Tim Scarfe

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.

• REFINES · 96%24 Jun → 21 SeptThen The speaker says future AGI might use LLMs because they are good at distributional learning.Now SPEAKER_02 says language models used with RLV contain substantial base knowledge.The earlier claim says future AGI might use LLMs because they are good at “distributional learning.” The later claim says language models used with RLV have “a lot of base knowledge.” These claims are compatible; the later statement adds a reason or capability rather than reversing the earlier position.
• REFINES · 96%24 Jun → 21 SeptThen The speaker says neurosymbolic systems make it possible to state explicit constraints, something very difficult with pure LLMs.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier claim says neurosymbolic systems allow explicit constraints that are “very difficult in pure LLMs,” and describes exploring that approach for safety-critical uses. Later, the speaker says “constraints are incredibly valuable,” defining them broadly to include knowledge. This broadens the earlier position rather than reversing it; the later statement is also clearly presented as the speaker’s own view.
• REFINES · 98%24 Jun → 21 SeptThen The speaker says neurosymbolic systems make it possible to state explicit constraints, something very difficult with pure LLMs.Now SPEAKER_02 describes constraints as operating at multiple levels of abstraction, including physical laws in the world and existing knowledge in minds.The earlier claim says neurosymbolic systems allow explicit constraints that are “very difficult in pure LLMs,” and that this avenue is worth exploring for safety-critical uses. The later claim broadens the idea, calling constraints “incredibly valuable” and describing them as knowledge at multiple levels of abstraction. This develops the earlier position rather than reversing it.
• REFINES · 78%20 Aug → 21 SeptThen The interviewer describes a prior conception of intelligence as traversing the combinational closure of core knowledge and abstractions in the brain.Now The interviewer believes that how a search space is reached, constrained, and shaped is critical to making the approach work.The earlier passage describes a prior conception of intelligence as “traversing the combinational closure” of core knowledge and abstractions. The later claim that shaping and constraining the search space is “critical” adds a more specific point about how the process works; it does not deny the earlier conception, and the surrounding earlier passage also frames that conception as incomplete rather than as the speaker’s current view.
• REFINES · 86%24 Jun → 8 SeptThen The speaker characterizes the chess system's knowledge as task-specific rather than general chess knowledge.Now The speaker argues that apparent general capability is partly an illusion: models can be shaped into specialized experts in any domain, but that expertise is latent rather than explicit.The earlier claim says the chess system had “purpose-built” knowledge that “wasn't general knowledge of chess.” The later claim says models can be shaped into experts, while that expertise is “a latent capability rather than an explicit capability.” These are compatible: specialized expertise can be latent in a model without being explicit, general chess knowledge in the system being discussed.
• REFINES · 89%24 Jun → 8 SeptThen The speaker believes it is probably technically achievable to automate the full AI research-and-development process, though it may require technology beyond what is currently available.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says automating the full AI research-and-development process is “probably” technically achievable, while noting it may require technology beyond what is currently available. The later claim says an automated AI scientist is possible “in the near future.” That adds a more specific timeframe and formulation, but does not reject the earlier position.
• REFINES · 78%24 Jun → 8 SeptThen The speaker predicts that AI systems will continue improving regularly and provide increasing value.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says AI systems “will continue to improve regularly” and provide “more and more value”; the later claim says an “automated AI scientist” may be possible “in the near future.” The latter is a more specific, ambitious possibility consistent with continued improvement, not a clear reversal of the earlier position.
• REFINES · 90%24 Jun → 8 SeptThen The speaker expects AI systems near AGI to perform about as well as, or at least approach, domain-specific systems.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says systems near AGI should be “basically the equivalent of their domain-specific counterparts, or at least be close.” The later claim says expert-guided AI can produce “magical” results in well-specified domains, while “there’s still a bit of a space of ambiguity.” That adds a limitation and context; it does not deny the earlier expectation of comparable performance.
• REFINES · 91%24 Jun → 8 SeptThen The speaker allows that discovery of new ideas might be automated.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says automated discovery of new ideas “could be” possible; the later claim says an “automated AI scientist” could be possible “in the near future.” The later statement is more specific and adds a timeframe, but it does not reverse the earlier openness to the possibility.
• REFINES · 88%24 Jun → 8 SeptThen The speaker sees little, though possibly some, progress in AI systems' ability to match the flexibility of human expert knowledge.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says there is “relatively little progress” in matching the flexibility of human expert knowledge. The later claim acknowledges that expert-guided AI can produce “magical things” in well-specified domains, while noting that ambiguity remains. That adds a favorable, bounded case rather than reversing the earlier assessment of flexibility.
• REFINES · 94%24 Jun → 8 SeptThen The speaker considers the semantics represented by current models very superficial, describing them as primarily distributions of words.Now The speaker describes neural-network representations as fractured and entangled, somewhat janky, and not fully robust.The earlier claim says current models’ semantics are “very superficial” and “really about distributions of words.” The later claim calls neural-network representations “fractured,” “janky,” and “not completely robust.” Both express a criticism of current representations; the later wording adds detail rather than reversing the earlier position.
• REFINES · 94%24 Jun → 8 SeptThen The speaker argues that human experts can readily solve unfamiliar tasks within their domain because of their broader understanding, whereas AI systems still struggle with them.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says AI systems “still do” poorly on unfamiliar tasks that human experts can solve through broader domain understanding. The later claim says expert-guided AI can produce “magical” results in “well-specified domains,” while ambiguity remains. That adds a condition about when AI performs impressively; it does not retract the earlier claim about difficulty with unfamiliar tasks.
• REFINES · 94%24 Jun → 8 SeptThen The speaker allows that discovery of new ideas might be automated.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says it “could be that we automate the discovery of new ideas.” The later claim imagines a specific form of that possibility: an efficient recursive loop in which the system adapts itself, chooses what to explore, and evolves. Calling that “a different type of technology” adds characterization, but does not deny the earlier possibility.
• REFINES · 94%24 Jun → 8 SeptThen The speaker sees no sound argument that automating the full AI research-and-development process is impossible.Now The speaker believes an automated AI scientist could be possible in the near future.The later claim adds a timeframe and states the possibility more directly, but does not reverse the earlier position. Seeing “no sound argument that this can’t be done” is consistent with later saying an automated AI scientist “is possible in the near future.”
• REFINES · 94%24 Jun → 8 SeptThen The speaker and Jan LeCun are dissatisfied with current AI capabilities described as reasoning because they consider them insufficiently robust.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier statement says current AI reasoning is not “robust enough.” The later statement acknowledges that expert-guided AI can produce “magical things” in “well-specified domains,” while still noting ambiguity. That adds a favorable, condition-specific qualification; it does not retract the earlier concern about robustness.
• REFINES · 91%24 Jun → 8 SeptThen The speaker says AI systems can approximate user requests without reliably fulfilling them as intended.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier claim says systems can “approximate what we ask for” but “don’t really do what we ask them” reliably. The later claim is conditional: with “enough of a specification,” a system “can actually converge towards the solution.” That qualification does not reverse the earlier general concern; it describes a case where convergence is possible.
• REFINES · 88%24 Jun → 8 SeptThen The speaker says AI has made substantial progress in interpolation but has not solved extrapolation, and characterizes much of the progress as interpolation rather than extrapolation.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says AI has made “huge progress” in interpolation but “hasn't solved the extrapolation problem.” The later claim says expert-guided AI can produce “magical” results in well-specified domains while “there's still a bit of a space of ambiguity.” Both describe strong performance in bounded settings alongside unresolved limitations; the later statement does not reverse the earlier position.
• REFINES · 88%24 Jun → 8 SeptThen The speaker says current AI systems have not been good at generating genuinely new ideas.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says current AI has not been good at generating “genuinely new ideas.” The later claim credits models with creating “novel combinations of things that we already know about” to solve tasks, while the surrounding discussion distinguishes this kind of innovation from human-like creativity. That adds a more specific account of what models can do, rather than clearly reversing the earlier claim about genuinely new ideas.
• REFINES · 96%24 Jun → 8 SeptThen The speaker says AI systems can approximate user requests without reliably fulfilling them as intended.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says AI systems “don’t really do what we ask them” reliably. The later claim adds that expert guidance can make them perform impressively in “well-specified domains,” while “there’s still a bit of a space of ambiguity.” This qualifies the earlier concern rather than reversing it.
• REFINES · 84%24 Jun → 8 SeptThen The relevant form of AI recursion involves a system generating fundamentally new ideas previously produced only by humans and carrying out the broader development process.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier statement describes the relevant kind of recursion as a system finding “a fundamentally new idea” and “leveraging” it in a broader development process. The later statement imagines a system “adapting itself,” choosing what is interesting, and “evolving by itself,” which adds detail about how such a loop might work rather than rejecting the earlier position. The later claim is explicitly framed as a possible future (“I can imagine a future”), not as a contrary assertion.
• REFINES · 98%24 Jun → 8 SeptThen Humans have limitations in memory.Now Humans can expand their memory.The earlier claim says humans “have limitations in memory,” while the later says humans “can expand their memory.” These are compatible: expanding memory can mitigate limitations without denying that they exist.
• REFINES · 89%24 Jun → 20 AugThen The speaker says AI systems can approximate user requests without reliably fulfilling them as intended.Now The speaker argues that AI systems follow the instructions they are given; in the example, a basic prompt to do something interesting with video frames would produce nothing interesting.The earlier claim says AI systems “don’t really do what we ask them” reliably, citing failures such as ignoring instructions not to hallucinate. The later claim that a basic “do something interesting” prompt would produce “nothing interesting” concerns whether systems independently generate goals or act with agency; the surrounding context says a specific goal must be supplied by people. That is a distinction about initiative, not a clear reversal of the earlier claim about reliably fulfilling intended instructions.
• REFINES · 84%24 Jun → 20 AugThen The speaker expects an AI system to follow a specified algorithm for a task.Now The speaker says AI can be directed toward specific goals or problems when given a suitable program and domain, but humans supply the framing.The earlier claim says an AI system is expected to “follow” a specified algorithm. The later claim says a system can do a specific thing or pursue a specific goal when given a suitable program, while “the framing always comes from us.” That adds conditions and a distinction between executing a specified task and independently choosing what to do; it does not clearly reverse the earlier position.
• REFINES · 84%24 Jun → 20 AugThen The speaker considers the semantics represented by current AI models superficial and primarily based on word distributions.Now AI models acquire some form of coarse-grained abstraction or internal structure during training.The later statement adds that models acquire “some kind of coarse grained abstraction” or structure during training. That can coexist with the earlier view that their semantics are “very superficial” and “really about distributions of words”; the later claim does not say the representations are deep or no longer distribution-based.
• REFINES · 88%24 Jun → 20 AugThen The speaker argues that human experts can readily solve unfamiliar tasks within their domain because of their broader understanding, whereas AI systems still struggle with them.Now The speaker says AI can be directed toward specific goals or problems when given a suitable program and domain, but humans supply the framing.The earlier claim says AI systems “still do” struggle with unfamiliar tasks that human experts can readily solve. The later claim says AI can solve “a specific problem” when given a specific program, while “the framing always comes from us.” These claims are compatible: the later statement describes performance on a specified task and adds that humans frame it; it does not retract the earlier point about difficulty with unfamiliar tasks.
• REFINES · 88%24 Jun → 20 AugThen The speaker expects an AI system to follow a specified algorithm for a task.Now The speaker argues that AI systems follow the instructions they are given; in the example, a basic prompt to do something interesting with video frames would produce nothing interesting.The earlier claim concerns following a specified algorithm: “we would expect the system to be able to follow it.” The later statement says systems “will basically do what you tell it to do,” while distinguishing that from a vague prompt to “do something interesting,” which would produce “nothing interesting.” These positions are compatible; the later claim adds a distinction between specific instructions and open-ended prompting rather than reversing the earlier position.
• REFINES · 83%24 Jun → 20 AugThen The speaker predicts that AI systems will continue improving regularly and provide increasing value.Now With a memory system and repeated user corrections, an AI system can build an increasingly accurate simulacrum of its user and perform the desired actions more often over time.Both claims describe AI systems improving over time. Earlier, the speaker says they will “continue to improve regularly” and provide “more and more value”; later, he gives a specific mechanism and outcome: memory plus corrections can make a system “increasingly” resemble the user and “increasingly” do the right thing. The later statement elaborates on the earlier prediction rather than reversing it.
• REFINES · 96%24 Jun → 20 AugThen The speaker says AI systems can approximate user requests without reliably fulfilling them as intended.Now With a memory system and repeated user corrections, an AI system can build an increasingly accurate simulacrum of its user and perform the desired actions more often over time.The earlier claim says systems “don’t really do what we ask them” reliably. The later claim describes a way performance can improve: with memory and repeated corrections, the system “increasingly does the right thing.” That adds a condition and a trajectory, but does not deny the earlier limitation.
• CONSISTENT · 88%24 Jun → 26 SeptThen The speaker considers 50% performance an unusual and problematic measure for interpreting the graph.Now Goodhart's Law states that a measure ceases to be a good measure when it becomes a target.The earlier claim says that “50% performance” is a “very weird measure” for interpreting a particular graph. The later statement describes Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure. These claims are compatible, but they address different points; the later statement does not reverse the earlier assessment.
• CONSISTENT · 96%24 Jun → 26 SeptThen The relevant form of AI recursion involves a system generating fundamentally new ideas previously produced only by humans and carrying out the broader development process.Now People often imagine recursive self-improvement as a system repeatedly rewiring its own brain.The earlier claim describes what the speaker considers the relevant form of AI recursion: a system generating fundamentally new ideas and carrying out broader development. The later claim describes what “most people” imagine—rewiring a system’s own brain—not the speaker’s own position. These statements can both be true; the later one does not revise the earlier view.
• CONSISTENT · 94%24 Jun → 21 SeptThen The speaker believes that generalization outside the training distribution remains a major problem for current AI systems.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier claim says generalization outside the training distribution remains a major problem. The later claim says constraints, broadly including knowledge, are valuable in AI systems. These are compatible positions, but the later statement does not say whether constraints solve the generalization problem.
• CONSISTENT · 96%24 Jun → 21 SeptThen The speaker estimates that developing substantially better reasoning, planning, and stable world models will take five to ten years.Now The speaker says vision models and flow models perform sophisticated inference processes.The earlier claim estimates that developing “substantially better reasoning, planning, and stable world models” will take “five to ten years.” The later claim says vision and flow models are already doing “quite sophisticated inferencing processes” and expresses excitement about inference-time computation. These address different capabilities and do not establish that the speaker has reversed the earlier timeline estimate.
• CONSISTENT · 96%24 Jun → 21 SeptThen The speaker says future AGI might use LLMs, which he considers good at distributional learning.Now SPEAKER_02 says language models used with RLV contain substantial base knowledge.The earlier claim says future AGI might use LLMs because they are good at “distributional learning.” The later claim says language models used with RLV have “a lot of base knowledge.” These are compatible observations about LLMs, not opposing positions or evidence of a changed mind.
• CONSISTENT · 99%24 Jun → 21 SeptThen The speaker says the participants' forecasts are closer than people may assume: one concentrates probability at five years out and later, while the other assigns some probability to two or three years out, and both assign some probability to 2045.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier claim concerns how the participants distribute probability across forecast timelines (for example, “all of my probability mass to five years out”). The later claim concerns excitement about “work at inference time” and how much it is discussed. These address different questions, so the later statement does not indicate a change in the earlier position.
• CONSISTENT · 91%24 Jun → 21 SeptThen Because new data is generated every day, researchers can continue to obtain somewhat more data.Now Machine-learning practitioners work with finite data samples and may need to interpolate between observed data points.The earlier claim says researchers can “continue to obtain somewhat more data” because new data is generated daily; it does not say their datasets are unlimited. The later claim says practitioners work with a “finite sample” and may need to interpolate. Both can be true: more data may become available over time while any particular learning task uses a finite sample.
• CONSISTENT · 93%24 Jun → 21 SeptThen The speaker has advocated neurosymbolic AI since at least 2001.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier statement says, “I've always advocated neurosymbolic AI”; the later says, “I'm a big believer that constraints are incredibly valuable,” including knowledge as a kind of constraint. These are compatible positions, not opposing claims about the same question. The later passage also contrasts the speaker’s view with someone else’s preference to “just scale” models, but does not indicate a change from the earlier position.
• CONSISTENT · 93%24 Jun → 21 SeptThen Because new data is generated every day, researchers can continue to obtain somewhat more data.Now Because machine-learning practitioners have finite data and face the curse of dimensionality, they may need to interpolate between data points and smooth and amortize the process to support prediction.The earlier claim says researchers can obtain “somewhat more data” because “new data is generated every day.” The later claim says practitioners have “finite data” and may need to interpolate and smooth for prediction. These are compatible: data can continue to accrue while remaining finite for a particular modeling task or point in time.
• CONSISTENT · 88%24 Jun → 21 SeptThen The speaker argued by 2001 that AI systems need symbols to perform abstract operations over variables.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier claim says AI systems need symbols for “abstract operations over variables.” The later passage says “I’m a big believer that constraints are incredibly valuable,” including knowledge. These are compatible positions, not a reversal: the later statement does not deny the need for symbols or variable operations.
• CONSISTENT · 96%20 Aug → 21 SeptThen The interviewer argues that intelligence is situated rather than as abstract and pure as people had thought.Now The interviewer argues that a theory is needed to explain phenomena including double descent, apparent overcoming of the curse of dimensionality, systematicity, productivity, and models' ability to produce novel syntactically competent sentences.The earlier claim says intelligence is “very situated” and “not quite as abstract and pure as we thought.” The later claim says a theory is needed to explain several model phenomena, including systematicity and novel syntactically competent sentences. These claims address different aspects of the topic and can both be true; the later statement does not retract the earlier one.
• CONSISTENT · 94%24 Jun → 21 SeptThen The speaker treats obeying a simple constraint such as not making illegal chess moves as a small-scale alignment problem that AI systems still struggle to solve.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier claim says AI systems “still” struggle to follow constraints, such as not making illegal chess moves. The later claim says constraints and knowledge are “incredibly valuable.” These are compatible: valuing constraints does not imply that current systems reliably follow them.
• CONSISTENT · 97%20 Aug → 21 SeptThen The interviewer argues that people with different perspectives can use their intelligence to reach different answers about the same phenomena.Now The interviewer argues that a theory is needed to explain phenomena including double descent, apparent overcoming of the curse of dimensionality, systematicity, productivity, and models' ability to produce novel syntactically competent sentences.These claims address different points rather than a changed position. Earlier, the interviewer says intelligence is situated and that people with different perspectives can reach different answers. Later, he says a theory is needed to explain several observed phenomena in models. The later claim does not deny or materially alter the earlier one.
• CONSISTENT · 91%20 Aug → 21 SeptThen The interviewer says the prior conception held that greater intelligence enables faster traversal of the combinational closure of knowledge and abstractions.Now The interviewer considers constraints extremely valuable and defines them broadly as knowledge operating at multiple levels of abstraction.The earlier passage describes a conception they previously held—that intelligence involves traversing knowledge and abstractions—and then says their findings make intelligence seem more situated. The later claim that constraints, broadly understood as knowledge, are valuable does not deny that earlier conception; it concerns the role of knowledge and constraints in systems, and is compatible with it.
• CONSISTENT · 91%24 Jun → 21 SeptThen The speaker considers the semantics represented by current AI models superficial and primarily based on word distributions.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier claim says current models’ semantics are “very superficial” and “really about distributions of words.” The later claim says the speaker is “a big believer that constraints are incredibly valuable,” including knowledge. These positions can coexist: valuing knowledge and constraints does not deny that current models’ semantics are superficial.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker says that the only general-purpose AI systems currently available are based on large language models.Now The speaker says reinforcement-learning-trained language models have substantial base knowledge and can perform things that were not possible before.The earlier claim says that the only general-purpose systems currently available are LLM-based. The later claim says RL-trained language models have substantial base knowledge and can do things that were not possible before. The later statement does not deny or materially revise the earlier one; it describes capabilities of language models.
• CONSISTENT · 93%24 Jun → 8 SeptThen The speaker characterizes the chess system's knowledge as task-specific rather than general chess knowledge.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim describes a chess system as having “sterile,” task-specific knowledge rather than “general knowledge of chess.” The later claim says an AI can be given a mathematics task or a knowledge question. That broader statement does not say the chess system had general chess knowledge, so it does not reverse the earlier position.
• CONSISTENT · 94%24 Jun → 8 SeptThen The speaker says AI systems may automate conventional, inside-the-box work and may outperform people on such work, citing the Go move 37 example.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier statement says AI may automate or outperform on “inside-the-box” work, while reaching AGI will require “Einstein level innovations.” The later statement says models can make “novel combinations” to solve tasks. That does not say they can produce the kind of outside-the-box innovations the earlier statement said AGI would require; the claims can both be true.
• CONSISTENT · 93%24 Jun → 8 SeptThen The speaker says that he and Yann LeCun consider current AI systems' purported reasoning insufficiently robust.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement concerns current systems, saying their purported reasoning is not “robust enough.” The later statement says an automated AI scientist may be possible “in the near future.” A future possibility does not contradict the earlier assessment of present capabilities; the later context also frames this as an emerging possibility prompted by research, not a claim that current reasoning is already robust.
• CONSISTENT · 89%24 Jun → 8 SeptThen The speaker cautions that explaining AI failures as failures on tasks the systems were not trained on is an oversimplification, while discussing AI risk.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier statement says that describing AI failures as failures on tasks the systems were not trained on is an “oversimplification,” in a discussion of AI risk. The later statement says an AI given a specified task, such as a mathematics problem, will “hill climb towards it.” These claims address different aspects of AI behavior and do not conflict: task-directed optimization does not negate the earlier caution about generalizing from training to untrained tasks.
• CONSISTENT · 82%24 Jun → 8 SeptThen The speaker says the same gaps remain despite quantitative improvement, with no principled solution to any of them.Now The speaker believes that extensive optimization in semi-specific objective domains still leaves something important missing.Both statements express that AI still has an unresolved shortcoming despite progress: earlier, the speaker says they see “exactly the same gaps” and “no principled solution”; later, they say that even with extensive optimization, “there’s something missing.” The later statement does not indicate a change in position.
• CONSISTENT · 88%24 Jun → 8 SeptThen The speaker says that much of the near-term experimentation people conduct on LLMs can be automated.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says that “a lot” of near-term LLM experimentation “can be automated.” The later claim says an automated AI scientist “could be possible in the near future.” The later statement is a broader possibility, but it does not retract or deny the earlier point; the earlier context even notes that generating genuinely new ideas was not the system’s forte, which is compatible with expressing interest in more capable AI scientists later.
• CONSISTENT · 94%24 Jun → 8 SeptThen The speaker thinks most AI safety arguments focus on AGI that is at least as intelligent as most smart people.Now The speaker says composable AI agents would make them much more worried.The earlier claim says most AI safety arguments concern AGI “at least as smart as” most smart people. The later claim says composable agents “would make me much more worried.” These concern different possible sources of concern—intelligence level versus composability—and the later statement does not deny the earlier one.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker describes AlphaFold as a carefully engineered hybrid neurosymbolic system that greatly outperforms a pure chatbot.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim describes AlphaFold as a carefully engineered hybrid system that “far outperforms” a pure chatbot. The later claim says models can “create novel combinations” of existing knowledge to solve tasks. These claims address different aspects and can both be true; the later statement does not reverse the earlier one.
• CONSISTENT · 94%24 Jun → 8 SeptThen The speaker says AGI must be able to solve original problems, not merely perform well through teaching to the test.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim says AGI must “solve things that are original,” beyond “teaching to the test.” The later claim describes a model hill-climbing toward a specified task, such as a mathematics problem; it does not say that this is enough to solve original problems or contradict the earlier requirement.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker thinks AI progress could accelerate quickly once critical ideas are found, analogous to the rapid progress in molecular biology after Watson and Crick figured out DNA.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim is that AI progress could speed up once critical ideas are found (“they could go fast”). The later claim imagines a particular future recursive system and says it would be “a different type of technology.” That describes a possible mechanism and its character; it does not deny that AI progress could accelerate.
• CONSISTENT · 91%24 Jun → 8 SeptThen The speaker sees no sound argument that automating the full AI research-and-development process is impossible.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says there is “no sound argument” that fully automating AI research and development “can't be done.” The later claim describes a possible recursive system and says it would be “a different type of technology”; it does not deny that such automation is possible or reverse the earlier position.
• CONSISTENT · 88%24 Jun → 8 SeptThen AI organizations differ to some extent in their reinforcement-learning methods and datasets.Now The RL systems being discussed use new training methods and apply architectures differently.The earlier claim says organizations differ somewhat in their reinforcement-learning methods and datasets. The later claim mentions new training methods and architectures being used differently. These claims are compatible: the later statement adds examples of technical variation rather than denying the earlier assessment.
• CONSISTENT · 97%24 Jun → 8 SeptThen The speaker says future AGI might use LLMs, which he considers good at distributional learning.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says future AGI “might use LLMs”; it does not say there would be one large AI. The later claim imagines “diverse AIs” with different expertise working together “rather than one big AI.” Those positions can coexist: multiple AIs could use LLMs. The later statement changes the proposed system’s organization, not the earlier claim about possible technology.
• CONSISTENT · 91%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have much better reasoning and planning methods.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim predicts “much better ways of doing reasoning and planning.” The later claim describes how future AI might be organized—“diverse AIs” with different expertise working from multiple angles, “rather than one big AI.” That architectural preference does not deny the earlier prediction about improved reasoning and planning.
• CONSISTENT · 94%24 Jun → 8 SeptThen The speaker believes appropriately designed software should let machines outperform humans in areas suited to machines and at least achieve human-level performance overall.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says appropriately designed software could let machines outperform people in suitable areas and reach human-level performance. The later claim says an automated AI scientist may be possible in the near future. These claims are compatible; the later one concerns a particular application and does not retract the earlier position.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker thinks AI progress could accelerate quickly once critical ideas are found, analogous to the rapid progress in molecular biology after Watson and Crick figured out DNA.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says progress could accelerate once critical ideas are found (“they could go fast”). The later claim says models can combine existing knowledge in novel ways to solve tasks. These statements concern different questions—future pace of AI progress versus a current model capability—and do not indicate a change of mind.
• CONSISTENT · 91%24 Jun → 8 SeptThen The speaker predicts that neurosymbolic AI will begin to take off in three or four years.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim predicts that “neurosymbolic AI starts to take off” in three or four years. The later claim says an automated AI scientist “is possible in the near future.” These concern different outcomes—neurosymbolic AI gaining traction versus the possibility of an automated AI scientist—and can both be true. The statements do not establish a change of mind.
• CONSISTENT · 98%24 Jun → 8 SeptThen The speaker says that newer AI systems enable tasks that were previously unavailable and that these capabilities are already affecting labor markets.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says newer AI systems enable previously unavailable tasks and are already affecting labor markets. The later claim imagines future AI as multiple specialized systems collaborating rather than one large AI. These address different questions, and neither statement denies the other.
• CONSISTENT · 98%24 Jun → 8 SeptThen The speaker regards an Apple paper involving Tower of Hanoi as an example of AI systems struggling with distribution shift.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim says the Tower of Hanoi paper illustrates AI systems struggling with “distribution shift.” The later claim says an AI can be given a mathematics task and “hill climb towards it.” These address different capabilities and do not conflict; the later statement does not retract the earlier concern.
• CONSISTENT · 91%24 Jun → 8 SeptThen The speaker believes that LLMs are not really the route to AGI and that achieved AGI will look quite different; he thinks Jan LeCun shares this view.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim says LLMs are “not really the route to AGI” and that AGI will “look pretty different.” The later claim imagines future AI as “diverse AIs” working from different angles rather than “one big AI.” That is compatible with, and may illustrate, the earlier view; it does not say LLMs are the route to AGI.
• CONSISTENT · 96%24 Jun → 8 SeptThen The interviewer says neurosymbolic AI is his personal favorite among possible alternative AI technologies.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement calls neurosymbolic AI the speaker’s “personal favorite” among possible alternative technologies. The later statement describes a preferred future arrangement of “diverse AIs” collaborating rather than “one big AI.” These claims do not contradict each other: the first identifies a favored technology, while the second describes how future AI systems might be organized.
• CONSISTENT · 94%24 Jun → 8 SeptThen The speaker says that the only general-purpose AI systems currently available are based on large language models.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim is about what general-purpose AI systems are currently available: “The only general ones we have are LLM-based.” The later claim imagines a possible future of “diverse AIs” with different expertise, rather than “one big AI.” A future possibility does not contradict a statement about current systems, so these claims do not establish a change of mind.
• CONSISTENT · 91%24 Jun → 8 SeptThen The speaker does not currently see a technical solution for coping with the problems posed by intelligence recursion.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says the speaker does not see a technical way to cope with the problems “right now.” The later claim imagines a future recursive system and calls it “a different type of technology”; it does not claim that this system solves those problems or contradict the earlier assessment.
• CONSISTENT · 94%24 Jun → 8 SeptThen The speaker expects a future state change in AI but says its timing is very difficult to predict.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says a future AI “state change” is expected, while its timing is “very hard” to know. The later claim says an automated AI scientist is possible in the near future. That does not deny the earlier position; the statements concern different claims about AI progress.
• CONSISTENT · 98%24 Jun → 8 SeptThen The speaker says current AI companies are not entirely trustworthy.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says current AI companies are “not entirely trustworthy.” The later claim says an automated AI scientist may be possible “in the near future.” These concern different questions and can both be true; the later statement does not revise the earlier assessment of the companies.
• CONSISTENT · 91%24 Jun → 8 SeptThen The speaker says the same gaps remain despite quantitative improvement, with no principled solution to any of them.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim says there is “no principled solution” to the gaps despite quantitative improvement. The later claim describes “magical” results in well-specified domains while acknowledging “a bit of a space of ambiguity.” Those points can both be true: useful performance in bounded settings does not assert that the broader gaps have been principledly solved.
• CONSISTENT · 98%24 Jun → 8 SeptThen The speaker says that system prompts against producing copyrighted material or hallucinating do not reliably prevent those behaviors.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim says system prompts do not reliably prevent failures such as producing copyrighted material or hallucinating. The later claim says expert use can produce “magical things” in “well-specified domains,” while acknowledging “a space of ambiguity.” It does not say those failures have been solved; the statements concern different conditions and are compatible.
• CONSISTENT · 94%24 Jun → 8 SeptThen The speaker argues that robust understanding of domains such as tic-tac-toe or chess should make related tasks relatively easy for AI systems.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim is conditional: if an AI had a robust understanding of domains like chess, related tasks “wouldn't be too hard.” The later claim describes giving an AI a specified math or knowledge task and says it will “hill climb towards it.” These statements address different aspects of AI capability and do not conflict; the later statement does not retract the earlier condition.
• CONSISTENT · 91%24 Jun → 8 SeptThen The speaker thinks current AI approaches are already encountering a bottleneck in at least one of their raw resources.Now The speaker suggests that a distributed collective AI system might face additional bottlenecks due to communication among its agents.The earlier claim says current AI approaches are already encountering a bottleneck in at least one raw resource. The later claim says distributed collective systems might face additional bottlenecks from message passing. These claims are compatible: the later statement identifies a possible bottleneck for a particular system design rather than denying the earlier one.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker's preferred conception of AGI is a system at least as capable as humans across many tasks and superior in some areas.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim describes what capabilities would qualify as AGI: “at least as good as humans in a bunch of things and better in certain ones.” The later claim describes a preferred form for future AI: “diverse AIs” working from different angles “rather than one big AI.” These claims address capability and system design, respectively, and can both be true; the later statement does not retract the earlier standard.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker says computers can memorize billions of digits of pi, far more than people; he gives his daughter’s memorization of 105 digits as a comparison.Now Humans can expand their memory.The earlier claim says computers can memorize far more than people (“billions of digits” versus 105). The later statement that “we can expand our memory,” in a context about using tools, does not deny that comparison; it describes extending human memory, not matching computers’ unaided storage capacity.
• CONSISTENT · 98%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have substantially better reasoning and planning methods and more stable world models.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim predicts that future AI will have “much better ways” of reasoning and planning and “more stable world models.” The later claim imagines future AI as “diverse AIs” tackling problems from different angles rather than “one big AI.” These address different aspects of future AI—capabilities versus organization—and can both be true; the later claim does not retract the earlier prediction.
• CONSISTENT · 93%24 Jun → 8 SeptThen The speaker describes AlphaFold as a carefully engineered hybrid neurosymbolic system that greatly outperforms a pure chatbot.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says AlphaFold is a carefully engineered hybrid system that “far outperforms” a pure chatbot. The later claim says expert-guided AI can produce “magical things” in well-specified domains while some ambiguity remains. These claims are compatible; the later statement does not deny AlphaFold’s comparative performance.
• CONSISTENT · 93%24 Jun → 8 SeptThen The speaker and Jan LeCun both approach AI through architecture or cognition.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim says the speaker and Jan LeCun take “an architectural approach or a cognitive approach” to AI. The later claim imagines future AI as “diverse AIs” examining problems from different angles rather than “one big AI.” These claims concern different aspects of AI and do not conflict: an architectural or cognitive approach can be compatible with a collective of specialized systems.
• CONSISTENT · 98%24 Jun → 8 SeptThen The speaker argued that current AI approaches cannot keep increasing their data by 100 times because the available data is insufficient.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says you “can't just keep 100x-ing” training data because “there just isn't enough data.” The later claim says models can “create novel combinations” of existing knowledge to solve tasks. These address different questions—data-scaling limits versus models’ task-solving abilities—and do not show a change in position.
• CONSISTENT · 96%24 Jun → 8 SeptThen A key bottleneck for tool-using AI systems is that they do not always call the appropriate code.Now The speaker suggests that a distributed collective AI system might face additional bottlenecks due to communication among its agents.The earlier claim identifies a bottleneck in tool use: systems “don't always call the right code.” The later claim discusses a different possible bottleneck in distributed systems: “message passing between all the different agents.” These claims can both be true and do not indicate a change of mind.
• CONSISTENT · 97%24 Jun → 8 SeptThen The speaker predicts that AI systems will continue improving regularly and provide increasing value.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AI systems will “continue to improve regularly” and provide “more and more value.” The later claim imagines future AI as “diverse AIs” with different expertise working together rather than “one big AI.” These address different aspects of AI’s future and are compatible; the later statement does not reverse the earlier prediction.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have much better reasoning and planning methods.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim predicts “much better ways of doing reasoning and planning.” The later claim describes a possible organization for future AI: “diverse AIs” examining problems from different angles “rather than one big AI.” These claims concern different aspects of future AI and are compatible; the later statement does not retract the prediction about reasoning and planning.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have much better reasoning and planning methods.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim predicts “much better ways of doing reasoning and planning.” The later claim says an automated AI scientist is “possible in the near future.” These are compatible: the later statement does not retract or deny the prediction about improved reasoning and planning.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker considers learning models of distributions a very useful cognitive skill, but not the only important cognitive skill.Now The speaker views intelligence as perspectival and expects it to be uneven and fragmented across capabilities.The earlier claim says learning models of distributions is “a super helpful cognitive skill” but “not the only one.” The later claim says intelligence is “perspectival” and “fractured and jagged,” describing unevenness across capabilities. These claims are compatible; the later statement does not retract the earlier one.
• CONSISTENT · 99%24 Jun → 8 SeptThen The speaker says driverless-car technology has made progress every year.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says driverless-car technology has “made progress every year.” The later claim says an efficiently self-adapting AI loop would be “a different type of technology.” These concern different technologies and can both be true; the later statement does not revise the earlier position.
• CONSISTENT · 95%24 Jun → 8 SeptThen Machines will eventually exceed human performance, but when and how this will happen remain uncertain to the speaker.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says machines “will exceed” human performance eventually; the later says an automated AI scientist “could be possible in the near future.” The later claim is compatible with the earlier one, not a reversal. They address different levels of specificity—general performance versus a particular AI-scientist capability.
• CONSISTENT · 97%24 Jun → 8 SeptThen The speaker says human cognition is impaired by confirmation bias and motivated reasoning.Now The speaker believes human intuitions about collective behavior are often unreliable and that such behavior could emerge in principle.The earlier claim says human cognition can be impaired by “confirmation bias and motivated reasoning.” The later claim concerns a different point: “a lot of our intuitions are broken” when considering collective behavior, and that such behavior could emerge in principle. Neither statement reverses the other; they address different aspects of cognition and collective behavior.
• CONSISTENT · 98%24 Jun → 8 SeptThen The speaker expects future periods of AI progress to exceed the progress of the last few years and expects AI to be enormously more advanced in 30 years.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim predicts that AI will make substantial progress—saying future periods may “exceed the last few years” and that AI will be “enormously ahead” in 30 years. The later claim describes a possible form of future AI: “diverse AIs” working from different angles rather than “one big AI.” These address different aspects of AI’s future and are compatible; the later statement does not retract the earlier prediction.
• CONSISTENT · 87%24 Jun → 8 SeptThen Despite quantitative progress in AI, the speaker believes that none of the key problems a cognitive creature must solve have been solved.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says AI has not solved the key problems a cognitive creature must solve. The later excerpt concerns whether AI capabilities are composable: “that doesn't really marry with my experience,” followed by the view that agents are “not yet composable.” These are different, compatible questions, so the later statement does not show a change in the earlier position.
• CONSISTENT · 93%24 Jun → 8 SeptThen The speaker believes appropriately designed software should let machines outperform humans in areas suited to machines and at least achieve human-level performance overall.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says appropriately designed software should let machines outperform humans in suitable areas and reach human-level performance overall. The later claim says expert-guided AI can perform impressively in well-specified domains, while ambiguity remains. These claims are compatible; the later statement does not deny the earlier position.
