become capable of tasks that top human workers can perform.The earlier statement conditionally predicts that AGI would be “totally transformative” because of its knowledge and ability to make connections. The later statement says AI will eventually do tasks top humans can do. That speaks to capability, not whether AGI’s effects will be transformative, so it does not establish a change of mind.
• NOT COMPARABLE · 96%26 Jun → 3 AugThen The labs’ research bet is that training AIs on millions of verifiable tasks across thousands of diverse reinforcement-learning environments will produce AGI capable of making progress on open-ended tasks for weeks despite errors and ambiguity.Now AI labs do not want to be cloud providers rather than build AGI, which they regard as a less compelling business.The earlier statement describes the labs’ research bet: that training across many verifiable tasks will produce AGI capable of sustained open-ended progress. The later statement says labs prefer building AGI to being cloud providers. These are different claims about the labs’ aims and business incentives, not opposing positions on the same question, so they do not establish a change of mind.
• NOT COMPARABLE · 99%17 Apr → 3 AugThen The speaker considers bandwidth and information sharing more important than incentive alignment in explaining the relevant organizational changes.Now The speaker is concerned that strong economies of scale in intelligence could concentrate power.The earlier claim compares bandwidth and information sharing with incentive alignment as explanations for organizational change. The later claim concerns economies of scale in intelligence and the risk of power concentration. They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%17 Apr → 3 AugThen Building enterprise features consumes substantial engineering time and capital that the speaker believes should instead go toward the core product.Now The speaker is uncertain whether a sudden increase in software-engineer supply would reduce the marginal value of software engineers and H100-generated revenue as described.These address different questions. Earlier, the speaker says enterprise features take substantial engineering time and capital that should go to the core product. Later, they say, “I actually don't know if this is true” about whether more software engineers would reduce their marginal value and H100-generated revenue. The later uncertainty does not revise the earlier view.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Although AIs are broadly deployed across the economy and encounter domain- and organization-specific tacit knowledge, they are not able to use that knowledge.Now AI labs do not want to be cloud providers rather than build AGI, which they regard as a less compelling business.The claims address different questions. The earlier statement says deployed AIs cannot use the tacit knowledge they encounter; the later says labs prefer building toward AGI over operating as cloud providers. The later claim does not state whether AI systems can use that knowledge, so it provides no evidence of a changed position on the earlier issue.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen The interviewer doubts that the simultaneous increases in compute and AI progress demonstrate that compute caused the progress.Now The speaker is uncertain how AI compute can sustain 3x year-over-year growth for the next few years and doubts it can grow faster.The earlier claim says, “I'm not sure I buy” that simultaneous increases in compute and AI progress establish a causal relationship. The later claim says the speaker does not know how compute scaling could continue at 3× year over year. One concerns whether compute caused AI progress; the other concerns whether compute growth can be sustained. The later statement does not establish a change in the earlier causal view.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now The H100 rental-value calculation does not account for AI being able to work nights and weekends.These statements address different questions. The earlier claim says “30 to 50% of a lab’s compute goes to inference”; the later says an H100 rental-value calculation does not account for AI working “nights and weekends.” Neither confirms nor reverses the other.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen The speaker has claimed that AI models are one millionth as sample-efficient as humans.Now AI labs believe they will build models within a year that make today's models look extremely poor.The earlier statement says the models are “one one millionth as sample efficient as humans,” while the later statement reports labs’ expectation that they will build models within a year that make current ones “look extremely shitty.” These concern different questions—sample efficiency versus anticipated future model capability—and can both be true; the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen If an AI can build a useful simulation of reality for rehearsal and strategy testing, dreaming could provide orders of magnitude more simulated samples per unit of wall-clock time.Now The hypothetical annual rental value of an H100 running a human-level software engineer would exceed 15 times its current spot price.The earlier claim is about whether AI “dreaming” through a useful reality simulation could generate “orders of magnitude more simulated samples.” The later claim concerns the hypothetical rental value of an H100 running a human-level software engineer. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%26 Jun → 3 AugThen The speaker says that adding enough compute to large language models caused fundamental natural-language-processing research problems to collapse.Now The speaker attributes approximately a 1.4x increase in annual compute capacity to Moore's law.The earlier statement says that adding enough compute caused fundamental NLP research problems to collapse. The later statement attributes an approximately 1.4x increase in annual compute capacity to Moore’s law. These address different questions—what scaling compute did to research problems versus what accounts for a particular rate of compute growth—and do not establish a change in position.
• NOT COMPARABLE · 97%26 Jun → 3 AugThen Inference compute is currently not used productively to improve the model.Now AI labs would prefer not to allocate a larger share of their compute to inference.The earlier claim says inference compute is “not used productively to improve the model.” The later claim says labs “would prefer not” to increase the share of compute allocated to inference. These address different questions—whether inference compute improves models versus whether labs want to allocate more compute to inference—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 3 AugThen The speaker says that adding enough compute to large language models caused fundamental natural-language-processing research problems to collapse.Now The speaker attributes about 1.2x annual compute growth to new fabs and says fab expansion will be bottlenecked through 2030 or later by the construction of new ASML UV machines.The earlier claim says that adding enough compute caused fundamental NLP research problems to “collapse.” The later claim concerns the physical limits on future compute growth, saying expansion will be “bottlenecked” by building fabs and ASML machines. These claims address different questions and can both be true; the later statement does not retract the earlier one.
• NOT COMPARABLE · 97%17 Oct → 3 AugThen Serving an additional instance of an AI model incurs additional cost.Now Smarter AI models will be able to generate more revenue from the same amount of compute.The earlier claim says that serving an additional model instance has an additional cost. The later claim says smarter models can generate more revenue from a fixed amount of compute. These claims address different aspects of AI economics and can both be true; the later statement does not deny or revise the cost claim.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen The speaker thinks Apple, with its current organization and computing resources, might be able to centrally plan an economy like Uruk's, but could not manage today's world economy.Now The speaker is concerned that strong economies of scale in intelligence could concentrate power.The earlier claim concerns whether Apple could centrally plan economies of different sizes and complexity: it might manage “the economy of Uruk” but “cannot manage the economy of the world economy as it exists today.” The later claim is about the consequences of “strong economies of scale for intelligence,” specifically concern about “power concentration.” These positions can coexist, but they address different questions, so the later statement does not establish a change of mind.
• NOT COMPARABLE · 98%17 Oct → 3 AugThen The speaker says that, excluding ordinary chat revenue and considering API revenue, the revenues of the companies being discussed are dominated by coding.Now The speaker says Anthropic's 10x annual revenue growth alongside 3x compute growth illustrates strong economies of scale in the AI model business.The earlier claim is about API revenue being “dominated by coding.” The later claim is about Anthropic’s revenue growing faster than its compute and what that suggests about “economies of scale.” These statements address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 96%17 Apr → 3 AugThen When demand for a skill is highly elastic, the ability to deploy many workers in parallel can matter more than salary.Now If standard economics is correct, the marginal value of labor and consequently compute should remain extremely high.The earlier claim says that for a skill with highly elastic demand, deploying many workers in parallel can matter more than salary. The later claim says that if standard economics applies, the marginal value of labor and compute should remain high, in a discussion of whether a large, fast labor-supply shock might defeat the usual heuristic. These claims address different economic points and do not establish that the speaker reversed the earlier position.
• NOT COMPARABLE · 98%3 Apr → 3 AugThen SPEAKER_00 argues that highly successful work can attract rapid attention from relevant audiences without years of prior fan-base building.Now The speaker says there is no deep reason that Anthropic's rapid revenue-growth trend must continue.The earlier claim concerns how quickly successful work can attract attention; the later claim concerns whether Anthropic's revenue-growth trend must continue. They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%17 Apr → 3 AugThen When demand for a skill is highly elastic, the ability to deploy many workers in parallel can matter more than salary.Now The speaker suggests that a sudden addition of 10 million software engineers might lower the marginal value of software-engineering labor and thus reduce the revenue an H100 could generate below 15 times its current level.The earlier claim is that for a skill with “high elasticity of demand,” deploying many workers in parallel can matter more than salary. The later claim concerns a specific supply shock—“10 million extra software engineers”—that might lower marginal value and reduce an H100’s revenue below 15 times its current level. These claims do not establish a change of position on the same question; the later speaker also says, “I actually don't know if this is true.”