• CONSISTENT · 96%24 Jun → 8 SeptThen The speaker allows that discovery of new ideas might be automated.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says automated discovery of new ideas is possible, while noting that most LLMs exploit existing ideas. The later claim is that models make “novel combinations” of things already known to solve tasks. That describes recombining existing knowledge, not a contrary claim about discovering genuinely new ideas.
• CONSISTENT · 95%24 Jun → 8 SeptThen The speaker and Jan LeCun both approach AI through architecture or cognition.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says the speaker and Jan LeCun take an “architectural approach or a cognitive approach.” The later claim describes a future with “diverse AIs” collaborating rather than “one big AI.” These address different aspects of AI—how to approach it versus how future systems might be organized—and the claims can both be true.
• CONSISTENT · 94%24 Jun → 8 SeptThen The speaker says current AI systems, including O3, do not reliably follow chess rules and sometimes make illegal moves.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says current systems “can't even play chess reliably according to the rules” and may “make illegal moves.” The later claim concerns expert-guided AI producing impressive results in “well-specified domains,” while still leaving “a space of ambiguity.” These claims address different settings and can both be true; the later statement does not retract the earlier one.
• CONSISTENT · 96%24 Jun → 20 AugThen The speaker says current AI systems have not been good at generating genuinely new ideas.Now The speaker argues that AI systems follow the instructions they are given; in the example, a basic prompt to do something interesting with video frames would produce nothing interesting.Both claims express skepticism about AI’s ability to originate ideas independently. Earlier, the speaker says current systems have “not been the forte” at generating genuinely new ideas; later, he says a broad prompt to “do something interesting” would produce “nothing interesting,” while specific goals and framing come from people. The later statement reinforces the earlier position rather than changing it.
• CONSISTENT · 96%24 Jun → 20 AugThen The speaker predicts that AI systems will continue improving regularly and provide increasing value.Now AI models can perform some tasks that humans do extremely well, and they appear to be improving continually.Both statements assert ongoing improvement: earlier, the speaker says AI systems “will continue improving regularly” and provide “more and more value”; later, he says models “seem to be getting better.” The later statement specifies that they can perform some tasks humans do extremely well, but does not reverse the earlier view.
• CONSISTENT · 94%24 Jun → 20 AugThen The speaker believes AI systems could become orders of magnitude more data-efficient than they currently are.Now The speaker says the AI technology is improving as they use it every day.The earlier claim says there is “orders of magnitudes more data efficiency to be carved out”; the later says the technology is “getting better” with daily use. These claims are compatible, and the later statement does not revise the earlier view about how much efficiency could eventually be gained.
• CONSISTENT · 94%24 Jun → 20 AugThen The speaker predicts that AI systems will improve programmer productivity, while noting caveats concerning the security and maintainability of generated code.Now The speaker argues that current AI coding tools generate large amounts of poor-quality, tangled code and create more problems than they solve.The earlier statement predicts that AI will improve programmer productivity, while flagging caveats about security and maintainability. The later statement criticizes current tools as generating “spaghetti garbage code” and creating “more problems than they solve.” The claims concern different scopes: prospective productivity gains with caveats versus the tools’ current performance. The later criticism is compatible with the earlier caveat, not evidence of a changed position.
• CONSISTENT · 94%24 Jun → 20 AugThen Reliably getting AI systems to use tools is difficult.Now With a memory system and repeated user corrections, an AI system can build an increasingly accurate simulacrum of its user and perform the desired actions more often over time.The earlier claim says reliably getting AI systems to use tools is difficult. The later claim says memory and repeated corrections can make a system perform the user’s desired actions more often over time. That describes a way performance may improve, not a reversal of the earlier claim that reliable tool use is difficult.
• CONSISTENT · 94%24 Jun → 20 AugThen The speaker believes AI systems are not close to matching a top software engineer, because that requires understanding the problem, the human goals behind it, and multiple domains.Now The speaker says AI can be directed toward specific goals or problems when given a suitable program and domain, but humans supply the framing.The earlier claim is that AI is “not close” to matching a top software engineer, given the need to understand human goals and multiple domains. The later claim concerns AI pursuing a specific goal when given a program, while “the framing always comes from us” and overcoming that limitation is “kind of doubtful.” That does not assert the broader engineering capability the earlier claim said was lacking.
• CONSISTENT · 91%24 Jun → 20 AugThen The speaker says people tend to take scenarios seriously, and vivid details can overwhelm their ability to assess them.Now Because the world is complicated, people tend to select information that favors their existing way of understanding it.Both claims describe ways people can be led to interpret information in line with their existing mental models: the earlier claim says vivid scenarios can overwhelm assessment, while the later says people select information that privileges how they already think about the world. The later statement does not retract or materially alter the earlier one.
• CONSISTENT · 91%24 Jun → 20 AugThen The speaker predicts that AI systems will continue improving regularly and provide increasing value.Now The speaker says the AI technology is improving as they use it every day.Both statements describe ongoing improvement: earlier, the speaker says the systems “will continue improving regularly,” and later says the technology “is getting better” with daily use. The later statement is consistent with the earlier prediction, not a different position.
• CONSISTENT · 93%24 Jun → 20 AugThen The speaker predicts that AI systems will continue improving regularly and provide increasing value.Now AI models are becoming increasingly adaptive and capable.The earlier statement predicts AI systems “will continue improving regularly” and provide “more and more value.” The later statement says the models “are becoming quite adaptive” and have “a lot of capabilities.” These are compatible claims about improvement and capability, not opposing positions. The later passage adds a concern about how deceptive apparent intelligence can be, but does not reverse the earlier view.
• CONSISTENT · 91%24 Jun → 20 AugThen The speaker sees little, though possibly some, progress in AI systems' ability to match the flexibility of human expert knowledge.Now AI models can perform some tasks that humans do extremely well, and they appear to be improving continually.The earlier claim says there is “little, though possibly some, progress” in matching the flexibility of human expert knowledge. The later claim says models do some things humans do “extremely well” and “seem to be getting better.” Progress on particular tasks does not contradict limited progress in flexible expert knowledge, so the statements can both be true.
• CONSISTENT · 97%24 Jun → 20 AugThen The speaker says that newer AI systems enable tasks that were previously unavailable and that these capabilities are already affecting labor markets.Now AI is already automating a meaningful amount of human labor and many jobs.Both claims assert that AI capabilities are affecting human labor: the earlier statement says newer systems' capabilities are “already affecting labor markets,” and the later says AI is “automating meaningfully human labor and lots of jobs.” The later wording is more specific, but does not reverse the earlier position.
• NOT COMPARABLE · 99%24 Jun → 26 SeptThen An early attempt to combine an LLM with a symbolic system by integrating it into Wolfram Alpha produced inconsistent results.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim concerns an early LLM–symbolic-system integration producing “hit or miss” results, specifically because reliable tool access was difficult. The later claim reports what critics say about LLMs used for research—loss of context and obscured origins. These are different claims about different issues; the later statement does not establish the speaker’s own changed position on the earlier question.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The speaker predicts that by 2035 people will regard LLMs as an incomplete answer, although they will continue using them for some tasks.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim is the speaker’s prediction that by 2035 LLMs will be seen as “not really the answer,” though still useful for some tasks. The later passage says “some critics” argue that LLM research removes information from context; it reports their view rather than clearly stating the speaker’s own position, and it does not directly address the earlier prediction.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The speaker says LLMs might be replaced because they are very inefficient in their use of energy and data.Now In autonomous settings, an LLM needs explanatory framing for a heuristic such as UCB to use it coherently; simply instructing it to use UCB is insufficient.The earlier claim says LLMs “might be replaced” because they are inefficient in energy and data. The later claim concerns how to make LLMs use heuristics coherently in autonomous settings. It neither revises nor denies the earlier view about possible replacement; the claims address different questions.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The speaker says LLMs might be replaced because they are very inefficient in their use of energy and data.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim says LLMs “might be replaced” because they are inefficient in energy and data. The later passage attributes a different criticism to “a lot of people” and does not establish that the speaker personally endorses it. These statements therefore do not show a change in the speaker’s own position.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen An early attempt to combine an LLM with a symbolic system by integrating it into Wolfram Alpha produced inconsistent results.Now In autonomous settings, an LLM needs explanatory framing for a heuristic such as UCB to use it coherently; simply instructing it to use UCB is insufficient.The earlier claim concerns an early LLM–symbolic-system integration producing “hit or miss” results, attributed to reliably accessing tools. The later claim concerns how an LLM should be framed to use a heuristic like UCB coherently in an autonomous setting. These address different questions, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 97%24 Jun → 26 SeptThen Performance on agent tasks is still poor.Now The team configured the agent to optimize itself by improving its harness, including its prompts and tools.The earlier claim is an assessment that agent-task performance is “still pretty poor.” The later claim describes a team’s method—“hill climbed towards improving the agent itself” by changing its harness, prompts, and tools—and does not say whether agent-task performance is now good or poor. It therefore does not establish a changed position on the same question.
• NOT COMPARABLE · 99%24 Jun → 26 SeptThen The speaker believes LLMs alone are inadequate for the nearer-term technical alignment problem, while neurosymbolic methods might offer a chance.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim is the speaker’s view that LLMs alone are inadequate for nearer-term technical alignment and that neurosymbolic methods might help. The later claim reports what “some critics argue” about using LLMs for research; it does not state the speaker’s own position on that issue, so there is no comparable asserted belief to assess.
• NOT COMPARABLE · 99%24 Jun → 26 SeptThen The speaker thinks it is possible to make LLMs play chess without illegal moves by using a tool call.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim concerns whether LLMs could play chess without illegal moves by using a tool call. The later claim reports critics’ concern that LLMs remove research information from context. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen A separate agent benchmark graph used a seconds-to-tens-of-seconds time scale, unlike the coding-task graph's minutes-to-hours-or-days scale.Now The research agent takes a metric and a problem as inputs and hill-climbs toward solving the problem.The earlier claim describes the time scale on a separate agent benchmark graph (“seconds to tens of seconds”), while the later claim describes how a research agent operates (“takes a metric and a problem” and “hill-climbs”). These are different attributes, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 26 SeptThen The speaker considers it unlikely that LLMs will play chess without illegal moves and without tool use within the next six months.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim is about whether LLMs can play chess without illegal moves and without tools within six months. The later claim reports what critics say about LLMs being used for research. These concern different questions, and the later statement is attributed to critics rather than presented as the speaker's own position.
• NOT COMPARABLE · 98%20 Aug → 26 SeptThen The interviewer describes the recursive self-improvement argument: after a long initial development, an AGI could help build successors in progressively less time, and each successor would improve.Now People often imagine recursive self-improvement as a system repeatedly rewiring its own brain.The earlier passage describes an interviewer’s recursive self-improvement argument (“we can use the AGI to help us build the next one”), while the later passage describes what “most people think” recursive self-improvement means. Neither extracted claim clearly states Tim Scarfe’s own position, and the passages do not establish that he changed his view.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The speaker believes LLMs alone are inadequate for the nearer-term technical alignment problem, while neurosymbolic methods might offer a chance.Now In autonomous settings, an LLM needs explanatory framing for a heuristic such as UCB to use it coherently; simply instructing it to use UCB is insufficient.The earlier claim concerns whether LLMs alone can solve nearer-term technical alignment, saying they are “just not up to the task” and neurosymbolic methods “might have a chance.” The later claim concerns how to frame a heuristic such as UCB for coherent use in autonomous settings. It does not address the earlier comparison or establish a changed view about LLMs’ alignment capabilities.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The speaker characterizes most LLM accomplishments as exploiting existing ideas rather than discovering new ones.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim asserts that most LLM accomplishments exploit existing ideas rather than discover new ones. The later claim reports what critics say about LLM-assisted research—“you’re taking the information out of its context” and “you don’t understand why it came about.” It does not state the speaker’s own view about whether LLMs discover ideas, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 26 SeptThen The speaker doubts that LLMs can infer causal laws well and does not expect them anytime soon to produce Einstein-level, radically different solutions to problems.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier statement is the speaker’s view that LLMs are poor at inferring causal laws and unlikely soon to produce radically new solutions. The later claim reports what critics say about using LLMs for research; it does not state whether the speaker agrees, and it addresses a different issue. There is no clear later position to compare.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The speaker says someone might discover a better method for learning models of distributions than current LLMs provide.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.These are not claims about the same question. The earlier passage speculates that “somebody might find a better way” to learn models of distributions; the later passage reports what “some critics argue” about LLMs removing information from context. Neither establishes a change in the speaker’s own position.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The speaker says LLMs might be replaced because they are highly inefficient in their use of energy and data.Now In autonomous settings, an LLM needs explanatory framing for a heuristic such as UCB to use it coherently; simply instructing it to use UCB is insufficient.The earlier statement is about LLMs potentially being replaced because they are inefficient in energy and data use. The later statement is about how to frame heuristics so an LLM can use them coherently in an autonomous setting. These address different questions, and the later claim does not say whether LLMs should or should not be replaced.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The early LLM–Wolfram Alpha approach was limited by the interface through which the model accessed tools.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim is Tim Scarfe’s own assessment that the early LLM–Wolfram Alpha approach was limited by reliable tool access. The later claim says “Some critics argue” that LLM research removes information from context; it reports others’ criticism rather than stating Scarfe’s own position. The claims therefore do not establish a change in his view.
• NOT COMPARABLE · 98%20 Aug → 26 SeptThen The interviewer says AI shows hints of adaptive self-improvement, especially during hill-climbing toward a particular task.Now In July, a London-based AI company posted a tweet that received about 1.8 million views and claimed to demonstrate the first evidence of recursive self-improvement.The earlier claim, attributed to Tim Scarfe, says there are “hints” of adaptive self-improvement during hill-climbing toward a task. The later claim only reports that a company’s tweet “claim[ed] to demonstrate” evidence of recursive self-improvement; it does not establish that the speaker endorses or rejects the claim. The statements also concern different kinds of self-improvement, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen The speaker says LLMs might be replaced because they are highly inefficient in their use of energy and data.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim is the speaker’s prediction that LLMs might be replaced because they are inefficient in energy and data. The later claim reports what critics say about using LLMs for research; it does not state whether the speaker agrees, nor address whether LLMs will be replaced. These are not clearly opposing positions.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen Many reported prompt-specific odd answers from LLMs have been fixed.Now Some critics argue that using LLMs for research removes information from its context and obscures why that information or idea arose.The earlier claim says many prompt-specific odd answers have been fixed (“a lot of them have been solved”). The later claim reports critics’ concern that LLM research removes information from context and obscures why ideas arose. These address different issues, and the later passage reports others’ criticism rather than stating a contrary position of the speaker’s own.
• NOT COMPARABLE · 98%24 Jun → 26 SeptThen AI actors have incentives to accelerate development and secure as many resources as possible for that purpose.Now The team configured the agent to optimize itself by improving its harness, including its prompts and tools.The earlier claim says AI actors are incentivized to accelerate development and obtain resources. The later claim describes a particular experiment in which a team let an agent improve its harness, prompts, and tools. That description does not state a changed view about actors’ incentives; the claims address different questions.
• NOT COMPARABLE · 96%24 Jun → 26 SeptThen The early LLM–Wolfram Alpha approach was limited by the interface through which the model accessed tools.Now In autonomous settings, an LLM needs explanatory framing for a heuristic such as UCB to use it coherently; simply instructing it to use UCB is insufficient.The earlier claim says the approach was limited by “the interface” and whether the model could “reliably get to the tools.” The later claim says that in autonomous settings, an LLM needs explanatory framing to use a heuristic like UCB coherently. These concern different obstacles and can both be true; the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 99%24 Jun → 26 SeptThen A separate agent benchmark graph used a seconds-to-tens-of-seconds time scale, unlike the coding-task graph's minutes-to-hours-or-days scale.Now The team configured the agent to optimize itself by improving its harness, including its prompts and tools.The earlier claim describes an agents benchmark graph’s time scale as “seconds to tens of seconds,” contrasting it with a coding-task graph. The later claim describes a team improving an agent by changing its “prompts and tooling.” These address different topics, so they provide no evidence that the person changed their position on the same question.
• NOT COMPARABLE · 98%20 Aug → 26 SeptThen The interviewer describes the recursive self-improvement argument: after a long initial development, an AGI could help build successors in progressively less time, and each successor would improve.Now In July, a London-based AI company posted a tweet that received about 1.8 million views and claimed to demonstrate the first evidence of recursive self-improvement.The earlier extracted claim is the interviewer’s description of a recursive self-improvement argument (“use the AGI to help us build the next one”), not clearly Tim Scarfe’s own position. The later claim reports that a company’s tweet claimed evidence of recursive self-improvement; it does not state whether Scarfe accepts that claim. The passages therefore do not establish a change in Scarfe’s beliefs.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The interviewer considers the idea of a durable advantage in large language models to be a myth, assuming LLMs are the relevant technology.Now The interviewer says language models used in RLV have substantial base knowledge.The earlier statement is Tim Scarfe’s view that “the notion of a durable advantage in LLMs … is a myth.” The later passage, “They have a lot of base knowledge,” concerns language models’ knowledge aiding exploration in RLV; it neither affirms nor denies whether LLMs can have a durable advantage. The extracted later text also attributes the claim to the interviewer, while the surrounding transcript identifies Tim as speaking, so it does not establish a comparable position on the same question.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker says someone might develop a better method for learning distributions; he describes distribution learning as a very helpful, but not exclusive, cognitive skill.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier claim concerns distribution learning as a useful cognitive skill and the possibility of finding a better method for it. The later claim expresses excitement about inference-time computation and adaptation. Neither claim states a position that reverses the other, and they do not clearly address the same underlying question.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker estimates that most of the timeline probability distribution lies beyond 10 years.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier claim is a timeline estimate: “most of the distribution is past 10 years.” The later claim expresses enthusiasm for inference-time computation and adaptation, saying “not that many people are talking about it.” It does not state a different timeline estimate or otherwise address the earlier question, so there is no basis to conclude the speaker changed their mind.
• NOT COMPARABLE · 96%24 Jun → 21 SeptThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now The speaker argues that a theory is needed to explain the observed phenomena in deep learning.The earlier claim says AI systems need “a deeper” form of semantics; the later says “we need to have a theory” explaining phenomena in deep learning. These could be related concerns, but they address different questions, and the excerpts do not establish whether the later theory is meant to provide the deeper semantics previously called for.
• NOT COMPARABLE · 97%24 Jun → 21 SeptThen The speaker considers the graph difficult to interpret reliably.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim concerns difficulty interpreting a particular graph (“weird to read anything off of this graph”). The later claim agrees that people may be overinterpreting similarities in model representations (“we’re reading too much into that”). These are related cautions, but the passages do not establish opposing positions or a change of mind; the later statement may be compatible with the earlier one.
• NOT COMPARABLE · 97%20 Aug → 21 SeptThen The speaker characterizes the opposing view as expecting scaling alone, including a future model such as GPT-7, to read and refactor problematic code, perhaps through reinforcement learning.Now SPEAKER_02 characterizes Rich Sutton’s “Bitter Lesson” as arguing against building knowledge and constraints into systems and in favor of scaling models of computation.The earlier claim describes an opposing position—expecting to “keep scaling” and for a future model to fix code—rather than clearly stating the speaker’s own view. In the later claim, the speaker explicitly states their own position: “constraints are incredibly valuable,” while attributing the scaling-focused position to Rich Sutton. These passages do not establish that the speaker changed their own belief.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker suspects the field is making mistaken assumptions and says it is difficult to predict when researchers will move beyond them.Now The speaker believes people give too little attention to inference because human activity does not often resemble inference.The earlier claim says the field may be making “bad assumptions” and that it is hard to predict when researchers will move past them. The later claim concerns people giving too little attention to inference. These address different issues, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker considers learning models of distributions a very useful cognitive skill, but not the only important cognitive skill.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier statement says learning models of distributions is “a super helpful cognitive skill” but “not the only one.” The later statement agrees that people may be “reading too much into” similarities between model representations. These address different questions, so the later remark does not establish a change in the speaker’s view of cognitive skills.
• NOT COMPARABLE · 96%24 Jun → 21 SeptThen Human reasoning is a low benchmark for machines because humans have cognitive biases, illusions, and memory problems.Now The speaker believes people give too little attention to inference because human activity does not often resemble inference.The earlier claim says human reasoning is a “low bar” for machines because of “cognitive biases and illusions” and memory problems. The later claim says people give too little attention to inference because human activity “does not resemble inference” much. These are different points about human cognition, and the later statement does not reverse the earlier one.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker estimates that O3 chooses a legal chess move with roughly 80% reliability.Now SPEAKER_02 believes that the method used to reach, constrain, and shape the search space is critical to making the approach work.The earlier claim estimates that O3 chooses a legal chess move with “maybe 80% reliability.” The later claim says that how one “constrain[s] and shape[s] the search space” is critical to making an approach work. These address different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker is sure he can devise non-orthodox chess variations, while acknowledging that some may already be known.Now The speaker considers it exciting that different model families may converge on similar geometric representations of abstract structural invariants in the world.The earlier claim is about the speaker's ability to devise non-orthodox chess variations (“sure that I will be able to come up with variations on chess”). The later claim concerns whether different model families converge on similar geometric representations. These are different questions, so the later statement does not indicate a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker says his book had three core chapters covering operations over variables, structured representations, and types and tokens (or kinds of individuals).Now SPEAKER_02 characterizes Rich Sutton’s “Bitter Lesson” as arguing against building knowledge and constraints into systems and in favor of scaling models of computation.These statements address different things. The earlier claim describes the speaker’s book as having chapters on “operations over variables,” “structured representations,” and “types and tokens.” The later claim characterizes Rich Sutton’s view as opposing “building knowledge and constraints into the system.” It does not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker says some instances of the problems can be solved, but the problems he observes appear to reflect the same underlying core problems.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim concerns whether longstanding problems have been solved: “some instances ... can be solved” but observed problems reflect “the same underlying core problems.” The later claim concerns whether researchers are overinterpreting similarities in model representations. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%24 Jun → 21 SeptThen The speaker believes the problems he described in his book remain unsolved.Now The speaker says that people no longer know what to believe because claims that were confidently made a few years ago have become uncertain.The earlier claim is specifically that the problems described in his book “have [not] been solved.” The later passage says “we don't even know what to believe anymore” about claims made confidently a few years ago, then lists broader questions. It does not say whether the specific problems from the book have since been solved, so there is no clear change of position to compare.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker says he identified a set of real gaps in 2001.Now SPEAKER_02 believes that the method used to reach, constrain, and shape the search space is critical to making the approach work.The earlier claim concerns whether the speaker identified gaps that were real in 2001. The later claim says that how a search space is reached and shaped is critical to making an approach work. These address different questions, and the later claim does not affirm or deny the earlier one.
• NOT COMPARABLE · 96%24 Jun → 21 SeptThen Some critical ideas needed for further AI progress are currently missing, according to the speaker.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier statement says “we’re missing some critical ideas” for further AI progress. The later statement says “constraints are incredibly valuable,” with knowledge given as an example. These can both be true; the later claim does not say those ideas are no longer missing or that constraints alone resolve the gap.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker says GPT-3 has produced an illegal chess move in which a queen jumps over a knight.Now The speaker describes AlphaGo's inference process as structured: it used Monte Carlo tree search to update visited nodes algorithmically, feed the results back, and retrain the model.The earlier claim reports GPT-3 making an illegal chess move (“a queen jumps over a knight”). The later claim describes AlphaGo’s structured inference process using Monte Carlo tree search. These concern different systems and different questions, so the later statement does not show a change in the speaker’s position about GPT-3 or chess reasoning.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker and Jan LeCun both emphasize world models, but have somewhat different ideas about what world models are.Now The speaker considers it exciting that different model families may converge on similar geometric representations of abstract structural invariants in the world.The earlier claim concerns the speaker and Jan LeCun both emphasizing world models while having different ideas about them. The later claim says convergent geometric representations across model families are “very very exciting,” but does not state whether the speaker’s or LeCun’s views on world models changed. These claims do not establish a change in position on the same question.
• NOT COMPARABLE · 96%24 Jun → 21 SeptThen The speaker and Jan LeCun are dissatisfied with current AI capabilities described as reasoning because they consider them insufficiently robust.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier claim says current AI systems’ reasoning is not “robust enough.” The later claim says constraints, broadly including knowledge, are “incredibly valuable.” These address different questions and can both be true; the statements do not show a change in position on the same underlying question.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker says AI has made substantial progress in interpolation but has not solved extrapolation, and characterizes much of the progress as interpolation rather than extrapolation.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier claim concerns the extent of AI's progress on interpolation versus extrapolation: it says progress has been substantial on interpolation, while extrapolation remains unsolved. The later claim expresses excitement about inference-time computation and adaptation, but does not say whether this solves extrapolation or change the earlier assessment. These are not clearly opposing positions.
• NOT COMPARABLE · 98%20 Aug → 21 SeptThen The interviewer describes a prior conception of intelligence as traversing the combinational closure of core knowledge and abstractions in the brain.Now The interviewer describes physical constraints as potentially consisting of the laws of physics and mental constraints as knowledge already acquired.The claims address different questions. The earlier claim describes a prior conception of intelligence as “traversing the combinational closure” of knowledge and abstractions. The later claim describes constraints, including “the laws of physics” and “things that we already know.” These statements do not establish a change in the speaker’s position on the same underlying question.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker says O3 is unable to avoid making illegal chess moves.Now SPEAKER_02 believes that the method used to reach, constrain, and shape the search space is critical to making the approach work.The earlier claim is that O3 “can't avoid illegal moves” in chess. The later claim says the method used to “constrain and shape the search space” is critical to making an approach work. It does not say whether O3 can avoid illegal moves, and the later statement concerns a different, unspecified approach, so these claims do not establish a change in position.
• NOT COMPARABLE · 93%20 Aug → 21 SeptThen The interviewer describes a prior conception of intelligence as traversing the combinational closure of core knowledge and abstractions in the brain.Now The interviewer considers constraints extremely valuable and defines them broadly as knowledge operating at multiple levels of abstraction.The earlier passage describes a prior conception of intelligence—“we thought” it involved traversing core knowledge and abstractions—and then says intelligence is “very situated.” The later passage asserts that constraints are “incredibly valuable,” broadly including knowledge. These concern related ideas but do not establish opposing positions on the same question; the earlier claim is about how intelligence was previously understood, while the later one is about the value of constraints in systems.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker says Herb Simon's 1957 classical technique for Tower of Hanoi generalizes to arbitrary lengths, while LLMs do not.Now The speaker says Clement Bonet's paper on searching latent spaces involved searching a decoder to find the best programs to generate for ARC.The earlier claim concerns whether LLMs generalize to arbitrary lengths on Tower of Hanoi; the later claim describes searching a decoder for programs on ARC. These are different tasks and claims, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen A separate agent benchmark graph used a seconds-to-tens-of-seconds time scale, unlike the coding-task graph's minutes-to-hours-or-days scale.Now The speaker says Goodfire researchers use PCA projections and Bézier curves on model activations for days of the week and months of the year, finding geometric structures in activation space.The earlier claim compares time scales in two benchmark graphs: the coding-task graph used “minutes to hours or days,” while the agent graph used “seconds to tens of seconds.” The later claim describes Goodfire’s geometric analyses of model activations using PCA and Bézier curves. It does not address the benchmark comparison or express a conflicting position, so no change of mind can be inferred.
• NOT COMPARABLE · 98%20 Aug → 21 SeptThen The interviewer says the prior conception held that greater intelligence enables faster traversal of the combinational closure of knowledge and abstractions.Now The interviewer argues that a theory is needed to explain phenomena including double descent, apparent overcoming of the curse of dimensionality, systematicity, productivity, and models' ability to produce novel syntactically competent sentences.The earlier claim describes a prior conception—“we thought that intelligence was” faster traversal of knowledge and abstractions. The later claim says a theory is needed to explain several model phenomena. That does not establish a different current position on the same question, so a change of mind cannot be inferred.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker doubts that LLMs can infer causal laws well and does not expect them anytime soon to produce Einstein-level, radically different solutions to problems.Now The speaker believes people give too little attention to inference because human activity does not often resemble inference.The earlier claim concerns whether LLMs can infer causal laws or produce radically different, Einstein-level solutions. The later claim says people give too little attention to inference because human activity does not often resemble it. These address different questions, so the later statement does not establish a change in the speaker’s view of LLMs or AGI.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker suggests that a Go system trained on a standard 21-by-21 board could play on rectangular boards and handle other variations.Now The speaker describes AlphaGo's inference process as structured: it used Monte Carlo tree search to update visited nodes algorithmically, feed the results back, and retrain the model.The earlier claim concerns whether a Go system trained on a standard board can generalize to rectangular boards and other variations. The later claim describes AlphaGo’s inference process—“Monte Carlo tree search” updating visited nodes and feeding results back for retraining. These address different questions, and the later statement does not affirm or deny the earlier claim.
• NOT COMPARABLE · 98%20 Aug → 21 SeptThen The interviewer argues that people with different perspectives can use their intelligence to reach different answers about the same phenomena.Now The interviewer finds it exciting that different model families might learn similar geometric representations of abstract structural invariances in the world.The earlier claim is an extracted summary of the interviewer’s point about people with different perspectives reaching different answers. The later claim concerns whether different model families learn similar geometric representations. These address different questions, and the later passage expresses excitement while cautioning that “we’re reading too much into that”; it does not retract the earlier point.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker says Herb Simon's 1957 classical technique for Tower of Hanoi generalizes to arbitrary lengths, while LLMs do not.Now The speaker considers it exciting that different model families may converge on similar geometric representations of abstract structural invariants in the world.The statements address different questions. The earlier claim concerns whether LLMs generalize to arbitrary-length Tower of Hanoi problems; the later claim concerns whether different model families learn similar geometric representations of abstract structure. The later statement does not revise or deny the earlier one.
• NOT COMPARABLE · 98%20 Aug → 21 SeptThen The interviewer argues that intelligence is situated rather than as abstract and pure as people had thought.Now The interviewer describes physical constraints as potentially consisting of the laws of physics and mental constraints as knowledge already acquired.The earlier passage is attributed to Tim Scarfe and describes intelligence as “very situated” rather than “abstract and pure.” The later extracted claim is also about Tim Scarfe, but it describes constraints at different levels—physical laws and knowledge in our minds—not whether intelligence is situated or abstract. These claims address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker says humans are extremely data-efficient on many, but not all, problems.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier statement concerns humans’ data efficiency on some problems (“fabulously data efficient on many problems”), while the later statement expresses excitement about inference-time computation and adaptation and says few people discuss it. These address different questions, so the later claim does not indicate a change in the earlier view.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The early LLM–Wolfram Alpha approach was limited by the interface through which the model accessed tools.Now According to a colleague quoted by the interviewer, RLV helps with exploration because the base language model has substantial conceptual understanding.The earlier claim is Tim Scarfe’s own assessment that the early approach was limited by reliable tool access. The later claim is explicitly attributed to a colleague and concerns RLV helping exploration through a language model’s conceptual understanding. It does not establish a change in Scarfe’s own position on the earlier claim.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker has heard, but cannot verify, that OpenAI staff search social media—including his posts—for examples of model failures.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The claims address different questions. Earlier, the speaker says he was told—while noting he cannot verify it—that OpenAI staff look on social media for examples. Later, he agrees that observers may be “reading too much into” similarities in model representations. The later claim does not confirm or deny the earlier one.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker recalls a video of a child completing a seven-disc Tower of Hanoi puzzle in about two minutes and believes the child could probably complete eight discs because the solution is recursive.Now The speaker says that in the spline-theory account, a small input change can cause the representation to jump to a different honeycomb region elsewhere in the space.The earlier claim concerns a child’s ability to solve a Tower of Hanoi puzzle, while the later claim concerns how small input changes affect representations in a model’s activation space. They address unrelated questions, so neither statement indicates a change in the other position.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker believes it is difficult to forecast when new ideas will emerge.Now The speaker believes people give too little attention to inference because human activity does not often resemble inference.The earlier claim is about how difficult it is to forecast when new ideas will emerge (“new ideas are hard to project”). The later claim is about people giving too little attention to inference because human activity does not resemble inference. These concern different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen Augmented data is not a universal solution; it works better for tasks such as mathematics, where the quality of the augmented data can be verified.Now Machine-learning practitioners work with finite data samples and may need to interpolate between observed data points.The earlier claim concerns whether augmented data is a universal solution and where it works best. The later claim concerns finite samples and interpolation between data points. These address different questions, and the later statement does not establish a changed view about augmented data.
• NOT COMPARABLE · 94%24 Jun → 21 SeptThen The speaker considers the semantics represented by current AI models superficial and primarily based on word distributions.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim says current models’ semantics are “very superficial” and “really about distributions of words.” The later claim is a caution that observers may be “reading too much into” similarities in the geometry of model representations. These claims address related but distinct points and can both be true; the later statement does not clearly revise the earlier view of model semantics.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker considers 50% performance an unusual and problematic measure for interpreting the graph.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim concerns whether “50% performance” is a useful measure for interpreting a graph. The later claim concerns whether people are reading too much into similarities between models’ learned representations. These are different questions, so the later statement does not establish a change in the speaker’s view about the measure.
• NOT COMPARABLE · 91%24 Jun → 21 SeptThen The speaker’s tentative view is that human reinforcement learning has provided only limited help with AI alignment.Now SPEAKER_02 believes constraints, broadly understood to include knowledge, are extremely valuable in AI systems.The earlier claim says, tentatively, that on alignment “all we have is maybe a human reinforcement learning,” suggesting limited available help. The later claim says “constraints are incredibly valuable” and includes knowledge among those constraints. That is an endorsement of a particular approach, but it does not clearly retract the earlier assessment; the statements address different aspects of alignment and do not establish a changed position.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The interviewer predicts that no durable advantage will persist if AI development remains based on LLMs.Now The interviewer says language models used in RLV have substantial base knowledge.The earlier claim concerns whether LLMs will have a durable advantage; the later claim concerns language models’ base knowledge in RLV. These address different questions, so the later statement does not establish a changed position on durability.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The interviewer considers the idea of a durable advantage in large language models to be a myth, assuming LLMs are the relevant technology.Now According to a colleague quoted by the interviewer, RLV helps with exploration because the base language model has substantial conceptual understanding.The earlier claim says durable advantage in LLMs is “a myth.” The later passage attributes the point about RLV and a base LLM’s conceptual understanding to “this guy Chris,” rather than clearly stating the speaker’s own position. It also addresses exploration, not whether durable advantage exists, so it does not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker says machine performance on the LiveCodeBench Pro benchmark was 0%, and that he does not know the human performance.Now The interviewer considers inference-time computation very exciting and believes relatively few people are discussing it.The earlier statement concerns machine performance on LiveCodeBench Pro, while the later statement expresses enthusiasm about inference-time computation and says few people discuss it. They address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker is concerned that the extent of data contamination and augmentation in the coding benchmarks is unknown.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier statement says the speaker is concerned that contamination and augmentation in coding benchmarks are not well understood. The later statement says inference-time computation and adaptation are exciting and under-discussed. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker says human cognition is impaired by confirmation bias and motivated reasoning.Now The speaker believes people give too little attention to inference because human activity does not often resemble inference.The earlier statement says human cognition is affected by “confirmation bias and motivated reasoning.” The later statement says people “don’t really do things that resemble inference that much,” so they give inference too little attention. These address different questions and can both be true; the later claim does not show a change in the speaker’s view about cognitive bias.
• NOT COMPARABLE · 96%24 Jun → 21 SeptThen The speaker considers the semantics represented by current models very superficial, describing them as primarily distributions of words.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim says current models’ semantics are “very superficial” and “really about distributions of words.” The later claim only agrees that people may be “reading too much into” similarities in geometric representations learned by different models. That does not clearly revise the earlier view about how semantic understanding works; the claims address related but distinct points.
• NOT COMPARABLE · 96%24 Jun → 21 SeptThen The speaker suspects the field is making mistaken assumptions and says it is difficult to predict when researchers will move beyond them.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim says the field is making “some bad assumptions” and that it is hard to predict when researchers will move past them. The later claim agrees that people may be “reading too much into” similarities in model representations. These concern different specific questions, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker considers voluntary regulation and industry self-regulation probably insufficient to ensure disclosure of AI model capabilities.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier statement concerns whether voluntary and industry self-regulation can ensure disclosure of AI capabilities; the later concerns whether observers are overinterpreting similarities in model representations. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker estimated, with uncertainty, that GPT-2 used about 5%–10% of the internet.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier claim estimates GPT-2's training-data coverage (“maybe 10% ... or 5%”), while the later claim expresses excitement about inference-time computation and adaptation. They address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%20 Aug → 21 SeptThen The interviewer argues that people with different perspectives can use their intelligence to reach different answers about the same phenomena.Now The interviewer believes that how a search space is reached, constrained, and shaped is critical to making the approach work.The earlier extracted claim attributes a view to the interviewer about people reaching different answers from different perspectives. The later claim attributes to the interviewer a view about shaping a search space. These statements address related but distinct points, and the later one does not clearly affirm or deny the earlier position, so a change of mind cannot be established.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker says Herb Simon's 1957 classical technique for Tower of Hanoi generalizes to arbitrary lengths, while LLMs do not.Now The speaker says that double descent, apparent overcoming of the curse of dimensionality, syntactic competence, novel sentence production, and abstract representations are phenomena that need explanation in deep learning.The earlier claim concerns a specific comparison: Herb Simon’s Tower of Hanoi technique “generalizes to arbitrary length,” while “LLMs do not.” The later claim says several deep-learning phenomena, including novel sentence production and apparent overcoming of the curse of dimensionality, “need explanation.” It does not state whether LLMs generalize to arbitrary-length Tower of Hanoi problems, so the claims do not establish a change in position.