• NOT COMPARABLE · 94%3 Apr → 3 AugThen Producing GPUs at scale would require solving problems in physics, chemistry, and materials science, building computing fabs, and improving them over decades to achieve Moore's-law progress.Now The speaker attributes about 1.2x annual compute growth to new fabs and says fab expansion will be bottlenecked through 2030 or later by the construction of new ASML UV machines.The earlier statement describes what it would take to produce GPUs at scale from scratch—“solving all kinds of interesting physics and chemistry problems” and improving fabs over decades. The later statement discusses current compute growth, saying it is bottlenecked by building new fabs and ASML UV machines through 2030 or later. These claims are compatible, but address different timeframes and questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now Compute price increases are stronger for the large-scale capacity frontier labs need to accumulate, since they cannot rely on buying spot instances.The earlier statement estimates what share of an AI lab’s compute is used for inference (“Around 30 to 50%”). The later statement concerns the effect of price increases on the large-scale compute capacity frontier labs need to accumulate. These address different questions, and neither claim establishes a changed position on the other.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen The speaker thinks Apple, with its current organization and computing resources, might be able to centrally plan an economy like Uruk's, but could not manage today's world economy.Now The speaker worries that the argument about compute scarcity may repeat historical mistakes in predicting scarcity.The earlier claim is about whether Apple could centrally manage economies of different scales: it might manage Uruk but not the modern world economy. The later claim—“worried that this kind of analysis” repeats past errors about scarcity—is about predictions of compute scarcity. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%17 Apr → 3 AugThen The speaker considers highly elastic demand, alongside transformative AI, probably the most underrated tangible factor for understanding the future AI society.Now The speaker frames continuation of the revenue-growth trend as depending on whether AI capabilities become sufficiently useful by the end of next year.The earlier claim calls highly elastic demand an underrated factor in understanding future AI society. The later claim says continuation of a revenue-growth trend depends on whether AI becomes sufficiently useful by the end of next year. These address different questions and can both be true; the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%17 Oct → 3 AugThen The interviewer speculates that radiologists’ wages may have increased because they provide scarce human work that remains a bottleneck to wider automation.Now The speaker thinks a labor-supply shock from AI might be so large and rapid that the usual heuristic about labor supply and wages cannot be relied on.The earlier passage is described as the interviewer speculating that radiologists’ wages rose because they are a scarce bottleneck; it does not establish the person’s own position. The later claim says a rapid, large AI labor-supply shock may defeat the usual wage heuristic. These concern different claims, and the excerpts do not show a clear reversal by the same speaker.
• NOT COMPARABLE · 98%17 Oct → 3 AugThen The speaker says population growth has driven economic growth over the last 50 years, while population and growth in frontier countries have stagnated; he contrasts this with earlier periods when population was not rapidly increasing.Now Paul Ehrlich bet that a basket of commodities would increase in price during the decade preceding 1990, amid his concerns about population growth.The earlier claim discusses population growth as a driver of economic growth. The later claim says Paul Ehrlich “was this famous doomer about population growth” and describes his bet; it does not state the speaker’s own position on whether population growth drives economic growth. These statements do not establish a change of mind.
• NOT COMPARABLE · 96%26 Jun → 3 AugThen The speaker argues that leaving inference compute unused for learning is wasteful because the most valuable learning information emerges during deployment.Now AI labs would prefer not to allocate a larger share of their compute to inference.The earlier claim says leaving inference compute unused for learning is “a huge waste” because useful learning information emerges during deployment. The later claim says labs would prefer not to increase the share of compute spent on inference. These concern related compute allocation issues, but the later statement reports the labs’ preference and does not establish that the speaker has changed their own view about learning from inference.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Successful online-learning models so far have learned a shared objective across millions of users rather than learning different things for each user.Now Unlike model training, each instance of human labor must be retrained from scratch.The earlier claim concerns online-learning models, saying they learn “a shared objective across millions of users” rather than different things for each user. The later claim compares human labor with model training, saying each instance of human labor “must be retrained from scratch.” These address different questions and do not show a change in the speaker’s position on the earlier claim.
• NOT COMPARABLE · 91%17 Oct → 3 AugThen The speaker says the peak cost of an AI session is relatively low and that AI costs may be amortized into the training run, depending on inference scaling.Now The speaker understands that AI lab margins, compute prices, and the share of compute allocated to inference are all already increasing.The earlier claim is about the relatively low cost of an AI session and the possibility of amortizing AI costs into training. The later claim says that lab margins, compute prices, and inference’s share of compute are already increasing. These address different economic measures, and the later statement does not clearly accept or reject the earlier claim. The surrounding turns also do not establish that both extracted claims are the same speaker’s own positions.
• NOT COMPARABLE · 94%26 Jun → 3 AugThen Inference compute is currently not used productively to improve the model.Now AI labs see inference revenue as a means of persuading investors to fund training of the next, larger and better model.The earlier claim says inference compute is “not used productively to improve the model,” while the later claim says inference revenue can help persuade investors to fund training a “larger and better model.” These concern different mechanisms—using inference compute itself to improve a model versus using revenue from inference to finance future training—so the later statement does not establish a change in the speaker’s position on the earlier question.
• NOT COMPARABLE · 99%17 Oct → 3 AugThen The speaker says economic growth often results from people generating ideas and others executing them to create valuable output.Now The speaker says high margins in a market economy arise when a product is substantially better than available substitutes.The earlier claim concerns sources of economic growth—people generating ideas and others executing them. The later claim concerns why a product can sustain high margins—because it is substantially better than available substitutes. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now If higher lab margins are the main outcome, margins on some top AI models could rise from 80% to over 90%, requiring the leading model to be far better than competitors.The earlier statement reports that “30–50%” of a lab’s compute is used for inference. The later statement discusses a conditional possibility about model margins rising “from 80% to over 90%” if higher lab margins dominate. These concern different measures and questions; the later claim does not revise the earlier estimate.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen In the speaker’s scenario, deployed experience across the economy eventually becomes the main source of AI improvement, surpassing pre-release training.Now Unlike model training, each instance of human labor must be retrained from scratch.The earlier claim concerns what will drive AI improvement over time: deployed experience will become more important than pre-release training. The later claim concerns how skills are shared across instances, contrasting human labor that must be retrained “from scratch” with model training. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now AI labs see inference revenue as a means of persuading investors to fund training of the next, larger and better model.The earlier statement gives an estimate of compute allocation: “Around 30 to 50% of a lab's compute goes to inference.” The later statement describes how labs view inference revenue—as a way to persuade investors to fund training—and does not state what share of compute is used for inference. These are different questions, so the later claim neither revises nor contradicts the earlier estimate.
• NOT COMPARABLE · 98%3 Apr → 3 AugThen The interviewer argues that nationalization could increase arms-race risk because China may be more likely to enter an arms race if the United States does.Now The speaker is concerned that strong economies of scale in intelligence could concentrate power.The earlier claim is explicitly an argument by the interviewer: nationalization could raise arms-race risk because China might respond. The later claim is the speaker’s own concern that economies of scale in intelligence could concentrate power. These do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now AI labs can charge a larger premium for models that use scarce compute more efficiently.The earlier claim estimates how much lab compute is used for inference (“30–50%”). The later claim says labs can charge more for models that use compute efficiently. These address different questions—compute allocation versus pricing—and neither reverses the other.
• NOT COMPARABLE · 99%17 Apr → 3 AugThen The interviewer considers the guest’s argument overly reliant on a macroeconomic outside view and proposes consulting AI researchers directly.Now The speaker thinks a labor-supply shock from AI might be so large and rapid that the usual heuristic about labor supply and wages cannot be relied on.These statements address different questions. The earlier claim says the guest’s argument is “too outside view” and suggests asking AI researchers; the later claim says a labor-supply shock may be “so big and so fast” that a general wage heuristic cannot be relied on. They do not establish a change of position on the same underlying issue, and the later speaker is identified only as “SPEAKER_00.”
• NOT COMPARABLE · 94%3 Apr → 3 AugThen SPEAKER_00 advises against radical policies that work only under one scenario and are counterproductive under others, suggesting nationalization might be such a policy.Now The speaker is concerned that strong economies of scale in intelligence could concentrate power.The earlier claim cautions against “big radical moves” that depend on one specific scenario, and suggests “nationalization might be one” such move. The later claim says the speaker is worried that economies of scale in intelligence could cause “power concentration.” That expresses a concern about an outcome, not a changed position on whether nationalization or other radical policies are advisable.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen The speaker says that adding enough compute to large language models caused fundamental natural-language-processing research problems to collapse.Now The speaker expects that many popular AI applications will probably become uneconomical as compute prices rise.These claims address different questions. The earlier statement says that adding enough compute made fundamental NLP research problems “collapse”; the later predicts that popular AI applications may “get priced out” as compute becomes more valuable. One concerns research progress from scaling, the other application economics, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 91%26 Jun → 3 AugThen Continual learning requires incorporating learned information into model weights.Now A model's training cost is incurred once, while the skills learned can be shared across all users.The earlier claim is about how continual learning incorporates new information: it says this requires going back to the weights. The later claim is about amortizing training costs and sharing already-learned skills across users. These can both be true; the later statement does not say that models learn continually without updating weights or otherwise deny the earlier claim.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen The speaker’s scenario assumes future context lengths will let AIs collaborate with a user for an entire week.Now The speaker is uncertain how AI compute can sustain 3x year-over-year growth for the next few years and doubts it can grow faster.The earlier statement introduces a scenario in which “future context lengths” let AIs collaborate for a week. The later statement concerns whether AI compute can sustain “3x compute scaling year over year.” These address different capabilities, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen A hypothetical dream command would spend substantial compute building and training in a game-like simulation of the model’s real-world observations.Now The hypothetical annual rental value of an H100 running a human-level software engineer would exceed 15 times its current spot price.The statements address different questions. The earlier describes a hypothetical “dream” command using compute to train in a game-like simulation; the later estimates the rental value of an H100 running a human-level software engineer. Neither establishes a changed position on the same underlying question.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now If lab compute grows 3x year over year while revenue grows 10x, either lab margins must increase to capture the surplus or compute prices must rise so suppliers below the labs capture it.The earlier claim is that “30 to 50%” of a lab’s compute goes to inference. The later claim concerns how a gap between compute growth and revenue growth is captured—through “lab margins” or “compute prices.” It neither affirms nor denies the earlier inference estimate, so there is no basis to identify a change in position.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Attention-generated fast weights enable sample-efficient in-context learning but have poor memory scaling.Now Frontier labs need large-scale compute that provides efficiency and flexibility and protects their model weights and customers' information.The claims address different questions. The earlier claim says attention-generated fast weights enable sample-efficient in-context learning but scale poorly in memory; the later claim says frontier labs need large-scale compute with efficiency, flexibility, and security. The later statement neither revises nor denies the earlier position.
• NOT COMPARABLE · 97%3 Apr → 3 AugThen Producing GPUs at scale would require solving problems in physics, chemistry, and materials science, building computing fabs, and improving them over decades to achieve Moore's-law progress.Now The speaker attributes approximately a 1.4x increase in annual compute capacity to Moore's law.The earlier claim concerns the difficulty of creating GPU production and fab capabilities from scratch, including decades of Moore’s-law progress. The later claim says Moore’s law accounts for “1.4x” of annual compute-capacity growth. That attribution does not establish a different view about how difficult it would be to build GPU production at scale.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen According to the speaker, all of the most successful and durable software companies of the last decade transitioned to enterprise customers.Now The speaker says Anthropic's 10x annual revenue growth alongside 3x compute growth illustrates strong economies of scale in the AI model business.The earlier statement says that successful, durable software companies transitioned to enterprise customers. The later statement is about Anthropic’s revenue growing faster than its compute and what that suggests about economies of scale. These address different questions, so the later claim does not establish a changed position on enterprise customers.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen AI models require many samples during training and are highly sample-inefficient.Now AI labs believe they will build models within a year that make today's models look extremely poor.The earlier claim says models are “incredibly sample inefficient during training.” The later claim reports that labs expect to build much better models within a year. These claims address different questions—training efficiency versus anticipated model progress—and can both be true, so they do not establish a change of mind.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen The speaker raises the possibility of fitting six months of on-the-job experience into an AI model’s context window.Now The speaker says AI's share of leading-edge wafer capacity cannot continue rising once it has absorbed all available capacity.The statements address different questions. The earlier claim asks hypothetically, “what if you could just fit those six months into the context window?”, in a discussion of online learning and model context. The later claim says AI cannot keep increasing its share of leading-edge wafer capacity after absorbing all available capacity. Neither statement establishes a change in the speaker’s position on the same issue.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now The speaker is uncertain whether AI lab margins or compute prices will rise to resolve the gap between revenue and compute growth.The earlier statement says that “Around 30 to 50% of a lab's compute goes to inference.” The later statement says, “It's not clear to me which world we end up in,” referring to whether lab margins or compute prices will rise. These address different questions, and the later statement does not establish a changed view about inference’s share of compute.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen In the speaker’s scenario, deployed experience across the economy eventually becomes the main source of AI improvement, surpassing pre-release training.Now The speaker says economists generally believe high immigration does not lower long-run wages because innovation and specialization raise the value of labor.The earlier claim concerns what drives AI improvement after deployment; the later claim concerns the long-run wage effects of immigration. They do not address the same underlying question, so they do not establish a change in position.
• NOT COMPARABLE · 99%17 Oct → 3 AugThen The speaker argues that the Industrial Revolution was not caused by a single magical key invention; examining 1770 or 1870 would not reveal one such invention.Now The speaker says other analyses suggest Ehrlich might have won the commodity-price bet had it been made in another decade.The claims address different questions. The earlier statement says the Industrial Revolution was not driven by one “magical key invention”; the later statement reports that Ehrlich might have won a commodity-price bet in a different decade. The later claim does not revise the earlier position.
• NOT COMPARABLE · 99%17 Oct → 3 AugThen The speaker says the Industrial Revolution shifted the economy into a much faster-progress regime, with exponential growth increasing tenfold.Now The speaker says Anthropic's 10x annual revenue growth alongside 3x compute growth illustrates strong economies of scale in the AI model business.The earlier claim says the Industrial Revolution moved the economy into a faster-progress regime, with growth increasing tenfold. The later claim says Anthropic’s revenue growing 10x while compute grew 3x illustrates economies of scale in the model business. These concern different economic mechanisms and do not establish a changed position on the same question.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen AI agents are able to solve more ambitious problems over longer periods of time.Now AI labs do not want to be cloud providers rather than build AGI, which they regard as a less compelling business.The earlier claim says AI agents are solving more ambitious problems over longer periods. The later claim concerns AI labs’ business priorities—preferring to build toward AGI rather than operate as cloud providers. These address different questions, so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 97%17 Apr → 3 AugThen Some companies with substantially less compute have beaten incumbents with more compute by having a more coherent vision and more concentrated research effort.Now If leading AI labs improve their monetization of compute while compute prices rise, other competitors will find it harder to compete for compute against labs that use it more productively.The earlier claim says that companies with less compute have sometimes beaten better-resourced incumbents through a more coherent vision and concentrated research effort. The later claim says that rising compute costs and stronger monetization by leading labs can make it harder for other competitors to compete for compute. These claims concern different mechanisms and do not establish a changed position on the same underlying question.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen The speaker’s podcast has generated more operational overhead as it has grown.Now The speaker doubts that the three sources of 3x annual compute growth can be accelerated substantially or even sustained.These claims address unrelated questions: the earlier statement says the podcast has required more operational work as it has grown, while the later statement expresses doubt that the components of annual compute-capacity growth can be accelerated or sustained. Neither claim bears on the other's position, so this is not evidence of a change of mind.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen Large organizations have difficulty maintaining a single coherent vision.Now The speaker is concerned that strong economies of scale in intelligence could concentrate power.The earlier claim says “Large organizations have difficulty maintaining a single coherent vision.” The later claim says the speaker is worried that “strong economies of scale for intelligence” could lead to “power concentration.” These concern related but distinct effects of scale, and the later statement does not establish a changed view about whether large organizations can maintain a coherent vision.
• NOT COMPARABLE · 97%26 Jun → 3 AugThen A hypothetical dream command would spend substantial compute building and training in a game-like simulation of the model’s real-world observations.Now AI labs need to spend the majority of their compute on training and experiments for the next model.The earlier claim describes a hypothetical “dream” command that would use substantial compute to simulate and train on the model’s observations, in response to a question about continual learning. The later claim says labs should spend “the majority” of compute on training and experiments for the next model, in a discussion about compute allocation. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen The speaker’s scenario assumes future context lengths will let AIs collaborate with a user for an entire week.Now The speaker expects robots eventually to turn silica sand and copper ore into computer chips.The earlier claim describes a scenario where future context lengths let AIs “co-work with you for a full week.” The later claim predicts robots will “convert shores of silica sand and mines of copper into new computer chips.” These concern different predictions, so the later statement does not show whether the speaker changed their view on the earlier one.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now If an AI lab's revenue continues growing 10x year over year while its compute grows 3x, its margins must increase, compute prices must rise, the share of compute used for inference rather than training must increase, or some combination of these must occur.The earlier claim reports that inference accounts for “30 to 50%” of a lab’s compute. The later claim describes a possible condition for sustaining rapid revenue growth despite slower compute growth: the inference share “has to increase,” among other options. It does not state what the current share is or contradict the earlier estimate; the claims address different questions.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen A couple more years of progress might produce transformer context windows that feel infinitely large.Now The speaker thinks it would be remarkable if Moore's law continued for even a few more years.The earlier statement concerns whether architectural progress could make transformer context windows feel “infinitely large.” The later statement concerns whether Moore’s law can continue for “a few more years.” These are different questions, and the earlier does not assert that Moore’s law will continue.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Mercury provides core business functions in one place, which the speaker says it does extremely well.Now The speaker is concerned that strong economies of scale in intelligence could concentrate power.The claims address different topics: the earlier statement praises Mercury for putting core business functions in one place, while the later statement expresses concern that economies of scale in intelligence could concentrate power. Neither statement changes or contradicts the other.