• NOT COMPARABLE · 91%20 Aug → 21 SeptThen The interviewer says the prior conception held that greater intelligence enables faster traversal of the combinational closure of knowledge and abstractions.Now The interviewer describes physical constraints as potentially consisting of the laws of physics and mental constraints as knowledge already acquired.The earlier claim describes a conception of intelligence as faster traversal of knowledge and abstractions. The later claim asserts that “constraints are incredibly valuable” and treats knowledge as a kind of constraint. These positions can coexist; the later statement does not clearly reject the earlier conception or establish a change in the speaker’s view of intelligence.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker says the participants' forecasts are closer than people may assume: one concentrates probability at five years out and later, while the other assigns some probability to two or three years out, and both assign some probability to 2045.Now The speaker says that people no longer know what to believe because claims that were confidently made a few years ago have become uncertain.The earlier claim compares two participants’ forecasting timelines: one puts probability at five years or later, while the other assigns some to two or three years out. The later claim says “we don't even know what to believe anymore” about claims that were made confidently in the past. These address different questions, so the later statement does not establish a change in the speaker’s forecasting position.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker says his book had three core chapters covering operations over variables, structured representations, and types and tokens (or kinds of individuals).Now SPEAKER_02 describes constraints as operating at multiple levels of abstraction, including physical laws in the world and existing knowledge in minds.The earlier claim describes the contents of the speaker’s book—chapters on “operations over variables,” “structured representations,” and “types and tokens.” The later claim expresses a view about the value of constraints and knowledge in systems. These address different questions, so they do not show whether the speaker changed position.
• NOT COMPARABLE · 97%24 Jun → 21 SeptThen The speaker attributes most recent AI interpolation progress to increased data and compute.Now The speaker says that double descent, apparent overcoming of the curse of dimensionality, syntactic competence, novel sentence production, and abstract representations are phenomena that need explanation in deep learning.The claims concern different questions. The earlier statement attributes most recent progress in AI interpolation to “more data and more compute.” The later statement lists phenomena in deep learning—including double descent and syntactic competence—that need explanation; it does not say what caused interpolation progress or reject the earlier explanation.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker considers the graph's axis very arbitrary.Now The speaker considers it exciting that different model families may converge on similar geometric representations of abstract structural invariants in the world.The earlier statement criticizes a graph's axis as “very arbitrary,” while the later statement finds convergence in models’ geometric representations “very very exciting.” These address different aspects of visualized representations, and the later passage does not say the graph axis is no longer arbitrary or otherwise reverse that criticism.
• NOT COMPARABLE · 97%24 Jun → 21 SeptThen When Claude uses an interpreter, the resulting system performs operations over variables in the Python code Claude generates.Now The interviewer says vision and flow models can use sophisticated inference, including latent-space or decoder search for programs that solve ARC tasks, along with test-time computation and adaptation.The earlier claim concerns Claude using an interpreter to perform variable operations in generated Python, in contrast with what a pure neural network can do. The later excerpt discusses sophisticated inference and search in vision and flow models, but the extracted claim is attributed to the interviewer, not Tim Scarfe, and does not state a changed position about interpreter-based variable operations. These are different topics and the later passage does not establish Tim's own position.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker says machine performance on the LiveCodeBench Pro benchmark was 0%, and that he does not know the human performance.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier statement reports a specific benchmark result—machine performance on LiveCodeBench Pro was “zero percent”—and says the speaker does not know the human result. The later statement expresses enthusiasm for inference-time computation and adaptation. These address different questions, so the later view does not establish a change in the speaker’s position on the benchmark result.
• NOT COMPARABLE · 91%24 Jun → 21 SeptThen The speaker says he has not really seen the gaps he has raised since 1998 solved over roughly 27 years, and finds it strange to expect them all to be solved within three years.Now The speaker says that people no longer know what to believe because claims that were confidently made a few years ago have become uncertain.The earlier claim is specifically that longstanding gaps have not been solved and that solving them within three years seems unlikely. The later claim—“we don't even know what to believe anymore”—concerns uncertainty about past confident claims, and does not say those gaps have since been solved or that the earlier timeline judgment changed.
• NOT COMPARABLE · 96%24 Jun → 21 SeptThen The speaker and Jan LeCun both emphasize world models, but have somewhat different ideas about what world models are.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim says they “both emphasize world models” while having “slightly different ideas” about them. The later claim agrees that people may be “reading too much into” similarities between model representations. That is a comment on how to interpret representational similarities, not a denial of the earlier position about emphasizing world models.
• NOT COMPARABLE · 99%24 Jun → 21 SeptThen The speaker considers Grok's failure to understand the three-in-a-row condition in the tic-tac-toe variant a conceptual weakness.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim concerns Grok’s failure to recognize a three-in-a-row condition as a “conceptual weakness.” The later claim concerns whether observers are “reading too much into” similarities between model representations. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker thinks Jan LeCun's approach is closer to his own, though LeCun would probably use different litmus tests.Now The speaker is very excited about computation and adaptation at inference time and believes relatively few people discuss this area.The earlier statement says LeCun’s approach is “closer to mine,” while the later statement expresses excitement about “work at inference time.” These address different aspects of the speaker’s views, and the later claim does not establish whether his view of LeCun’s approach has changed.
• NOT COMPARABLE · 98%20 Aug → 21 SeptThen The interviewer argues that intelligence is situated rather than as abstract and pure as people had thought.Now The interviewer considers constraints extremely valuable and defines them broadly as knowledge operating at multiple levels of abstraction.The earlier claim is attributed to the interviewer as an argument that intelligence is “situated” and “not quite as abstract and pure” as previously thought. The later claim is also the interviewer’s view, that “constraints are incredibly valuable,” including knowledge in the mind. These address related but distinct questions, and the excerpts do not establish that the interviewer changed their position on either one.
• NOT COMPARABLE · 98%20 Aug → 21 SeptThen AI models acquire some form of coarse-grained abstraction or internal structure during training.Now The speaker describes AlphaGo's inference process as structured: it used Monte Carlo tree search to update visited nodes algorithmically, feed the results back, and retrain the model.The earlier claim says models acquire “some kind of coarse grained abstraction” during training. The later claim describes AlphaGo’s inference process—Monte Carlo tree search updating visited nodes and feeding results back for retraining. These address different aspects of a system, so the later statement does not establish a changed position about the earlier claim.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker estimates that developing much better reasoning and planning and more stable world models will take five or ten years.Now The speaker says vision models and flow models perform sophisticated inference processes.The earlier statement estimates a timeline for developing better reasoning, planning, and more stable world models (“five or ten years”). The later statement says current vision and flow models already perform “quite sophisticated inferencing processes.” These address different capabilities and do not establish that the speaker changed their view about the earlier timeline.
• NOT COMPARABLE · 98%24 Jun → 21 SeptThen The speaker thinks Jan LeCun's approach is closer to his own, though LeCun would probably use different litmus tests.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim says LeCun’s approach is “closer to mine,” while noting he may use different litmus tests. The later claim is about whether observers are “reading too much into” similarities in model representations. These address different questions, and the later passage does not establish a changed view about how closely LeCun’s approach matches the speaker’s.
• NOT COMPARABLE · 95%24 Jun → 21 SeptThen The speaker considers the graph's axis very arbitrary.Now The speaker agrees that observers may be overinterpreting similarities between model representations.The earlier claim says the graph’s axis is “very arbitrary,” while the later claim agrees that people “may be reading too much into” similarities between model representations. These statements concern different aspects of the analysis, and the later passage does not say whether the speaker’s view of the graph’s axis has changed.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker expects future periods of AI progress to exceed the progress of the last few years and expects AI to be enormously more advanced in 30 years.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim predicts that AI will make more progress in future periods and be far more advanced in 30 years. The later claim says an automated AI scientist may be possible in the near future. These are related but distinct claims; the later one does not establish a changed position on the earlier forecast.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says current AI systems may refuse an obvious biological-weapons request, but such refusals are easily bypassed through jailbreaking.Now The speaker says composable AI agents would make them much more worried.The earlier claim is that obvious safety refusals are easily jailbroken and alignment systems do not reliably do what users want. The later claim is conditional: “if” AI agents were composable, that would make the speaker “much more worried.” It does not revise or deny the earlier assessment; the statements address different aspects of AI safety, and the later one explores a hypothetical.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that safety-critical AI uses should wait for a better alignment solution if systems cannot reliably refuse harmful requests and prevent foreseeable human harm, even if the solution takes five or 500 years.Now The speaker says composable AI agents would make them much more worried.The earlier claim concerns delaying safety-critical AI uses until systems can reliably refuse harmful requests and prevent foreseeable harm. The later claim says that composable agents would make the speaker “much more worried.” These address different aspects of AI safety, and the later statement does not establish a changed position on the earlier question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker estimates that most of the timeline probability distribution lies beyond 10 years.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier claim estimates a timeline, saying “most of the distribution is past 10 years.” The later claim only says “There are just very complex dynamics,” about external intelligence systems. It does not state a different timeline estimate, so the claims do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Sam and Dario have hinted that they will soon be able to automate the full AI idea-to-implementation-and-scaling process, but the speaker is skeptical of those claims.Now The speaker recalls that the report may project 60 million AI agents operating at 20 times human speed by 2032.The earlier claim says the speaker is “not really buying” Sam and Dario’s claims about soon automating the AI idea-to-implementation process. The later claim only recalls a possible report projection—“60 million agents” by 2032—and does not say whether the speaker believes or endorses it. These are different claims, so the later statement does not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Sam discussed AI risks more often in the past than he does now.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim is about Sam's past and present frequency of discussing AI risks: he "used to talk more" about them and "talks less now." The later claim describes possible harmful behavior in a future wave of AI agents. It does not establish whether the speaker's view about Sam's discussion frequency changed, and the statements address different questions.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that AI systems will continue improving regularly and provide increasing value.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim says AI systems will “continue to improve regularly” and provide “more and more value.” The later claim describes a possible next wave of agents that communicate and specialize, while noting potential harms such as collusion. These claims do not directly conflict: the later statement does not retract the prediction of increasing value, and it addresses a more specific development and its risks.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says human cognition is constrained by a lack of location-addressable memory, a point he argued in his book Kludge.Now The speaker says current agents adapt through chain-of-thought processes and skill-surface adaptation, essentially using memory systems.The earlier claim concerns a human cognitive limitation—“a lack of location-addressable memory”—while the later claim describes how current AI agents adapt, “through chain of thoughts” and “skill surface adaptation,” using memory systems. These are different subjects and questions, so the later statement does not show a changed position about human cognition.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes Einstein-level innovations will be necessary to achieve AGI and, certainly, ASI.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says “Einstein-level innovations” will be needed for AGI and ASI. The later claim imagines an efficiently self-adapting recursive loop as “a different type of technology,” but does not say whether that would supply such innovations or make them unnecessary. The later statement therefore does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker predicts that by 2035 people will regard LLMs as an incomplete answer, although they will continue using them for some tasks.Now In the speaker's account, construct validity is checked by an LLM judge that ensures the system is not cheating and is doing the task correctly.The earlier statement predicts that by 2035 people will see LLMs as “not really the answer,” while still using them for some tasks. The later statement describes an LLM judge checking whether an experiment is conducted correctly; it does not state a view about LLMs’ future role or whether they are an adequate answer. These address different questions, so the later claim neither confirms nor changes the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker believes the unconstrained AGI they fear should not exhibit many of the edge-case failures being discussed.Now The speaker says composable AI agents would make them much more worried.The earlier claim concerns whether feared unconstrained AGI should exhibit many of the edge-case failures under discussion. The later claim concerns whether composable AI agents would increase the speaker’s worry. These are different concerns, and the statements do not establish that the speaker changed their position on the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes LLMs generally produce inside-the-box solutions, whereas achieving AGI will require outside-the-box innovations.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim concerns whether LLMs can produce “outside-the-box” innovations needed for AGI. The later claim concerns how expert guidance affects AI performance in “well-specified domains” and the ambiguity that remains. These address different questions, so the later statement does not establish a change in the speaker’s view about AGI or innovation.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker characterizes most LLM accomplishments as exploiting existing ideas rather than discovering new ones.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier claim concerns what LLMs have mostly accomplished—“exploiting existing ideas” rather than “discovery of new ideas.” The later claim concerns Elon Musk’s strength in recognizing future opportunities rather than engineering or optimization. These address different subjects and questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker estimates that developing much better reasoning and planning and more stable world models will take five or ten years.Now Fine-tuning a large model is very difficult, according to the speaker.The earlier statement estimates how long it will take to develop better reasoning, planning, and world models (“five or ten years”). The later statement says fine-tuning a large model is difficult. These address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker thinks a global treaty prohibiting work on these kinds of AI systems, with reporting requirements, would require substantial political will and is probably unlikely.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier statement concerns whether governments could establish a treaty prohibiting work on certain AI systems, and says it would take substantial political will and is probably unlikely. The later statement predicts that AI agents may communicate and exhibit harmful behaviors such as collusion. It does not say whether the speaker supports or opposes such a treaty, so the claims do not establish a change of position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The interviewer does not expect any actor to gain a durable advantage if the AI paradigm remains roughly unchanged.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier statement says that if “the paradigm stays roughly like it is,” the speaker does not expect anyone to gain “a durable advantage.” The later statement concerns expert-guided AI in “well-specified domains” and a “supervisor illusion”; it does not say whether any actor can gain a durable advantage. The claims address different questions, so the later passage does not establish a change of mind.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker says AI systems need a deeper kind of semantics than current models provide.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim is that AI needs “a deeper kind of semantics” than current models provide. The later claim concerns whether AI skills or agents are “composable,” and says that idea does not match the speaker’s experience. These are related topics, but the excerpts do not establish that the speaker changed their position on the same question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes OpenAI is not fully forthcoming about what it did.Now The speaker believes Elon Musk has less agency than the listener thinks he has.The claims address unrelated questions. The earlier statement says OpenAI is “not an entirely forthcoming” about what it did; the later says Elon Musk “doesn't have quite as much agency” as the listener thinks. Neither statement indicates a change in the other position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker does not currently see a technical solution for coping with the problems posed by intelligence recursion.Now The speaker says grounding is not necessarily an argument that intelligence requires physical instantiation or consciousness.The earlier claim is that the speaker does not see a technical way to cope with intelligence recursion “right now.” The later claim is that grounding does not necessarily imply intelligence requires “physical instantiation” or “consciousness.” These address different questions, and the later statement does not say whether a technical solution to recursion exists.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker does not currently see a technical solution for coping with the problems posed by intelligence recursion.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim is about whether there is currently a technical solution to problems posed by intelligence recursion (“I don't see any technical way of coping ... right now”). The later claim predicts that agents may communicate and exhibit harmful behaviors such as collusion; it does not say whether those problems have a technical solution. These statements do not establish a change in position on the same question.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says AI systems need “a deeper” form of semantics. The later claim says the speaker’s experience does not align with AI skills being composable by merging agents. These concern related but distinct questions, and the statements do not establish a change in the speaker’s position on the same underlying issue.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that much of the near-term experimentation people conduct on LLMs can be automated.Now In the speaker's account, construct validity is checked by an LLM judge that ensures the system is not cheating and is doing the task correctly.The earlier claim is that many near-term LLM experiments can be automated. The later passage explains how an LLM judge checks construct validity in a particular experimental setup. It does not state a changed view about whether experimentation can be automated, and the claims address different aspects of the topic.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker characterizes himself and Yann LeCun as taking architectural or cognitive approaches to AI.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement says the speaker and Yann LeCun take “an architectural approach or a cognitive approach” to AI. The later statement imagines future AI as “diverse AIs” working from different angles “rather than one big AI.” These statements address related but distinct questions, and the earlier claim does not establish a position that the later claim reverses.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen A more sophisticated AGI forecast represents probability across multiple possible arrival years rather than naming only one predicted year.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim concerns how to express an AGI forecast—using probability across multiple possible arrival years rather than naming one year. The later claim concerns expert-guided AI performance in well-specified domains and remaining ambiguity. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 89%24 Jun → 8 SeptThen The speaker says that some very old AI systems can outperform modern general-purpose AI systems.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The claims address different abilities. The earlier statement says some old AI systems can outperform modern general-purpose systems on a domain-specific task (Tower of Hanoi); the later says models can combine existing knowledge in novel ways to solve tasks. The later claim does not deny the earlier comparison, so these statements do not establish a change of mind.
• NOT COMPARABLE · 93%24 Jun → 8 SeptThen At the time of the conversation, AI systems were not very differentiated, with most actors using LLMs and reasoning models.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says AI systems were “not very differentiated,” with most using LLMs and reasoning models. The later statement concerns whether AI skills are composable, saying that the speaker’s experience “doesn’t really marry” with that idea. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks neurosymbolic methods are worth exploring for near-term safety-critical AI applications, while remaining unsure whether they help with superintelligence and saying they might offer some benefit.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier claim concerns whether neurosymbolic methods may help with near-term safety-critical AI applications, while the later claim concerns whether a mixed AI-and-human ecosystem can participate in a superorganism. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Since at least 1998, the speaker has focused on how AI systems generalize beyond their training; he says transparency about this is lacking.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement concerns a longstanding research question—how AI systems generalize beyond training and the lack of transparency. The later statement describes a preferred imagined future structure: “diverse AIs” examining problems from different angles “rather than one big AI.” These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says driverless cars have made progress.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier claim says only that driverless cars “have made progress.” The later claim concerns Elon Musk’s strengths—his ability to identify promising future areas rather than engineering or optimization. These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Open-weight models allow essentially anyone to participate in AI development to some degree.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim concerns access and participation: open-weight models let “essentially anyone” participate in AI development “to some degree.” The later claim concerns model capabilities: models can combine known things in novel ways to solve a task. These address different questions, so the later statement does not establish a change in the speaker’s position on the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen An early attempt to combine an LLM with a symbolic system by integrating it into Wolfram Alpha produced inconsistent results.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier statement describes an early LLM–symbolic-system integration as producing “hit or miss” results, specifically due to unreliable tool access. The later statement says expert use can produce “magical” results in well-specified domains, while ambiguity remains. These concern different setups and conditions, so the later claim does not establish a changed view about the earlier attempt.
• NOT COMPARABLE · 92%24 Jun → 8 SeptThen The speaker regards language models using tools as a form of neurosymbolic AI, though not the right form.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says language models using tools are “a species of neurosymbolic AI,” but “not the right one.” The later claim says models can combine existing knowledge in novel ways to solve tasks. These address different questions—how to classify a kind of AI versus what models can do—and the later claim does not establish a changed view about neurosymbolic AI.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says future AGI might use LLMs, which he considers good at distributional learning.Now The speaker thinks future weight adaptation might solve the problem of sharing knowledge among agents.The earlier claim says future AGI might use LLMs because they are good at “distributional learning.” The later claim speculates that “weight adaptation” might solve agents’ knowledge-sharing problem. These address different questions, and the later statement does not reverse the earlier view about LLMs.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes, and thinks Yann LeCun also believes, that LLMs are not the primary route to AGI and that achieved AGI will look substantially different.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is about whether LLMs are the route to AGI and whether AGI will look different. The later claim is that an automated AI scientist may be possible in the near future. That does not establish that the speaker now sees LLMs as the route to AGI or has changed their view of what AGI will look like.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says his book had three core chapters covering operations over variables, structured representations, and types and tokens (or kinds of individuals).Now The speaker says models appear to use representations at multiple levels of resolution along a spectrum of abstraction.The earlier claim says the speaker’s book had chapters on “operations over variables,” “structured representations,” and “types and tokens.” The later claim says models use representations at “multiple levels of resolution” along “a spectrum of representations.” These are not opposing positions on the same question: one describes the book’s contents, while the other makes a claim about how models represent information. The statements do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker considers the estimates for solving some of the discussed problems overly optimistic.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim concerns how quickly specific AI problems can be solved, calling the estimates “overly optimistic.” The later claim concerns whether a collective AI system would face limits on its scale. These are different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker believes AI will eventually acquire the missing capability for understanding concepts such as three-in-a-row, and says achieving it is not impossible.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says AI getting the missing capability to understand concepts like “three-in-a-row” is “not impossible.” The later claim says an “automated AI scientist” is possible “in the near future.” These concern related but distinct capabilities, and the later statement does not say whether such a scientist would have the earlier capability. The statements do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Human reasoning is a low benchmark for machines because humans have cognitive biases, illusions, and memory problems.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says humans are “a low bar” for machines because of human “cognitive biases and illusions” and memory problems. The later claim concerns expert-guided AI in “well-specified domains,” where results can seem “magical” while ambiguity remains. The later statement does not deny the earlier view; it addresses a more specific question about AI performance and its limits.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says human cognition is constrained by a lack of location-addressable memory, a point he argued in his book Kludge.Now The speaker believes intelligence circuits and motifs exist outside individuals.The earlier claim says human cognition is constrained by “a lack of location-addressable memory.” The later claim says intelligence circuits and motifs “exist outside” individuals. These address different questions, and the later statement does not deny or revise the earlier claim.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen A system in which a neural network creates the symbolic tools needed for a problem is a neurosymbolic solution.Now Models capture the symbolic residue of language when language is ingested into them.The earlier claim defines a neurosymbolic solution as one where a neural network creates symbolic tools for a problem. The later claim says models capture the “symbolic residue of language” during ingestion. These address different properties of systems, and neither confirms nor contradicts the other, so they do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker cautions that explaining AI failures as failures on tasks the systems were not trained on is an oversimplification, while discussing AI risk.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim concerns whether AI failures on untrained tasks explain AI risk, and explicitly calls that explanation an oversimplification. The later claim predicts that interacting agents may learn, specialize, and exhibit harmful behaviors such as collusion. These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers it unlikely that LLMs will play chess without illegal moves and without tool use within the next six months.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim concerns whether LLMs could play chess without illegal moves or tool use within six months; the later claim concerns whether an automated AI scientist could be possible in the near future. These are different capabilities, so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker expects transparency to be less likely as AI systems approach AGI or become meaningfully differentiated.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier statement is about whether transparency will decline as AI approaches AGI or becomes more differentiated. The later statement predicts that AI agents may communicate, learn, specialize, and exhibit harmful behaviors. It does not state a view about transparency, so the claims do not establish a change of mind on the same question.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker says AGI must be able to solve original problems, not merely perform well through teaching to the test.Now The speaker rejects intelligence as a magical quality and says it depends on the scope of tasks and on existing knowledge.The earlier claim says AGI must solve “things that are original,” rather than merely succeed through “teaching to the test.” The later claim says intelligence is “fractured and jagged” and depends on “a scope of tasks” and “what you already know.” That does not clearly deny the earlier capability claim; the statements address related but distinct questions, so a change of mind cannot be established.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The interviewer considers the idea of a durable advantage in large language models to be a myth, assuming LLMs are the relevant technology.Now The speaker says reinforcement-learning-trained language models have substantial base knowledge and can perform things that were not possible before.The earlier statement says a “durable advantage in LLMs” is “a myth.” The later statement discusses models’ capabilities, base knowledge, and representations; it does not say whether durable advantages exist. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers it an exceptionally rapid convergence if the issues he has worried about for 25 years were all resolved by 2027.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim concerns how rapidly the speaker expects long-standing problems to be resolved: reaching 2027 with those issues resolved would be “everything everywhere all at once.” The later claim concerns whether a proposed AI-human collective or “super organism” would have “limits on its scale.” These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker treats obeying a simple constraint such as not making illegal chess moves as a small-scale alignment problem that AI systems still struggle to solve.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim is that models still struggle to follow constraints, including “don't make illegal moves in chess.” The later claim is that a model given a specified task “will hill climb towards” a solution. These address different aspects of model behavior, and the later statement does not say the model reliably completes the task or follows constraints, so it does not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker considers it unlikely that LLMs will play chess without illegal moves and without tool use within the next six months.Now In the speaker's account, construct validity is checked by an LLM judge that ensures the system is not cheating and is doing the task correctly.The earlier claim concerns whether LLMs could play chess without illegal moves and without tool use within six months. The later claim concerns an LLM judge checking whether a system is cheating and performing a task correctly. These statements address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker thinks the form of AGI relevant to the risks under discussion would probably exceed humans on short-term-memory tasks while matching human versatility and flexibility.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns the capabilities of risk-relevant AGI relative to humans: it would probably exceed people on short-term-memory tasks while having human-like versatility and flexibility. The later claim imagines a future made up of diverse AIs collaborating rather than “one big AI.” That describes a possible system design, not whether AI matches or exceeds human abilities, so the statements do not establish a change of mind.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker has advocated neurosymbolic AI since at least 2001.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says he has “always advocated neurosymbolic AI,” while the later claim says an automated AI scientist “is possible in the near future.” The later statement does not show whether he still advocates neurosymbolic AI or whether his view of it changed; the claims address different questions.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says AGI must be able to solve original problems, not merely perform well through teaching to the test.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim concerns whether AGI must solve original problems rather than merely perform well on test-like tasks. The later claim says an AI can hill-climb toward a specified mathematics task or answer a knowledge question. It does not say whether those tasks are original, whether the AI can solve them reliably, or whether this is the same capability being discussed, so the statements do not establish a change of position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Open-weight models allow essentially anyone to participate in AI development to some degree.Now The speaker says an AI's weights can be copied so that the AI can be run in a thousand instances.The earlier claim says open-weight models let “essentially anyone” participate in AI development to some degree. The later claim says weights can be copied and run “a thousand times.” These are related observations, but the later statement does not establish a changed position on who can participate in AI development.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Automating the full AI research-and-development process would be destabilizing if technically feasible, but not if the claims are merely hype and the capability cannot be achieved.Now The interviewer considers an entirely AI-run economy plausible but suspects it might become degenerate in some way.The earlier claim concerns whether automating AI research and development would destabilize things if technically feasible. The later claim concerns whether an entirely AI-run economy is plausible and might become degenerate. These are different questions, and the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes LLMs generally produce inside-the-box solutions, whereas achieving AGI will require outside-the-box innovations.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim is about the kind of innovation needed for AGI: LLMs produce “inside the box” solutions, while AGI will require “outside the box” innovations. The later claim says that on a specified task, such as mathematics, a model will “hill climb towards it.” That describes task-directed problem solving and does not say whether the model can produce the outside-the-box innovations the earlier claim said AGI requires.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker assesses AI progress primarily through cognitive-science problems rather than quantitative progress.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim describes the speaker’s approach to assessing AI progress: “more from the cognitive science.” The later claim describes an AI capability—being given a mathematics task and “hill climb[ing] towards it.” These statements address different aspects and do not establish that the speaker changed their position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says newer AI systems interpolate substantially better than earlier systems.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says newer systems are “much better at interpolating” than previous systems. The later claim discusses expert-guided AI producing “magical” results in well-specified domains while leaving “a bit of a space of ambiguity.” These address different aspects of AI performance, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker does not see current AI systems solving the qualitative cognitive problems he considers necessary to address.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns current AI systems, saying they do not solve the qualitative cognitive problems the speaker thinks need addressing. The later claim is a speculation about a possible future arrangement of diverse AIs working together, not a claim that current systems already solve those problems. These statements address different timeframes and do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker attributes most recent AI interpolation progress to increased data and compute.Now The speaker suggests that a distributed collective AI system might face additional bottlenecks due to communication among its agents.The earlier claim attributes recent AI interpolation progress mostly to “more data and more compute.” The later claim says distributed collective systems “might have additional bottlenecks” from communication among agents. These address different questions—what drove interpolation progress versus what bottlenecks distributed systems may face—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker considers the machine-intelligence level depicted toward the end of AI 2027 unlikely within the next decade and not very likely in the following decade, while allowing that it is possible.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says the level of machine intelligence depicted toward the end of AI 2027 is “not very likely” in the next decade or the decade after, though “certainly possible.” The later statement says an automated AI scientist is “possible in the near future.” These concern related but not clearly identical milestones: the later claim does not establish that the speaker now expects the specific intelligence level discussed earlier to arrive soon.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker characterizes most LLM accomplishments as exploiting existing ideas rather than discovering new ones.Now The speaker believes shortcuts and validity problems are less problematic in this research than he had expected.The earlier claim concerns whether LLMs discover new ideas or mainly exploit existing ones. The later claim concerns whether shortcuts and validity problems undermine a particular line of research. These are different questions; the later passage does not say the speaker changed their view about LLMs' capacity for discovery.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says GPT-3 has produced an illegal chess move in which a queen jumps over a knight.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The earlier claim is that GPT-3 made an illegal chess move: “a queen jump over a knight.” The later claim concerns whether AlphaGo’s Move 37 was creative or innovative. These address different questions, so the later statement does not show a change in the speaker’s view about GPT-3’s chess errors.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes that presenting a distribution of possible scenarios would be helpful.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier claim says presenting “a distribution of possible scenarios” would be helpful. The later claim says Elon Musk’s strength is recognizing “areas that are interesting that might work in the future.” These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Sam discussed AI risks more often in the past than he does now.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier statement is about Sam’s pattern of discussing AI risks—“he used to talk more about this kind of risk stuff and talks less now.” The later statement is the speaker’s speculation about whether an efficiently self-adapting recursive AI loop would be “a different type of technology.” These address different questions, and the earlier claim is about Sam rather than the speaker’s own position on recursive AI.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen Sam and Dario appear to want to automate the full AI research-and-development process now.Now The speaker describes a model in which duplicated AI instances acquire different job skills and are then merged so their skills can be combined compositionally.The earlier claim is about Sam and Dario’s apparent goal—“to automate the full AI research-and-development process.” The later claim describes a technical possibility: duplicated model instances acquiring different skills and being merged. It does not state whether the speaker’s view of Sam and Dario’s goal has changed, and the two claims address different questions.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker is concerned that the extent of data contamination and augmentation in the coding benchmarks is unknown.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier statement is about uncertainty over contamination and augmentation in coding benchmarks (“we don't know how much data contamination there is”); the later statement is about expert use of a system in well-specified domains (“magical things happen” while “there's still a bit of a space of ambiguity”). They address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker expects AI systems near AGI to perform about as well as, or at least approach, domain-specific systems.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim concerns whether AI systems nearing AGI would perform comparably to domain-specific systems (“be basically the equivalent” or “be close”). The later claim says an automated AI scientist may be possible in the near future. These are related topics, but the claims address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker argues that AI systems should understand concepts such as three-in-a-row, believes current systems lack something needed for that understanding, and expects that capability to be achieved eventually, though it is not guaranteed to be easy.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is that AI will eventually gain the ability to understand concepts such as “three in a row.” The later claim is that an “automated AI scientist” could be possible in the near future. These concern related but distinct capabilities, and the later statement does not say whether such a scientist would have the kind of understanding described earlier.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker questions whether training at a scale of 10^45, even on everything, would solve distribution shift and yield principles for efficient and useful performance in new domains.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim asks whether training at 10^45 scale would solve distribution shift and support useful performance in new domains. The later claim concerns how expert use can produce “magical things” in well-specified domains while “a bit of a space of ambiguity” remains. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes that AI systems' failure to distinguish types from tokens can cause details about individuals in the same category to blend together and produce hallucinations.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim is that AI can blend details about similar individuals and hallucinate; the later claim is that AI skills or agents do not compose well. These concern different capabilities, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen AGI is an abbreviation for artificial general intelligence.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim, “AGI is an abbreviation for artificial general intelligence,” defines an acronym. The later claim concerns expert-guided AI producing “magical things” in well-specified domains while leaving “a bit of a space of ambiguity.” These address different questions, so the later statement does not show a change in the speaker’s position on the earlier claim.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker characterizes Sam's biggest concern at a Senate appearance a few weeks earlier as the company not making all the money or extracting all the value it could.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier statement describes what Sam's biggest concern was at a Senate appearance: that the company did not “make all the money and extract all the value” it could. The later statement is the speaker's assessment of Elon Musk's ability to spot promising future opportunities, as opposed to engineering or optimization. These concern different people and questions, so they do not establish a change in the speaker's own position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen AI organizations differ to some extent in their reinforcement-learning methods and datasets.Now Current reinforcement-learning-trained models differ from models that existed many years ago.The earlier claim is that AI organizations differ somewhat in their reinforcement-learning methods and datasets. The later claim is that current RL-trained models differ from models from many years ago. These address different comparisons—organizations versus models across time—so the statements do not establish a change of position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Despite quantitative progress in AI, the speaker believes that none of the key problems a cognitive creature must solve have been solved.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim says that key problems a cognitive creature must solve remain unsolved despite AI’s quantitative progress. The later claim only says that an AI can hill-climb toward a specified task, such as a mathematics problem; it does not say the task is solved or that the earlier problems have been solved. These statements do not establish a change in position.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker would not consider a system that cannot perform arithmetic to be satisfactory as AGI.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns a minimum capability for AGI: the speaker would not be satisfied with an AGI that “can't do arithmetic.” The later passage concerns whether AI skills or agents are composable, saying this “doesn't really marry with my experience.” These address different questions, so the later statement does not establish a change in the speaker’s AGI criterion.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes there are too many problems and gaps for all of them to be solved simultaneously.Now The speaker argues that highly complex, information-sharing collectives can build societies, causing some objections to this possibility to lose force.The earlier claim concerns whether numerous visual-intelligence problems could all be solved at once, saying there are “too many problems, too many gaps” for that. The later claim concerns whether complex, information-sharing collectives can build societies and make some objections “fall away.” These address different questions, and the later statement does not establish a changed position on the earlier prediction.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker says that the only general-purpose AI systems currently available are based on large language models.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim is about what kinds of general-purpose AI systems are available: “the only general ones we have are LLM-based.” The later claim is about what models can do—combine existing knowledge in novel ways to solve tasks. It does not say whether those models are general-purpose or whether they are LLM-based, so it does not establish a changed position on the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that some very old AI systems can outperform modern general-purpose AI systems.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim is about relative performance on particular tasks: “some very old AI systems can outperform modern general-purpose AI systems.” The later claim concerns how expert guidance can help AI produce impressive results in “well-specified domains,” while some ambiguity remains. These claims address different questions and can both be true; the later statement does not reverse the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker estimates that developing substantially better reasoning, planning, and stable world models will take five to ten years.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier statement estimates that developing substantially better reasoning and stable world models will take “five or ten years.” The later statement says only that external intelligence systems have “very complex dynamics.” It does not address the development timeline or clearly assert a position on the same question, so no change of mind can be established.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says AI systems can approximate user requests without reliably fulfilling them as intended.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier statement concerns whether AI systems reliably follow intended instructions (“they don't really do what we ask them”). The later statement concerns whether AI capabilities can be composed by combining agents (“doesn't really marry with my experience”). These are related topics but not the same underlying claim, so the later statement does not establish a change of mind about instruction-following.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker believes current alignment systems still fail to reliably do what users intend and exhibit sorcerer’s-apprentice-style failures.Now The speaker believes that extensive optimization in semi-specific objective domains still leaves something important missing.The earlier claim concerns alignment systems failing to follow user intent, including jailbreaks and unintended behavior. The later claim says optimization in “semi-specific” domains still leaves “something missing,” but does not specify that the missing element is reliable alignment or that the speaker’s earlier view has changed. The claims do not establish a changed position on the same question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker argues that novel programming tasks, unlike familiar website coding with many examples, create out-of-distribution challenges that are especially relevant to programming AGI.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim concerns how novel programming tasks create out-of-distribution challenges. The later claim predicts AI agents communicating, learning, and specializing, possibly with harmful behaviors. These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that an AI system should understand the concept of three-in-a-row when discussing it, and says current systems lack something needed for that understanding.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns whether an AI should understand a specific concept—“three in a row”—and whether current systems lack something needed for that. The later claim imagines a future architecture of “diverse AIs” collaborating rather than “one big AI.” These address different questions, and the later statement does not say whether such systems understand concepts like three-in-a-row.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that AI systems will remain highly vulnerable to outlier problems one year later.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AI systems will remain “very vulnerable to those outliers” one year later. The later claim imagines “diverse AIs” examining problems from different angles rather than “one big AI.” These address different questions—robustness to outliers versus how future AI systems might be organized—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen AI had already surpassed human ability in mathematics by roughly 60 to 80 years ago.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is about AI surpassing human ability in mathematics (“already did on like math”), while the later claim concerns whether an automated AI scientist could be possible in the near future. These address different capabilities, so the later statement does not establish a changed view about mathematical performance or contradict the earlier claim.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker thinks the form of AGI relevant to the risks under discussion would probably exceed humans on short-term-memory tasks while matching human versatility and flexibility.Now The speaker says current agents adapt through chain-of-thought processes and skill-surface adaptation, essentially using memory systems.The earlier statement concerns the capabilities of a risk-relevant form of AGI, saying it would probably exceed humans on some short-term-memory tasks while having human versatility and flexibility. The later statement describes how current agents adapt, through chain-of-thought, skill-surface adaptation, and memory systems. It does not affirm or deny the earlier AGI-capability claim, so there is not enough evidence to determine whether the speaker’s position changed.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen A key bottleneck for tool-using AI systems is that they do not always call the appropriate code.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim identifies a tool-use bottleneck: systems “don't always call the right code.” The later claim says a model given a specified task “will hill climb towards it.” These address different capabilities, and the later statement does not say that tool-calling is no longer a bottleneck.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker tentatively identifies the agents in the separate graph as possibly web-browsing agents.Now The speaker says composable AI agents would make them much more worried.The earlier statement tentatively identifies agents in a separate graph as “maybe like web browsing agents.” The later statement says composable agents would make the speaker “much more worried.” These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker considers the machine-intelligence level depicted toward the end of AI 2027 unlikely within the next decade and not very likely in the following decade, while allowing that it is possible.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim is a forecast about when the intelligence depicted in AI 2027 might emerge: it is “not very likely” in the next two decades, though “certainly possible.” The later passage discusses expert-guided AI in well-specified domains and ambiguity, then turns to the meaning of intelligence. It does not state a comparable timeline forecast, so it does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects neurosymbolic AI eventually to undergo a state change but says its timing is very hard to predict.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier claim says neurosymbolic AI will eventually undergo a “state change,” though its timing is hard to predict. The later claim only says that external intelligence systems have “very complex dynamics”; it does not state whether the speaker still expects that state change, so the claims do not establish a change of mind.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen Sam and Dario appear to want to automate the full AI research-and-development process now.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says Sam and Dario “want to automate the full AI research-and-development process”; the later says an automated AI scientist “could be possible in the near future.” One concerns their intentions, the other the feasibility of the idea, so the later claim does not show a change in the speaker’s position on the same question.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says it is possible that his view that AI needs a fix is correct, although proving it is very difficult and his supporting evidence is qualitative.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says AI may need a “fix” and that LLMs mostly exploit existing ideas rather than discover new ones. The later claim says AI composability “doesn't really marry with my experience.” These concern different aspects of AI, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 92%24 Jun → 8 SeptThen The speaker argues that human experts can readily solve unfamiliar tasks within their domain because of their broader understanding, whereas AI systems still struggle with them.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says AI systems “still do” struggle with a particular kind of unfamiliar chess task that human experts can solve. The later claim says models can make “novel combinations” of existing knowledge to solve tasks, but does not assert that they can solve those same unfamiliar tasks as reliably as experts. The statements concern different capabilities and examples, so the later claim does not clearly reverse the earlier one.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes solving visual-intelligence problems, including video-related problems, within three or four years would be very lucky.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is about the timeline for solving visual-intelligence and video-related problems: the speaker says three or four years would be “very lucky” and estimates most of the distribution is beyond ten years. The later claim is about whether an automated AI scientist could be possible in the near future. These are different questions, and the statements alone do not establish that the speaker changed their view on the same capability or timeline.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker characterizes Sam's biggest concern at a Senate appearance a few weeks earlier as the company not making all the money or extracting all the value it could.Now The speaker perceives a Silicon Valley culture of reducing or reifying things.These statements address different questions. The earlier claim characterizes Sam’s concern at a Senate appearance as the company not making or extracting all the value it could; the later claim says there is “a bit of a culture in Silicon Valley to reduce things or reify things.” They do not establish that the speaker changed their own position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that scientific evaluation of AI systems is difficult because their training data is not transparent.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says AI systems are difficult to evaluate because their training data is not transparent. The later claim says expert-guided AI can produce impressive results in well-specified domains while some ambiguity remains. These address different aspects of AI evaluation and use, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that AI systems will not replace the best coders or reach Jeff Dean's coding level within the next decade.Now The speaker recalls that the report may project 60 million AI agents operating at 20 times human speed by 2032.The earlier claim is the speaker’s prediction that AI will not replace the best coders or match Jeff Dean within a decade. The later passage only recalls a report’s projection of “60 million agents” running “at 20 times human speed”; it does not state whether those agents will replace top coders or reach that level. These claims do not establish a changed position on the same question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen AGI is an abbreviation for artificial general intelligence.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier statement only defines “AGI” as “artificial general intelligence.” The later statement imagines future AI as “a collective of diverse” systems rather than “one big AI.” These address different questions, so the later claim does not show a change in the speaker’s position on the earlier claim.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says it would be bad for humanity if AI alignment were solved only 200 years from now while AGI or ASI arrived within the next 10 to 20 years.Now The speaker says composable AI agents would make them much more worried.The earlier claim concerns the timing of alignment relative to AGI/ASI: it would be bad if alignment took 200 years while advanced AI arrived within 10–20 years. The later claim concerns whether composable AI agents would increase the speaker’s worry. These are related to AI risk but address different questions, and the later statement does not revise the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen A system in which a neural network creates the symbolic tools needed for a problem is a neurosymbolic solution.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier claim defines a neurosymbolic solution as one where a neural network creates symbolic tools. The later claim concerns whether an AI system can converge toward a solution when a problem has enough specification. These statements address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says the chess system could not solve a contrived position requiring three white queens and a Black win on the next move.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The claims address different questions. The earlier statement says the chess system could not solve a particular contrived position involving three white queens and a Black win on the next move. The later statement characterizes AlphaGo’s Move 37 as innovative rather than inherently creative, while noting why it appeared creative to observers. Neither claim establishes a changed position on the same underlying question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen In a tic-tac-toe variant where winning required three in a row on the edges, Grok suggested moves that let the speaker make three in a row, showing a failure to develop the correct strategy.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The earlier claim is about Grok's performance in a tic-tac-toe variant: it suggested moves that allowed a winning line, which the speaker called a strategic failure. The later claim discusses how to characterize AlphaGo's Move 37 and creativity. These address different examples and questions, so they do not establish a change in the speaker's position.