• NOT COMPARABLE · 98%17 Oct → 3 AugThen If the current AI paradigm is on a path to AGI, the progression is not expected to consist of gradually automating small tasks across knowledge-work occupations.Now The speaker expects AI eventually to become capable of tasks that top human workers can perform.The earlier claim is about the *pattern* of progress toward AGI: it says progress will not happen by gradually automating small tasks across knowledge-work occupations. The later claim says AI will eventually be capable of tasks top human workers can perform. That does not establish a different view about how automation progresses, so the two claims are not directly comparable.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now A lab that trains the best and most efficient model will be able to charge higher margins than it does today.The earlier statement says what “people who are in favor of the current training paradigm will say”: that training’s sample-efficiency disadvantage is amortized across many sessions. It is not clearly the speaker’s own position. The later statement predicts that a lab with the “best, most efficient model” can charge higher margins, which addresses a different question and does not establish a change in the speaker’s view.
• NOT COMPARABLE · 98%3 Apr → 3 AugThen The speaker argues that invention often proceeds through experimentation before the relevant theories or machinery are understood, citing the steam engine's development before thermodynamics and the Wright brothers' airplane experiments.Now The speaker worries that the argument about compute scarcity may repeat historical mistakes in predicting scarcity.The earlier claim is that invention often advances through experimentation before its theory is understood, while the later claim is concern that analysis of compute scarcity may repeat past errors about scarcity. These positions address different questions and do not show that the speaker reversed their view about how invention proceeds.
• NOT COMPARABLE · 97%3 Apr → 3 AugThen SPEAKER_00 believes classical liberalism has not been a helpful guide in conditions where small changes can lead to very different outcomes.Now The speaker is concerned that strong economies of scale in intelligence could concentrate power.The earlier statement says classical liberalism has not been a helpful guide under conditions of epistemic uncertainty. The later statement expresses concern that economies of scale in intelligence could concentrate power. These address different questions, so the later concern does not establish a change in the earlier position.
• NOT COMPARABLE · 99%17 Apr → 3 AugThen The speaker considers bandwidth and information sharing more important than incentive alignment in explaining the relevant organizational changes.Now The speaker finds it surprising that intelligence could sustain margins above 90% without competition eroding them.The earlier claim compares bandwidth and information sharing with incentive alignment as explanations for organizational change. The later claim expresses surprise about whether intelligence can sustain margins above 90% without competition eroding them. These address different questions, so the statements provide no evidence of a change in position.
• NOT COMPARABLE · 94%17 Oct → 3 AugThen The speaker says economic growth often results from people generating ideas and others executing them to create valuable output.Now The speaker says economists generally believe high immigration does not lower long-run wages because innovation and specialization raise the value of labor.The earlier claim says growth often comes from people generating ideas and others executing them. The later claim reports economists’ view that innovation and specialization can raise labor’s value and prevent immigration from lowering long-run wages. These are related ideas, but they address different questions; the later claim also attributes the position to economists rather than clearly asserting it as the speaker’s own.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen In the speaker’s scenario, deployed experience across the economy eventually becomes the main source of AI improvement, surpassing pre-release training.Now The speaker is uncertain whether a sudden increase in software-engineer supply would reduce the marginal value of software engineers and H100-generated revenue as described.The earlier claim concerns how AI improves over time, saying deployed experience becomes the main source of improvement rather than pre-release training. The later claim—“I actually don't know if this is true”—concerns whether more software engineers would reduce their marginal value and H100-generated revenue. These address different questions, so the later uncertainty does not show a change in the earlier view.
• NOT COMPARABLE · 97%26 Jun → 3 AugThen Successful online-learning models so far have learned a shared objective across millions of users rather than learning different things for each user.Now A model's training cost is incurred once, while the skills learned can be shared across all users.The earlier claim is about online-learning models learning a shared objective rather than different things for each user. The later claim is about the one-time cost of training a model and sharing learned skills across users. These concern different aspects of model learning, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%17 Apr → 3 AugThen AI compute is expected to be about 10 times its current level in 2027.Now If an AI lab's revenue continues growing 10x year over year while its compute grows 3x, its margins must increase, compute prices must rise, the share of compute used for inference rather than training must increase, or some combination of these must occur.The earlier claim predicts that AI compute overall will be about “10x more” by 2027. The later claim discusses AI lab compute growing “3x year over year” relative to revenue growing 10x, and what economic changes would follow. These concern different measures and claims; the later statement does not establish a changed view about the earlier forecast.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen The speaker’s podcast has generated more operational overhead as it has grown.Now The speaker says Anthropic's 10x annual revenue growth alongside 3x compute growth illustrates strong economies of scale in the AI model business.The earlier statement is about the speaker’s podcast: it has required “more and more operational overhead” as it grew. The later statement is about economies of scale in the AI model business, inferred from Anthropic’s revenue and compute growth. These concern different subjects, so they do not establish a change of mind.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now The H100 rental-value calculation does not account for AI being able to work nights and weekends.These statements address different questions. The earlier passage reports what “people who are in favor of the current training paradigm will say” about amortizing training costs; the later says an H100 rental-value calculation omits AI working “nights and weekends.” Neither claim establishes a change in the speaker’s own position on the same issue.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen Building enterprise features consumes substantial engineering time and capital that the speaker believes should instead go toward the core product.Now The speaker suggests that a sudden addition of 10 million software engineers might lower the marginal value of software-engineering labor and thus reduce the revenue an H100 could generate below 15 times its current level.The earlier claim says enterprise features consume engineering time and capital that should go to the core product. The later claim considers a hypothetical increase in software engineers and its possible effect on the revenue an H100 could generate; it does not state a position about whether enterprise features are worth building. These address different questions, so the later statement does not show a change of mind.
• NOT COMPARABLE · 94%26 Jun → 3 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now A model's training cost is incurred once, while the skills learned can be shared across all users.The earlier claim says that “people who are in favor of the current training paradigm will say” the training cost is amortized across sessions, so it reports their argument rather than clearly stating the speaker’s own view. The later claim says the training cost is “one-time” and skills are “shared across all your users,” but the earlier passage does not establish whether the speaker held or rejected that view.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen The speaker’s podcast has generated more operational overhead as it has grown.Now The speaker is uncertain how AI compute can sustain 3x year-over-year growth for the next few years and doubts it can grow faster.The earlier claim is about the speaker’s podcast creating more operational overhead as it grows (“more and more operational overhead”). The later claim is uncertainty about whether AI compute can sustain 3x year-over-year growth. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 97%26 Jun → 3 AugThen The speaker’s scenario assumes future context lengths will let AIs collaborate with a user for an entire week.Now The speaker says AI's share of leading-edge wafer capacity cannot continue rising once it has absorbed all available capacity.The earlier passage describes a scenario in which future AI context lengths allow week-long collaboration; it does not assert a position about wafer capacity or compute scaling. The later passage says AI cannot keep increasing its share after absorbing leading-edge wafer capacity. These address different questions, so there is no comparable earlier position from which to infer a change of mind.
• NOT COMPARABLE · 98%17 Oct → 3 AugThen Serving an additional instance of an AI model incurs additional cost.Now AI labs can charge a larger premium for models that use scarce compute more efficiently.The earlier statement says that serving an additional model instance has an additional cost. The later statement says labs can charge a premium for models that use scarce compute more efficiently. These claims concern different aspects of AI economics and do not indicate a change in position about whether additional instances incur costs.
• NOT COMPARABLE · 98%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now If leading AI labs improve their monetization of compute while compute prices rise, other competitors will find it harder to compete for compute against labs that use it more productively.The earlier claim says that “30 to 50%” of a lab’s compute goes to inference. The later claim says that better monetization and rising compute costs could make it harder for competitors to obtain compute. These address different questions, and the later statement does not revise the earlier estimate.
• NOT COMPARABLE · 98%17 Oct → 3 AugThen The speaker says the peak cost of an AI session is relatively low and that AI costs may be amortized into the training run, depending on inference scaling.Now The speaker is uncertain whether AI lab margins or compute prices will rise to resolve the gap between revenue and compute growth.The earlier statement says serving another model instance is relatively inexpensive and that AI costs may be amortized into training. The later statement—“It’s not clear to me which world we end up in”—expresses uncertainty about whether lab margins or compute prices will rise. These address related but different economic questions, and the later claim does not directly revise the earlier position.
• NOT COMPARABLE · 94%17 Oct → 3 AugThen The speaker says population growth has driven economic growth over the last 50 years, while population and growth in frontier countries have stagnated; he contrasts this with earlier periods when population was not rapidly increasing.Now The speaker says the Simon-Ehrlich bet is commonly presented as showing that Ehrlich's Malthusian worldview was wrong because he underestimated how markets and human ingenuity can conserve scarce resources.The earlier claim concerns population growth as a driver of economic growth and stagnation in frontier countries. The later claim describes how the Simon–Ehrlich bet is commonly interpreted—its wording says it is “supposed to illustrate” that Ehrlich’s worldview was wrong—and the surrounding turn says the analogy may be wrong. These are not clearly opposing positions on the same question, and the later passage does not establish the speaker’s own view of population-driven growth.
• NOT COMPARABLE · 99%3 Apr → 3 AugThen If GPUs were necessary to create a virtual cell, curing cancer in the 1960s would have required upgrading the entire economy.Now The speaker attributes about 1.8x annual compute growth to AI absorbing wafer capacity previously allocated to smartphones and PCs.The earlier statement is conditional—“If GPUs were necessary”—and concerns whether curing cancer in the 1960s would have required upgrading the entire economy. The later statement concerns AI absorbing wafer capacity and contributing to compute growth. They address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 97%26 Jun → 3 AugThen Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.Now AI labs can charge a larger premium for models that use scarce compute more efficiently.The earlier claim reports what advocates say: the sample-efficiency disadvantage applies only during training and is amortized across sessions. The later claim says labs can charge more for models that use scarce compute efficiently. These address different points and do not establish that the speaker changed their own position.
• NOT COMPARABLE · 99%17 Apr → 3 AugThen Some companies with substantially less compute have beaten incumbents with more compute by having a more coherent vision and more concentrated research effort.Now A model that produces the same result with less compute effectively expands available compute and increases compute's value.The earlier statement says that companies with less compute can beat incumbents through a more coherent vision and concentrated research effort. The later statement says that models getting the same result with less compute effectively expand available compute and increase its value. These address different questions—competitive research strategy versus the effect of compute efficiency—so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 3 AugThen Using AI to find ways to improve TSMC's manufacturing process, efficiency, and yield would count as R&D.Now The speaker expects AI's share of TSMC's leading-edge N3 wafer capacity to rise from 60% to 86% and probably hit a limit by the end of next year.The earlier claim says that using AI to improve TSMC’s manufacturing process, efficiency, and yield would count as “R&D.” The later claim predicts that AI’s share of leading-edge wafer capacity will rise and “hit a limit.” These address different questions—how to classify AI-driven process improvements versus how much wafer capacity AI demand may use—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%26 Jun → 3 AugThen Approximately 30–50% of an AI lab’s compute is used for inference.Now If standard economics is correct, the marginal value of labor and consequently compute should remain extremely high.The earlier claim says that “30 to 50%” of a lab’s compute is used for inference; the later claim says that, “if standard economics is correct,” the marginal value of labor and compute should remain high. These concern different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen The speaker finds Gemini 2.5 Pro’s reasoning traces interesting to read while it processes complex requests and tasks.Now Mercury provides core business functions in one place, which the speaker says it does extremely well.The earlier claim is about finding Gemini 2.5 Pro’s reasoning traces interesting; the later claim praises Mercury’s business functions. They concern different products and questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen The speaker considers Google Gemini 2.5 Pro the best model they have tried for their post-production tools.Now Mercury provides core business functions in one place, which the speaker says it does extremely well.The claims address unrelated topics: the earlier statement says Gemini 2.5 Pro is the best model the speaker tried for post-production tools, while the later statement praises Mercury for providing core business functions in one place. Neither claim changes or comments on the other position.
• NOT COMPARABLE · 98%17 Oct → 26 JunThen The speaker says that LLMs and VLMs provide basic representations and generalization without requiring those foundational representation problems to be solved separately.Now The speaker says that adding enough compute to large language models caused fundamental natural-language-processing research problems to collapse.The earlier claim concerns LLMs/VLMs supplying representations and generalization that reduce foundational work in self-driving deployment. The later claim concerns compute causing “fundamental research problems in natural language processing” to collapse. These address different underlying questions, so the later statement does not establish a change in the earlier position. The surrounding context also does not clearly identify who said the later quoted sentence.
• NOT COMPARABLE · 97%17 Apr → 26 JunThen Humans and other animals differ greatly in their ability to survive in unfamiliar environments, partly because of differences in knowledge.Now Human continual learning primarily consolidates useful intuitions and big-picture knowledge rather than preserving every observation for immediate recall.The earlier claim concerns how humans differ from other animals in surviving unfamiliar environments, citing differences in knowledge. The later claim concerns how human continual learning works—“chiseling the right intuitions and big picture knowledge” rather than retaining every observation for recall. These claims address different questions, and the later one does not revise the earlier position.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen The speaker says they used Gemini 2.5 Pro to write code for their post-production scripts.Now Mercury provides core business functions in one place, which the speaker says it does extremely well.The earlier statement is about using Gemini 2.5 Pro to write post-production scripts; the later statement praises Mercury's business functions. These are about different products and questions, so the claims do not establish a change in the speaker's position.
• NOT COMPARABLE · 97%17 Apr → 26 JunThen ChatGPT illustrates that post-training using about 1% additional compute can make a model pretrained on internet data highly competent at chatbot interaction.Now One reason for slower AI computer-use progress is that models receive much less high-quality multimodal data during pretraining.The earlier statement is about post-training making a pretrained model competent at chatbot interaction; the later statement explains slower progress in computer use by pointing to less high-quality multimodal pretraining data. They address different capabilities and questions, so the later claim does not revise or contradict the earlier one.
• NOT COMPARABLE · 98%17 Apr → 26 JunThen Reinforcement learning currently accounts for only a small fraction of the total compute used to train the discussed models.Now The speaker says that more reinforcement-learning training has improved models’ in-session intelligence, generality, and sample efficiency.The passages appear to be spoken by different people: the earlier quote continues Dwarkesh Patel’s turn, while the later quote appears to be Tamay Besiroglu’s turn. That makes it unclear whether these are claims by the same person. In any case, they address different points: the earlier says RL is “a small fraction of total compute,” while the later says more RL training has improved models’ capabilities; neither statement by itself contradicts the other.
• NOT COMPARABLE · 97%17 Oct → 26 JunThen The speaker says the peak cost of an AI session is relatively low and that AI costs may be amortized into the training run, depending on inference scaling.Now Advocates of current AI training argue that its sample-efficiency disadvantage applies only during training, whose one-time cost is amortized across billions of sessions.The earlier claim is the speaker's own point that deployment costs are relatively favorable and may be amortized into training. The later claim is explicitly attributed to advocates of the current training paradigm: they argue that training is a one-time cost amortized across many sessions. That reported argument does not establish the speaker's own position or show a change of mind.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen Preparing software for enterprise customers is difficult because features such as single sign-on, role-based access controls, and comprehensive audit logs are complex and tedious to build and prone to bugs and edge cases.Now Deployment reveals organization-specific context, users’ purposes, and the model’s real-world failure patterns.The earlier claim says enterprise features such as SSO and audit logs are “complex and tedious to build” and prone to bugs. The later claim says deployment reveals useful information about organizations, users’ purposes, and model failures. These address different questions—implementation difficulty versus the value of deployment feedback—so the later statement does not show a changed position on the earlier one.
• NOT COMPARABLE · 96%17 Apr → 26 JunThen The speaker describes a view that models are not trained to find certain connections, but that an RL environment incentivizing this behavior might require relatively little additional compute.Now The speaker likens training AI only in RL environments to educating a brilliant student with classroom case studies but no real internship.The earlier passage reports another view: models lack certain connections because they are not trained for them, and additional RL incentives might help with relatively little compute. The later passage offers an analogy about RL training as classroom case studies without a real internship. These are related observations, but they do not establish opposing positions on the same question or show that the speaker changed their mind.
• NOT COMPARABLE · 95%17 Oct → 26 JunThen If the current AI paradigm is on a path to AGI, the progression is not expected to consist of gradually automating small tasks across knowledge-work occupations.Now The labs’ research bet is that training AIs on millions of verifiable tasks across thousands of diverse reinforcement-learning environments will produce AGI capable of making progress on open-ended tasks for weeks despite errors and ambiguity.The earlier statement asserts that, if the current paradigm is on a path to AGI, progress will not look like gradually automating small tasks across knowledge-work jobs. The later statement describes what “the labs” think their training approach will achieve; it does not establish the speaker’s own position on that earlier claim, nor directly address how automation will progress.