• NOT COMPARABLE · 93%24 Jun → 8 SeptThen The speaker predicts that by 2035, people will still use LLMs for some tasks but will regard them as not having been the answer to achieving advanced AI.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim predicts that by 2035 LLMs will still be used for some tasks but will not have been “the answer” to achieving advanced AI. The later claim says an automated AI scientist may be possible in the near future. That possibility does not establish that AI scientists will achieve advanced AI or contradict the earlier prediction; the claims address different questions.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes that generalization outside the training distribution remains a major problem for current AI systems.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says generalization outside the training distribution “remains a huge problem.” The later claim imagines future AI as “diverse AIs” examining problems from different angles rather than “one big AI.” These address different questions, and the later statement does not say whether the generalization problem has changed or been solved.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Since at least 1998, the speaker has focused on how AI systems generalize beyond their training; he says transparency about this is lacking.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says the speaker has long asked how AI generalizes beyond training and that “we just don't have transparency on that.” The later claim discusses expert-guided AI producing results in “well-specified domains” while leaving “a bit of a space of ambiguity.” That does not clearly revise or deny the earlier position about transparency or generalization; the statements address related but different aspects of AI.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes that hallucinations caused by information blending remain a problem for AI systems, alongside unresolved problems in reasoning and planning.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns ongoing AI problems—hallucinations from information blending, reasoning, and planning. The later claim imagines a future architecture of “diverse AIs” collaborating “from different angles,” rather than “one big AI.” It does not say those systems would eliminate the earlier problems or otherwise revise the speaker’s view of them; the statements address different questions.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen At the time of the conversation, AI systems were not very differentiated, with most actors using LLMs and reasoning models.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement is about the present: AI systems are “not very differentiated” and most actors use similar LLMs and reasoning models. The later statement imagines a possible future with “diverse AIs” approaching problems from different angles, rather than one big AI. A current-state observation and a future-oriented vision do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker considers the estimates for solving some of the discussed problems overly optimistic.Now The speaker believes shortcuts and validity problems are less problematic in this research than he had expected.The earlier claim says estimates for solving some previously discussed problems were “generous,” while the later claim says shortcuts and validity problems are “not as much of a problem as we thought it would be.” These concern different specific issues, and the later statement does not clearly revise the earlier assessment of timelines.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker believes that hallucinations caused by information blending remain a problem for AI systems, alongside unresolved problems in reasoning and planning.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says hallucinations from information blending “remain a problem,” along with issues in reasoning and planning. The later claim discusses expert-guided AI producing “magical” results in well-specified domains while retaining “a space of ambiguity.” These address different aspects of AI performance, and the later statement does not retract or deny the earlier concerns.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes AI systems are not close to matching a top software engineer, because that requires understanding the problem, the human goals behind it, and multiple domains.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AI is not close to matching a top software engineer, citing the need to understand problems, human goals, and multiple domains. The later claim imagines future AIs with different expertise collaborating across problems, rather than one large AI; it does not say whether such systems would match top software engineers. These address different questions, so the later statement does not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says current AI companies are not entirely trustworthy.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim concerns whether “current companies are not entirely trustworthy.” The later claim concerns expert-guided AI producing impressive results in “well-specified domains” while some ambiguity remains. These address different questions, so the later statement does not show a change in the speaker’s view about company trustworthiness.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker expects a future state change in AI but says its timing is very difficult to predict.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim says a future AI “state change” is expected, but its timing is hard to predict. The later claim describes how an AI works on a specified task (“hill climb towards it”); it does not address whether or when that future state change will occur, so the claims do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Despite quantitative progress in AI, the speaker believes that none of the key problems a cognitive creature must solve have been solved.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim says that quantitative progress has not solved the key problems a cognitive creature must solve. The later claim imagines future AI as a diverse collective rather than one large AI. These address different questions, and the later claim does not say those cognitive problems have been solved or otherwise revise the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen AI has made some progress toward AGI and very little progress toward alignment, according to the speaker.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim assesses progress on two distinct dimensions: “some progress towards AGI” and “very little towards alignment.” The later claim predicts agents communicating and specializing, with possible harmful behaviors such as collusion. It does not state how much progress has been made toward AGI or alignment, so the claims do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Human reasoning is a low benchmark for machines because humans have cognitive biases, illusions, and memory problems.Now The human brain is not Turing complete.The earlier claim says human reasoning is a low benchmark for machines because of cognitive biases, illusions, and memory problems. The later claim says only that “the human brain is not Turing complete.” These are different questions—comparative reasoning ability versus computational completeness—so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says there are already signs that neurosymbolic AI is gaining momentum.Now The speaker says many startups are pursuing recursive-improvement superintelligence.The earlier claim concerns signs that neurosymbolic AI is gaining momentum (“already see signs” and conferences “getting bigger”). The later claim concerns many startups pursuing recursive-improvement superintelligence. These are different claims about different developments, so the later statement does not establish a changed view about the earlier one.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker considers voluntary regulation and industry self-regulation probably insufficient to ensure disclosure of AI model capabilities.Now The speaker argues that apparent general capability is partly an illusion: models can be shaped into specialized experts in any domain, but that expertise is latent rather than explicit.The earlier claim concerns whether voluntary regulation and self-regulation are sufficient to ensure disclosure of AI model capabilities. The later claim concerns whether models have latent general capabilities that can be shaped into domain-specific expertise. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker argues that safety-critical AI uses should wait for a better alignment solution if systems cannot reliably refuse harmful requests and prevent foreseeable human harm, even if the solution takes five or 500 years.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim concerns whether to delay safety-critical AI uses until alignment can prevent harmful outputs (“wait until we have a better solution”). The later claim describes autonomous recursive adaptation as “a different type of technology.” These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker suggests that the weakest plausible AGI could perform at the level of an ordinary person, who may reason poorly and exhibit substantial confirmation bias.Now The speaker believes human intuitions about collective behavior are often unreliable and that such behavior could emerge in principle.The earlier claim concerns the capability threshold for the weakest plausible AGI: it could perform at an ordinary-person level, despite ordinary people being “not very good at reasoning” and having “substantial confirmation bias.” The later claim concerns whether behavior could emerge from collectives and whether intelligence, capability, and power can be distinguished. These statements do not establish a changed position on the same question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes that AI systems' failure to distinguish types from tokens can cause details about individuals in the same category to blend together and produce hallucinations.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns a cause of hallucinations: AI can blur details among individuals in the same category. The later claim imagines diverse AI systems collaborating on problems from different angles, rather than one large AI. These address different questions and neither affirms nor rejects the other.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker is concerned that the extent of data contamination and augmentation in the coding benchmarks is unknown.Now The speaker believes that extensive optimization in semi-specific objective domains still leaves something important missing.The earlier claim concerns uncertainty about “data contamination and augmentation” in coding benchmarks. The later claim concerns whether optimization in “semi-specific objective domains” can address what the speaker feels is “something missing.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes that hallucinations caused by information blending remain a problem for AI systems, alongside unresolved problems in reasoning and planning.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says hallucinations and problems in reasoning and planning remain issues. The later claim says, “that doesn't really marry with my experience,” in a discussion of whether AI skills can be composed by merging agents. These concern different aspects of AI, so the later statement does not establish a changed position on the earlier claim.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen AI has made some progress toward AGI and very little progress toward alignment, according to the speaker.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier statement says AI has made “some progress toward AGI” but “very little progress toward alignment.” The later passage concerns whether skills or agents are composable, saying that this “doesn't really marry with my experience.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker characterizes himself and Yann LeCun as taking architectural or cognitive approaches to AI.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says the speaker and Yann LeCun take an “architectural approach or a cognitive approach” to AI. The later claim discusses expert-guided AI, a “supervisor illusion,” and ambiguity in well-specified domains. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker estimates that choosing a legal chess move takes a human about 100 milliseconds, while expressing uncertainty about the estimate.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim estimates how long a human takes to choose a legal chess move (“about 100 milliseconds”). The later claim concerns expert use of AI in well-specified domains (“magical things happen”) while noting ambiguity. These address different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects a future state change in AI but says its timing is very difficult to predict.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier statement says, “eventually there will be a state change,” while noting that its timing is hard to predict. The later statement discusses how expert guidance can make AI work well in specified domains while ambiguity remains. It does not address whether or when a future state change will occur, so the claims do not establish a change of position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says the participants agree that current efforts on AI alignment are inadequate and need substantial improvement.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier statement is about alignment efforts being inadequate and needing improvement; the later statement predicts that future AI agents may communicate and exhibit harmful behaviors. These address different questions, and the later prediction does not indicate a changed view about alignment.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says he has not really seen the gaps he has raised since 1998 solved over roughly 27 years, and finds it strange to expect them all to be solved within three years.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim says the speaker had not seen long-standing gaps solved and doubted they would all be solved within three years. The later claim says expert use can produce “magical things” in “well-specified domains,” while ambiguity remains. That does not establish that the earlier gaps have since been solved; the statements address different claims about progress and current usefulness.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen A key bottleneck for tool-using AI systems is that they do not always call the appropriate code.Now For tasks on the long tail, an expert can guide an AI model by making its path of least resistance roughly correct, increasing the chance it produces the correct result.The earlier claim concerns tool-using systems failing to call the appropriate code. The later claim concerns an expert guiding an AI on long-tail tasks so its “path of least resistance” is “roughly correct.” These address different aspects of AI performance, and the later statement does not say whether the code-calling bottleneck still holds.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker says someone might discover a better method for learning models of distributions than current LLMs provide.Now The speaker says reinforcement-learning-trained language models have substantial base knowledge and can perform things that were not possible before.The earlier claim is speculative: someone might find a better method for learning distribution models than current LLMs provide. The later claim reports that RL-trained language models already have substantial base knowledge and can do things that were not possible before. These address different questions—possible alternative methods versus current model capabilities—so the later statement does not establish a change in the speaker’s position.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker predicts that neurosymbolic AI will begin to take off in three or four years.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim predicts that “neurosymbolic AI starts to take off” in three or four years. The later claim says an “automated AI scientist” is possible in the near future. These are related ideas, but the statements do not establish opposing positions on the same question; the later statement does not retract or revise the earlier forecast.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes AI systems are not close to matching a top software engineer, because that requires understanding the problem, the human goals behind it, and multiple domains.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim is about whether AI is close to matching a top software engineer, while the later claim imagines a possible future architecture of multiple specialized AIs working together rather than one large AI. The later statement does not say whether such a collective would match a top software engineer, so it does not establish a change in position on the earlier question.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker argued by 2001 that AI systems need symbols to perform abstract operations over variables.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim is about whether pure neural networks can perform abstract operations over variables without symbols. The later claim says an AI can be given a mathematics or knowledge task and hill-climb toward an answer, but does not say whether it uses symbols or contradict the earlier limitation. These statements address different questions, so the later claim does not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that if the necessary ideas are missing, progress depends on when new ideas are developed.Now The speaker rejects intelligence as a magical quality and says it depends on the scope of tasks and on existing knowledge.The earlier claim says progress may depend on when new ideas are developed if the necessary ideas are missing. The later claim says intelligence depends on task scope and existing knowledge and is “fractured and jagged.” These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen A neurosymbolic system can work if its neural networks reliably call the tools appropriate to the problem.Now The speaker says fractured, entangled neural-network representations have some degree of composability.The earlier claim is conditional on neural networks reliably calling the appropriate tools: that would let a neurosymbolic system work. The later claim says fractured, entangled representations are “composable to some degree.” These address different aspects of neurosymbolic systems, and the later statement neither affirms nor denies the earlier condition.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects current systems to fail on many chess variants and problems, and says he can devise tic-tac-toe variants such as a five-by-five board where wins are allowed only on the edges.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim concerns current systems failing on many chess and tic-tac-toe variants; the later claim says an AI can “hill climb” toward a solution to a specified task such as a math problem. These statements address different tasks and do not establish a changed position about the same systems or abilities.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that an AI system should understand the concept of three-in-a-row when discussing it, and says current systems lack something needed for that understanding.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.These statements address different questions. Earlier, the speaker says AI should understand “what a three in a row is” and that current systems are “missing something” for that understanding. Later, they imagine future AI as “diverse AI’s” looking at problems “from different angles” rather than “one big AI.” The later claim does not establish a changed position about AI’s ability to understand concepts.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers learning models of distributions a very useful cognitive skill, but not the only important cognitive skill.Now The speaker argues that apparent general capability is partly an illusion: models can be shaped into specialized experts in any domain, but that expertise is latent rather than explicit.The earlier claim says learning models of distributions is “a super helpful cognitive skill” but “not the only one.” The later claim concerns whether models have general capabilities or latent expertise that can be activated through specialization. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says current AI systems, including O3, do not reliably follow chess rules and sometimes make illegal moves.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns whether AI systems reliably follow chess rules and says they “sometimes make illegal moves.” The later claim concerns whether AI skills can be composed by merging agents, and says that idea “doesn't really marry with my experience.” These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers it an exceptionally rapid convergence if the issues he has worried about for 25 years were all resolved by 2027.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier statement describes reaching a rapid convergence on long-standing concerns by 2027 (“everything everywhere all at once”). The later statement is about a narrower optimization condition: an AI system can converge when there is “enough of a specification.” These do not establish opposing positions on the same question.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker says someone might develop a better method for learning distributions; he describes distribution learning as a very helpful, but not exclusive, cognitive skill.Now The speaker thinks future weight adaptation might solve the problem of sharing knowledge among agents.The earlier claim says learning distributions is “a super helpful” cognitive skill, but “not the only one,” and that someone might find a better method. The later claim speculates that weight adaptation “maybe” could solve agents’ knowledge-sharing problem. These address different questions, and the later statement is tentative rather than a clear reversal of the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that newer AI systems enable tasks that were previously unavailable and that these capabilities are already affecting labor markets.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim is that newer systems enable previously unavailable tasks and are already affecting labor markets. The later claim is that models can make “novel combinations” of existing knowledge to solve tasks. These claims address different questions, and the later statement does not affirm or deny the earlier one, so there is no clear evidence of a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says frontier-math evaluations may have contamination issues because OpenAI had access to the problems.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier statement alleges possible contamination in frontier-math evaluations because OpenAI had access to the problems. The later statement describes a model working toward a specified task, such as a math problem. It does not address whether evaluations are contaminated or deny the earlier allegation, so a change in position cannot be established.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen According to the speaker's characterization of Jan LeCun's view, current AI is not close to the goal because it lacks genuine world models.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim reports Jan LeCun’s view that AI is “not close” because it lacks “world models”; it is not clearly the speaker’s own position. The later passage says the idea “doesn't really marry with my experience,” but the claim it responds to is not included, and it concerns composability rather than whether AI has world models. These passages do not establish a change of the speaker’s own view on the same question.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen AI had already surpassed human ability in mathematics by roughly 60 to 80 years ago.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim says AI had surpassed human ability in mathematics by roughly 60–80 years ago. The later claim says an AI can be given a mathematics task and will “hill climb towards it.” That describes a way of approaching a task, not whether AI outperforms humans, so it neither confirms nor reverses the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen AI could surpass human abilities in some dimensions within 10 years; the speaker thinks this will probably occur in some dimensions but not others.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns whether AI will surpass human abilities on some dimensions and when: “it could happen in 10 years” and “probably will happen on some of those dimensions.” The later claim concerns how future AI might be organized: “diverse AIs” collaborating rather than “one big AI.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker believes it is difficult to forecast when new ideas will emerge.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier claim is about the difficulty of forecasting when new ideas will emerge (“new ideas are hard to project”). The later claim is about Elon Musk’s ability to recognize promising future areas, not a direct statement about whether new ideas can generally be forecast. The claims concern different questions, so the later remark does not establish a change in the earlier position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Machines will eventually exceed human performance, but when and how this will happen remain uncertain to the speaker.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says machines “will someday” exceed human performance, while the later claim discusses how expert-guided AI performs in “well-specified domains” and notes “a space of ambiguity.” The later statement does not say whether the speaker still expects machines to eventually exceed human performance; it addresses a different question about current AI use and its limits.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says frontier-math evaluations may have contamination issues because OpenAI had access to the problems.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier statement concerns possible contamination in frontier-math evaluations because OpenAI had access to the problems. The later statement concerns expert-guided AI performance in well-specified domains and remaining ambiguity. These address different questions, so the later claim does not establish a changed position on evaluation contamination.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Performance on agent tasks is still poor.Now The speaker says memory-based agent adaptation is not very composable, and that organizations currently face this as a major problem.The earlier claim says agent-task performance is “still pretty poor.” The later claim says memory-based agent adaptation is “not very composable” and presents that as an organizational problem. These address different aspects of agents, and the later claim does not say performance has improved or contradict the earlier assessment.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker says AI systems may automate conventional, inside-the-box work and may outperform people on such work, citing the Go move 37 example.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says AI may outperform people on “inside-the-box” work, while “Einstein level innovations” may be needed for AGI. The later statement says an “automated AI scientist” could be possible in the near future. That does not establish whether the speaker thinks such a scientist would produce the outside-the-box innovations they previously said AGI requires, or whether they mean a more limited kind of automated science.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes the graph has several flaws, including flaws relevant to the argument about AI reliability.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim concerns flaws in a particular graph—“a lot of problems” that are “relevant to this argument.” The later claim concerns expert-guided AI in “well-specified domains” and remaining ambiguity. These address different subjects, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker predicts that two years before ASI is achieved, AI systems will have full solutions to all the cognitive problems he listed.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim predicts that AI will have “full solutions to all of those things” two years before ASI. The later claim says an automated AI scientist is “possible in the near future.” These concern related capabilities, but do not establish opposing positions on the same question; the timelines and claims differ.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that some very old AI systems can outperform modern general-purpose AI systems.Now The speaker believes an automated AI scientist could be possible in the near future.These claims address different questions. The earlier statement says some “very old AI systems can outperform modern general-purpose AI systems”; the later says an “automated AI scientist” may be possible in the near future. The possibility of an AI scientist does not conflict with the comparative performance claim.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker does not expect automation of current AI systems to generate the Einstein-level innovations needed for AGI or ASI.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says automating current machines will not produce the Einstein-level innovations needed for AGI or ASI. The later claim says the speaker’s experience does not align with the idea that AI skills can be composed by merging agents. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier statement says current models’ semantics are “very superficial” and that “we need a deeper one.” The later statement concerns intelligence as adaptivity, claiming models can “create novel combinations” of existing knowledge to solve tasks. The later claim does not establish that the speaker now thinks models have the deeper semantics called for earlier; the two claims address related but distinct capabilities.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker estimates that choosing a legal chess move takes a human about 100 milliseconds, while expressing uncertainty about the estimate.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier claim estimates how long a human takes to choose a legal chess move (“100 milliseconds or something like that”). The later claim says external intelligence systems have “very complex dynamics.” These address different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker estimated, with uncertainty, that GPT-2 used about 5%–10% of the internet.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement estimates how much of the internet GPT-2 used (“maybe 10% … or 5%”). The later statement concerns whether an automated AI scientist could be possible in the near future. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen AI actors have incentives to accelerate development and secure as many resources as possible for that purpose.Now The interviewer considers an entirely AI-run economy plausible but suspects it might become degenerate in some way.The earlier claim says AI actors are incentivized to accelerate development and secure resources. The later claim says an entirely AI-run economy “seems plausible” but might “become degenerate.” These address different questions—development incentives versus the possible character of an AI-run economy—so they do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker has heard, but cannot verify, that OpenAI staff search social media—including his posts—for examples of model failures.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.These statements address different questions. Earlier, the speaker says he was told—while noting he cannot verify it—that OpenAI staff look on social media for examples of model failures. Later, he discusses expert-guided AI producing impressive results in well-specified domains while some ambiguity remains. The later claim does not revise or deny the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker thinks Claude and Gemini may have acceptable frontier-math scores, while current systems are being optimized for the test through augmentation whose nature is unknown.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim concerns whether Claude and Gemini have acceptable frontier-math scores and whether current systems are optimized for tests through augmentation. The later claim concerns whether a collective AI-human system would have limits on its scale. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says that merely fairly reasonable instruction-following is inadequate in some high-stakes situations, including systems guiding weapons.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim says that “fairly reasonable” instruction-following may be inadequate when systems are “guiding weapons.” The later claim predicts that AI agents may communicate and specialize, potentially leading to “collusion and lots of bad things.” These address different questions, so the later statement does not establish a change in the speaker’s view about high-stakes instruction-following.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says that merely fairly reasonable instruction-following is inadequate in some high-stakes situations, including systems guiding weapons.Now The speaker says grounding is not necessarily an argument that intelligence requires physical instantiation or consciousness.The earlier claim says that “fairly reasonable” instruction-following is not enough in some high-stakes situations, such as systems guiding weapons. The later claim says intelligence does not necessarily require “physical instantiation” or consciousness. These address different questions—reliability in high-stakes use versus what intelligence requires—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Because new data is generated every day, researchers can continue to obtain somewhat more data.Now Current reinforcement-learning-trained models differ from models that existed many years ago.The earlier claim is about the ongoing availability of some additional data, while the later claim is about how current RL-trained models differ from older models. They do not address the same underlying question, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes transparency has limited value to an AI organization until it develops a unique capability that competitors cannot rapidly reproduce.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.These claims address different questions. The earlier claim concerns when an AI organization should be transparent—saying transparency has limited value once it has a capability competitors cannot rapidly reproduce. The later claim concerns limits on the scale of a proposed AI collective. Neither claim establishes a changed position on the other's subject.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker believes LLMs generally produce inside-the-box solutions, whereas achieving AGI will require outside-the-box innovations.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AGI will require “outside the box solutions.” The later claim describes a possible future with “diverse AIs” examining problems from different angles, rather than “one big AI.” That does not say whether such collaboration would provide the outside-the-box innovations the speaker earlier said AGI requires, so the claims cannot be compared as a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says there are already signs of neurosymbolic AI beginning to take off.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says there are “already signs” of neurosymbolic AI “beginning to take off,” citing AlphaFold and growing conferences. The later claim says models can “create novel combinations” of existing knowledge to solve tasks. These address different questions—neurosymbolic AI’s uptake versus models’ adaptivity—so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says the chess system could not solve a contrived position requiring three white queens and a Black win on the next move.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier claim concerns a chess system’s inability to solve a contrived position (“wouldn't be able to do that at all”). The later claim says external intelligence systems have “very complex dynamics.” These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Humans have limitations in memory.Now The human brain is not Turing complete.The earlier claim says humans have memory limitations; the later claim says the human brain is not Turing complete. These concern different properties, and the later claim does not deny the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says there are already signs of neurosymbolic AI beginning to take off.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says there are already signs of neurosymbolic AI “beginning to take off”; the later claim says an automated AI scientist “could be possible in the near future.” These concern related AI developments, but they are not opposing positions on the same underlying question, so the later statement does not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker suggests that a Go system trained on a standard 21-by-21 board could play on rectangular boards and handle other variations.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier statement concerns whether a Go system trained on a standard board can handle rectangular boards and other variations. The later statement concerns expert use of AI in well-specified domains and remaining ambiguity. These address different capabilities and questions, so they do not establish a change of mind.