• NOT COMPARABLE · 88%3 Apr → 26 JunThen The speaker argues that language models should be judged more negatively for failing to make discoveries of this kind, since humans have demonstrated the ability to do so.Now The speaker says that adding enough compute to large language models caused fundamental natural-language-processing research problems to collapse.The earlier claim says language models should be judged more negatively for failing to make discoveries that humans can make. The later claim says that enough compute caused fundamental NLP research problems to “collapse.” That may express optimism about scaling, but it does not establish that the speaker changed their view about LLMs’ ability to make discoveries or how they should be judged.
• NOT COMPARABLE · 98%17 Apr → 26 JunThen Humans differ from other animals in their ability to formulate and execute long-term plans; other animals often rely on instinct or evolutionarily familiar environments.Now Human continual learning primarily consolidates useful intuitions and big-picture knowledge rather than preserving every observation for immediate recall.The earlier claim concerns how humans differ from animals in long-term planning and adapting to unfamiliar environments. The later claim concerns what human continual learning preserves—“useful intuitions and big-picture knowledge” rather than every observation. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen Reinforcement learning currently accounts for only a small fraction of the total compute used to train the discussed models.Now Because on-the-job data is scarce, AI models need sample-efficient learning to benefit from it.The earlier statement is about how much compute reinforcement learning uses to train models (“a small fraction of total compute”). The later statement is about the scarcity of data available to models while working and the need for sample-efficient learning. They address different resources and questions, so the later claim does not revise or contradict the earlier one.
• NOT COMPARABLE · 99%3 Apr → 26 JunThen If cancer treatment is bottlenecked by creating a virtual cell, a million superintelligences in the 1960s might still have been unable to cure cancer without first developing the computing technology needed to simulate that cell.Now If an AI can build a useful simulation of reality for rehearsal and strategy testing, dreaming could provide orders of magnitude more simulated samples per unit of wall-clock time.These claims address different questions. The earlier claim says that curing cancer could have been impossible without the computing technology needed to simulate a virtual cell. The later claim says AI could use simulations for rehearsal and strategy testing to obtain more simulated experience. The later statement does not revise the earlier claim about cancer treatment or virtual-cell bottlenecks.
• NOT COMPARABLE · 98%17 Apr → 26 JunThen The speaker says podcast post-production requires understanding delivery, context, and other information unavailable from a text-only transcript.Now Perfectly reproducing a session transcript is a poor learning target because job improvement depends on more than recalling every event.These claims address different questions. The earlier claim says podcast post-production needs delivery and context that a text-only transcript lacks. The later claim says perfect transcript recall is a poor learning target for improving at a job. Neither claim affirms or denies the other.
• NOT COMPARABLE · 98%17 Apr → 26 JunThen WorkOS says it has helped Vercel, Plaid, Vanta, OpenAI, and hundreds of other companies become enterprise-ready by providing APIs for common enterprise features.Now Cloning common applications and websites, including Slack and Gmail, can provide environments for computer-use training.The earlier statement describes WorkOS helping companies become enterprise-ready with APIs for common features; the later says cloning applications can provide environments for computer-use training. These are different topics, and neither claim establishes a changed position on the same question.
• NOT COMPARABLE · 97%17 Apr → 26 JunThen Next-token prediction training does not incentivize models to discover the connections being discussed.Now A naive SFT method for continual learning trains the base model to predict all tokens observed in a session.The earlier claim says next-token prediction does not incentivize models to discover particular connections. The later claim describes a naive SFT method as predicting all tokens observed during a session; it does not say whether that objective incentivizes discovery. These address different questions, so the later statement does not establish a change in position.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen The speaker says Gemini’s multimodal capabilities allow them to input a four-hour raw audio file directly.Now Mercury provides core business functions in one place, which the speaker says it does extremely well.These statements concern different products and capabilities. The earlier claim is that Gemini can accept a four-hour raw audio file because of its multimodal capabilities; the later claim is that Mercury provides core business functions in one place. Neither statement takes a position on the other’s claim, so there is no evidence of a change of mind.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen The speaker says Gemini can generate high-quality transcripts and identify useful clip snippets from raw audio.Now Perfectly reproducing a session transcript is a poor learning target because job improvement depends on more than recalling every event.The earlier claim is about Gemini’s ability to transcribe audio and identify clip snippets; the later claim is about whether perfectly reproducing a work-session transcript is a good learning target. These address different questions, so the later statement does not revise or contradict the earlier one.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen Many B2B software companies begin by building self-serve, consumer-grade products.Now Cloning common applications and websites, including Slack and Gmail, can provide environments for computer-use training.The earlier claim says many B2B software companies begin with self-serve, consumer-grade products. The later claim says clones of common apps can provide environments for computer-use training. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%3 Apr → 26 JunThen The speaker found that current AI models can produce larger productivity improvements in unfamiliar domains, where their assistance is more novel than autocomplete, surprising.Now In the speaker’s scenario, deployed experience across the economy eventually becomes the main source of AI improvement, surpassing pre-release training.The earlier statement is an observation about current models: their productivity benefit can be larger when they make a “novel contribution” than when acting like autocomplete. The later statement concerns a future continual-learning scenario, saying deployed experience will become the “main way” AIs improve. These address different questions, so the later claim does not show a changed view about the earlier productivity observation.
• NOT COMPARABLE · 98%17 Apr → 26 JunThen Preparing software for enterprise customers is difficult because features such as single sign-on, role-based access controls, and comprehensive audit logs are complex and tedious to build and prone to bugs and edge cases.Now In many real-world tasks, verification may require months or years of action, and outcomes cannot be efficiently isolated by running thousands of slightly varied parallel rollouts.The earlier claim says enterprise-readiness features are complex, tedious, and prone to bugs. The later claim concerns how difficult it is to verify outcomes in real-world tasks over long periods. These are different topics, and neither statement establishes a changed position on the same question.
• NOT COMPARABLE · 99%3 Apr → 26 JunThen The transition from small rockets to very large rockets took two decades of varied, unexpected experimentation, even with highly capable people.Now A couple more years of progress might produce transformer context windows that feel infinitely large.The earlier statement describes the historical transition from small to large rockets as taking two decades of experimentation. The later statement speculates that transformer context windows might feel infinitely large after a couple more years of progress. These concern different technologies and questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%3 Apr → 26 JunThen Daniel Kokotajlo originally intended his forecast to cover developments beyond 2026, but removed its final chapter when he reached 2027 because AI developments and the accelerating automation loop made the situation confusing and highly uncertain.Now The speaker offers a speculative scenario for continual learning by 2027 or 2028.The earlier statement describes why Daniel removed a chapter from a forecast: he says the situation was “so confusing” and there was “so much uncertainty.” The later claim is explicitly framed as “one scenario” for what continual learning might look like in 2027 or 2028. It does not establish a different asserted position on the same question.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen Humans sometimes find the kinds of logical connections under discussion, and such examples are not uncommon.Now Humans compress information when learning, and this compression supports generalization and reasoning.The earlier claim says humans sometimes find the kinds of logical connections under discussion, while the later claim says learning involves compression that supports generalization and reasoning. These address different questions; the later statement neither revises nor denies the earlier one.
• NOT COMPARABLE · 98%17 Apr → 26 JunThen The speaker considers Dario's forecasts poorly calibrated and cites a 2018 expectation that something would happen by 2020.Now The speaker offers a speculative scenario for continual learning by 2027 or 2028.The earlier claim says Dario’s forecasts were “not been that well calibrated,” citing an expectation from 2018 about 2020. The later claim introduces “one scenario” for what continual learning might look like by 2027 or 2028. That speculative scenario does not state a position about Dario’s forecasting accuracy, so the claims do not establish a change of mind on the same question.
• NOT COMPARABLE · 99%17 Apr → 26 JunThen The speaker says Gemini can generate high-quality transcripts and identify useful clip snippets from raw audio.Now Mercury provides core business functions in one place, which the speaker says it does extremely well.The earlier claim is about Gemini producing transcripts and identifying clips from audio; the later claim is about Mercury providing business functions in one place. They concern different products and capabilities, so neither claim establishes a changed position on the same question.
• NOT COMPARABLE · 95%17 Apr → 26 JunThen More efficient AI copies increase the number of AI researchers and enable more parallel searches for optimizations.Now If AIs become capable of building highly faithful application clones, computer-use progress will become faster than it is now.The earlier claim says that more efficient AI copies increase the number of AI researchers and allow more parallel searches for optimizations. The later claim is a conditional prediction about computer-use progress: if AIs can build highly faithful application clones, progress will become faster. These concern different mechanisms and outcomes, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%17 Apr → 26 JunThen AI systems conducting software R&D can discover ways to run copies of themselves more efficiently.Now If AIs become capable of building highly faithful application clones, computer-use progress will become faster than it is now.The claims concern different mechanisms and outcomes. The earlier claim says AI systems doing software R&D can find ways to run copies more efficiently; the later claim says that if AIs can build highly faithful application clones, computer-use progress will become faster. The later statement neither reverses nor clearly updates the earlier one.
• NOT COMPARABLE · 98%17 Apr → 26 JunThen Preparing software for enterprise customers is difficult because features such as single sign-on, role-based access controls, and comprehensive audit logs are complex and tedious to build and prone to bugs and edge cases.Now Currently, creating cloned application and website environments for computer-use training is labor-intensive and does not scale.The earlier claim concerns the difficulty of making software enterprise-ready, citing SSO, access controls, and audit logs. The later claim concerns the labor and scalability of creating cloned applications and websites for computer-use training. These are different tasks, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%17 Apr → 26 JunThen Computer use or long-term computer agency might, like reasoning, be unlocked with a relatively small fraction of total compute.Now Computer-use tasks can be verified clearly.The earlier claim is a tentative prediction that computer use or long-term computer agency might be unlocked with relatively little compute, by analogy with reasoning. The later claim says computer-use tasks are clearly verifiable. Verifiability is a different property from how much compute is needed or whether agency will be unlocked, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 91%17 Apr → 17 OctThen The observed growth in AI task length suggests that AI is getting closer to intelligence understood as long-term coherence in task execution.Now If the current AI paradigm is on a path to AGI, the progression is not expected to consist of gradually automating small tasks across knowledge-work occupations.The earlier claim says the observed growth in AI task length suggests AI is getting closer to intelligence as “long-term coherence in task execution.” The later claim is about how progress toward AGI would unfold across knowledge-work occupations, saying it is “very much not” a gradual process of removing small tasks from each occupation. These claims address different aspects of progress and can both be true; the later statement does not clearly retract the earlier one.
• NOT COMPARABLE · 94%17 Apr → 17 OctThen The speaker describes a view that models are not trained to find certain connections, but that an RL environment incentivizing this behavior might require relatively little additional compute.Now The speaker says that LLMs and VLMs provide basic representations and generalization without requiring those foundational representation problems to be solved separately.The earlier passage presents a view about models needing RL or scaffolding to find certain connections, using the words “their view.” The later passage is Karpathy’s own claim about LLMs/VLMs getting basic representations and generalization “for free.” These are not clearly opposing positions on the same question, and the earlier passage does not clearly establish his own belief.
• NOT COMPARABLE · 98%3 Apr → 17 OctThen The speaker argues that China's growth miracle could not have occurred without China's ability to copy Western technology.Now The speaker says population growth has driven economic growth over the last 50 years, while population and growth in frontier countries have stagnated; he contrasts this with earlier periods when population was not rapidly increasing.The earlier claim says China’s growth miracle depended on its ability to copy Western technology. The later claim discusses population growth as a driver of economic growth over the last 50 years. These address different explanations and do not establish that the speaker changed their view about either claim.
• NOT COMPARABLE · 98%17 Apr → 17 OctThen According to the speaker, all of the most successful and durable software companies of the last decade transitioned to enterprise customers.Now The speaker says that, excluding ordinary chat revenue and considering API revenue, the revenues of the companies being discussed are dominated by coding.The earlier claim concerns successful software companies transitioning to enterprise customers; the later claim concerns API revenues being “dominated by coding.” These are different questions, and neither statement establishes a changed position on the other.
• NOT COMPARABLE · 99%3 Apr → 17 OctThen The speaker considers UBI preferable to a bespoke social program that continues providing the same dialysis machine in 2050 despite the availability of artificial superintelligence.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier claim concerns whether UBI is preferable to a bespoke social program, while the later claim concerns AI accelerating economic progress without a single decisive invention. They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 17 OctThen Tesla FSD benefits from driving data collected across millions of miles at the periphery and from improvements developed using that data.Now The speaker argues that serving another AI model instance costs far less than building another car, making the economics of wider AI deployment more favorable than deployment of physical systems such as self-driving cars.The earlier claim says Tesla FSD can benefit from driving data gathered across many miles and improvements based on that data. The later claim compares the marginal cost of serving another AI model instance with building another car, and says wider AI deployment has more favorable economics. These address different questions—data-driven improvement versus the cost of deploying additional instances—and do not establish a changed position.
• NOT COMPARABLE · 96%3 Apr → 17 OctThen The speaker says progress at every layer of AI development has been substantially more difficult than the strongest optimists anticipated.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier claim says AI development has been harder than optimists expected (“every single layer ... has been much harder”). The later claim predicts a faster-progress economic regime (“you did move the economy to a regime where the progress was much faster”). These address different questions—difficulty of building AI versus its expected economic effects—so they do not establish a change of mind.
• NOT COMPARABLE · 96%3 Apr → 17 OctThen The speaker argues that China's growth miracle could not have occurred without China's ability to copy Western technology.Now The speaker argues that the Industrial Revolution was not caused by a single magical key invention; examining 1770 or 1870 would not reveal one such invention.The earlier claim is that China’s growth miracle depended on copying Western technology. The later claim is that the Industrial Revolution was not driven by a single magical invention. These address different historical questions, and the later statement does not deny the earlier one.
• NOT COMPARABLE · 98%17 Apr → 17 OctThen The speaker considers highly elastic demand, alongside transformative AI, probably the most underrated tangible factor for understanding the future AI society.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier claim identifies “highly elastic demand” as an underrated factor in understanding future AI society. The later claim is about AI accelerating economic progress without a single decisive invention. These concern different aspects of AI’s effects, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%3 Apr → 17 OctThen SPEAKER_00 says that small changes to scenario parameters can lead to a very different resulting world, analogous to a hash function's sensitivity to its input.Now The speaker thinks a return to exponential population growth could produce hyper-exponential growth in output.The earlier claim says small changes to scenario parameters can yield a very different world (“almost like a hash function”). The later claim concerns a possible relationship between renewed exponential population growth and hyper-exponential output growth. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%17 Apr → 17 OctThen One AI researcher reported that AI models save four to eight hours per week in familiar domains, where they function more like autocomplete.Now The speaker argues that serving another AI model instance costs far less than building another car, making the economics of wider AI deployment more favorable than deployment of physical systems such as self-driving cars.The earlier statement reports an AI researcher’s estimate that AI saves “four to eight hours a week” in familiar domains; it does not establish the speaker’s own position on AI’s overall usefulness or deployment economics. The later statement addresses a different question: the cost of serving additional model instances and the economics of wider deployment. These claims do not show a change of mind.
• NOT COMPARABLE · 96%3 Apr → 17 OctThen A possible policy response to automation is to protect existing jobs instead of distributing the wealth generated by automation.Now The interviewer suggests that when automation covers 99% of a job, the remaining human 1% may become highly valuable if it bottlenecks the overall process.The earlier claim presents protecting existing jobs as a possible policy response to automation. The later passage concerns whether the remaining human portion of a highly automated job could become a valuable bottleneck; it does not establish a changed view about that policy response. The later claim is also attributed to the interviewer in the extraction, so it is unclear whether it reflects the same person's position.
• NOT COMPARABLE · 99%3 Apr → 17 OctThen The speaker argues that China's growth miracle could not have occurred without China's ability to copy Western technology.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier statement concerns China's growth miracle and the role of copying Western technology. The later statement concerns the pace and nature of AI-driven economic progress. They do not address the same underlying question, so there is no basis here to determine whether the speaker changed their mind.
• NOT COMPARABLE · 96%17 Apr → 17 OctThen The interviewer suggests that in an AI-driven economy, many humans might contribute zero or negative value to output.Now The speaker says consultants and accountants do not appear to be receiving large productivity improvements from AI.The earlier passage is framed as the interviewer suggesting that humans might contribute “zero or even minus” in an AI-driven economy, so it is not clear that this was the speaker’s own position. The later claim says consultants and accountants do not appear to be getting “huge productivity improvements” from AI; that observation does not by itself deny the earlier suggestion.
• NOT COMPARABLE · 98%3 Apr → 17 OctThen The speaker says call-center workers have not yet been fired in the context of AI's economic impact.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier statement says call-center workers “haven’t been fired yet,” describing an observed economic impact at that time. The later statement says AI may move the economy into a regime of “much faster” progress, while not arriving through one sudden invention. These address different aspects of AI’s economic impact, and the later claim does not say whether call-center workers have been or will be displaced.