• NOT COMPARABLE · 87%24 Jun → 8 SeptThen The speaker does not expect automation of current AI systems to generate the Einstein-level innovations needed for AGI or ASI.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says automating current AI systems would not produce the Einstein-level innovations needed for AGI or ASI. The later claim says an automated AI scientist may be possible in the near future. Those claims are not necessarily inconsistent: an automated scientist could be possible without producing those innovations or achieving AGI/ASI. The excerpts do not establish that the speaker changed their view on the same question.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes the graph has several flaws, including flaws relevant to the argument about AI reliability.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says the graph has “several flaws,” some relevant to an argument about AI reliability. The later claim says an idea “doesn't really marry with my experience,” apparently in a discussion of whether AI agents are composable. These statements do not establish opposing positions on the same underlying question, so a change of mind cannot be inferred.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Gary Marcus believes more people probably read AI 2027's darker scenario than its positive scenario.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns Gary Marcus’s view about which AI 2027 scenario more people read; the later claim concerns the speaker’s imagined design of future AI systems. These address different questions, so they do not establish a change in the same position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The relevant form of AI recursion involves a system generating fundamentally new ideas previously produced only by humans and carrying out the broader development process.Now The speaker says many startups are pursuing recursive-improvement superintelligence.The earlier claim defines the kind of recursion being discussed: a system generating “a fundamentally new idea” and carrying out broader development. The later claim says startups are pursuing “recursive improving superintelligence.” That reports what startups are doing; it does not establish a changed view about the earlier definition or whether such systems can achieve it.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says polls indicate that 75% of Americans are already worried about AI and expects public concern to rise, seeing no active measure likely to reduce it in time.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim concerns public sentiment: polls show Americans are worried about AI, and the speaker expects that worry to rise. The later claim concerns technological feasibility: an automated AI scientist may be possible in the near future. These positions address different questions and do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker describes a general driving solution as level-five autonomy that requires neither geofencing nor specialized maps.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier claim concerns what a general driving solution would require: “level five” autonomy without “geofencing” or specialized maps. The later claim concerns Elon Musk’s strengths—identifying promising future opportunities rather than engineering or optimization. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The interviewer says that people are already pushing against the proposed AI red lines, although they may not yet have crossed them.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.These claims address different questions. The earlier extracted claim describes the interviewer’s observation that people are pushing against AI red lines, while the later claim predicts that future AI agents may communicate, specialize, and exhibit harmful behaviors. Neither claim establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker believes the problems he described in his book remain unsolved.Now The speaker believes that extensive optimization in semi-specific objective domains still leaves something important missing.The earlier claim is specifically that the problems described in his book “have [not] been solved.” The later claim is that even extensive optimization in “semi-specific objective domains” leaves “something missing.” These may be related views about AI, but the excerpts do not establish that they address the same problems or that his position on the book’s problems changed.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that smart people can solve tasks such as the Tower of Hanoi and that six-year-olds can learn chess rules.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim is about what smart people and six-year-olds can do; the later claim is about how an AI model approaches a specified task. They do not establish different positions on the same question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The term AGI has been defined in different ways.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier statement says that “the term AGI has been defined in different ways”; the later says an AI system can “converge towards the solution” when it has enough specification. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes transparency has limited value to an AI organization until it develops a unique capability that competitors cannot rapidly reproduce.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The claims concern different questions. The earlier statement says transparency has limited value once an AI organization develops a hard-to-reproduce capability; the later statement describes Elon Musk’s perceived strength as recognizing promising future opportunities rather than engineering or optimization. The later claim does not address transparency or the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen A neurosymbolic system can work if its neural networks reliably call the tools appropriate to the problem.Now The speaker says current agents adapt through chain-of-thought processes and skill-surface adaptation, essentially using memory systems.The earlier claim concerns whether a neurosymbolic system can work if neural networks reliably call the appropriate tools. The later claim describes how current agents adapt—through chain-of-thought, skill-surface adaptation, and memory systems. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 83%24 Jun → 8 SeptThen The speaker argues that robust understanding of domains such as tic-tac-toe or chess should make related tasks relatively easy for AI systems.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says that if AI had “a robust understanding” of domains such as chess, related tasks “wouldn't be too hard.” The later claim discusses expert-guided AI producing “magical” results in well-specified domains while retaining “a space of ambiguity.” These statements can coexist, but they address different aspects of AI performance; the later passage does not clearly accept or reject the earlier claim.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen Performance on agent tasks is still poor.Now The speaker thinks current agents are not yet composable.The earlier claim says agent-task “performance ... is still pretty poor.” The later claim says agents “are not yet composable.” These concern different capabilities; the later statement does not establish whether the speaker’s view of agent-task performance changed.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker expects AI organizations to develop unique intellectual property, increasing their incentive to limit transparency.Now The speaker views intelligence as externalized and collective.The earlier claim concerns organizations’ incentives to limit transparency as they develop unique intellectual property (“more reason to be less transparent”). The later claim concerns the nature of intelligence (“intelligence is externalized, it’s collective”). These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker suggests that a Go system trained on a standard 21-by-21 board could play on rectangular boards and handle other variations.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The earlier claim concerns whether a Go system trained on a standard board can handle rectangular boards and other variations. The later claim concerns whether AlphaGo’s Move 37 was creative or merely innovative. These address different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes LLMs alone are inadequate for the nearer-term technical alignment problem, while neurosymbolic methods might offer a chance.Now The speaker says fractured, entangled neural-network representations have some degree of composability.The earlier claim concerns whether LLMs alone can solve near-term technical alignment and whether neurosymbolic methods may help. The later claim says fractured, entangled representations are “composable to some degree.” It does not say whether this changes the speaker’s view about LLMs’ adequacy or the value of neurosymbolic methods.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker says that he and Yann LeCun consider current AI systems' purported reasoning insufficiently robust.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says current AI systems’ purported reasoning is not “robust enough.” The later claim says the speaker’s experience does not support the idea that AI skills are composable. These are related concerns about AI capabilities, but they address different questions; the later statement does not establish a changed position on reasoning robustness.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker believes Einstein-level innovations will be necessary to achieve AGI and, certainly, ASI.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says Einstein-level innovations will be needed for AGI and ASI. The later claim says an automated AI scientist could be possible in the near future. That does not establish that the speaker now believes AI can achieve AGI without those innovations; the claims address related but distinct questions.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker argues that human experts can readily solve unfamiliar tasks within their domain because of their broader understanding, whereas AI systems still struggle with them.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says human experts can handle unfamiliar tasks within their domain while “these systems still do” struggle. The later claim imagines future AI as “diverse AIs” with different expertise examining problems from multiple angles rather than “one big AI.” These address different questions—current AI performance versus a possible future AI arrangement—and the later statement does not say that AI has overcome the limitation described earlier.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says the same gaps remain despite quantitative improvement, with no principled solution to any of them.Now In ambiguous problem domains, what specification is needed remains mysterious, according to the speaker.The earlier statement says that the same gaps remain and that there is “no principled solution” to them. The later statement says that in “the ambiguity regime” a specification is needed, but what that means is “really mysterious.” These may concern related issues, but the claims do not establish that the speaker now has a different view about whether the earlier gaps have principled solutions.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker has heard, but cannot verify, that OpenAI staff search social media—including his posts—for examples of model failures.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim is that the speaker was told OpenAI staff search social media for model-failure examples, while the later claim says their experience does not align with AI skills being composable. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker argues that AI systems should understand concepts such as three-in-a-row, believes current systems lack something needed for that understanding, and expects that capability to be achieved eventually, though it is not guaranteed to be easy.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns whether AI will eventually gain a particular kind of understanding: the speaker says, “We will get it” and “I don't think it's impossible.” The later claim describes a preferred or imagined future architecture—“diverse AIs” working together “rather than one big AI.” It does not say whether that arrangement would achieve the earlier kind of understanding, so these statements do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker regards language models using tools as a form of neurosymbolic AI, though not the right form.Now The speaker says reinforcement-learning-trained language models understand things differently from humans but address the exploration problem by learning how things fit together and exploring novel trajectories.The earlier claim says tool-using language models are “a species of neurosymbolic AI” but “not the right one.” The later claim discusses how RL-trained language models “understand how things fit together” and explore “novel trajectories.” It does not address whether tool use is neurosymbolic AI or retract the earlier qualification, so these statements do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says O3 is unable to avoid making illegal chess moves.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier claim is about O3’s ability to avoid illegal chess moves (“It can't avoid illegal moves”). The later claim describes the complexity of dynamics among external intelligence systems (“There are just very complex dynamics”). These address different questions, so the later statement does not establish a change in the speaker’s view about O3 or illegal moves.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker expects the discussed AI capability to eventually be achieved.Now The speaker recalls that the report may project 60 million AI agents operating at 20 times human speed by 2032.The earlier claim is the speaker’s expectation that the discussed capability will eventually be achieved (“I think we all think that's going to happen”). The later statement is framed as recalling what someone else said (“I think you said by 2032…”), not as the speaker asserting or rejecting that projection. It therefore does not establish a change in the speaker’s own position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that AI systems will improve programmer productivity, while noting caveats concerning the security and maintainability of generated code.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim concerns AI improving programmer productivity, with caveats about code security and maintainability. The later claim concerns whether an automated AI scientist could be possible in the near future. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker believes current alignment systems still fail to reliably do what users intend and exhibit sorcerer’s-apprentice-style failures.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim says alignment systems “still don't really do what we want them to do,” citing failures such as jailbreaks and other misalignment. The later claim says expert use can produce “magical” results in well-specified domains, while “there's still a bit of a space of ambiguity.” These claims address different aspects and conditions of system performance, so the later statement does not establish that the speaker changed their view about alignment failures.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker estimates that developing much better reasoning and planning and more stable world models will take five or ten years.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier claim estimates that better reasoning, planning, and more stable world models will take “five or ten years to develop.” The later claim only says there are “very complex dynamics.” It does not revise or contradict the earlier timeline estimate; the claims address different questions.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says that system prompts against producing copyrighted material or hallucinating do not reliably prevent those behaviors.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim is that system prompts such as “don’t produce copyrighted material” or “don’t hallucinate” do not reliably prevent those behaviors. The later claim concerns expert-guided AI producing “magical” results in well-specified domains while leaving “a space of ambiguity.” These claims can both be true; the later statement does not say the system prompts reliably work or otherwise retract the earlier criticism.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen A more sophisticated AGI forecast represents probability across multiple possible arrival years rather than naming only one predicted year.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns how to express an AGI forecast: a probability curve is more sophisticated than naming one year. The later claim concerns how future AI systems might be organized: diverse AIs collaborating rather than one large AI. These address different questions, so they do not show a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says the participants use somewhat different forecasting methodologies.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The claims address different subjects: the earlier statement says the participants use “slightly different methodologies” for forecasting, while the later statement identifies Elon Musk’s strength as recognizing “areas that are interesting” and might work in the future. Neither claim establishes a position that conflicts with the other.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes there are too many problems and gaps for all of them to be solved simultaneously.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim concerns whether many visual-intelligence problems could all be solved within a few years: Scarfe says there are “too many problems, too many gaps” for that to happen “all at once.” The later claim concerns whether a proposed AI-and-human “super organism” would have limits on its scale. These address different questions, and neither claim establishes a changed position on the other.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Since at least 1998, the speaker has focused on how AI systems generalize beyond their training; he says transparency about this is lacking.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns whether AI systems generalize beyond training and the lack of transparency about that. The later claim says the speaker’s experience does not align with AI skills being composable. These address related but distinct questions, and neither claim clearly reverses the other.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker does not think AI will surpass human abilities across the relevant dimensions within two years.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier statement is about whether AI will surpass human abilities across relevant dimensions within two years (“I don't think it can happen in two”). The later statement concerns whether AI agents’ skills can currently be composed by combining them (“doesn't really marry with my experience”). These address different questions, so the later claim does not establish a change in the earlier prediction.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects current systems to fail on many chess variants and problems, and says he can devise tic-tac-toe variants such as a five-by-five board where wins are allowed only on the edges.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement concerns whether current systems can solve chess variants and a proposed tic-tac-toe variant; the later statement imagines a future architecture of diverse AIs rather than one large AI. Neither claim reverses the other, and they address different questions.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The interviewer says neurosymbolic AI is his personal favorite among possible alternative AI technologies.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is about neurosymbolic AI being a personal favorite among alternative technologies; the later claim is that an automated AI scientist may be possible in the near future. These address different questions, and neither statement establishes a changed position on the other.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says humans are extremely data-efficient on many, but not all, problems.Now The speaker believes that extensive optimization in semi-specific objective domains still leaves something important missing.The earlier statement says humans are “fabulously data efficient on many problems,” with the qualification “not all.” The later statement says optimization in “objective” and “semi-specific domains” still leaves “something missing.” These address different questions—human data efficiency versus whether optimization captures some aspect of understanding—and the later claim does not deny the earlier one.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The interviewer says that people are already pushing against the proposed AI red lines, although they may not yet have crossed them.Now The speaker says composable AI agents would make them much more worried.The earlier claim concerns people pushing against proposed AI red lines (“They may not be breaking through them”). The later claim concerns whether composable AI agents would make the speaker more worried. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker expects AI organizations to develop unique intellectual property, increasing their incentive to limit transparency.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The claims address different questions. The earlier claim says unique AI intellectual property could increase organizations’ incentive to limit transparency (“more reason to be less transparent”); the later claim concerns whether a proposed AI collective-intelligence system would have limits on its scale. The later statement does not revise the earlier position on transparency.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker expects neurosymbolic AI eventually to undergo a state change but says its timing is very hard to predict.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says a state change in neurosymbolic AI will happen eventually, while its timing is hard to predict. The later statement says an automated AI scientist could be possible in the near future. These concern related but not clearly identical questions, and the later claim does not say that a state change has occurred or give a contrary view about its timing.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects many distribution-shift cases to continue exposing failures in AI systems next year.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim predicts that distribution-shift cases will continue to expose AI failures next year. The later claim says an automated AI scientist may be possible in the near future. These address different questions, so the later statement does not show a changed position on the earlier prediction.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes AI will eventually acquire the missing capability for understanding concepts such as three-in-a-row, and says achieving it is not impossible.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim says, “We will get it” and “I don't think it's impossible,” referring to a missing capability for understanding concepts such as three-in-a-row. The later claim says an AI can “hill climb” toward a specified task, such as a mathematics problem. These address different capabilities, and the later statement does not establish a changed view about the earlier one.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker places the most optimistic timeline at 2030, including the very fastest case.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier statement concerns the earliest optimistic timeline—“the very fastest case” being 2030. The later statement concerns whether an ecosystem of AIs and humans can participate in a “super organism.” These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen A more sophisticated AGI forecast represents probability across multiple possible arrival years rather than naming only one predicted year.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim is about how to express an AGI forecast—using probability across possible arrival years rather than one date. The later claim is about the form of future AI intelligence: “a collective of diverse” AIs rather than “one big AI.” These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen At the time of the conversation, AI systems were not very differentiated, with most actors using LLMs and reasoning models.Now The speaker describes a model in which duplicated AI instances acquire different job skills and are then merged so their skills can be combined compositionally.The earlier claim is about the state of AI systems “right now,” saying they are “not very differentiated.” The later passage describes a compositional setup in which duplicated instances acquire different skills and are merged; it does not clearly assert that this is the current state of AI systems or revise the earlier assessment.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker considers the semantics represented by current models very superficial, describing them as primarily distributions of words.Now The speaker says models appear to use representations at multiple levels of resolution along a spectrum of abstraction.The earlier claim says current models’ semantics are “very superficial” and “really about distributions of words.” The later claim says models use representations “at multiple levels of resolution” along a spectrum of abstraction. These statements may suggest a more nuanced view, but the later claim does not clearly retract or contradict the earlier assessment, and it addresses representations rather than explicitly revising the earlier claim about semantics.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker suggests that the weakest plausible AGI could perform at the level of an ordinary person, who may reason poorly and exhibit substantial confirmation bias.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim concerns the capability threshold for “the weakest AGI,” saying it could be “as good as Joe Sixpack” and noting ordinary people may reason poorly and show confirmation bias. The later claim concerns expert-guided AI in “well-specified domains,” where “magical things happen” despite some ambiguity. These address different conditions and do not establish a change in the speaker’s view of the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes AI systems could become orders of magnitude more data-efficient than they currently are.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is that AI systems could become “orders of magnitudes more data efficient.” The later claim is that an automated AI scientist could be possible “in the near future.” These concern different questions, and neither statement indicates a change in the other position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects many distribution-shift cases to continue exposing failures in AI systems next year.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim predicts that distribution-shift cases will continue to expose AI failures next year. The later claim discusses expert-guided AI producing strong results in well-specified domains while ambiguity remains. These address different questions, and the later statement does not establish a changed view about the earlier prediction.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers voluntary regulation and industry self-regulation probably insufficient to ensure disclosure of AI model capabilities.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim concerns whether voluntary or industry self-regulation can ensure disclosure of AI model capabilities. The later claim concerns expert-guided AI performance and ambiguity in well-specified domains; it does not address regulation or disclosure, so these statements do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Many reported prompt-specific odd answers from LLMs have been fixed.Now In the speaker's account, construct validity is checked by an LLM judge that ensures the system is not cheating and is doing the task correctly.The earlier claim says many prompt-specific odd answers have been fixed. The later statement describes an LLM judge checking whether an experiment is conducted correctly. They address different topics, so the later statement does not show a change in the speaker’s view about whether those odd answers were fixed.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen AI has made some progress toward AGI and very little progress toward alignment, according to the speaker.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says AI has made “some progress toward AGI and very little progress toward alignment.” The later claim discusses expert-guided AI producing “magical” results in well-specified domains while ambiguity remains. These address different questions, so the later statement does not show a change in the speaker’s view about progress toward AGI or alignment.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says the participants agree that current efforts on AI alignment are inadequate and need substantial improvement.Now The speaker says composable AI agents would make them much more worried.The earlier statement says current AI alignment efforts are inadequate and need improvement. The later statement is about a different issue: the speaker says composable agents “would make me much more worried,” and frames that concern conditionally with “if they were” composable. It does not establish a changed position on alignment.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker predicts that neurosymbolic AI will begin to take off in three or four years.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim predicts that neurosymbolic AI will “start[] to take off” in three or four years. The later claim describes a next wave involving agents “talking to each other and learning and specializing.” These are not clearly opposing predictions: the claims concern different AI developments, and the later excerpt does not establish whether the speaker still expects neurosymbolic AI to take off.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker characterizes the chess system's knowledge as task-specific rather than general chess knowledge.Now The speaker rejects intelligence as a magical quality and says it depends on the scope of tasks and on existing knowledge.The earlier claim says the chess system had “sterile,” task-specific knowledge rather than “general knowledge of chess.” The later claim is about intelligence more broadly, arguing that it depends on “a scope of tasks” and “what you already know.” These statements concern different questions and do not establish a change in the speaker’s view of the chess system’s knowledge.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says he identified a set of real gaps in 2001.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier statement is about whether the speaker identified real gaps in 2001. The later statement only describes external intelligence systems as having “very complex dynamics.” It does not address those gaps or establish a changed position on the same question.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says current AI systems have not been good at generating genuinely new ideas.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says current AI systems have not been good at generating “genuinely new ideas.” The later claim imagines a future in which diverse AIs examine problems from different angles, rather than asserting that current systems can generate genuinely new ideas. These statements address different timeframes and do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker estimates that choosing a legal chess move takes a human about 100 milliseconds, while expressing uncertainty about the estimate.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The earlier claim estimates how long a human takes to choose a legal chess move («about 100 milliseconds»). The later claim discusses whether AlphaGo’s Move 37 was «innovative but not creative» in an inherent sense. These address different questions, so they provide no evidence of a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that AI systems will improve programmer productivity, while noting caveats concerning the security and maintainability of generated code.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.These claims address different questions. The earlier claim predicts AI will improve programmer productivity, with caveats about code security and maintainability. The later claim anticipates AI agents communicating, learning, and specializing, with possible harmful behaviors such as collusion. The later statement does not revise or deny the earlier productivity prediction.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker believes the problems he described in his book remain unsolved.Now In ambiguous problem domains, what specification is needed remains mysterious, according to the speaker.The earlier claim says the problems described in the speaker’s book “have [not] been solved.” The later claim says that, in “the ambiguity regime,” the needed specification is “really mysterious.” The excerpts do not establish that these are the same problem or that the later statement reverses the earlier assessment.
• NOT COMPARABLE · 93%24 Jun → 8 SeptThen Pure neural networks lack operations over variables and fail at tasks that require such operations.Now The speaker says fractured, entangled neural-network representations have some degree of composability.The earlier claim is that pure neural networks lack operations over variables and therefore fail at tasks requiring them. The later claim is only that fractured, entangled representations are “composable to some degree.” That does not establish that they perform the variable operations discussed earlier or overturn the earlier claim.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker reports that AI performance was poor on a chess task involving a position with three white queens and an immediate win for black.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says performance “was not great” on an unusual chess position involving three white queens and an immediate win for black. The later claim says expert-guided AI can produce “magical things” in “well-specified domains,” while acknowledging ambiguity. These concern different task conditions, so the statements do not establish a change in position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker thinks current AI approaches are already encountering a bottleneck in at least one of their raw resources.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says current AI approaches are encountering a bottleneck in at least one raw resource. The later claim says an automated AI scientist could be possible in the near future. These address different questions, and neither claim establishes whether the bottleneck prevents that possibility, so they do not show a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker does not think AI will surpass human abilities across the relevant dimensions within two years.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is that AI will not surpass human abilities across the relevant dimensions “within two years.” The later claim is that an “automated AI scientist” could be possible “in the near future.” Those are not necessarily the same threshold: an AI scientist could be possible without surpassing humans across those dimensions, and “near future” does not establish a two-year timeframe. The statements therefore do not show a clear change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says the participants' forecasts are closer than people may assume: one concentrates probability at five years out and later, while the other assigns some probability to two or three years out, and both assign some probability to 2045.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim concerns how close two participants’ long-term forecasts are (“probability mass to five years out” versus some probability at two or three years out). The later claim concerns expert use of a system in well-specified domains (“magical things happen”) while ambiguity remains. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says newer AI systems interpolate substantially better than earlier systems.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says newer systems are “much better at interpolating” than previous systems. The later claim imagines future AI as “diverse AIs” examining problems from different angles rather than “one big AI.” These address different questions—comparative interpolation ability versus a possible future AI arrangement—so they do not establish a change of mind.
• NOT COMPARABLE · 84%24 Jun → 8 SeptThen The speaker says AI systems perform worse when evaluated on problems outside the typical cases.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns AI systems performing worse on problems outside typical cases (“move these things off the typical problem, they're even worse”). The later passage concerns composability or combining AI skills, but the fragment “that doesn't really marry with my experience” does not clearly identify what “that” refers to. These are not clearly the same underlying question, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Some critical ideas needed for further AI progress are currently missing, according to the speaker.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says, “We're missing some critical ideas” for further AI progress; the later says an automated AI scientist “is possible in the near future.” These can both be true: the later claim does not say those ideas are no longer missing or that progress has already solved the gap.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says GPT-3 has produced an illegal chess move in which a queen jumps over a knight.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier statement is about GPT-3 making an illegal chess move (“a queen jump over a knight”); the later statement is about an AI hill-climbing toward a specified math or knowledge task. These are different capabilities and questions, so the later claim does not establish a changed position on chess reasoning.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker predicts that two years before ASI is achieved, AI systems will have full solutions to all the cognitive problems he listed.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier statement predicts that AI would have “full solutions” to listed cognitive problems two years before ASI. The later statement describes a conditional optimization result: an AI can “converge towards the solution” when there is “enough of a specification.” These concern different levels of claim, and the later statement neither confirms nor denies the earlier prediction.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker argues that AGI should not be expected to share human memory limitations.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.These claims address different questions. The earlier claim says AGI should not be expected to share human memory limitations; the later claim says an ecosystem of AIs and humans can participate in a superorganism. The later statement does not revise or contradict the earlier one.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says current AI companies are not entirely trustworthy.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier statement is about whether current AI companies are trustworthy (“not entirely trustworthy”). The later statement concerns whether AI skills or agents can be composed (“doesn't really marry with my experience”). These are different questions, so the claims do not establish a change in position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker assesses AI progress primarily through cognitive-science problems rather than quantitative progress.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement says the speaker approaches AI progress through “cognitive science.” The later statement imagines future AI as “diverse AIs” examining problems from different angles rather than “one big AI.” These address different questions—how to assess AI progress versus how future AI might be organized—and the later view does not negate the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker estimates that O3 chooses a legal chess move with roughly 80% reliability.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier statement estimates O3's chess-move reliability as “maybe 80%.” The later statement discusses expert use in “well-specified domains,” saying results can seem magical while “a bit of a space of ambiguity” remains. These address different questions, so the later claim does not establish a change in the speaker’s view of chess reliability.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes current AI systems are already close to performing at the level of an apprentice software engineer.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says AI is “close” to performing like an “apprentice engineer.” The later claim describes a possible future recursive system as “a different type of technology.” These address different questions, and neither establishes a changed position on the other.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker believes that LLMs are not really the route to AGI and that achieved AGI will look quite different; he thinks Jan LeCun shares this view.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says LLMs are “not really the route to AGI” and that AGI will “look pretty different.” The later claim describes expert-guided AI producing impressive results in well-specified domains, but does not say whether LLMs are a route to AGI or what AGI will look like. These statements do not establish a change in position on the same question.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker considers GPT-3's chess-rule failures striking and detectable without expert knowledge.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim is about GPT-3 making conspicuous chess-rule errors (“have a queen jump over a knight”) that do not require expert knowledge to spot. The later claim is about expert-guided AI producing “magical” results in well-specified domains while ambiguity remains. These claims concern different conditions and do not establish that the speaker reversed their view of the chess failures or of AI capability overall.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Gary Marcus believes more people probably read AI 2027's darker scenario than its positive scenario.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier claim reports Gary Marcus’s belief about which AI 2027 scenario more people read; it does not establish the speaker’s own position on the later question of whether AIs and humans can participate in a superorganism. These claims address different questions, so no change of mind can be inferred.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker estimates that developing substantially better reasoning, planning, and stable world models will take five to ten years.Now The speaker says current agents adapt through chain-of-thought processes and skill-surface adaptation, essentially using memory systems.The earlier claim estimates that developing substantially better reasoning, planning, and stable world models will take “five or ten years.” The later claim describes how current agents adapt using chain-of-thought, skills, and memory systems. It does not say whether those systems amount to the better world models the earlier claim anticipated, so the statements do not establish a change in position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker predicts that by 2035, people will still use LLMs for some tasks but will regard them as not having been the answer to achieving advanced AI.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim predicts that by 2035, LLMs will still be used for some tasks but “weren’t really the answer” to achieving advanced AI. The later claim imagines future AI as “diverse AIs” collaborating rather than “one big AI.” These address different aspects of AI’s future, and the later statement does not say whether LLMs will or will not be the answer to advanced AI.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Symbolic systems can perform operations over variables that pure neural networks fail to perform.Now Models capture the symbolic residue of language when language is ingested into them.The earlier claim is that symbolic systems can perform operations over variables that pure neural networks fail to perform. The later claim is that models capture the “symbolic residue of language” when language is ingested. These address different questions, and the later statement neither confirms nor denies the earlier position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker considers 50% performance an unusual and problematic measure for interpreting the graph.Now The speaker believes shortcuts and validity problems are less problematic in this research than he had expected.The earlier claim calls “50% performance” a “very weird measure” for interpreting a graph. The later claim says “shortcuts” and “validity problems” are “not as much of a problem as we thought.” These concern different aspects of the research, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says that he and Yann LeCun consider current AI systems' purported reasoning insufficiently robust.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says current AI systems’ purported reasoning is “not ... robust enough.” The later claim says expert-guided AI can do “magical things” in well-specified domains while ambiguity remains. These statements address related but distinct points, and the later passage does not say that AI reasoning has become robust or retract the earlier assessment.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The term AGI has been defined in different ways.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The statements address different questions. Earlier, the speaker says AGI has “been defined in different ways”; later, he imagines future AI as “a collective of diverse” AIs rather than “one big AI.” The later view does not establish a change in his position on how AGI is defined.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks current AI approaches are already encountering a bottleneck in at least one of their raw resources.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says current AI approaches are encountering a bottleneck in a raw resource. The later claim imagines future AI as diverse systems collaborating rather than one large AI. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker says domain-specific AI systems currently perform much better than general-purpose AI systems.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says domain-specific systems are “much better than the general ones right now.” The later claim says expert-guided AI can produce “magical” results in well-specified domains while some ambiguity remains. The later statement does not compare domain-specific and general-purpose systems, so it does not establish a change in that position.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker expects an AI system to follow a specified algorithm for a task.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier extracted claim concerns following a specified algorithm, while the later claim concerns converging on a solution in optimization problems with enough specification. The later statement does not clearly reverse the earlier one; the passages address related but distinct capabilities.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says polls indicate that 75% of Americans are already worried about AI and expects public concern to rise, seeing no active measure likely to reduce it in time.Now The interviewer asks whether reduced human participation in the economy could cause economic mode collapse, in the context of an AI dividend.The earlier statement says that 75% of Americans are already worried about AI and that their concern is expected to rise. The later passage is a question about whether reduced human participation could cause “mode collapse of the economy”; it does not state the speaker’s own position. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The interviewer predicts that no durable advantage will persist if AI development remains based on LLMs.Now The interviewer considers an entirely AI-run economy plausible but suspects it might become degenerate in some way.The earlier claim concerns whether any durable advantage will persist in LLM-based AI development; the later concerns whether an entirely AI-run economy might become degenerate. These are different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says current AI systems struggle to multiply large numbers unless they can use tools.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says current AI systems struggle with large-number multiplication unless they can use tools. The later claim imagines future AI systems as diverse AIs collaborating across areas of expertise rather than one large AI. These address different questions, so the later statement does not show a change in the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes solving visual-intelligence problems, including video-related problems, within three or four years would be very lucky.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The earlier claim is about the timeline for solving visual-intelligence problems: solving them in “three or four years” would be “really lucky.” The later claim only says that external intelligence systems have “very complex dynamics.” It does not state a different timeline or position on solving visual intelligence, so there is no comparable later claim from which to infer a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Symbolic systems can perform operations over variables that pure neural networks fail to perform.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier statement is about the capabilities of symbolic systems versus pure neural networks: “Symbolic systems can perform operations over variables that pure neural networks fail to perform.” The later statement concerns expert use of “the system” in well-specified domains and remaining ambiguity. It does not address that earlier comparison, so there is not enough evidence to infer a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker sees a possible case for intervention on safety-critical AI issues, potentially through incentives rather than delaying development.Now The speaker says composable AI agents would make them much more worried.The earlier claim says the speaker sees a possible case for intervention on safety-critical AI issues, while the later claim says composable agents would make them “much more worried.” These statements address different questions and can both be true; the later conditional concern does not show a change in the speaker’s view on intervention.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says there are already signs of neurosymbolic AI beginning to take off.Now The speaker says many startups are pursuing recursive-improvement superintelligence.The earlier claim says there are already signs of neurosymbolic AI beginning to take off, while the later claim says many startups are pursuing recursive-improvement superintelligence. These concern different developments, and the later statement does not say whether neurosymbolic AI is taking off or not.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker considers it an exceptionally rapid convergence if the issues he has worried about for 25 years were all resolved by 2027.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement describes resolving a set of long-standing concerns by 2027 as “everything everywhere all at once”; it does not state a position on automated AI scientists. The later claim says an automated AI scientist is “possible in the near future.” These address different questions, so the later statement does not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker sees no sound argument that automating the full AI research-and-development process is impossible.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier claim concerns whether there is a sound argument that full AI research-and-development automation is impossible. The later claim concerns whether an ecosystem of humans and AIs can participate in a superorganism. These are different questions, and the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen Reliably getting AI systems to use tools is difficult.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim is specifically that getting AI systems to use tools reliably is difficult. The later claim concerns expert-guided AI producing impressive results in well-specified domains while some ambiguity remains. It does not state whether reliable tool use is difficult, so these claims do not establish a change of position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes that AI progress forecasting cannot be done in a fully closed-form way.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim concerns forecasting AI progress: it says it cannot be done “in a closed form way.” The later claim concerns expert-guided AI in well-specified domains and says there is still “a space of ambiguity.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes solving visual-intelligence problems, including video-related problems, within three or four years would be very lucky.Now The speaker views intelligence as perspectival and expects it to be uneven and fragmented across capabilities.The earlier claim is a forecast about timing: solving visual-intelligence problems in “three or four years” would be “really lucky.” The later claim describes the structure of intelligence as “perspectival” and “fractured and jagged.” These address different questions, so the later statement does not show a change in the earlier forecast.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker says some instances of the problems can be solved, but the problems he observes appear to reflect the same underlying core problems.Now The speaker believes shortcuts and validity problems are less problematic in this research than he had expected.The earlier claim says that some problems can be solved but that the problems the speaker observes reflect the same underlying core problems. The later claim says that validity problems in a particular research direction are less problematic than he had expected. The excerpts do not establish that these are the same problems or the same question, so they do not show a clear change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes OpenAI is not fully forthcoming about what it did.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The claims concern different questions. The earlier statement says OpenAI is “not an entirely forthcoming” about what it did; the later statement says the speaker’s experience does not align with an idea about composing AI skills by merging agents. Neither claim establishes a position on the other topic, so there is no basis to assess a change of mind.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker cautions that explaining AI failures as failures on tasks the systems were not trained on is an oversimplification, while discussing AI risk.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says that explaining AI failures as failures on tasks they were not trained on is “an oversimplification.” The later claim concerns expert-guided AI producing impressive results in “well-specified domains” while leaving “a bit of a space of ambiguity.” These address related but distinct points, and the later statement does not clearly reverse the earlier one.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says LLMs might be replaced because they are very inefficient in their use of energy and data.Now In the speaker's account, construct validity is checked by an LLM judge that ensures the system is not cheating and is doing the task correctly.The earlier claim is that LLMs might be replaced because they are inefficient in energy and data. The later claim describes using an LLM judge to check whether an experiment is valid; it does not state whether LLMs should be replaced or whether their efficiency has changed. These concern different questions, so there is no evidence of a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says humans are extremely data-efficient on many, but not all, problems.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim says humans are “fabulously data efficient on many problems,” while the later claim concerns expert use of a system producing “magical” results in well-specified domains, with “a space of ambiguity” remaining. These address different questions, so the later statement does not show a changed view about human data efficiency.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Reliably getting AI systems to use tools is difficult.Now For tasks on the long tail, an expert can guide an AI model by making its path of least resistance roughly correct, increasing the chance it produces the correct result.The earlier claim is that getting AI systems to use tools reliably is difficult. The later claim is about experts guiding models on long-tail tasks so their likely path is more correct; it does not say whether tool use has become reliable or difficult. These address different capabilities, so the later statement does not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects transparency to be less likely as AI systems approach AGI or become meaningfully differentiated.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim concerns expectations about transparency, saying the speaker is “less optimistic about transparency” as systems approach AGI or become differentiated. The later claim says an efficiently self-adapting recursive loop would be “a different type of technology”; it does not state whether such a system would be more or less transparent. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker characterizes Sam's biggest concern at a Senate appearance a few weeks earlier as the company not making all the money or extracting all the value it could.Now The speaker characterizes Amazon as an adaptive ecosystem that continually restructures itself, develops capabilities, and then optimizes those capabilities.The earlier statement describes what the speaker says Sam’s concern was—“not making all the money” or extracting all the value possible. The later statement describes Amazon as an adaptive ecosystem that restructures and optimizes itself. These address different subjects, so the later claim does not establish a change in the speaker’s own position on the earlier question.
• NOT COMPARABLE · 87%24 Jun → 8 SeptThen The speaker expects an AI system to follow a specified algorithm for a task.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim concerns whether a system can follow a specified algorithm, and the surrounding passage immediately notes that current systems cannot reliably follow even chess rules. The later claim says a model will “hill climb toward” a specified task’s solution; that does not establish that it can reliably follow an algorithm or reach the solution, so the statements do not clearly show a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects future periods of AI progress to exceed the progress of the last few years and expects AI to be enormously more advanced in 30 years.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim concerns the pace and extent of future AI progress: the speaker expects periods of progress to “exceed the last few years” and AI to be “enormously ahead” in 30 years. The later claim concerns how future AI might be organized: “diverse AIs” collaborating “rather than one big AI.” These claims address different questions and are compatible; the later statement does not show a change in the earlier forecast.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker sees a possible case for intervention on safety-critical AI issues, potentially through incentives rather than delaying development.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The statements address different questions. The earlier claim says the speaker can “see an argument for some kind of intervention” on safety-critical AI issues; the later claim imagines a self-adapting recursive AI loop and says it would be “a different type of technology.” The later statement does not indicate whether the speaker’s view on intervention changed.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says humans reach theoretical limits on some psychometric tasks, such as detecting whether a photon is present or absent.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier statement concerns human performance limits on psychometric tasks, such as detecting a photon's presence or absence. The later statement concerns Elon Musk's strengths in recognizing promising future opportunities rather than engineering or optimization. These address different questions, so they do not establish a change in the speaker's position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker predicts that by 2035, people will still use LLMs for some tasks but will regard them as not having been the answer to achieving advanced AI.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim predicts that LLMs will be seen as “not really the answer” to achieving advanced AI, while still being used for some tasks. The later claim says expert-guided AI can produce “magical” results in well-specified domains but leaves ambiguity. It does not address whether LLMs are the answer to achieving advanced AI, so these statements do not establish a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that if the necessary ideas are missing, progress depends on when new ideas are developed.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier claim concerns how progress depends on developing missing ideas; the later claim concerns Elon Musk’s particular strengths. Neither statement clearly takes a position on the same underlying question, so they do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes transparency has limited value to an AI organization until it develops a unique capability that competitors cannot rapidly reproduce.Now The speaker argues that highly complex, information-sharing collectives can build societies, causing some objections to this possibility to lose force.The earlier claim concerns when an AI organization should be transparent about its capabilities: transparency has limited value until it has something competitors cannot quickly reproduce. The later claim concerns whether complex, information-sharing collectives can build societies. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen In a science-fiction scenario set around 2045, the speaker imagined that a complete brain simulation would run slower than a human brain and fail to explain how the brain works.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement describes a science-fiction scenario in which a brain simulation would not explain how the brain works; it does not state a position about whether AI could conduct scientific research. The later claim explicitly says an automated AI scientist may be possible in the near future. These address different questions, so the claims do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says that smart people can solve tasks such as the Tower of Hanoi and that six-year-olds can learn chess rules.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim says “smart people” can do tasks such as Tower of Hanoi and learn chess rules. The later claim says an AI can be given a mathematics task and “hill climb towards it.” These concern different subjects and examples, so they do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes AI systems could become orders of magnitude more data-efficient than they currently are.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AI could become “orders of magnitudes more data efficient”; the later claim imagines “diverse AIs” collaborating rather than “one big AI.” These address different questions—data efficiency versus how future AI systems might be organized—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker tentatively identifies the agents in the separate graph as possibly web-browsing agents.Now The speaker thinks current agents are not yet composable.The earlier claim tentatively identifies agents in a separate graph as “possibly web-browsing agents.” The later claim is about whether current agents are composable: “I don't think they are yet composable.” These address different questions, so the later statement does not show a change in the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says someone might develop a better method for learning distributions; he describes distribution learning as a very helpful, but not exclusive, cognitive skill.Now The speaker argues that apparent general capability is partly an illusion: models can be shaped into specialized experts in any domain, but that expertise is latent rather than explicit.The earlier claim says distribution learning is a helpful cognitive skill and that someone might find a better method for it. The later claim concerns whether models have general capabilities or latent, domain-specific expertise. These do not establish opposing positions on the same question, so there is no clear evidence of a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker does not expect automation of current AI systems to generate the Einstein-level innovations needed for AGI or ASI.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says the speaker does not expect “automating the current machines” to produce the innovations needed for AGI or ASI. The later claim imagines a future made up of “diverse AIs” collaborating rather than “one big AI.” The later statement does not say whether that arrangement would achieve AGI or ASI, so it does not establish a change in the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Augmented data is not a universal solution; it works better for tasks such as mathematics, where the quality of the augmented data can be verified.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim addresses augmented data: it is “not a universal solution” and works better when its quality can be verified. The later claim addresses expert use of a system in “well-specified domains,” while acknowledging “a bit of a space of ambiguity.” It does not say whether augmented data is or is not a universal solution, so these claims do not establish a change of position on the same question.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says the gaps he identified in 2001 concerned multilayer perceptrons rather than transformers, which had not yet been invented.Now The speaker describes neural-network representations as fractured and entangled, somewhat janky, and not fully robust.The earlier claim is about which architectures his 2001 observations concerned: multilayer perceptrons, not transformers. The later claim characterizes neural-network representations as “fractured” and “not completely robust.” These address different questions, so they do not establish a change in his position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker predicts that AI systems will improve programmer productivity, while noting caveats concerning the security and maintainability of generated code.Now The speaker says composable AI agents would make them much more worried.The earlier claim predicts that AI will improve programmer productivity, with caveats about code security and maintainability. The later claim says composable AI agents would make the speaker “much more worried.” These address different questions—productivity effects versus concern about agent composability—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks most AI safety arguments focus on AGI that is at least as intelligent as most smart people.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim is about what capability level most AI safety arguments focus on: AGI “at least as smart as most smart people.” The later claim is about what form future AI might take: “diverse AIs” with different expertise rather than “one big AI.” These address different questions, so the later statement does not show a change in the speaker’s view about the earlier claim.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says the chess system could not solve a contrived position requiring three white queens and a Black win on the next move.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim concerns a specific chess system failing a contrived position: it “wouldn't be able to do that at all.” The later claim describes an AI model making progress toward a specified task, such as a mathematics problem: it “will hill climb towards it.” These concern different tasks and possibly different systems, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argued by 2001 that AI systems need symbols to perform abstract operations over variables.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AI systems need symbols for abstract operations over variables; the later claim imagines multiple specialized AIs collaborating rather than one large AI. These address different questions, and the later statement does not say whether the speaker still holds the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes current AI systems are already close to performing at the level of an apprentice software engineer.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says current AI is “close” to apprentice-engineer performance. The later claim imagines future AI as “diverse AIs” working from different angles rather than “one big AI.” These address different questions—present-day capability versus future system design—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker thinks the form of AGI relevant to the risks under discussion would probably exceed humans on short-term-memory tasks while matching human versatility and flexibility.Now The speaker says composable AI agents would make them much more worried.The earlier statement concerns what capabilities risk-relevant AGI would likely have: it could exceed people at short-term-memory tasks while matching human versatility. The later statement concerns whether AI agents are composable and says that composability would make the speaker more worried. These address different questions, and the later claim does not revise the earlier capability assessment.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker says AI systems need a deeper kind of semantics than current models provide.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says current models’ semantics are “very superficial” and that AI needs “a deeper” semantics. The later claim says models can create “novel combinations” of existing knowledge to solve tasks. That describes adaptivity, not whether their semantics are deep, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker regards an Apple paper involving Tower of Hanoi as an example of AI systems struggling with distribution shift.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim says the Tower of Hanoi paper illustrates AI systems struggling with distribution shift. The later claim says a model given a specified mathematics task will hill climb toward it. These address different claims and do not establish that the speaker changed their view about distribution shift.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker says AI has made substantial progress in interpolation but has not solved extrapolation, and characterizes much of the progress as interpolation rather than extrapolation.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says AI has made progress in “interpolation” but “hasn't solved the extrapolation problem.” The later claim concerns whether AI skills can be composed across agents, and the quoted passage is fragmentary. These are related capability questions, but the statements do not establish a change on the same underlying question.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The interviewer predicts that no durable advantage will persist if AI development remains based on LLMs.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier passage concerns whether LLMs will produce a durable advantage: it says “nothing’s going to be durable.” The later passage concerns expert-guided AI in well-specified domains and collective intelligence; it does not assert that a durable LLM advantage exists or otherwise reverse the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says the scenario and reality were roughly aligned during the first couple of months, but have diverged in their pace.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier statement concerns whether events matched a predicted timeline: it says there was agreement during the first few months but “a divergence in speed.” The later statement concerns whether AI agents’ skills can be composed, saying that idea “doesn't really marry with my experience.” These are different underlying questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen AI had already surpassed human ability in mathematics by roughly 60 to 80 years ago.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim says AI had already surpassed human ability in mathematics “60 to 80 years ago.” The later statement says an AI will “hill climb” toward a solution to a specified mathematics problem; it does not say whether AI is better or worse than humans. The claims address different questions, so the later statement does not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says domain-specific AI systems currently perform much better than general-purpose AI systems.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says domain-specific systems perform “much better than the general ones right now,” comparing current performance. The later claim imagines a future with “diverse AIs” collaborating “rather than one big AI,” describing a possible future architecture. These positions can both be true; the later statement does not establish a change in the speaker’s view of current systems.