• NOT COMPARABLE · 99%3 Apr → 17 OctThen SPEAKER_00 says blogging can help people build networks and develop their ideas.Now The speaker says economic growth often results from people generating ideas and others executing them to create valuable output.The earlier claim is that blogging can help people build networks and develop ideas; the later claim is that economic growth often comes from people generating ideas and others executing them. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 96%3 Apr → 17 OctThen Blogging can help people build networks and develop ideas even when they do not succeed financially.Now The speaker says economic growth often results from people generating ideas and others executing them to create valuable output.The earlier statement concerns blogging as a career path and its benefits even without financial success: it “helps you build up a network” and “develop your ideas.” The later statement describes a general account of economic growth, involving people who “come up with ideas” and others who “execute those ideas.” These address different questions, and the later claim neither affirms nor denies the earlier view about blogging.
• NOT COMPARABLE · 98%17 Apr → 17 OctThen The speaker expects many aspects of the future to differ from the present and is comfortable with those differences.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier claim is about being comfortable with many future differences from the present. The later claim predicts that AI will accelerate economic progress and that this shift will not come from one sudden invention. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%3 Apr → 17 OctThen The speaker considers Robin Hanson to have been too pessimistic about AI progress or deployment.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier claim says Robin Hanson was “too pessimistic,” while the later claim predicts a faster-progress economic regime from AI without a single decisive invention. These address different propositions; the later statement does not establish whether the speaker’s view of Hanson changed.
• NOT COMPARABLE · 99%17 Apr → 17 OctThen SPEAKER_00 says it was possible in principle for society to have remained a slave society through the present day.Now The speaker argues that the Industrial Revolution was not caused by a single magical key invention; examining 1770 or 1870 would not reveal one such invention.The earlier claim concerns whether society could have remained a slave society (“could ... have been a slave society to this day”). The later claim concerns whether the Industrial Revolution depended on a single key invention (“not ... some key invention”). These address different questions, so they provide no evidence of a change in position.
• NOT COMPARABLE · 91%3 Apr → 17 OctThen The speaker says call-center workers have not yet been fired in the context of AI's economic impact.Now The speaker says consultants and accountants do not appear to be receiving large productivity improvements from AI.The earlier claim is that call-center workers “haven’t been fired yet”; the later claim is that consultants and accountants do not seem to be getting “huge productivity improvements.” These are different indicators and worker groups, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%3 Apr → 17 OctThen One proposed response to automation is to protect jobs instead of distributing the wealth automation creates.Now The interviewer suggests that when automation covers 99% of a job, the remaining human 1% may become highly valuable if it bottlenecks the overall process.The earlier statement describes a proposed response to automation: protecting jobs rather than distributing the wealth automation creates. The later statement is attributed to the interviewer, not the person whose views are being compared, and raises a different point about the value of the remaining human work when a job is mostly automated. It does not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%17 Apr → 17 OctThen SPEAKER_00 says it was possible in principle for society to have remained a slave society through the present day.Now The speaker says the Industrial Revolution shifted the economy into a much faster-progress regime, with exponential growth increasing tenfold.The earlier claim concerns whether society could have remained a slave society (“could ... have been a slave society to this day”). The later claim concerns the Industrial Revolution accelerating economic progress (“the progress was much faster”). These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%17 Apr → 17 OctThen Lower costs for running or developing AI copies enable more parallel experiments, including GPT-4-scale training runs at much lower cost than in 2023 or 2024.Now The speaker says the peak cost of an AI session is relatively low and that AI costs may be amortized into the training run, depending on inference scaling.The earlier claim concerns lower costs of running or developing AI copies enabling more parallel experiments, including cheaper GPT-4-scale training runs. The later claim concerns the relatively low peak cost of serving an AI session and possibly amortizing those costs into training. These address different cost contexts, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%17 Apr → 17 OctThen Lower costs for running or developing AI copies enable more parallel experiments, including GPT-4-scale training runs at much lower cost than in 2023 or 2024.Now The speaker argues that serving another AI model instance costs far less than building another car, making the economics of wider AI deployment more favorable than deployment of physical systems such as self-driving cars.The earlier claim concerns how lower training or development costs enable more parallel AI experiments. The later claim concerns the cost of serving additional model instances compared with building cars. These are different economic questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%17 Apr → 17 OctThen The interviewer suggests that in an AI-driven economy, many humans might contribute zero or negative value to output.Now The interviewer speculates that radiologists’ wages may have increased because they provide scarce human work that remains a bottleneck to wider automation.The earlier passage concerns whether humans may contribute “zero or even minus” value in an AI-driven economy. The later passage concerns whether radiologists’ wages rise when scarce human work remains a bottleneck to automation. These address different questions, and neither establishes a change in the speaker’s position on the other.
• NOT COMPARABLE · 88%3 Apr → 17 OctThen The speaker found that current AI models can produce larger productivity improvements in unfamiliar domains, where their assistance is more novel than autocomplete, surprising.Now The speaker says consultants and accountants do not appear to be receiving large productivity improvements from AI.The earlier statement says AI help was surprisingly productive in less familiar domains, where it made a novel contribution. The later statement says consultants and accountants do not seem to be getting huge productivity improvements. These concern different levels of specificity and potentially different occupations or settings; the later claim does not establish that the earlier observation was withdrawn or reversed.
• NOT COMPARABLE · 91%17 Apr → 17 OctThen The interviewer suggests that in an AI-driven economy, many humans might contribute zero or negative value to output.Now The interviewer suggests that when automation covers 99% of a job, the remaining human 1% may become highly valuable if it bottlenecks the overall process.The earlier claim says many humans in an AI-driven economy might contribute zero or negative value to output. The later claim describes a different possibility: a remaining human task can be highly valuable when it bottlenecks an automated job. These claims concern different conditions and can both be true; neither clearly retracts the other.
• NOT COMPARABLE · 99%3 Apr → 17 OctThen The speaker considers universal basic income a better approach than a bespoke social program that would continue procuring the same dialysis machine in 2050 even if artificial superintelligence existed.Now The speaker expects AI to bring a comparable shift toward faster economic progress, but not through one sudden moment when a crucial invention is made.The earlier claim concerns whether UBI is preferable to a bespoke social program that might keep procuring the same dialysis machine; the later claim concerns AI accelerating economic progress without one sudden decisive invention. They address different questions, so the later statement does not establish a change in the earlier position. The surrounding turns also identify different speakers (Scott Alexander and Andrej Karpathy), making it unclear that these are claims by the same person.
• NOT COMPARABLE · 97%17 Apr → 17 OctThen The speaker suggests AI systems may be more compatible and easier to coordinate with one another than with humans, because they are copies of one another and may face lower transaction costs in interaction.Now The speaker argues that serving another AI model instance costs far less than building another car, making the economics of wider AI deployment more favorable than deployment of physical systems such as self-driving cars.The earlier claim concerns whether AI systems may coordinate more easily with one another than with humans, citing that they are “copies of each other” and may have “lower transaction costs.” The later claim concerns the economics of serving additional model instances, saying deployment is “much more favorable” than building another car. These claims address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%17 Apr → 17 OctThen The speaker says current AI systems are not yet capable enough to be onboarded as full workers, but expects their becoming capable enough to be a major unlock for the roles the speaker hires for.Now The speaker says that AI systems presented as general-purpose and capable of any knowledge work are, in practice, overwhelmingly being used for coding.The earlier claim is about capability: current systems are “aren't smart enough to be onboarded as full workers,” with becoming capable enough expected to be “such an unlock.” The later claim is about actual usage, saying AI is “overwhelmingly doing only coding.” Usage patterns do not establish whether the systems are capable enough for the roles the speaker hires for, so the statements do not show a change of mind.
• NOT COMPARABLE · 96%17 Apr → 17 OctThen The speaker considers highly elastic demand, alongside transformative AI, probably the most underrated tangible factor for understanding the future AI society.Now The speaker says that the speed of AI deployment and its early-stage value may be the most important question in the world right now.The earlier claim concerns highly elastic demand as an underrated factor in understanding future AI society; the later claim concerns how quickly AI is deployed and how valuable it is early on. These are related but not the same question, and the later statement does not retract the earlier one.
• NOT COMPARABLE · 99%3 Apr → 17 OctThen SPEAKER_00 argues that highly successful work can attract rapid attention from relevant audiences without years of prior fan-base building.Now The speaker says economic growth often results from people generating ideas and others executing them to create valuable output.The earlier claim concerns whether successful work can quickly attract an audience without years of fan-base building. The later claim concerns how ideas and execution contribute to economic growth. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%17 Apr → 17 OctThen The speaker imagines a future company spending $100 million annually on inference for replicated Jensen Huang-like AI systems that perform organizational tasks and monitor the company for consistency with a central vision.Now The speaker says the peak cost of an AI session is relatively low and that AI costs may be amortized into the training run, depending on inference scaling.The earlier statement imagines a future company using many Jensen-like AI copies for organizational work; the later statement says the cost of serving another model instance is relatively low and deployment economics are favorable. These concern different questions, and the earlier passage is explicitly hypothetical, so they do not establish a change in the speaker’s position.