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says the gaps he identified in 2001 concerned multilayer perceptrons rather than transformers, which had not yet been invented.Now The speaker says fractured, entangled neural-network representations have some degree of composability.The earlier statement is about which architectures the speaker’s 2001 observations concerned (“multilayer perceptions, not … transformers”). The later statement is about whether fractured, entangled representations are composable (“to some degree”). These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says current AI systems may refuse an obvious biological-weapons request, but such refusals are easily bypassed through jailbreaking.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim concerns whether current systems’ refusals of biological-weapons requests can be bypassed by jailbreaking. The later claim imagines a future system that adapts and evolves autonomously, calling it “a different type of technology.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 93%24 Jun → 8 SeptThen The speaker believes AI systems are not close to matching a top software engineer, because that requires understanding the problem, the human goals behind it, and multiple domains.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim is that AI is “not close” to matching a top software engineer, citing understanding problems, human goals, and multiple domains. The later excerpt says the speaker’s experience “doesn't really marry” with an idea about composable AI skills. These address different capabilities, and the later excerpt does not establish a changed position on whether AI can match a top software engineer.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker sees little, though possibly some, progress in AI systems' ability to match the flexibility of human expert knowledge.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement assesses current progress, saying there is “little, though possibly some,” progress in matching the flexibility of human expert knowledge. The later statement imagines a possible future approach—“diverse AIs” examining problems from different angles “rather than one big AI.” It does not say that this approach has already achieved that flexibility or revise the earlier assessment.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen According to the speaker's characterization of Jan LeCun's view, current AI is not close to the goal because it lacks genuine world models.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier statement is explicitly framed as Jan LeCun’s view (“he would say”), not the speaker’s own position. The later statement is the speaker’s own claim that models can “create novel combinations” to solve tasks. These claims do not establish a change in the speaker’s view about whether AI has world models.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks AI systems probably received substantial chess-related training data from publicly available sources.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is that AI systems probably received substantial chess-related training data from public sources. The later claim is that an automated AI scientist could be possible in the near future. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker expects AI systems near AGI to perform about as well as, or at least approach, domain-specific systems.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says near-AGI systems should perform “about as well as” domain-specific systems. The later claim imagines “diverse AIs” collaborating “rather than one big AI.” These address different questions—expected capability versus system organization—and the later statement does not deny the earlier expectation.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker predicts that AI systems will remain highly vulnerable to outlier problems one year later.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim predicts that AI systems will remain “very vulnerable to those outliers” a year later. The later claim concerns expert-guided AI in “well-specified domains” and says there is still “a bit of a space of ambiguity.” It does not say whether systems remain vulnerable to outlier problems, so these statements do not establish a change in position.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have substantially better reasoning and planning methods and more stable world models.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim predicts “much better ways of doing reasoning and planning” and “much more stable world models.” The later claim says an automated AI scientist could be possible “in the near future.” These are related, but the statements do not establish that the later claim reverses the earlier one; the earlier claim gives no clear timeframe for those improvements, and the two claims concern different capabilities.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen In a science-fiction scenario set around 2045, the speaker imagined that a complete brain simulation would run slower than a human brain and fail to explain how the brain works.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier passage describes a science-fiction scenario in which a brain simulation would run slower than humans and fail to explain how the brain works. The later passage imagines an efficiently self-adapting AI loop as “a different type of technology.” These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says AI systems perform worse when evaluated on problems outside the typical cases.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim is about AI systems performing worse on problems outside typical cases. The later claim says a model will “hill climb towards” a specified task, such as a mathematics problem; it does not say the model succeeds equally well on unusual or atypical problems. These statements address different aspects of performance, so they do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes appropriately designed software should let machines outperform humans in areas suited to machines and at least achieve human-level performance overall.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier claim concerns whether appropriately designed software could let machines outperform humans in some areas and reach human-level performance overall. The later claim says an ecosystem of AIs and humans can participate in a “super organism.” These address different questions, and the later statement does not revise or deny the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers Grok's failure to understand the three-in-a-row condition in the tic-tac-toe variant a conceptual weakness.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The statements address different questions. The earlier claim calls Grok’s failure to recognize the tic-tac-toe condition a “conceptual weakness”; the later claim distinguishes AlphaGo’s Move 37 as “innovative but not creative” in an account of creativity and understanding. Neither statement establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes that AI progress forecasting cannot be done in a fully closed-form way.Now The speaker believes an automated AI scientist could be possible in the near future.These claims address different questions. The earlier claim says forecasting AI progress cannot be done “in a closed form way”; the later claim says an automated AI scientist could be possible “in the near future.” The latter does not establish a changed view about forecasting.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker thinks Jan LeCun's approach is closer to his own, though LeCun would probably use different litmus tests.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim says LeCun’s approach is “closer to mine,” while the later claim says a proposed collective-intelligence view “would place some kind of limits on its scale.” These address different questions, and the later statement is part of a discussion of a proposed scenario; it does not establish a changed position on the earlier comparison.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen Pure neural networks lack operations over variables and fail at tasks that require such operations.Now The speaker describes neural-network representations as fractured and entangled, somewhat janky, and not fully robust.The earlier claim is about whether pure neural networks can perform operations over variables and handle tasks requiring them. The later claim describes neural-network representations as “fractured,” “janky,” and “not completely robust.” That does not establish a changed position on variable operations or task performance.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker objects to the claim that continued scaling will automatically solve AI problems.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim objects to scaling automatically solving AI problems; the later claim says the speaker’s experience does not align with AI skills being composable. These concern different propositions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker thinks Claude and Gemini may have acceptable frontier-math scores, while current systems are being optimized for the test through augmentation whose nature is unknown.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim concerns whether Claude and Gemini have acceptable frontier-math scores and whether current systems are optimized for tests. The later claim concerns how expert use can produce strong results in well-specified domains despite ambiguity. These statements do not address the same underlying question, so they provide no clear evidence of a change in position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker proposes treaties establishing AI red lines and avoiding crossing them until the relevant questions have answers.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says we “just shouldn't do it until we have answers” and proposes treaties around AI red lines. The later claim only imagines an autonomous recursive loop and describes it as “a different type of technology”; it does not say whether the speaker now supports or opposes developing it. The later statement therefore does not establish a change in position.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker thinks AI progress could accelerate quickly once critical ideas are found, analogous to the rapid progress in molecular biology after Watson and Crick figured out DNA.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is that AI progress could accelerate once critical ideas are found (“they could go fast”). The later claim is that an automated AI scientist could be possible in the near future. These concern related but distinct questions, and the later statement does not say whether the speaker’s view about the pace of AI progress has changed.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker argues that AGI should not be expected to share human memory limitations.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns whether AGI should be expected to share human memory limitations (“should not be expected”), while the later claim concerns whether AI skills can be composed by merging agents (“doesn't really marry with my experience”). These address different questions, so the later statement does not establish a change of mind about the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker reports that AI performance was poor on a chess task involving a position with three white queens and an immediate win for black.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim concerns poor performance on a particular chess position (“performance was not great”), while the later claim describes a model’s tendency to “hill climb” on a specified task, with mathematics as an example. The later statement does not say it will solve the task successfully, so these claims do not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes that AI systems' failure to distinguish types from tokens can cause details about individuals in the same category to blend together and produce hallucinations.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim says that failing to distinguish types from tokens can cause hallucinations about individuals. The later claim imagines future AI as a collective of specialized systems. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker estimated, with uncertainty, that GPT-2 used about 5%–10% of the internet.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim estimates how much of the internet GPT-2 used (“maybe 10% … or 5%”). The later claim concerns expert-guided AI performance and remaining ambiguity (“magical things happen” in well-specified domains, but “there’s still a bit of a space of ambiguity”). These address different questions, so they do not show a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says future AGI might use LLMs because they are good at distributional learning.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The claims address different questions. Earlier, the speaker says future AGI might use LLMs because they are “great at kind of distributional learning.” Later, he describes expert-guided AI producing “magical things” in “well-specified domains” while ambiguity remains. The later claim does not say whether he still thinks AGI might use LLMs, so no change of mind can be established.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker says someone might discover a better method for learning models of distributions than current LLMs provide.Now The speaker says reinforcement-learning-trained language models understand things differently from humans but address the exploration problem by learning how things fit together and exploring novel trajectories.The earlier statement speculates that someone might find a better method for learning distribution models than current LLMs. The later statement discusses how RL-trained language models explore novel trajectories by understanding how things fit together. These concern different questions, and the later claim neither confirms nor denies the possibility raised earlier.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen AGI is an abbreviation for artificial general intelligence.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim defines the abbreviation (“AGI is an abbreviation for artificial general intelligence”), while the later claim describes an AI’s ability to work toward a specified mathematics task or answer a knowledge question. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The interviewer does not expect any actor to gain a durable advantage if the AI paradigm remains roughly unchanged.Now The interviewer asks whether reduced human participation in the economy could cause economic mode collapse, in the context of an AI dividend.The earlier statement asserts that no actor would gain a durable advantage if the AI paradigm stayed roughly unchanged. The later passage is a question about whether reduced human participation could cause economic mode collapse; it does not state the interviewer’s own position, and it addresses a different issue.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says there are already signs that neurosymbolic AI is gaining momentum.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is that there are already signs neurosymbolic AI is gaining momentum; the later claim is that an automated AI scientist could be possible in the near future. These claims are compatible, but they address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker expects neurosymbolic AI eventually to undergo a state change but says its timing is very hard to predict.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier statement says neurosymbolic AI may eventually undergo a “state change,” while the later statement imagines an efficiently self-adapting recursive loop as “a different type of technology.” The later claim does not say whether that loop is the state change the speaker previously anticipated, or whether it changes their view about neurosymbolic AI.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says that if AI never becomes able to solve original problems, this limitation would provide an Achilles heel.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier statement concerns whether AI can solve original problems: “if AI never can do that,” the limitation would be an Achilles heel. The later statement says an AI will “hill climb” toward a specified task, such as a mathematics problem. That does not establish whether it can solve original problems, so the claims do not show a change in position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker reports that AI performance was poor on a chess task involving a position with three white queens and an immediate win for black.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns poor performance on a specific chess task involving an unusual board position. The later claim says that AI skills do not yet seem “composable,” apparently in response to a discussion of combining skills or agents. These address different capabilities, so the statements do not establish a change in position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says it is possible that his view that AI needs a fix is correct, although proving it is very difficult and his supporting evidence is qualitative.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim is that AI may need a “fix” because LLMs mostly exploit existing ideas rather than discover new ones. The later claim concerns how expert guidance can make AI perform impressively in well-specified domains while ambiguity remains. That observation does not establish a different view about whether AI needs a fix or can discover new ideas.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that scientific evaluation of AI systems is difficult because their training data is not transparent.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.These claims address different questions. The earlier statement says AI systems are hard to evaluate scientifically because their training data is not transparent. The later statement predicts that AI agents will communicate, learn, and specialize, possibly with harmful behaviors. Neither claim affirms or denies the other.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The interviewer says that people are already pushing against the proposed AI red lines, although they may not yet have crossed them.Now The interviewer asks whether reduced human participation in the economy could cause economic mode collapse, in the context of an AI dividend.The earlier passage says people are “pushing against” proposed AI red lines; the later passage asks whether reduced human participation could cause “mode collapse of the economy.” These concern different questions, and both are framed as the interviewer’s statement or question rather than a clear assertion of the same position. They do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that novel programming tasks, unlike familiar website coding with many examples, create out-of-distribution challenges that are especially relevant to programming AGI.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says novel programming tasks create “out-of-distribution” challenges relevant to programming AGI. The later claim says an efficiently self-adapting recursive loop would be “a different type of technology.” These address different questions, and neither statement clearly reverses the other.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Reliably getting AI systems to use tools is difficult.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim says AI systems can be hard to get to use tools reliably. The later claim says a model will “hill climb” toward a solution on a specified task, such as mathematics. That describes a different capability and does not establish whether tool use is reliable, so the statements do not show a change in position on the same question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says that system prompts against producing copyrighted material or hallucinating do not reliably prevent those behaviors.Now The speaker believes human intuitions about collective behavior are often unreliable and that such behavior could emerge in principle.The earlier claim concerns whether system prompts reliably prevent specific model behaviors (“they still do”); the later claim concerns whether complex collective behavior could emerge and whether human intuitions about it are reliable. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that novel programming tasks, unlike familiar website coding with many examples, create out-of-distribution challenges that are especially relevant to programming AGI.Now The speaker suggests that a distributed collective AI system might face additional bottlenecks due to communication among its agents.The earlier statement concerns how well AI handles novel programming tasks compared with familiar ones; the later concerns possible communication bottlenecks in distributed collective systems. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says distribution shift remains a problem for driverless cars.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says that “the distribution shift problem still remains” for driverless cars. The later claim says the speaker’s experience “doesn't really marry with” an idea about composing AI skills or merging agents. These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes current alignment systems still fail to reliably do what users intend and exhibit sorcerer’s-apprentice-style failures.Now The speaker thinks current agents are not yet composable.The earlier claim is that alignment systems “still don't really do what we want them to do” and have “sorcerer's apprentice style problems.” The later claim is that agents “are not yet composable.” These concern different properties, and neither statement establishes a changed view about the other.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker thinks AI systems probably received substantial chess-related training data from publicly available sources.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement concerns what chess-related training data AI systems probably received; the later statement imagines a future with diverse AIs collaborating rather than one large AI. These address different questions, so the later view does not indicate a change in the earlier belief.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes, and thinks Yann LeCun also believes, that LLMs are not the primary route to AGI and that achieved AGI will look substantially different.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim is that “LLMs are not really the route to AGI” and that AGI will “look pretty different.” The later excerpt says the speaker’s experience “doesn’t really marry” with an idea, while the surrounding passage concerns whether AI agents’ skills are composable. That is a different, narrower question; it does not establish a changed position on LLMs as a route to AGI.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says AI systems need a deeper kind of semantics than current models provide.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AI systems need “a deeper” kind of semantics than current models provide. The later claim imagines future AI as “diverse AIs” collaborating rather than “one big AI.” These concern different questions—semantic capability versus system organization—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes it is probably technically achievable to automate the full AI research-and-development process, though it may require technology beyond what is currently available.Now The speaker recalls that the report may project 60 million AI agents operating at 20 times human speed by 2032.The earlier claim asserts that automating the full AI R&D process is probably technically achievable, perhaps with future technology. The later claim merely recalls a report's projection about the number and speed of agents; it does not state whether the speaker now believes full AI R&D automation is achievable. These claims do not establish a change in the speaker's position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker considers the estimates for solving some of the discussed problems overly optimistic.Now The speaker believes that extensive optimization in semi-specific objective domains still leaves something important missing.The earlier claim says estimates for solving particular problems within a few years are overly optimistic; the later says optimization in semi-specific objective domains still leaves something missing. These are related but not clearly the same claim: the later statement does not establish whether the earlier timeline estimate has changed.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker thinks neurosymbolic methods are worth exploring for near-term safety-critical AI applications, while remaining unsure whether they help with superintelligence and saying they might offer some benefit.Now The speaker says composable AI agents would make them much more worried.The earlier claim is about whether neurosymbolic methods could help with near-term safety-critical AI applications; the later claim is that composable AI agents would make the speaker more worried. These concern different questions, and the later statement does not revise the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argued that current AI approaches cannot keep increasing their data by 100 times because the available data is insufficient.Now The speaker suggests that a distributed collective AI system might face additional bottlenecks due to communication among its agents.The earlier claim concerns limits on scaling data—“you can't just keep 100x-ing that because there just isn't enough data.” The later claim concerns communication bottlenecks in distributed collective systems—“additional bottlenecks” from message passing. These are different constraints, and the later statement does not revise or deny the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have much more stable world models.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim predicts that future AI systems will have “much more stable world models.” The later claim imagines future AI as “diverse AIs” collaborating rather than “one big AI.” These address different aspects of future AI—world-model stability versus system organization—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 82%24 Jun → 8 SeptThen Sam and Dario have hinted that they will soon be able to automate the full AI idea-to-implementation-and-scaling process, but the speaker is skeptical of those claims.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says, “I'm not really buying those claims” that Sam and Dario will soon automate the full “idea-to-implementation-and-scaling process.” The later statement says an “automated AI scientist” is possible in the near future, in a discussion about recreating experiments and taking creative next steps. These capabilities overlap, but the later claim does not clearly endorse the specific full-process claims the speaker previously doubted; the statements alone do not establish a change of mind.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker believes the unconstrained AGI they fear should not exhibit many of the edge-case failures being discussed.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim concerns whether an unconstrained AGI would have many of the specific edge-case failures under discussion; the later claim anticipates that interacting AI agents may exhibit harmful behaviors, including collusion. These are not clearly the same question or the same systems, and the later statement does not establish that the speaker changed their view about the earlier edge cases.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The term AGI has been defined in different ways.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier statement says, “The term AGI has been defined in different ways,” which is about varying definitions of AGI. The later statement says the speaker’s experience “doesn’t really marry with” the idea that AI skills can be composed by merging agents. These address different questions, so the later claim does not establish a change in the speaker’s view about AGI definitions.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen Sam and Dario have hinted that they will soon be able to automate the full AI idea-to-implementation-and-scaling process, but the speaker is skeptical of those claims.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim is skepticism about Sam and Dario's prediction that AI will soon automate the full idea-to-implementation-and-scaling process: the speaker says, “I'm not really buying those claims.” The later claim concerns expert-guided AI in “well-specified domains,” where “magical things happen,” while ambiguity remains. That narrower observation does not establish that the speaker has changed their view about full-process automation.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Sam discussed AI risks more often in the past than he does now.Now The speaker says composable AI agents would make them much more worried.These statements address different questions. The earlier claim says Sam “used to talk more” about AI risks than he does now; the later claim is the speaker’s own view that composable agents “would make me much more worried.” The later statement does not establish a change in the speaker’s position on the earlier claim.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says current AI systems have not been good at generating genuinely new ideas.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says AI has not been good at generating “genuinely new ideas.” The later claim says the speaker’s experience does not align with an idea about composing AI skills or merging agents. These address related but distinct capabilities, so the later statement does not establish a change in the speaker’s view about AI generating novel ideas.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen Automating the full AI research-and-development process would be destabilizing if technically feasible, but not if the claims are merely hype and the capability cannot be achieved.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says full AI research-and-development automation would be destabilizing if technically feasible, while the later statement says an automated AI scientist may be possible soon. The later claim does not establish that full-process automation is feasible or state a view about its effects, so these do not clearly show a change in position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen A neurosymbolic system can work if its neural networks reliably call the tools appropriate to the problem.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier claim says a neurosymbolic system can work if its neural networks reliably call the tools appropriate to the problem. The later claim concerns a different condition and capability: an AI system converging toward a solution when an optimization problem has enough specification. Neither statement directly changes or denies the other.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers humans fairly data-efficient but doubts that human data efficiency has reached theoretical limits across the board.Now The speaker believes human intuitions about collective behavior are often unreliable and that such behavior could emerge in principle.The earlier claim concerns whether humans have reached theoretical limits in data efficiency; the later claim concerns whether complex behavior could emerge in collectives and whether human intuitions about it are reliable. They address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker expects the discussed AI capability to eventually be achieved.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says, “I think we all think that's going to happen,” but its referent is not specified in the supplied passage. The later statement specifically says, “It is possible in the near future to have an automated AI scientist.” Without knowing what capability the earlier claim referred to, it is not possible to tell whether these are positions on the same question or whether the speaker’s view changed.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen Pure neural networks lack operations over variables and fail at tasks that require such operations.Now The human brain is not Turing complete.The earlier claim is that “pure neural networks lack operations over variables” and therefore fail at certain tasks. The later claim is that “the human brain is not Turing complete.” These address related but distinct properties, and the excerpts do not establish whether the speaker sees them as equivalent or whether the later claim changes the earlier position.
• NOT COMPARABLE · 93%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have substantially better reasoning and planning methods and more stable world models.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim predicts improved AI reasoning, planning, and world models; the later claim describes a possible future organization of AI as “diverse AIs” rather than “one big AI.” These address different aspects of future AI and can both be true, so the later statement does not establish a changed position on the earlier prediction.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks Waymo may be ahead in driverless cars, but says its New York City deployment will use a safety driver and geofencing.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier claim concerns Waymo’s New York City deployment and its use of a safety driver and geofencing. The later claim concerns Elon Musk’s ability to identify promising future opportunities rather than his engineering and optimization skills. These address different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker sees a possible case for intervention on safety-critical AI issues, potentially through incentives rather than delaying development.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier statement says the speaker sees a possible case for “some kind of intervention” on safety-critical AI issues. The later statement predicts that AI agents may show “bad behaviors” such as collusion; it does not state a position on whether intervention is warranted. These address different questions, so the later claim does not show a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that by the end of the year, progress on AI agents will be less advanced than the AI 2027 scenario hypothesizes.Now The speaker recalls that the report may project 60 million AI agents operating at 20 times human speed by 2032.The earlier statement is the speaker's prediction that by year-end, agent progress would be less advanced than the AI 2027 scenario hypothesized. The later statement merely recalls a possible 2032 projection—“60 million agents running at 20 times human speed”—and does not state whether the speaker now agrees with it or how their forecast has changed. These are not clearly comparable asserted positions.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Open-weight models allow essentially anyone to participate in AI development to some degree.Now The speaker describes a model in which duplicated AI instances acquire different job skills and are then merged so their skills can be combined compositionally.The earlier claim says open-weight models let “essentially anyone” participate in AI development. The later claim describes duplicated model instances acquiring different skills and being merged; it does not state a position on who can participate in AI development. These address different questions, so the later statement does not establish a change of mind.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker argued that current AI approaches cannot keep increasing their data by 100 times because the available data is insufficient.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says current AI approaches cannot keep increasing their data by 100 times because there is not enough available data. The later statement says an automated AI scientist may be possible in the near future. These address different questions, and the later claim does not establish that the speaker has changed their view about data scaling.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that AI systems will not replace the best coders or reach Jeff Dean's coding level within the next decade.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim predicts that AI will not replace the best coders or reach Jeff Dean’s level within a decade. The later claim describes AI agents communicating, learning, and specializing, with possible harmful behaviors. It does not say whether those agents will replace top coders or attain that level, so these statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes, and thinks Yann LeCun also believes, that LLMs are not the primary route to AGI and that achieved AGI will look substantially different.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim says LLMs are “not really the route to AGI” and that AGI will “look pretty different.” The later claim describes expert-guided AI producing strong results in “well-specified domains,” while leaving ambiguity. It does not say whether LLMs are a route to AGI, so it neither reverses nor clearly updates the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker attributes most recent AI interpolation progress to increased data and compute.Now The interviewer recalls a projection that by 2032 there might be 60 million AI agents running at 20 times human speed.The earlier claim is the speaker's view that recent AI interpolation progress is mostly due to increased data and compute. The later passage is the speaker quoting the interviewer’s recollection of a projection about the number and speed of future AI agents; it does not state a view about what caused interpolation progress. These are different questions, and the later passage is explicitly a recalled projection rather than a comparable claim by the speaker.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen In a science-fiction scenario set around 2045, the speaker imagined that a complete brain simulation would run slower than a human brain and fail to explain how the brain works.Now The human brain is not Turing complete.The earlier statement describes a science-fiction scenario in which a brain simulation would run slower than a human brain and fail to explain the brain. The later statement says, “The human brain is not Turing complete.” These address different questions, so the claims do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says the participants acknowledge value in some forecasting techniques used by the other, despite their differences.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier statement concerns whether the participants are far apart and whether each recognizes value in the other's forecasting techniques. The later statement concerns Elon Musk's strengths—identifying promising future opportunities rather than engineering or optimization. These do not address the same underlying question, so they provide no evidence of a change in position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Gary Marcus believes the three participants share values around wanting AI to benefit humanity rather than harm it.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier statement concerns shared values about AI benefiting humanity (“what is best for humanity” and “positive for humanity”); the later statement concerns whether an ecosystem of humans and AIs can participate in a “super organism.” They address different questions, so the claims do not establish a change in position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker predicts that two years before ASI is achieved, AI systems will have full solutions to all the cognitive problems he listed.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim predicts capability—“two years before we achieved ASI, we would have full solutions” to the listed cognitive problems. The later claim imagines an architecture of “diverse AIs” examining problems from different angles “rather than one big AI.” These positions can coexist; the later statement does not revise the earlier prediction.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker says AI systems may automate conventional, inside-the-box work and may outperform people on such work, citing the Go move 37 example.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier claim concerns AI automating or outperforming people on conventional, “inside-the-box” work, while the later claim concerns expert-guided AI producing impressive results in “well-specified domains” despite some ambiguity. These are compatible observations, but they do not establish whether the speaker’s view on the same underlying question changed.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The interviewer does not expect any actor to gain a durable advantage if the AI paradigm remains roughly unchanged.Now The interviewer considers an entirely AI-run economy plausible but suspects it might become degenerate in some way.The earlier claim is that if the AI paradigm stays roughly as it is, no actor will gain a durable advantage. The later claim is that an entirely AI-run economy seems plausible, though it might become degenerate. These address different questions and do not establish a change in the speaker’s position on the same issue.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker has advocated neurosymbolic AI since at least 2001.Now The speaker says fractured, entangled neural-network representations have some degree of composability.The earlier claim says, “I've always advocated neurosymbolic AI,” while the later claim says fractured, entangled representations are “composable to some degree.” The later statement describes a technical property; it does not establish whether the speaker’s view of neurosymbolic AI changed.
• NOT COMPARABLE · 90%24 Jun → 8 SeptThen The speaker thinks Claude and Gemini may have acceptable frontier-math scores, while current systems are being optimized for the test through augmentation whose nature is unknown.Now The speaker believes that extensive optimization in semi-specific objective domains still leaves something important missing.The earlier claim says current systems are being optimized for tests and that “we don't know what the augmentation is.” The later claim says that even when “we can hill climb and we can optimize,” “there's something missing.” These statements can coexist: the earlier one does not say optimization is sufficient, and the later one addresses a broader question about what AI lacks. There is no clear reversal on the same specific claim.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker regards language models using tools as a form of neurosymbolic AI, though not the right form.Now The speaker says models appear to use representations at multiple levels of resolution along a spectrum of abstraction.The earlier claim says language models using tools are “a species of neurosymbolic AI” but “not the right one.” The later claim discusses models using representations at “multiple levels of resolution” and intelligence relating to a “scope of tasks.” It does not revisit or deny the earlier view about tool use and neurosymbolic AI, so no change of mind is established.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker does not think AI will surpass human abilities across the relevant dimensions within two years.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier claim says, “I don't think it can happen in two,” referring to AI surpassing human abilities within two years. The later claim says a mixed AI-human ecosystem can participate in a “super organism.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says domain-specific AI systems currently perform much better than general-purpose AI systems.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says domain-specific systems are “much better than the general ones right now,” a comparison of current performance. The later claim says models can “create novel combinations” to solve tasks, a claim about capability. These statements address different questions, so the later one does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Fairly reasonable behavior is insufficient in many circumstances where AI systems guide weapons.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim says that “fairly reasonable behavior is insufficient” in some circumstances where AI guides weapons. The later claim predicts that future AI agents may communicate, specialize, and exhibit harmful behaviors such as collusion. It does not state a different view about acceptable behavior in weapons systems, so these claims do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen At a Senate appearance, Sam described substantial harm to humanity as his biggest worry.Now The speaker says composable AI agents would make them much more worried.These statements do not establish a change in one person’s position. The earlier claim reports that “Sam” said substantial harm to humanity was his biggest worry; the later claim is Tim Scarfe’s own statement that composable AI agents would make him “much more worried.” The claims are attributed to different people and address different concerns.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says newer AI systems interpolate substantially better than earlier systems.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim is that newer systems are “much better at interpolating” than previous systems. The later claim is that models can “create novel combinations” of existing knowledge to solve tasks. These concern different capabilities, and the later statement does not retract or deny the earlier one.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker considers 2030 the earliest plausible date in the most bullish scenario.Now The speaker considers Elon Musk's distinctive strength to be identifying promising areas of future opportunity, rather than engineering or optimization.The earlier claim is about timing: the earliest plausible date for a scenario is “2030.” The later claim is about Elon Musk’s strengths, saying his “ability to recognize areas that are interesting” matters more than engineering or optimization. These address different questions, so the statements provide no evidence of a change in position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker predicts that AI systems will remain highly vulnerable to outlier problems one year later.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim concerns AI systems’ continued vulnerability to outlier problems (“still be very vulnerable”), while the later claim concerns whether an automated AI scientist could be possible. These are different questions, and the later statement does not say that the outlier vulnerability has gone away.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says current AI systems struggle to multiply large numbers unless they can use tools.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim is specifically that current systems struggle to multiply large numbers unless they can use tools. The later claim is that models can combine existing knowledge in novel ways to solve tasks. These address different capabilities, and the later statement does not say that the multiplication limitation has changed.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker thinks AI systems probably received substantial chess-related training data from publicly available sources.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim is that AI systems probably had substantial chess-related training data (“a lot of data on chess”). The later claim is that models can combine existing knowledge in novel ways to solve tasks (“create novel combinations”). These address different questions—training-data exposure versus capability—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says current AI systems struggle to multiply large numbers unless they can use tools.Now The speaker says an AI's weights can be copied so that the AI can be run in a thousand instances.The earlier claim concerns whether systems struggle with large-number multiplication without tools; the later claim concerns copying AI weights to run many instances. These address different questions, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker's preferred conception of AGI is a system at least as capable as humans across many tasks and superior in some areas.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier statement describes a capability threshold for AGI: it should be “at least as good as humans in a bunch of things and better in certain ones.” The later statement describes a possible form of future AI: “diverse AIs” working from different angles “rather than one big AI.” These claims address capability and system organization, respectively, so the later view does not show a change in the earlier position.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker says future AGI might use LLMs because they are good at distributional learning.Now The speaker thinks future weight adaptation might solve the problem of sharing knowledge among agents.The earlier claim says future AGI might use LLMs because they are good at distributional learning. The later claim speculates that weight adaptation might solve knowledge sharing among agents. These address different questions—what technology AGI might use versus how agents might share knowledge—so they do not establish a change in position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes that LLMs are not really the route to AGI and that achieved AGI will look quite different; he thinks Jan LeCun shares this view.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns whether LLMs are a route to AGI: the speaker says they are “not really the route to AGI.” The later claim concerns whether AI skills or agents can be composed, saying that this “doesn't really marry with my experience.” These are related AI topics, but the statements do not establish opposing positions on the same question.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker and Jan LeCun both emphasize world models, but have somewhat different ideas about what world models are.Now The speaker says models appear to use representations at multiple levels of resolution along a spectrum of abstraction.The earlier statement says, “We would both emphasize world models” while noting different ideas about them. The later statement says models use “representations at multiple levels of resolution” along a spectrum of abstraction. These claims are related to representations, but they address different points; the later claim does not establish a changed position on the earlier one.
• NOT COMPARABLE · 87%24 Jun → 8 SeptThen The speaker and Jan LeCun both approach AI through architecture or cognition.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier statement is about the speaker and Jan LeCun sharing an “architectural approach or a cognitive approach.” The later statement says models can make “novel combinations” of existing knowledge to solve tasks. These concern related AI topics, but the later claim does not say whether the speaker’s own approach—or view of LeCun’s approach—has changed.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen Symbolic systems can perform operations over variables that pure neural networks fail to perform.Now The speaker says fractured, entangled neural-network representations have some degree of composability.The earlier claim says “Symbolic systems can perform operations over variables that pure neural networks fail to perform.” The later claim says fractured, entangled neural-network representations “are sort of composable to some degree.” These concern different capabilities, and the later statement does not say that neural networks can perform the variable operations discussed earlier. The claims therefore do not establish a change of position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says frontier-math evaluations may have contamination issues because OpenAI had access to the problems.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim concerns possible contamination in frontier-math evaluations because OpenAI had access to the problems. The later claim concerns expert use of AI in well-specified domains and residual ambiguity. These address different questions, so the later statement does not establish a changed view about evaluation contamination.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker sees little, though possibly some, progress in AI systems' ability to match the flexibility of human expert knowledge.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim says there is “relatively little progress” in AI matching the “flexibility to human expert knowledge.” The later claim concerns expert use producing “magical things” in “well-specified domains,” while acknowledging “a space of ambiguity.” That does not establish a changed view about AI matching human flexibility; the statements address different aspects of AI capability.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen AI organizations differ to some extent in their reinforcement-learning methods and datasets.Now The speaker says reinforcement-learning-trained language models understand things differently from humans but address the exploration problem by learning how things fit together and exploring novel trajectories.The earlier claim says AI organizations differ somewhat in their reinforcement-learning methods and datasets. The later claim says RL-trained language models learn how things fit together and explore novel trajectories. These address different questions—organizational differences versus models’ problem-solving—and the later statement does not revise the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says machine performance on the LiveCodeBench Pro benchmark was 0%, and that he does not know the human performance.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier statement concerns machine performance on the LiveCodeBench Pro benchmark (“zero percent”), while the later statement concerns expert use of a system in well-specified domains (“magical things happen”) and remaining ambiguity. They address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 88%24 Jun → 8 SeptThen The speaker expects an AI system to follow a specified algorithm for a task.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim concerns following a specified algorithm, and the surrounding passage says current systems cannot reliably follow rules. The later claim says an AI can be given a mathematics or knowledge task and will “hill climb towards it.” That describes pursuing a task, not necessarily following an algorithm or reliably completing the task, so the claims do not establish a change of position on the same specific question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The interviewer says OpenAI was originally open but is no longer open.Now The interviewer asks whether reduced human participation in the economy could cause economic mode collapse, in the context of an AI dividend.The earlier statement concerns whether OpenAI is still open: “It is not open anymore.” The later passage asks whether reduced human participation could cause “mode collapse of the economy”; it does not state the speaker’s position on that question. These address different issues, so they do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says AI systems perform worse when evaluated on problems outside the typical cases.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AI systems perform worse on problems outside typical cases. The later claim imagines future AI as diverse systems collaborating from different angles rather than one large AI. These address different questions, and the later statement does not revise the earlier performance claim.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker and Jan LeCun are dissatisfied with current AI capabilities described as reasoning because they consider them insufficiently robust.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says current AI systems’ “reasoning” is not “robust enough.” The later claim concerns whether AI skills or agents are composable, saying the speaker’s experience “doesn't really marry” with that idea. These are related topics but not the same underlying question, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker tentatively identifies the agents in the separate graph as possibly web-browsing agents.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim tentatively identifies agents in a separate graph as “maybe like web browsing agents.” The later claim discusses a possible future wave of agents “talking to each other and learning and specializing,” including potential collusion. These statements address different aspects of AI agents, and the later statement does not establish a changed view about the earlier agents.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker objects to the claim that continued scaling will automatically solve AI problems.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier statement objects to scaling *automatically* solving AI problems. The later statement concerns whether a proposed AI collective-intelligence system would have limits on its scale. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that merely fairly reasonable instruction-following is inadequate in some high-stakes situations, including systems guiding weapons.Now The interviewer argues that war games can provide information even when incorrect, and that repeated war games should reveal abstract recurring motifs.The earlier claim says that merely “fairly reasonable” instruction-following can be inadequate in some high-stakes situations, such as systems guiding weapons. The later passage is the interviewer’s point about war games yielding information and recurring motifs; it does not state the same position or show that the earlier speaker changed their view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker is sure he can devise non-orthodox chess variations, while acknowledging that some may already be known.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim is about being able to devise non-orthodox chess variations (“sure that I will be able to come up with variations on chess”). The later claim concerns expert use of AI in well-specified domains and remaining ambiguity. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers 2030 the earliest plausible date in the most bullish scenario.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim says the “very fastest case” for the event under discussion is 2030. The later claim says a proposed AI “super organism” would have “some kind of limits on its scale.” These address different questions, so the later statement does not show a change in the earlier timeline view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The interviewer says estimates for recursive self-improvement depend on its definition and suggests his own timeline may be longer than the guest's.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier passage concerns how estimates of recursive self-improvement timelines depend on its definition (“my timeline may be a little longer”). The later passage imagines a recursive system and says efficient operation would be “a different type of technology.” These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes the unconstrained AGI they fear should not exhibit many of the edge-case failures being discussed.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns whether the feared unconstrained AGI would exhibit many edge-case failures (“there really shouldn't be a lot of edge cases”). The later claim concerns whether current AI skills or agents are composable (“doesn't really marry with my experience”). These are different questions, so the later statement does not establish a change of mind on the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen A system in which a neural network creates the symbolic tools needed for a problem is a neurosymbolic solution.Now The speaker says fractured, entangled neural-network representations have some degree of composability.The earlier claim defines a neurosymbolic solution as one where “a neural network creates the symbolic tools needed for a problem.” The later claim says fractured, entangled representations are “composable to some degree.” These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says he has not really seen the gaps he has raised since 1998 solved over roughly 27 years, and finds it strange to expect them all to be solved within three years.Now The speaker believes shortcuts and validity problems are less problematic in this research than he had expected.The earlier statement says he had not seen longstanding gaps solved and that expecting them all to be solved within three years seemed strange. The later statement says shortcut-related validity problems in a particular line of research are less problematic than he expected. That narrower assessment does not establish that he changed his broader view about solving the gaps.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers the machine-intelligence level depicted toward the end of AI 2027 unlikely within the next decade and not very likely in the following decade, while allowing that it is possible.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says the intelligence level depicted in AI 2027 is “not very likely” within specified timeframes, though “certainly possible.” The later claim imagines future AI as “diverse AIs” collaborating rather than “one big AI.” These address likelihood and system design, respectively, so the later statement neither reverses nor clearly updates the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker questions whether training at a scale of 10^45, even on everything, would solve distribution shift and yield principles for efficient and useful performance in new domains.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim questions whether training at a particular scale would solve distribution shift and produce useful generalization. The later claim says a proposed AI collective-intelligence system might have limits on its scale. These concern different questions, and the later statement does not establish a changed view about the earlier training claim.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker expects AI organizations to develop unique intellectual property, increasing their incentive to limit transparency.Now The speaker argues that highly complex, information-sharing collectives can build societies, causing some objections to this possibility to lose force.The earlier claim concerns AI organizations’ incentives to limit transparency as they develop unique intellectual property. The later claim concerns whether complex, information-sharing collectives can build societies. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes that AI progress forecasting cannot be done in a fully closed-form way.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns whether AI progress forecasting can be done in a fully closed-form way. The later claim says the speaker’s experience does not align with an idea about composing AI skills by merging agents. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says human cognition is impaired by confirmation bias and motivated reasoning.Now The speaker believes intelligence circuits and motifs exist outside individuals.The earlier claim says human cognition is impaired by “confirmation bias and motivated reasoning.” The later claim says intelligence circuits and motifs “exist outside” individuals. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker’s tentative view is that human reinforcement learning has provided only limited help with AI alignment.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier statement concerns whether human reinforcement learning has helped with AI alignment (“all we have is maybe a human reinforcement learning”). The later statement predicts that future AI agents may communicate, learn, specialize, and exhibit harmful behaviors such as collusion. These address different questions, so the later claim does not show a change in the earlier alignment position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says that much of the near-term experimentation people conduct on LLMs can be automated.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim concerns automating near-term LLM experiments (“a lot of that” experimentation can be automated). The later claim concerns expert use in well-specified domains producing “magical” results while some ambiguity remains. These address different aspects of AI, and neither clearly reverses the other, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker believes OpenAI is not fully forthcoming about what it did.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier claim is about whether “OpenAI is not fully forthcoming about what it did.” The later claim is about whether an ecosystem of “AIs and humans together” can participate in a superorganism. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says computers can memorize billions of digits of pi, far more than people; he gives his daughter’s memorization of 105 digits as a comparison.Now The speaker says an AI's weights can be copied so that the AI can be run in a thousand instances.The earlier claim is about computers’ capacity to memorize information (“billions of digits of pi”), while the later claim is about copying AI weights and running multiple instances (“run it a thousand times”). These address different capabilities and do not show a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker assesses AI progress primarily through cognitive-science problems rather than quantitative progress.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim says the speaker’s approach to assessing AI progress is “more from the cognitive science.” The later claim imagines future AI as “diverse AIs” examining problems from different angles rather than “one big AI.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker believes AI will eventually acquire the missing capability for understanding concepts such as three-in-a-row, and says achieving it is not impossible.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim concerns whether AI will eventually gain a particular missing capability—understanding what a concept such as “three in a row” is. The later claim says models can combine existing knowledge in novel ways to solve tasks. That describes adaptivity, not whether they will acquire the earlier capability, so it does not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker describes AlphaFold as a carefully engineered hybrid neurosymbolic system that greatly outperforms a pure chatbot.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says AlphaFold is a carefully engineered hybrid system that outperforms a pure chatbot. The later claim imagines future AI as diverse systems collaborating rather than one large AI. These address different questions—AlphaFold’s design and performance versus how future AI might be organized—so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that neurosymbolic AI will begin to take off in three or four years.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim predicts that “neurosymbolic AI starts to take off” in three or four years. The later claim imagines future AI as “diverse AIs” tackling problems from different angles rather than “one big AI.” These address different questions—when a particular approach may gain traction versus how future AI might be organized—so the later statement does not establish a change in the earlier prediction.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says O3 is unable to avoid making illegal chess moves.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The earlier claim is about whether O3 can avoid illegal chess moves (“It can't avoid illegal moves”). The later claim discusses whether AlphaGo’s Move 37 was creative or innovative (“innovative but not creative”). These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes Einstein-level innovations will be necessary to achieve AGI and, certainly, ASI.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says Einstein-level innovations will be needed to achieve AGI/ASI. The later claim describes a possible future architecture—“diverse AIs” examining problems from different angles “rather than one big AI.” It does not say whether that arrangement would produce the innovations the speaker previously considered necessary, so the statements do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker does not see current AI systems solving the qualitative cognitive problems he considers necessary to address.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim says the speaker does not see current AI systems solving certain qualitative cognitive problems. The later claim says their experience does not align with the idea that AI skills can be composed by merging agents. These are not clearly the same underlying question, and the later extracted claim is difficult to interpret without the surrounding question or more of the quote. There is not enough evidence to establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes it is difficult to forecast when new ideas will emerge.Now The speaker believes human intuitions about collective behavior are often unreliable and that such behavior could emerge in principle.The earlier claim is about forecasting when new ideas will emerge (“new ideas are hard to project”). The later claim is about whether behavior can emerge from collectives (“that kind of behavior could emerge”); it does not state a view about forecasting new ideas. These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Humans have limitations in memory.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim is about human memory limitations (“Humans have limitations in memory”). The later claim concerns whether a collective-intelligence system would face limits on its scale (“it would place some kind of limits on its scale”). These address different questions, so the later statement does not establish a change in the speaker’s view about human memory or the same underlying issue.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that newer AI systems enable tasks that were previously unavailable and that these capabilities are already affecting labor markets.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement says newer AI systems enable previously unavailable tasks and are already affecting labor markets. The later statement says an automated AI scientist may be possible in the near future. These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker predicts that by 2035 people will regard LLMs as an incomplete answer, although they will continue using them for some tasks.Now The speaker believes shortcuts and validity problems are less problematic in this research than he had expected.The earlier claim predicts that by 2035 LLMs will be seen as an incomplete answer, while the later claim says shortcut-related validity problems in a particular line of research are less problematic than expected. These address different questions, and the later passage does not say that LLMs are a complete answer or retract the earlier prediction.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The interviewer says neurosymbolic AI is his personal favorite among possible alternative AI technologies.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier statement identifies neurosymbolic AI as a personal favorite among alternative technologies. The later statement imagines future AI as “diverse AIs” with different expertise rather than “one big AI.” These claims address different aspects of AI and can both be true; the later claim does not say whether the speaker still favors neurosymbolic AI.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says the scenario and reality were roughly aligned during the first couple of months, but have diverged in their pace.Now The speaker says developers adapt agents' skill surfaces, resulting in agents that function like different people.The earlier claim says the scenario and reality were initially aligned but later “diverg[ed] in speed.” The later claim says developers “adapt their skill surfaces” so agents are “different people.” These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers GPT-3's chess-rule failures striking and detectable without expert knowledge.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The earlier claim says GPT-3’s chess-rule failures are “quite striking” and detectable without expertise. The later claim characterizes AlphaGo’s Move 37 as “innovative but not creative” in an account of creativity and exploration. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that if the necessary ideas are missing, progress depends on when new ideas are developed.Now The speaker believes human intuitions about collective behavior are often unreliable and that such behavior could emerge in principle.The earlier claim concerns whether progress depends on developing missing ideas, noting that LLMs mostly exploit existing ideas. The later claim concerns whether complex collective behavior could emerge and whether human intuitions about it are reliable. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says polls indicate that 75% of Americans are already worried about AI and expects public concern to rise, seeing no active measure likely to reduce it in time.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier statement concerns Americans’ worry about AI and whether that concern will increase; the later statement concerns whether an ecosystem of humans and AIs can participate in a superorganism. They address different questions, and neither claim establishes a changed position on the other.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that AI systems should understand concepts such as three-in-a-row, believes current systems lack something needed for that understanding, and expects that capability to be achieved eventually, though it is not guaranteed to be easy.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier statement says AI will eventually acquire the missing capability for understanding concepts like three-in-a-row, while the later statement imagines an efficient recursive self-improvement loop as a different type of technology. These address different questions; the later claim neither retracts nor denies the earlier expectation.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says distribution shift remains a problem for driverless cars.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says that “distribution shift remains a problem for driverless cars.” The later claim says an “automated AI scientist” may be possible in the near future. These address different questions, so the later statement does not establish a change in the speaker’s view about driverless cars or distribution shift.
• NOT COMPARABLE · 91%24 Jun → 8 SeptThen The speaker believes it is probably technically achievable to automate the full AI research-and-development process, though it may require technology beyond what is currently available.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says full AI R&D automation is “probably” technically achievable, perhaps only with later technology. The later claim says the speaker can “imagine a future” with an efficient self-adapting recursive loop and calls that “a different type of technology.” It does not say whether that loop is achievable or retract the earlier view, so there is no clear change of position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says neurosymbolic systems make it possible to state explicit constraints, something very difficult with pure LLMs.Now The speaker says fractured, entangled neural-network representations have some degree of composability.The earlier claim is about neurosymbolic systems allowing explicit constraints that are difficult to state in pure LLMs. The later claim is about fractured, entangled neural-network representations being “composable to some degree.” These address different properties, and the later statement does not revise or deny the earlier one.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker regards an Apple paper involving Tower of Hanoi as an example of AI systems struggling with distribution shift.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says the Apple Tower of Hanoi paper is an example of AI systems struggling with “distribution shift.” The later claim imagines “diverse AIs” with different expertise examining problems from different angles, rather than “one big AI.” These address different questions—AI performance under distribution shift versus a possible future AI arrangement—so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Machines will eventually exceed human performance, but when and how this will happen remain uncertain to the speaker.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim says machines “will exceed” human performance eventually, while the later claim describes how expert use can produce “magical” results in well-specified domains, with “a bit of” ambiguity remaining. The later statement does not say whether machines will eventually surpass human performance, so it does not establish a change in position.
• NOT COMPARABLE · 95%24 Jun → 8 SeptThen The speaker proposes treaties establishing AI red lines and avoiding crossing them until the relevant questions have answers.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim proposes treaties and avoiding AI red lines until relevant questions have answers. The later claim predicts that agents may communicate, learn, specialize, and exhibit harmful behavior such as collusion. That prediction neither rejects the earlier proposal nor states a different position on safeguards; the claims address different questions.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that by the end of the year, progress on AI agents will be less advanced than the AI 2027 scenario hypothesizes.Now The interviewer recalls a projection that by 2032 there might be 60 million AI agents running at 20 times human speed.The earlier statement is Tim Scarfe’s own prediction that by year-end progress on agents would be “less far ahead” than the AI 2027 scenario hypothesized. The later passage says, “I think you said” there might be 60 million agents by 2032; it recalls another person’s projection rather than stating Tim’s own position, so it does not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker argues that robust understanding of domains such as tic-tac-toe or chess should make related tasks relatively easy for AI systems.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim is conditional: if AI systems have a “robust understanding” of a domain like chess, related tasks “wouldn't be too hard.” The later claim says a model given a specified task, such as mathematics, “will hill climb towards it.” That describes task-directed behavior, not whether the model understands the domain or how easy the task is, so the statements do not establish a change in position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker suspects the field is making mistaken assumptions and says it is difficult to predict when researchers will move beyond them.Now The speaker believes human intuitions about collective behavior are often unreliable and that such behavior could emerge in principle.The earlier statement says, “we're making some bad assumptions” and that it is hard to predict when researchers will move past them. The later statement says, “a lot of our intuitions are broken” about collective behavior and that such behavior “could emerge.” These concern related but distinct questions, so they do not establish a change in the speaker’s position on the same underlying issue.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Automating the full AI research-and-development process would be destabilizing if technically feasible, but not if the claims are merely hype and the capability cannot be achieved.Now The speaker recalls that the report may project 60 million AI agents operating at 20 times human speed by 2032.The earlier claim asserts that automating the full AI R&D process would be destabilizing if feasible. The later claim only recalls a possible report projection—“60 million agents” at “20 times human speed” by 2032—and does not state whether that scenario would be destabilizing. These address different questions, so the later statement does not establish a change in the speaker’s position.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker argues that an AI system should understand the concept of three-in-a-row when discussing it, and says current systems lack something needed for that understanding.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns whether AI understands a specific concept: it says systems are “missing something” to understand three-in-a-row. The later claim concerns whether AI agents’ skills can be composed, saying the speaker’s experience “doesn’t really marry” with that idea. These are related topics but not clearly the same underlying question, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Fairly reasonable behavior is insufficient in many circumstances where AI systems guide weapons.Now The speaker says composable AI agents would make them much more worried.The earlier claim says that “fairly reasonable” behavior is not enough when AI systems guide weapons. The later claim says the speaker would be “much more worried” if AI agents were composable. These address different concerns, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that scientific evaluation of AI systems is difficult because their training data is not transparent.Now The speaker says composable AI agents would make them much more worried.The earlier statement is about difficulty evaluating AI because its training data is opaque (“we don't know what's in the training set”). The later statement is about whether AI agents can be composed and whether that would increase the speaker’s concern (“if they were [composable], that would make me much more worried”). These address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that neurosymbolic AI will begin to take off in three or four years.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement predicts that “neurosymbolic AI starts to take off” in three or four years. The later statement imagines “diverse AIs” collaborating rather than “one big AI.” These address different questions—when a particular AI approach may gain traction versus how future AI systems might be organized—so the later claim does not establish a change in the earlier prediction.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen AI could surpass human abilities in some dimensions within 10 years; the speaker thinks this will probably occur in some dimensions but not others.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim concerns whether AI might surpass human abilities in some dimensions within 10 years (“it could happen in 10 years”); the later claim describes a possible future structure for AI (“diverse AIs” rather than “one big AI”). These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says it would be bad for humanity if AI alignment were solved only 200 years from now while AGI or ASI arrived within the next 10 to 20 years.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier statement says it would be bad for humanity if alignment were solved only "in 200 years" while AGI/ASI arrived "in the next decade or two." The later statement describes an efficiently self-adapting recursive loop as "a different type of technology." These address different questions, so the later claim neither changes nor clarifies the earlier position on alignment timing.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks it is possible to make LLMs play chess without illegal moves by using a tool call.Now In the speaker's account, construct validity is checked by an LLM judge that ensures the system is not cheating and is doing the task correctly.The earlier claim is that LLMs could play chess without illegal moves by calling a tool. The later claim describes an LLM judge checking whether a system is cheating and doing a task correctly. These concern different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that by the end of the year, progress on AI agents will be less advanced than the AI 2027 scenario hypothesizes.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim is a forecast about the pace of progress relative to the AI 2027 scenario: “by the end of the year, we will already be less far ahead” than that scenario hypothesizes. The later claim describes a possible future wave of agents that “talk[] to each other and learn[] and specialize[],” including possible harmful behavior. It does not say whether that development is ahead of, behind, or consistent with the earlier forecast, so these statements do not establish a change in position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker considers the semantics represented by current AI models superficial and primarily based on word distributions.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim is that current models’ semantics are “very superficial” and “really about distributions of words.” The later claim concerns whether AI skills or agents are composable, and the excerpt is fragmentary (“that doesn’t really marry with my experience”). These address different questions, so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that safety-critical AI uses should wait for a better alignment solution if systems cannot reliably refuse harmful requests and prevent foreseeable human harm, even if the solution takes five or 500 years.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim says safety-critical uses should wait for a better alignment solution if systems cannot reliably refuse harmful requests or prevent foreseeable harm. The later claim predicts that a future wave of AI agents may communicate and specialize, with possible harmful behaviors such as collusion. A prediction about possible behavior does not establish a changed position on whether deployment should wait for better alignment.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says computers can memorize billions of digits of pi, far more than people; he gives his daughter’s memorization of 105 digits as a comparison.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim is about computers’ memory capacity: they can “memorize billions of digits of pi,” far more than people. The later claim is about an AI’s ability to pursue a specified mathematics task or answer a knowledge question. These concern different capabilities, so the statements do not establish a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers Grok's failure to understand the three-in-a-row condition in the tic-tac-toe variant a conceptual weakness.Now The speaker believes that extensive optimization in semi-specific objective domains still leaves something important missing.The earlier claim concerns Grok’s failure to understand a specific tic-tac-toe condition, which the speaker calls a “conceptual weakness.” The later claim is a broader reflection that optimization in “objective kind of semi-specific domains” leaves “something missing.” These statements address different questions and are not inconsistent; the later statement does not retract the earlier judgment.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker suggests that the weakest plausible AGI could perform at the level of an ordinary person, who may reason poorly and exhibit substantial confirmation bias.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim concerns the capability level of the weakest plausible AGI: it could be “as good as Joe Sixpack.” The later claim concerns whether an “automated AI scientist” could be possible in the near future. These are different questions, and the statements do not establish that the speaker changed their position on either one.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes current AI systems are already close to performing at the level of an apprentice software engineer.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says AI systems are already “close” to an apprentice engineer’s level; the later claim says an automated AI scientist may be possible “in the near future.” These concern different capabilities and do not establish a change in position on the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says the gaps he identified in 2001 concerned multilayer perceptrons rather than transformers, which had not yet been invented.Now The RL systems under discussion are still self-attention transformers, but their structure differs from earlier systems.The earlier statement says the gaps identified in 2001 concerned multilayer perceptrons, not transformers. The later statement says current RL systems are self-attention transformers with a different structure. These address different claims and do not show a change in the speaker’s position about the gaps.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen AI could surpass human abilities in some dimensions within 10 years; the speaker thinks this will probably occur in some dimensions but not others.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim concerns AI surpassing human abilities on some dimensions, possibly within 10 years. The later claim is that an automated AI scientist could be possible in the near future. These are related but distinct questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says current AI systems may refuse an obvious biological-weapons request, but such refusals are easily bypassed through jailbreaking.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim says current AI systems’ refusals can be “easily bypassed” and that alignment systems “don’t really do what we want.” The later claim speculates that a future wave of communicating AI agents may have “bad behaviors” or “collusion.” These concern different systems and claims; the later statement does not establish a changed view about the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker proposes treaties establishing AI red lines and avoiding crossing them until the relevant questions have answers.Now The speaker says composable AI agents would make them much more worried.The earlier claim advocates treaties and avoiding crossing AI red lines until relevant questions are answered. The later claim says composable agents would make the speaker “much more worried.” These address different questions; the later concern neither affirms nor retracts the earlier position on treaties or red lines.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker considers 2030 the earliest plausible date in the most bullish scenario.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement concerns the earliest plausible date for a particular scenario (“the very fastest case, it'd be 2030”), while the later statement says an automated AI scientist is possible “in the near future.” The excerpts do not establish that these refer to the same milestone or timeframe, so they do not show a clear change in position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says there are already signs that neurosymbolic AI is gaining momentum.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier statement says there are “already signs” that neurosymbolic AI is gaining momentum, citing AlphaFold and growing conferences. The later statement says models can combine existing knowledge in novel ways to solve tasks. These address different questions—field momentum versus model capability—so they do not establish a change of mind.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker says current AI systems, including O3, do not reliably follow chess rules and sometimes make illegal moves.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier claim concerns whether current systems reliably obey chess rules and sometimes make illegal moves. The later claim says a model given a specified task will “hill climb towards” a solution. That describes progress toward a task, not whether the model reliably follows rules or succeeds, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker estimates that most of the timeline probability distribution lies beyond 10 years.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says “most of the distribution is past 10 years” for solving visual-intelligence problems, while the later claim says an automated AI scientist is “possible in the near future.” These address different capabilities, and saying something is possible does not establish that the earlier timeline estimate changed.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks a global treaty prohibiting work on these kinds of AI systems, with reporting requirements, would require substantial political will and is probably unlikely.Now The speaker says composable AI agents would make them much more worried.The earlier statement concerns the political feasibility of a global treaty prohibiting work on certain AI systems, saying it would take “a lot of willpower” and is “probably not likely.” The later statement concerns whether AI agents are composable and says that if they were, it would make the speaker “much more worried.” These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers GPT-3's chess-rule failures striking and detectable without expert knowledge.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim concerns GPT-3 making conspicuous chess-rule errors, such as having a queen jump over a knight, that need no expert to detect. The later claim concerns expert use producing impressive results in well-specified domains while ambiguity remains. These statements address different capabilities and conditions, so the later one does not establish a changed view about the earlier chess failures.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says some instances of the problems can be solved, but the problems he observes appear to reflect the same underlying core problems.Now The speaker describes human collective intelligence as diverse people with different intellectual lineages jointly examining problems and finding useful perspectives.The earlier claim says that some instances of the problems can be solved, while the problems he observes reflect recurring core problems. The later claim describes how diverse people can jointly examine problems and find useful angles. These address different points and do not establish a changed position on the same underlying question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker considers humans fairly data-efficient but doubts that human data efficiency has reached theoretical limits across the board.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim concerns whether humans have reached theoretical limits in data efficiency: the speaker says they are “pretty good” but doubts they are “at the theoretical limits” across the board. The later claim concerns the scale of a proposed collective-intelligence system, saying it would “place some kind of limits on its scale.” These address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks a global treaty prohibiting work on these kinds of AI systems, with reporting requirements, would require substantial political will and is probably unlikely.Now The interviewer asks whether reduced human participation in the economy could cause economic mode collapse, in the context of an AI dividend.The earlier claim concerns whether a global treaty prohibiting work on certain AI systems is politically possible. The later passage asks whether reduced human participation could cause “mode collapse of the economy”; it does not state a position on the treaty question or answer the economic question.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker treats obeying a simple constraint such as not making illegal chess moves as a small-scale alignment problem that AI systems still struggle to solve.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier claim says systems still struggle to obey simple constraints, using “don’t make illegal moves in chess” as a small-scale alignment problem. The later claim says an AI system “can actually converge towards the solution” when there is “enough of a specification.” These are not clearly opposing positions: the later statement is conditional and does not say that systems reliably obey such constraints. The claims also concern different formulations of the problem, so a change of mind cannot be established.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker expects many distribution-shift cases to continue exposing failures in AI systems next year.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says distribution-shift cases will continue to expose AI failures next year. The later claim imagines future AI as diverse systems collaborating across perspectives rather than one large AI. These address different questions, so the later statement does not establish a change in the earlier view.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker says he identified a set of real gaps in 2001.Now The speaker views intelligence as perspectival and expects it to be uneven and fragmented across capabilities.The earlier claim says the speaker identified “a set of real gaps in 2001,” but does not specify what those gaps were. The later claim concerns intelligence being “quite perspectival” and “fractured and jagged.” The available statements do not establish that these address the same question or that the later view reverses the earlier one.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker’s tentative view is that human reinforcement learning has provided only limited help with AI alignment.Now The speaker says reinforcement-learning-trained language models understand things differently from humans but address the exploration problem by learning how things fit together and exploring novel trajectories.The earlier statement concerns whether human reinforcement learning has helped with AI alignment (“maybe a human reinforcement learning”). The later statement concerns how reinforcement-learning-trained language models explore novel trajectories and solve an exploration problem. It does not establish a changed view about alignment.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says it would be bad for humanity if AI alignment were solved only 200 years from now while AGI or ASI arrived within the next 10 to 20 years.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim concerns the timing of AI alignment relative to AGI/ASI: it would be bad if alignment took 200 years while AGI/ASI arrived within 10–20 years. The later claim predicts that future AI agents may communicate and specialize, possibly with harmful behaviors such as collusion. It does not address alignment timing or reverse the earlier judgment, so the statements are not evidence of a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Gary Marcus believes the three participants share values around wanting AI to benefit humanity rather than harm it.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement is about shared values—wanting AI to benefit humanity—while the later statement is about whether an automated AI scientist could be possible in the near future. They address different questions, and neither claim establishes a change in the other position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen According to the speaker's characterization of Jan LeCun's view, current AI is not close to the goal because it lacks genuine world models.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The earlier statement is explicitly a characterization of Jan LeCun’s view (“he would say” current AI is not close because it lacks world models), not clearly the speaker’s own position. The later statement says expert-guided AI can produce “magical” results in well-specified domains while leaving ambiguity. These address different points, and the later claim does not establish that the speaker changed their view about whether AI has world models or is close to the goal.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects current systems to fail on many chess variants and problems, and says he can devise tic-tac-toe variants such as a five-by-five board where wins are allowed only on the edges.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim concerns systems’ ability to handle chess and tic-tac-toe variants; the later claim concerns expert use in “well-specified domains,” while acknowledging “a space of ambiguity.” These address different tasks and conditions, so the later statement does not show a changed position on the earlier claim.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says that if AI never becomes able to solve original problems, this limitation would provide an Achilles heel.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says that AI’s inability to solve original problems would be an “Achilles heel.” The later claim imagines diverse AIs collaborating from different angles “rather than one big AI.” These address different questions—AI’s ability to solve original problems versus how future AI might be organized—so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker doubts that LLMs can infer causal laws well and does not expect them anytime soon to produce Einstein-level, radically different solutions to problems.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim concerns whether LLMs can infer causal laws or produce “Einstein level” solutions needed for AGI. The later claim concerns expert use in “well-specified domains,” where results can seem “magical” despite ambiguity. These address different capabilities and settings, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Gary Marcus believes the three participants share values around wanting AI to benefit humanity rather than harm it.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement concerns shared values—wanting AI to benefit humanity rather than harm it. The later statement concerns a preferred future AI architecture: “diverse AIs” working from different angles “rather than one big AI.” These address different questions, so the later claim does not show a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says driverless-car technology has made progress every year.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says driverless-car technology has “made progress every year,” while the later says an automated AI scientist “could be possible in the near future.” These address different technologies and questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says human cognition is constrained by a lack of location-addressable memory, a point he argued in his book Kludge.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim concerns a constraint on human memory: the speaker says human cognition is limited by a “lack of location-addressable memory.” The later claim concerns possible limits on the scale of a proposed AI collective-intelligence system. These address different questions, and the later statement does not revise the earlier view.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker thinks neurosymbolic methods are worth exploring for near-term safety-critical AI applications, while remaining unsure whether they help with superintelligence and saying they might offer some benefit.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says neurosymbolic methods may help with near-term safety-critical applications, while expressing uncertainty about superintelligence. The later claim says an automated AI scientist may be possible in the near future. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker expects the discussed AI capability to eventually be achieved.Now The speaker expects a next wave of AI in which agents communicate, learn, and specialize, while possibly exhibiting harmful behaviors such as collusion.The earlier claim only says the speaker expects a capability discussed in the surrounding conversation to eventually be achieved. The later claim anticipates agents communicating, learning, and specializing, possibly with harmful behaviors. The excerpts do not establish that these are the same capability or reveal a conflicting position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker's preferred conception of AGI is a system at least as capable as humans across many tasks and superior in some areas.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim describes a preferred capability-based conception of AGI: it should be “at least as good as humans in a bunch of things” and “better in certain ones.” The later claim says efficient self-adaptation and autonomous evolution would be “a different type of technology.” These address different questions, and the later statement does not revise or deny the earlier conception.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have much more stable world models.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier statement predicts that AI systems will have “much more stable world models,” while the later one says an “automated AI scientist” may be possible in the near future. These concern different capabilities; the later statement does not establish a changed view about stable world models.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker considers the graph difficult to interpret reliably.Now The speaker believes shortcuts and validity problems are less problematic in this research than he had expected.The earlier statement says the graph is “weird to read” and difficult to interpret reliably. The later statement says shortcut-related validity problems are “not as much of a problem as we thought it would be.” These concern different issues, and the later passage does not establish a changed view about interpreting the graph.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes the graph has several flaws, including flaws relevant to the argument about AI reliability.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says the graph has “a lot of problems,” some “relevant to this argument.” The later claim says an automated AI scientist may be possible “in the near future.” These address different questions; the later statement does not show that the speaker changed their view of the graph’s flaws.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker thinks most AI safety arguments focus on AGI that is at least as intelligent as most smart people.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is about what intelligence level most AI safety arguments focus on: AGI “at least as smart as” most smart people. The later claim is about whether an automated AI scientist could be possible “in the near future.” These address different questions, and the later statement does not establish a changed view about the intelligence threshold for safety concerns.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen The speaker believes that generalization outside the training distribution remains a major problem for current AI systems.Now The speaker says that expert-guided AI can produce impressive results in well-specified domains, creating a supervisor illusion while leaving some ambiguity.The claims address different questions. The earlier statement says that generalization outside the training distribution remains a major problem; the later statement describes impressive results in well-specified domains while noting residual ambiguity. Success in well-specified domains does not show that the speaker has changed their view about out-of-distribution generalization.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker expects transparency to be less likely as AI systems approach AGI or become meaningfully differentiated.Now The speaker says composable AI agents would make them much more worried.The earlier claim concerns whether AI systems will be transparent as they approach AGI or become differentiated. The later claim concerns whether composable agents would make the speaker more worried. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker predicts that AI systems will not replace the best coders or reach Jeff Dean's coding level within the next decade.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim is about whether AI will replace the best coders or reach Jeff Dean’s coding level within a decade. The later claim is that an automated AI scientist may be possible in the near future. The latter does not directly say AI will replace top coders or match Jeff Dean, so these statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes AI systems could become orders of magnitude more data-efficient than they currently are.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim says there is more data efficiency to be gained: “orders of magnitudes more data efficiency.” The later claim says models can create “novel combinations” of existing knowledge to solve tasks. That describes adaptivity, not how data-efficient models are, so the statements do not establish a change in position on the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker estimates that O3 chooses a legal chess move with roughly 80% reliability.Now The speaker characterizes the dynamics of external intelligence systems as very complex.The claims concern different questions: the earlier estimates O3’s reliability at choosing a legal chess move (“maybe 80%”), while the later describes external intelligence systems’ dynamics as “very complex.” Neither claim updates or denies the other.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Gary Marcus believes more people probably read AI 2027's darker scenario than its positive scenario.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim reports Gary Marcus’s view about which AI 2027 scenario more people read: “more people probably read the darker scenario.” The later claim concerns whether “an automated AI scientist” could be possible in the near future. These address different questions, so they do not establish a change in the speaker’s position.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen Sam and Dario appear to want to automate the full AI research-and-development process now.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier statement says Sam and Dario appear to want to automate the full AI R&D process; the later statement describes the speaker’s imagined future as “diverse AIs” collaborating rather than “one big AI.” These address different questions, and the later view does not establish whether the speaker’s earlier assessment of Sam and Dario changed.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Some critical ideas needed for further AI progress are currently missing, according to the speaker.Now The speaker imagines future AI systems as diverse AIs with different expertise collaborating to examine problems from multiple angles, rather than relying on one large AI.The earlier claim says “we’re missing some critical ideas” needed for further AI progress. The later claim imagines future AI systems as diverse AIs collaborating, “rather than one big AI.” It addresses a possible form of future AI, not whether critical ideas are currently missing, so the statements do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says distribution shift remains a problem for driverless cars.Now The speaker thinks future weight adaptation might solve the problem of sharing knowledge among agents.The earlier claim says that distribution shift “still remains a problem for driverless cars.” The later passage discusses whether future weight adaptation might solve a different problem—sharing knowledge or skills among agents—and does not establish a changed position on distribution shift.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker objects to the claim that continued scaling will automatically solve AI problems.Now The speaker says that when given a specified task, such as a mathematics problem, an AI model will hill climb toward a solution.The earlier statement objects to the claim that scaling will “automatically solve AI problems.” The later statement says a model will “hill climb” toward a solution when given a specified task. That describes task-directed behavior, not whether scaling automatically solves AI problems, so it does not establish a change in position on the same question.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker questions whether training at a scale of 10^45, even on everything, would solve distribution shift and yield principles for efficient and useful performance in new domains.Now Fine-tuning a large model is very difficult, according to the speaker.The earlier claim asks whether training at “10 to 45” would solve distribution shift and support useful performance in new domains. The later claim says “it’s very difficult to fine tune a big fat model.” These address different questions, so the later statement does not establish a change in the speaker’s position on the earlier question.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says that smart people can solve tasks such as the Tower of Hanoi and that six-year-olds can learn chess rules.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim is about what people can do: “smart people can… do things like Tower of Hanoi” and “six-year-olds can learn to follow the rules of chess.” The later claim concerns AI models combining knowledge to solve tasks. These statements do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says LLMs might be replaced because they are highly inefficient in their use of energy and data.Now In the speaker's account, construct validity is checked by an LLM judge that ensures the system is not cheating and is doing the task correctly.The earlier claim says LLMs might be replaced because they are “very inefficient in terms of energy and data.” The later claim describes an “LLM judge” used to check construct validity. These address different questions, and the later statement does not indicate a changed view about whether LLMs should or might be replaced.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker characterizes himself and Yann LeCun as taking architectural or cognitive approaches to AI.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim says the speaker and Yann LeCun take “an architectural approach or a cognitive approach”; the later claim imagines future AI as “diverse AIs” working from different angles rather than “one big AI.” These concern different aspects of AI and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that future AI systems will have much more stable world models.Now The speaker says composable AI agents would make them much more worried.The earlier claim predicts that future AI systems will have “much more stable world models.” The later claim says that composable agents, if they existed, would make the speaker “much more worried.” These address different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker would not consider a system that cannot perform arithmetic to be satisfactory as AGI.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim concerns what capabilities an AGI must have to be satisfactory: it should be “at least as good as humans in a bunch of things” and able to do arithmetic. The later claim concerns possible limits on the scale of a collective-intelligence system. These address different questions, and the later statement does not indicate a changed view about AGI capabilities.
• NOT COMPARABLE · 96%24 Jun → 8 SeptThen The speaker says that if AI never becomes able to solve original problems, this limitation would provide an Achilles heel.Now An AI system can converge toward a solution when the problem is specified sufficiently, according to the speaker.The earlier claim concerns whether AI can solve original problems, saying it would be an “Achilles heel” if it never could. The later claim concerns convergence on an optimization problem “if you have enough of a specification.” These address different conditions and do not establish a changed position on the same question.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker argues that AGI should not be expected to share human memory limitations.Now The speaker supposes that the proposed collective-intelligence view would impose some limits on the system's scale.The earlier claim says AGI should not be expected to share human memory limitations. The later claim says a proposed collective-intelligence system may have limits on its scale. These address different questions, so the later statement does not establish a change in the speaker’s view about AGI and human memory limitations.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen Augmented data is not a universal solution; it works better for tasks such as mathematics, where the quality of the augmented data can be verified.Now The speaker says AI models can combine existing knowledge in novel ways to solve tasks.The earlier claim concerns when augmented data works well: it is “not a universal solution” and works better when its quality can be verified. The later claim concerns whether AI models can use existing knowledge in novel combinations to solve tasks. These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The interviewer says OpenAI was originally open but is no longer open.Now The speaker says their experience with AI does not align with the idea that AI skills can be composed by merging agents.The earlier claim concerns whether OpenAI is still open (“It is not open anymore”), while the later claim concerns whether AI skills or agents are composable (“doesn't really marry with my experience”). These are different questions, so the statements do not establish a change of mind on the same position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker says the scenario and reality were roughly aligned during the first couple of months, but have diverged in their pace.Now The speaker describes a model in which duplicated AI instances acquire different job skills and are then merged so their skills can be combined compositionally.The earlier claim concerns whether a forecast and reality were aligned in pace during the first months and then diverged. The later claim describes combining skills across duplicated AI instances. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker considers humans fairly data-efficient but doubts that human data efficiency has reached theoretical limits across the board.Now The speaker believes that an ecosystem combining AIs and humans can participate in a superorganism, making the ecosystem's mixed composition not especially important.The earlier claim concerns whether humans have reached theoretical limits in data efficiency: the speaker says humans are “pretty good” but doubts they are “at the theoretical limits” across the board. The later claim concerns whether an ecosystem of AIs and humans can participate in a “super organism.” These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 94%24 Jun → 8 SeptThen The speaker says it is possible that his view that AI needs a fix is correct, although proving it is very difficult and his supporting evidence is qualitative.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim says it is possible the speaker is right that AI “needs a fix,” noting that current LLMs mostly exploit existing ideas rather than discover new ones. The later claim says an automated AI scientist may be possible in the near future. That possibility does not itself deny the earlier concern; the claims address related but distinct questions, and the later passage does not establish that the speaker has revised his view about AI’s current limitations.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker considers the graph difficult to interpret reliably.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim says it is “weird to read anything off of this graph,” describing difficulty interpreting a graph reliably. The later claim says expert use can produce “magical things” in well-specified domains while “there’s still a bit of a space of ambiguity.” These concern different claims and do not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen Fairly reasonable behavior is insufficient in many circumstances where AI systems guide weapons.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier claim says that “fairly reasonable behavior is insufficient” when AI systems guide weapons. The later claim describes autonomous recursive self-improvement as “a different type of technology.” These address different questions, so the later statement does not establish a change in the speaker’s view about weapon-guiding AI.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says the participants agree that current efforts on AI alignment are inadequate and need substantial improvement.Now The speaker imagines that an efficiently operating recursive loop in which an AI system adapts itself, selects what to explore, and evolves autonomously would constitute a different kind of technology.The earlier statement says current AI alignment efforts are inadequate and need improvement. The later statement imagines an efficiently self-adapting recursive loop as “a different type of technology.” These address different questions; the later claim does not establish a changed view about alignment efforts.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker thinks it is possible to make LLMs play chess without illegal moves by using a tool call.Now The speaker says expert use of the system can produce seemingly magical results in well-specified domains, although some ambiguity remains.The earlier claim is that LLMs could play chess without illegal moves “by calling a tool.” The later claim concerns expert use producing “magical things” in well-specified domains while ambiguity remains. It does not address tool-assisted chess or deny that possibility, so these statements do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker is sure he can devise non-orthodox chess variations, while acknowledging that some may already be known.Now The speaker characterizes AlphaGo's Move 37 as innovative rather than inherently creative, though it appeared creative to people who understood the relevant creative space and its structure.The earlier claim concerns whether he can devise non-orthodox chess variations; the later claim concerns whether AlphaGo's Move 37 was creative or innovative. They address different questions, so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 83%24 Jun → 8 SeptThen The speaker places the most optimistic timeline at 2030, including the very fastest case.Now The speaker believes an automated AI scientist could be possible in the near future.The earlier claim puts the “very fastest case” at 2030, while the later claim says an automated AI scientist is possible “in the near future.” These may differ, but “near future” is not defined, and the claims do not establish whether they mean the same capability or timeline.
• NOT COMPARABLE · 97%24 Jun → 8 SeptThen AI actors have incentives to accelerate development and secure as many resources as possible for that purpose.Now The speaker suggests that a distributed collective AI system might face additional bottlenecks due to communication among its agents.The earlier claim says AI actors are incentivized to accelerate development and obtain resources. The later claim says distributed collective systems “might have additional bottlenecks” from communication among agents. These concern different issues—actors’ incentives versus a system’s technical constraints—so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker believes there are too many problems and gaps for all of them to be solved simultaneously.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier statement says there are “too many problems, too many gaps” for all the discussed problems to be solved “at once” in three or four years. The later statement imagines future AI as “diverse AIs” with different expertise working from different angles, “rather than one big AI.” These address different questions—near-term problem-solving timelines versus the possible organization of future AI—and do not establish a change in the speaker’s position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker predicts that neurosymbolic AI will begin to take off in three or four years.Now The speaker says composable AI agents would make them much more worried.The earlier claim predicts that “neurosymbolic AI starts to take off” in three or four years. The later claim says composable agents “would make me much more worried.” These address different questions—anticipated adoption or success versus concern about composability—so the later statement does not establish a changed position on the earlier prediction.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker does not see current AI systems solving the qualitative cognitive problems he considers necessary to address.Now The speaker imagines future AI intelligence as a collective of diverse, differently specialized AIs examining problems from different angles rather than as one large AI.The earlier claim says the speaker does not see current AI systems solving the qualitative cognitive problems he considers necessary. The later claim imagines a possible future architecture—“diverse AIs” examining problems from different angles “rather than one big AI.” A future design preference does not establish that he now believes AI solves those cognitive problems, so these statements do not show a change in position.
• NOT COMPARABLE · 98%24 Jun → 8 SeptThen The speaker would not consider a system that cannot perform arithmetic to be satisfactory as AGI.Now The speaker says an AI can be given a specified mathematics task and will hill-climb toward solving it, or be asked a knowledge question.The earlier claim is a standard for what would count as satisfactory AGI: it must not be unable to do arithmetic. The later claim describes an AI being given a mathematics task and hill-climbing toward a solution; it does not say whether the system can perform arithmetic or whether it meets the earlier AGI standard.
• NOT COMPARABLE · 99%24 Jun → 8 SeptThen The speaker says driverless-car technology has made progress every year.Now The speaker says many startups are pursuing recursive-improvement superintelligence.The earlier claim is that driverless-car technology has made progress every year (“progress made every year”), while the later claim is that many startups are pursuing recursive-improvement superintelligence (“many startups doing this”). These concern different topics, and the later statement does not indicate a changed view about driverless-car progress.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The speaker thinks most AI safety arguments focus on AGI that is at least as intelligent as most smart people.Now The interviewer argues that current AI has shown intelligence is not what people previously thought it was.The earlier claim is about what level of capability most AI safety arguments focus on: AGI “at least as smart as ... most smart people.” The later claim concerns what current AI has shown about the nature of intelligence (“intelligence is not this thing that we thought it was”), and its attribution to the speaker is also unclear from the surrounding turns. These do not establish a changed position on the same question.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker says people tend to take scenarios seriously, and vivid details can overwhelm their ability to assess them.Now The interviewer argues that these views increasingly abstract humans and their place within the systems they inhabit.The earlier statement says vivid scenarios can overwhelm people's ability to assess them. The later statement concerns views that increasingly abstract humans and their place in systems. These are different claims, and the later quote is attributed to Tim Scarfe rather than clearly asserting the extracted claim as the same person's position. There is not enough evidence of a change on the same underlying question.
• NOT COMPARABLE · 93%24 Jun → 20 AugThen Reliably getting AI systems to use tools is difficult.Now Current AI systems require human supervision.The earlier claim says AI tools are difficult to get working reliably; the later says current AI systems need human supervision. These claims can both be true, and the later statement does not establish that the speaker changed their view about tool reliability.
• NOT COMPARABLE · 96%24 Jun → 20 AugThen Reliably getting AI systems to use tools is difficult.Now Most people cannot see the problems in AI-generated code, making those problems deceptive.The earlier statement is about whether AI systems can use tools reliably; the later statement is about whether people can spot problems in AI-generated code. These are related AI topics, but they address different questions, so the later claim does not establish a change in the earlier position.
• NOT COMPARABLE · 91%24 Jun → 20 AugThen The speaker argued by 2001 that AI systems need symbols to perform abstract operations over variables.Now The speaker believes intelligence is a physical property of the universe and that AI models can create abstractions that work reasonably well in abstract domains such as chess.The earlier claim is that AI needs symbols for abstract operations over variables, and that pure neural networks lack those operations. The later claim is that intelligence is a physical property and that models can form abstractions that work reasonably well in domains like chess. The later statement does not say whether those models perform variable operations symbolically, so it neither reverses nor clearly restates the earlier position.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker expects future periods of AI progress to exceed the progress of the last few years and expects AI to be enormously more advanced in 30 years.Now The interviewer says energy use might continue increasing at its current growth pattern for another 3,700 years.The earlier claim is Tim Scarfe’s own forecast that AI will make substantial progress: “periods of AI progress” may exceed recent progress, and AI will be “enormously ahead” in 30 years. The later extracted claim concerns energy use continuing to increase, and the surrounding passage indicates Scarfe is summarizing what Adam Becker’s book says (“you said”). It does not establish a later position by Scarfe on the same question, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker predicts that future AI systems will have much more stable world models.Now The speaker says AI models’ capabilities offer some support for the functionalist hypothesis.The earlier claim predicts that future AI systems will have “much more stable world models.” The later claim says current models’ capabilities offer “some” support for the functionalist hypothesis. These address different questions, and the later statement does not say whether the speaker still expects more stable world models.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The speaker says that newer AI systems enable tasks that were previously unavailable and that these capabilities are already affecting labor markets.Now The speaker says AI models’ capabilities offer some support for the functionalist hypothesis.The earlier claim is that newer AI systems enable previously unavailable tasks and are already affecting labor markets. The later claim is that models’ capabilities provide some support for the functionalist hypothesis. These address different questions, so the statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 96%24 Jun → 20 AugThen The speaker considers the semantics represented by current AI models superficial and primarily based on word distributions.Now AI models use syntactic processing different from humans and arguably better in some respects; this appears to provide some validation for the functionalist hypothesis.The earlier claim says current models’ semantics are “very superficial” and “really about distributions of words.” The later claim says their syntactic processing differs from humans’ and may be “better” in some respects, lending support to functionalism. These address related but distinct questions, and the later statement does not clearly retract the earlier assessment of semantics.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The graph's benchmark covered coding tasks rather than tasks in general.Now Ray Kurzweil selected human-relevant historical examples rather than potentially more globally important factors and fitted the selected examples to a graph.The earlier claim is about the scope of a benchmark graph: it covered “coding tasks rather than tasks in general.” The later claim is about how Kurzweil chose historical examples for a different graph. These statements concern different graphs and selection criteria, so they do not establish a change in position on the same question.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The speaker believes current AI systems are already close to performing at the level of an apprentice software engineer.Now AI is already automating a meaningful amount of human labor and many jobs.The earlier claim says AI is “close” to apprentice software-engineer performance; the later claim says AI is “automating meaningfully human labor and lots of jobs.” These concern related but distinct questions, and the later statement does not establish a change in the earlier assessment.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The interviewer says estimates for recursive self-improvement depend on its definition and suggests his own timeline may be longer than the guest's.Now The interviewer describes a vision in which humanity collectively imprints its will on the universe, escapes its causal constraints, and continues reaching toward the stars indefinitely.The earlier claim concerns how estimates of recursive self-improvement depend on its definition and says the speaker's estimate may be longer. The later passage describes humanity's possible ability to shape the universe and continue indefinitely. These address different questions, and neither establishes a changed position on the other's subject.
• NOT COMPARABLE · 91%24 Jun → 20 AugThen The speaker expects an AI system to follow a specified algorithm for a task.Now Current AI systems require human supervision.The earlier excerpt says, “If we specified an algorithm for something,” and the surrounding turn discusses whether systems could follow it; the later claim says current AI “needs human supervision.” These do not establish opposing positions: following a specified algorithm and needing supervision can both be true, and the excerpts do not make clear they address the same capability or conditions.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker characterizes himself and Yann LeCun as taking architectural or cognitive approaches to AI.Now The speaker argues that people in Silicon Valley are increasingly abstracting humans and their place within the systems they inhabit.The earlier claim says the speaker and Yann LeCun take an “architectural approach or a cognitive approach” to AI. The later claim says Silicon Valley is “increasingly abstracting us and our place in the system.” These address different questions, and the later statement does not clarify the speaker’s own approach to AI or whether it has changed.
• NOT COMPARABLE · 95%24 Jun → 20 AugThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now The speaker says AI models use a different kind of syntactic processing from humans and argues it may be better in some respects.The earlier claim says AI needs “a deeper” form of semantics than current models provide. The later claim says models’ syntactic processing differs from humans’ and may be “better” in some respects. These claims concern related but distinct questions—semantic depth versus comparative syntactic performance—and the later statement does not clearly retract the earlier view.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen At the time of the conversation, AI systems were not very differentiated, with most actors using LLMs and reasoning models.Now People anthropomorphize AI models, much as people anthropomorphized the ELIZA system in the past.The earlier claim concerns how differentiated AI systems were at that time: “what’s there is not very differentiated.” The later claim concerns people anthropomorphizing AI models, comparing this to ELIZA. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen Reliably getting AI systems to use tools is difficult.Now AI models can perform some tasks that humans do extremely well, and they appear to be improving continually.The earlier claim is about the difficulty of getting AI systems to use tools reliably (“hard to get the tools to work reliably”). The later claim concerns models’ ability to perform some tasks humans do well and their apparent improvement (“can do things that we do extremely well” and “getting better”). These are different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker argued by 2001 that AI systems need symbols to perform abstract operations over variables.Now The speaker argues that people in Silicon Valley are increasingly abstracting humans and their place within the systems they inhabit.The earlier claim is about whether AI systems need symbols for abstract operations over variables. The later claim concerns Silicon Valley's increasing abstraction of humans and their place in systems. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now Some people seriously entertain the possibility that AI models have phenomenal experience rather than merely simulate it.The earlier claim is attributed to Tim Scarfe and says AI systems need “a deeper” semantics. The later passage is attributed to Adam Becker and reports that “some people” take seriously the possibility of phenomenal experience; it does not state Becker’s own position. These are not comparable assertions by the same person, so they do not establish a change of mind.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now The speaker argues that people in Silicon Valley are increasingly abstracting humans and their place within the systems they inhabit.The earlier statement says AI systems need “a deeper” form of semantics. The later statement says Silicon Valley is “increasingly abstracting” people and their place in systems. These address different questions, and the later claim does not show whether the speaker’s view of AI semantics changed.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker thinks most AI safety arguments focus on AGI that is at least as intelligent as most smart people.Now The speaker suggests that the current period may be the most deceptive in human history because AI appears intelligent.The earlier statement is about what most AI safety arguments focus on: AGI “at least as smart as” most smart people. The later statement says AI can seem intelligent and that this may be “the most deceptive time in human history.” It does not say what level of intelligence safety arguments focus on, or otherwise directly revise the earlier position; the claims address different questions.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker says current AI systems struggle to multiply large numbers unless they can use tools.Now The speaker argues that current AI coding tools generate large amounts of poor-quality, tangled code and create more problems than they solve.The earlier claim concerns AI systems’ difficulty multiplying large numbers without tools; the later claim concerns the quality and usefulness of AI-generated code. These are different capabilities, so the statements do not establish a change in position.
• NOT COMPARABLE · 95%24 Jun → 20 AugThen The speaker's preferred conception of AGI is a system at least as capable as humans across many tasks and superior in some areas.Now The interviewer describes the recursive self-improvement argument: after a long initial development, an AGI could help build successors in progressively less time, and each successor would improve.The earlier statement gives the speaker’s preferred conception of AGI: it should be “at least as good as humans in a bunch of things” and “better in certain ones.” The later extracted claim describes a recursive-improvement scenario, but its quote is framed as “the first AGI” helping build successors; it does not establish whether the speaker is asserting that scenario as their own view, nor does it revise their earlier criteria for AGI. These statements therefore do not clearly address the same underlying position.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker does not expect automation of current AI systems to generate the Einstein-level innovations needed for AGI or ASI.Now AI is already automating a meaningful amount of human labor and many jobs.The earlier claim says the speaker does not expect automating current AI systems to produce the Einstein-level innovations needed for AGI or ASI. The later claim says AI is automating meaningful amounts of human labor and many jobs. These address different outcomes, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker argued by 2001 that AI systems need symbols to perform abstract operations over variables.Now Some people in Silicon Valley view humans as causal patterns or programs that could be uploaded to computers.The earlier claim is about whether AI systems need symbols for abstract operations over variables. The later passage reports what “some people in Silicon Valley” think about humans as uploadable programs; it does not state the speaker’s own view or address the same question.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker believes appropriately designed software should let machines outperform humans in areas suited to machines and at least achieve human-level performance overall.Now The speaker says some people in Silicon Valley conceptualize humans as causal patterns or programs that could be uploaded to computers.The earlier claim asserts that appropriately designed software could let machines outperform humans in some areas and reach human-level performance overall. The later statement says that “some” people in Silicon Valley view humans as “causal patterns or as programs” that could be uploaded; it reports their view, not the speaker’s own position. These claims address different questions, so the later statement does not show a change in the speaker’s belief about machine performance.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The speaker considers the machine-intelligence level depicted toward the end of AI 2027 unlikely within the next decade and not very likely in the following decade, while allowing that it is possible.Now The speaker considers it doubtful, but not impossible, that AI systems will overcome their dependence on humans to frame their goals.The earlier claim concerns the timing and likelihood of reaching the level of machine intelligence depicted in AI 2027: it is “not really…very likely in the next decade” and “not…super likely in the decade after that,” though “certainly possible.” The later claim concerns whether AI can overcome its dependence on humans to frame goals, which the speaker calls “kind of doubtful” but allows “maybe.” These are related AI topics, but they are different questions, so the statements do not establish a change of mind.
• NOT COMPARABLE · 96%24 Jun → 20 AugThen The speaker predicts that future AI systems will have substantially better reasoning and planning methods and more stable world models.Now The speaker says AI models’ capabilities offer some support for the functionalist hypothesis.The earlier claim predicts that future AI will have better reasoning and planning and more stable world models. The later claim says current models’ capabilities provide some support for the functionalist hypothesis. These are different questions, so the statements do not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The interviewer does not expect any actor to gain a durable advantage if the AI paradigm remains roughly unchanged.Now The interviewer observes that some people seem obsessed with infinity and continued growth.The earlier claim concerns whether anyone can gain a durable advantage if the AI paradigm stays roughly the same (“I don't see anybody getting a durable advantage”). The later claim comments on an “obsession with infinity and keeping going.” These address different questions, so the later statement does not establish a changed position on the earlier one.
• NOT COMPARABLE · 94%24 Jun → 20 AugThen The speaker says newer AI systems interpolate substantially better than earlier systems.Now The speaker says the AI technology is improving as they use it every day.The earlier claim is specifically that newer systems are “much better at interpolating” than previous systems. The later claim is that the technology is “getting better” with daily use. Both suggest improvement, but the later statement does not address interpolation or make a directly comparable claim about earlier systems.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker suspects the field is making mistaken assumptions and says it is difficult to predict when researchers will move beyond them.Now The interviewer observes that some people seem obsessed with infinity and continued growth.The earlier claim is Tim Scarfe’s suspicion that researchers are making “bad assumptions,” while the later passage says there is “an obsession with infinity and keeping going.” The later claim is attributed to the interviewer, Adam Becker, not Scarfe, and concerns a specific tendency rather than clearly restating or revising Scarfe’s earlier view. These excerpts do not establish a change in Scarfe’s position.
• NOT COMPARABLE · 94%24 Jun → 20 AugThen The speaker objects to the claim that continued scaling will automatically solve AI problems.Now The speaker characterizes the opposing view as expecting scaling alone, including a future model such as GPT-7, to read and refactor problematic code, perhaps through reinforcement learning.The earlier statement is the speaker’s own objection to scaling *automatically* solving AI problems. The later passage frames the idea as an opposing view—“all we need to do is keep scaling it” and a future model “will kind of bring it back in check”—rather than clearly asserting that view or stating the speaker’s own position. It therefore does not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen A separate agent benchmark graph used a seconds-to-tens-of-seconds time scale, unlike the coding-task graph's minutes-to-hours-or-days scale.Now Ray Kurzweil selected human-relevant historical examples rather than potentially more globally important factors and fitted the selected examples to a graph.The earlier statement compares the time scales of an agents benchmark graph and a coding-task graph. The later statement describes how Ray Kurzweil selected examples for a graph. They concern different graphs and different questions, so they do not establish a change in position.
• NOT COMPARABLE · 94%24 Jun → 20 AugThen The speaker predicts that AI systems will improve programmer productivity, while noting caveats concerning the security and maintainability of generated code.Now The speaker characterizes the opposing view as expecting scaling alone, including a future model such as GPT-7, to read and refactor problematic code, perhaps through reinforcement learning.The earlier claim is Tim Scarfe’s own prediction that AI will improve programmer productivity, with caveats about code security and maintainability. The later passage—“just hold on boys just wait for a couple of years” and “the next version will kind of bring it back in check”—appears to be a rhetorical characterization of an opposing view, not a clear statement of Tim’s own position. It therefore does not establish that he changed his mind about productivity.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen A key bottleneck for tool-using AI systems is that they do not always call the appropriate code.Now Most people cannot see the problems in AI-generated code, making those problems deceptive.The earlier claim identifies a specific bottleneck: AI systems “don't always call the right code.” The later claim says people “can't see the problems” in AI-generated code. These concern different issues—tool selection versus recognizing flaws in generated code—and the later statement does not establish a changed view about the earlier bottleneck.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The speaker argues that robust understanding of domains such as tic-tac-toe or chess should make related tasks relatively easy for AI systems.Now The speaker believes intelligence is a physical property of the universe and that AI models can create abstractions that work reasonably well in abstract domains such as chess.The earlier claim says that if AI had a robust understanding of chess-like domains, related tasks would be relatively easy. The later claim says AI abstractions work reasonably well in abstract domains such as chess. These address different questions—what robust understanding would imply versus how well current models perform—and the later statement does not clearly assert that the models have robust understanding. No change in position is established.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker predicts that AI systems will not replace the best coders or reach Jeff Dean's coding level within the next decade.Now The speaker argues that current AI coding tools generate large amounts of poor-quality, tangled code and create more problems than they solve.The earlier claim predicts that AI will not replace the best coders or reach Jeff Dean's level “within the next decade.” The later claim says current tools produce “spaghetti garbage code” and “create[] more problems than [they] solve.” These claims address different questions and timeframes, so the later criticism does not show a change in the earlier prediction.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says AI systems need a deeper kind of semantics than current models provide.Now People anthropomorphize AI models, much as people anthropomorphized the ELIZA system in the past.The earlier claim says AI systems need “a deeper” kind of semantics than current models provide. The later claim says people “anthropomorphize these models,” comparing them to ELIZA. These address different questions, so the later statement does not show a change in the speaker’s view about AI semantics.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The speaker predicts that future AI systems will have much more stable world models.Now AI models are becoming increasingly adaptive and capable.The earlier claim predicts that AI systems will have “much more stable world models.” The later claim says models are “becoming quite adaptive” and have “a lot of capabilities,” while discussing how deceptive their apparent intelligence may be. These are related observations about AI, but the later statement does not say whether models have stable world models or contradict the earlier prediction; the claims address different aspects of capability.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The interviewer says estimates for recursive self-improvement depend on its definition and suggests his own timeline may be longer than the guest's.Now The interviewer observes that some people seem obsessed with infinity and continued growth.The earlier extracted text is not a clear statement of Tim Scarfe’s own timeline: it says the interviewer notes that estimates depend on the definition and that his estimate may be longer than the guest’s. The later statement, “there does seem to be an obsession with infinity and keeping going,” is an observation about people’s attitudes toward growth. These address different questions, so they do not establish a change in his position.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says people tend to take scenarios seriously, and vivid details can overwhelm their ability to assess them.Now The interviewer suggests that people may not originate their own actions and may instead function as causal conduits.The earlier statement is the speaker’s own claim about how people respond to scenarios: vivid details can overwhelm their ability to assess them. The later extracted claim concerns whether people originate their actions or act as causal conduits, but the quote frames this as a discussion of what “they imagine,” and the extraction says the interviewer suggests it. It does not establish the speaker’s own position on that question, so the statements cannot be compared as a change of mind.
• NOT COMPARABLE · 96%24 Jun → 20 AugThen The speaker says AI systems need a deeper kind of semantics than current models provide.Now The speaker says AI models use a different kind of syntactic processing from humans and argues it may be better in some respects.The earlier claim is that AI systems need “a deeper” semantics than current models provide. The later claim concerns models’ different kind of syntactic processing, which may be “better” in some respects and may support a functionalist hypothesis. The later statement does not say that current models have deeper semantics, so it does not clearly reverse the earlier position; the claims address different aspects of AI language abilities.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says that some very old AI systems can outperform modern general-purpose AI systems.Now Some people seriously entertain the possibility that AI models have phenomenal experience rather than merely simulate it.The earlier claim is Tim Scarfe’s view that some old AI systems can outperform modern general-purpose systems. The later claim reports that some people take seriously the possibility that AI models have phenomenal experience; it is not presented as the speaker’s own position. These claims address different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker characterizes Sam's biggest concern at a Senate appearance a few weeks earlier as the company not making all the money or extracting all the value it could.Now The interviewer characterizes Jeff Bezos's view as treating a lack of growth as death.The earlier claim describes Sam's concern about the company failing to make or extract as much money as it could; the later claim characterizes Jeff Bezos's view of growth. These concern different people and different underlying questions, so they do not establish a change in the speaker's own position.
• NOT COMPARABLE · 96%24 Jun → 20 AugThen The speaker believes the graph has several flaws, including flaws relevant to the argument about AI reliability.Now Ray Kurzweil selected human-relevant historical examples rather than potentially more globally important factors and fitted the selected examples to a graph.The earlier claim says the graph has “a lot of problems,” including some relevant to the argument. The later claim identifies a specific concern with Kurzweil’s graph: he selected “things that were relevant to humans” and fitted them to it. The passages do not establish that these are the same graph or the same underlying question, so they do not show a change in position.
• NOT COMPARABLE · 96%24 Jun → 20 AugThen AI organizations differ to some extent in their reinforcement-learning methods and datasets.Now AI models are becoming increasingly adaptive and capable.The statements address different questions. The earlier claim says organizations differ somewhat in their reinforcement-learning methods and datasets; the later says models are becoming “increasingly adaptive and capable.” The latter does not revise or deny the claim about how organizations differ.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The speaker says current AI companies are not entirely trustworthy.Now The speaker suggests that the current period may be the most deceptive in human history because AI appears intelligent.The earlier claim is that “current companies are not entirely trustworthy,” which concerns AI companies. The later claim says the current period “might be the most deceptive time in human history” because AI “seems like it is intelligent,” which concerns how AI appears. These are different questions, and the later statement does not establish a changed view about company trustworthiness.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker regards language models using tools as a form of neurosymbolic AI, though not the right form.Now AI models use syntactic processing different from humans and arguably better in some respects; this appears to provide some validation for the functionalist hypothesis.The earlier claim says language models using tools are “a species of neurosymbolic AI” but “not the right one.” The later claim says their different, sometimes arguably better, syntactic processing provides “some” validation for functionalism. These address related but distinct questions, and the later statement does not say the speaker has changed their view about tool use or neurosymbolic AI.
• NOT COMPARABLE · 96%24 Jun → 20 AugThen The speaker believes AI systems are not close to matching a top software engineer, because that requires understanding the problem, the human goals behind it, and multiple domains.Now The speaker considers it doubtful, but not impossible, that AI systems will overcome their dependence on humans to frame their goals.The earlier claim says AI is “not close” to matching a top software engineer because that requires understanding problems, human goals, and multiple domains. The later claim says it is “kind of doubtful” that AI will overcome humans’ role in framing goals. These ideas are related, but the later statement addresses goal-setting and agency, not whether AI can match a top software engineer, so it does not establish a change in position on the same question.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker estimates that most of the timeline probability distribution lies beyond 10 years.Now The interviewer says energy use might continue increasing at its current growth pattern for another 3,700 years.The earlier claim estimates that most of the timeline for solving visual-intelligence problems lies “beyond 10 years.” The later passage concerns energy use possibly increasing for “another 3700 years” and frames this as something Adam Becker said in his book. It does not state a different view about the AI timeline.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen Machines will eventually exceed human performance, but when and how this will happen remain uncertain to the speaker.Now Continued growth must end at some point, according to the interviewer.The earlier claim is attributed to Tim Scarfe and concerns whether machines will eventually exceed human performance. The later claim is attributed to the interviewer, Adam Becker, and concerns continued growth ending. These are different speakers and different questions, so they do not establish a change in Tim Scarfe’s position.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker believes that AI systems' failure to distinguish types from tokens can cause details about individuals in the same category to blend together and produce hallucinations.Now The speaker says some people believe individuals can disconnect themselves from the system they inhabit and go elsewhere.The earlier claim is about AI hallucinations caused by confusing types and tokens when representing people. The later claim reports what “some of them” believe about individuals disconnecting from a system. These concern different questions, and the later statement is attributed to others rather than asserted as the speaker’s own position.
• NOT COMPARABLE · 94%24 Jun → 20 AugThen The speaker thinks most AI safety arguments focus on AGI that is at least as intelligent as most smart people.Now The speaker believes intelligence is a physical property of the universe and that AI models can create abstractions that work reasonably well in abstract domains such as chess.The earlier claim is about what most AI safety arguments focus on: AGI “at least as smart as” most smart people. The later claim says intelligence is a physical property and that models can perform reasonably well in abstract domains such as chess. These address different questions, and the later claim does not state whether the speaker’s view of safety arguments or AGI-level intelligence has changed.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker considers voluntary regulation and industry self-regulation probably insufficient to ensure disclosure of AI model capabilities.Now The interviewer argues that AI model outcomes depend on question framing and the information available to the model.The earlier claim is about whether voluntary or industry self-regulation is enough to ensure disclosure of AI capabilities. The later passage discusses how question framing and available information affect model answers; it does not address regulation or disclosure. The statements concern different questions, so they do not establish a change of mind.
• NOT COMPARABLE · 94%24 Jun → 20 AugThen The speaker believes AI systems are not close to matching a top software engineer, because that requires understanding the problem, the human goals behind it, and multiple domains.Now The interviewer argues that AI models possess intelligence in some sense but still lack the capability or quality people want.The earlier claim is attributed to Tim Scarfe, while the later claim is described as the interviewer’s argument. The context does not establish that both claims are the same person’s own positions, so they cannot be compared as a change of mind.
• NOT COMPARABLE · 92%24 Jun → 20 AugThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now AI models use syntactic processing different from humans and arguably better in some respects; this appears to provide some validation for the functionalist hypothesis.The earlier claim says current models’ semantics are “very superficial” and that “we need a deeper one.” The later claim says their different syntactic processing may be “better than” humans’ in some respects and may validate functionalism. That does not say the models have deeper semantics or retract the earlier claim; the statements address related but distinct questions.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The interviewer says estimates for recursive self-improvement depend on its definition and suggests his own timeline may be longer than the guest's.Now Continued growth must end at some point, according to the interviewer.The earlier statement concerns recursive self-improvement timelines: “depending on your definition” the estimates may differ. The later concerns whether continued growth eventually ends: “at some point it must end.” These address different questions, so they do not establish a change in position.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The relevant form of AI recursion involves a system generating fundamentally new ideas previously produced only by humans and carrying out the broader development process.Now Some people in Silicon Valley view humans as causal patterns or programs that could be uploaded to computers.The earlier claim describes a form of AI recursion: a system generating ideas formerly produced only by humans and developing them further. The later claim reports what some Silicon Valley people think about humans as uploadable programs; it does not state the speaker's own view, nor address AI recursion. These are different questions, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 96%24 Jun → 20 AugThen The speaker says current AI systems have not been good at generating genuinely new ideas.Now The interviewer argues that current AI has shown intelligence is not what people previously thought it was.The earlier claim says current AI has not been good at generating “genuinely new ideas.” The later claim concerns whether AI has changed what people thought intelligence was, which is a different question. Also, the later extraction attributes that view to “the interviewer,” so it is unclear that it represents the same person’s own position.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker thinks AI systems probably received substantial chess-related training data from publicly available sources.Now Some people seriously entertain the possibility that AI models have phenomenal experience rather than merely simulate it.The earlier claim is that AI systems likely had substantial chess-related training data. The later claim reports that some people take seriously the possibility that models have phenomenal experience. These concern different questions, and the later statement is about other people's views rather than an asserted position by the speaker.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says that system prompts against producing copyrighted material or hallucinating do not reliably prevent those behaviors.Now The speaker says some people believe individuals can disconnect themselves from the system they inhabit and go elsewhere.The earlier statement is about AI systems failing to follow prompts such as “don’t produce copyrighted material” or “don’t hallucinate.” The later statement describes what “some people” believe about individuals disconnecting from a system. These are different subjects, and the later claim is attributed to others rather than asserted as the speaker’s own view, so it does not show a change of mind.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen AI could surpass human abilities in some dimensions within 10 years; the speaker thinks this will probably occur in some dimensions but not others.Now The interviewer describes a vision in which humanity collectively imprints its will on the universe, escapes its causal constraints, and continues reaching toward the stars indefinitely.The earlier claim is the speaker’s prediction that AI may surpass human abilities on some dimensions within 10 years. The later passage describes a vision of humanity escaping causal constraints and reaching the stars; it is attributed to the interviewer, not clearly asserted by the speaker, and addresses a different question. These statements do not establish a change in the speaker’s position.
• NOT COMPARABLE · 95%24 Jun → 20 AugThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now The interviewer argues that AI model outcomes depend on question framing and the information available to the model.The earlier claim says AI needs “a deeper” form of semantics than current models provide. The later claim says model outcomes are “very situated” and depend on framing and available information, but does not say whether models do or do not need deeper semantics. These claims address related but distinct points, so the later statement does not establish a change in position.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says human cognition is constrained by a lack of location-addressable memory, a point he argued in his book Kludge.Now The speaker says some people believe individuals can disconnect themselves from the system they inhabit and go elsewhere.The earlier claim concerns a limitation of human cognition—“a lack of location-addressable memory.” The later claim reports what “some people” believe about individuals disconnecting from a system and going elsewhere. These are different questions, and the later statement is attributed to others rather than asserted as the speaker’s own position.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker regards language models using tools as a form of neurosymbolic AI, though not the right form.Now People anthropomorphize AI models, much as people anthropomorphized the ELIZA system in the past.The earlier claim is about whether language models using tools count as a form of neurosymbolic AI, while the later claim is about people anthropomorphizing AI models. These address different questions, so the later statement does not establish a change in the earlier position.
• NOT COMPARABLE · 84%24 Jun → 20 AugThen The speaker sees little, though possibly some, progress in AI systems' ability to match the flexibility of human expert knowledge.Now The speaker considers it doubtful, but not impossible, that AI systems will overcome their dependence on humans to frame their goals.The earlier claim assesses progress to date in matching the flexibility of human expert knowledge (“little, though possibly some, progress”). The later claim is about whether AI will eventually overcome dependence on humans to frame its goals (“doubtful,” but “maybe” possible). These are related concerns, but the statements address different capabilities and timeframes, so they do not establish a change of mind.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says scientists try to self-correct motivated reasoning because they recognize it in themselves, but scientists still engage in it.Now The speaker says James Lovelock’s Gaia theory inspired them to think of Earth as a system rather than as a collection of individuals.The statements address different subjects. The earlier claim concerns scientists recognizing and trying to correct motivated reasoning, while the later claim concerns being inspired by Gaia theory to think of Earth as a system. There is no comparable position here from which to infer a change of mind.
• NOT COMPARABLE · 88%24 Jun → 20 AugThen The speaker argues that AI systems need a deeper form of semantics than current models provide.Now The interviewer argues that AI models possess intelligence in some sense but still lack the capability or quality people want.The earlier claim says current models have superficial semantics and that “we need a deeper one.” The later claim says models have “intelligence” in some sense but still lack “the thing we want.” These claims can coexist: having some intelligence does not establish that models have deeper semantics. The passages do not show a clear change in position on the same question, and the later claim’s attribution is not fully clear from the surrounding turns.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen Gary Marcus believes the three participants share values around wanting AI to benefit humanity rather than harm it.Now Some people in Silicon Valley view humans as causal patterns or programs that could be uploaded to computers.The earlier statement is about shared values and wanting AI to benefit humanity. The later statement reports what “some of them” in Silicon Valley think about humans as “causal patterns” or “programs”; it does not state the speaker’s own view. These concern different questions, so there is no basis to infer a change of mind.
• NOT COMPARABLE · 94%24 Jun → 20 AugThen The speaker believes current AI systems are already close to performing at the level of an apprentice software engineer.Now Current AI systems require human supervision.The earlier claim says AI systems are “close” to apprentice-engineer capability; the later says they “need[] human supervision.” These can both be true, and the claims address capability versus oversight rather than opposing positions on the same question.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says it would be bad for humanity if AI alignment were solved only 200 years from now while AGI or ASI arrived within the next 10 to 20 years.Now The interviewer describes a vision in which humanity collectively imprints its will on the universe, escapes its causal constraints, and continues reaching toward the stars indefinitely.The earlier claim says it would be bad for humanity if alignment were solved only after 200 years while AGI or ASI arrived in the next decade or two. The later passage describes a vision of humanity escaping causal constraints and reaching the stars indefinitely; it does not address when alignment is solved or whether that timing would be bad. The statements therefore do not establish a change in position on the same question.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says people took time to recognize that genes were not made of proteins.Now The speaker says James Lovelock’s Gaia theory inspired them to think of Earth as a system rather than as a collection of individuals.These statements address different topics. The earlier claim says it took time for people to recognize that genes were not made of proteins; the later says Gaia theory inspired the speaker to think of Earth as a system rather than as individuals. They do not establish a change in the speaker’s position on the same question.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen AI organizations differ to some extent in their reinforcement-learning methods and datasets.Now AI models acquire some form of coarse-grained abstraction or internal structure during training.The claims address different questions. The earlier statement says AI organizations differ somewhat in their reinforcement-learning methods and datasets; the later says models acquire “some kind of coarse grained abstraction” during training. Neither claim establishes a changed position on the same underlying issue.
• NOT COMPARABLE · 95%24 Jun → 20 AugThen The speaker predicts that AI systems will improve programmer productivity, while noting caveats concerning the security and maintainability of generated code.Now Most people cannot see the problems in AI-generated code, making those problems deceptive.The earlier claim says AI systems will improve programmer productivity, with caveats about code security and maintainability. The later claim says problems in AI-generated code are deceptive because most people cannot see them. These claims can both be true; the later statement does not say AI fails to improve productivity or directly revise the earlier prediction.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The speaker says there are already signs of neurosymbolic AI beginning to take off.Now The speaker says some people in Silicon Valley conceptualize humans as causal patterns or programs that could be uploaded to computers.The earlier claim concerns signs that neurosymbolic AI is beginning to take off. The later claim concerns what some people in Silicon Valley think humans are—“causal patterns or … programs” that could be uploaded. These address different questions, so they do not show a change in the speaker’s position.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The speaker characterizes himself and Yann LeCun as taking architectural or cognitive approaches to AI.Now The speaker says some people in Silicon Valley conceptualize humans as causal patterns or programs that could be uploaded to computers.The earlier claim is about an approach to AI: the speaker says he and Yann LeCun take “an architectural approach or a cognitive approach.” The later claim is about how some people conceptualize humans, and the excerpt does not clearly establish whether the speaker endorses that view or is describing it. These are not clearly claims about the same question, so a change of mind cannot be inferred.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker says current AI companies are not entirely trustworthy.Now The interviewer argues that current AI has shown intelligence is not what people previously thought it was.The earlier claim concerns trust in AI companies: “current companies are not entirely trustworthy.” The later claim concerns whether AI has changed how people understand intelligence, and is described as the interviewer’s argument. These are different questions, so the later statement does not establish a change in the speaker’s view about trust.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The interviewer does not expect any actor to gain a durable advantage if the AI paradigm remains roughly unchanged.Now The interviewer says AI shows hints of adaptive self-improvement, especially during hill-climbing toward a particular task.The earlier claim concerns whether an actor can gain a durable advantage if the AI paradigm “stays roughly like it is.” The later claim says there is “some kind of adaptive self-improvement” during task-specific hill-climbing. These are not clearly opposing positions: the later statement concerns a narrower capability and does not establish that any actor gains a durable advantage.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker regards language models using tools as a form of neurosymbolic AI, though not the right form.Now The speaker says AI models use a different kind of syntactic processing from humans and argues it may be better in some respects.The earlier claim says language models using tools are “a species of neurosymbolic AI” but “not the right one.” The later claim says their syntactic processing may be better than humans’ in some respects and may validate a functionalist hypothesis. These address different questions, so the later statement does not show a change in the earlier position.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker says there are already signs of neurosymbolic AI beginning to take off.Now The speaker suggests that the current period may be the most deceptive in human history because AI appears intelligent.The earlier claim is that there are signs of neurosymbolic AI beginning to take off, citing AlphaFold and growing conferences. The later claim is that AI can seem intelligent in a way the speaker describes as deceptive. These address different questions, and the later statement does not say whether neurosymbolic AI is taking off or not.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen According to the speaker's characterization of Jan LeCun's view, current AI is not close to the goal because it lacks genuine world models.Now The interviewer argues that AI models possess intelligence in some sense but still lack the capability or quality people want.These statements do not establish a change in the same person’s position. The earlier claim explicitly characterizes Jan LeCun’s view: “we’re not close because we don’t really have world models.” The later claim says models “do have intelligence” in some sense but “still don’t have the thing we want.” It neither directly addresses world models nor clearly contradicts the earlier claim, and the earlier statement is attributed to someone else’s view.
• NOT COMPARABLE · 99%24 Jun → 20 AugThen The interviewer says estimates for recursive self-improvement depend on its definition and suggests his own timeline may be longer than the guest's.Now The interviewer says energy use might continue increasing at its current growth pattern for another 3,700 years.The earlier claim concerns estimates for recursive self-improvement, while the later claim concerns how long energy use might keep increasing. They address different questions, and the later passage is also the interviewer summarizing something said in the guest’s book rather than clearly stating the interviewer’s own position.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The interviewer says neurosymbolic AI is his personal favorite among possible alternative AI technologies.Now The interviewer says AI shows hints of adaptive self-improvement, especially during hill-climbing toward a particular task.The earlier statement names neurosymbolic AI as the speaker’s personal favorite among alternative technologies. The later statement says AI shows some adaptive self-improvement during task-specific hill-climbing. These address different questions, and the later statement does not retract or qualify the preference for neurosymbolic AI.
• NOT COMPARABLE · 97%24 Jun → 20 AugThen The interviewer does not expect any actor to gain a durable advantage if the AI paradigm remains roughly unchanged.Now The interviewer argues that AI model outcomes depend on question framing and the information available to the model.The earlier claim is that, if the AI paradigm stays roughly the same, no actor will gain a durable advantage. The later claim is about how model answers depend on question framing and available information. It does not state a different view about durable advantage, so these claims address different questions.
• NOT COMPARABLE · 98%24 Jun → 20 AugThen The speaker objects to the claim that continued scaling will automatically solve AI problems.Now Continued growth must end at some point, according to the interviewer.The earlier claim says the speaker bristles at the idea that scaling will “automatically” solve AI problems. The later extracted claim says growth “must end at some point,” but explicitly attributes this position to the interviewer, not the speaker. It therefore does not establish the speaker’s later view or a change in their position.