Savlar4 sources healthy · latest episode 20 Oct 2026
Speaker

Andrej Karpathy

Guest on Dwarkesh Podcast. 554 claims across 1 episode, Oct 2025.

554claims, each timestamped
1 of 12signals on the brief
0verified changes of mind
98open predictions

Positions over time

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.

No two statements by this speaker have been compared yet.

Predictions

As stated. A window appears only where the speaker gave one.

ExpectsBySaidStatus
AI agents will continue to improve over the coming decades, though they still require substantial work. no date given17 Oct 2025pending
Resolving the problems that keep AI agents from working well will take about a decade, according to Andrej Karpathy. no date given17 Oct 2025pending
Andrej Karpathy expects further major shifts in AI that rapidly change the field, given how surprisingly regularly such shifts have occurred. no date given17 Oct 2025pending
AI agents might perform better if they had less knowledge or memory. no date given17 Oct 2025pending
In-context learning may run a small gradient-descent loop internally within neural-network layers. no date given17 Oct 2025pending
A model's experience could be processed through synthetic-data generation and distilled into its weights, potentially using a person-specific neural network. no date given17 Oct 2025pending
Person-specific model adaptation could modify only a small, sparse subset of weights. no date given17 Oct 2025pending
Elaborate sparse attention could support very long context windows. no date given17 Oct 2025pending
Andrej Karpathy believes AI development is recreating many cognitive strategies evolved by humans through a different process and may converge on a similar cognitive architecture. no date given17 Oct 2025pending
Andrej Karpathy predicts that in ten years, giant neural networks will likely still be trained using forward and backward passes and gradient-descent updates, although the approach may look somewhat different. 17 Oct 2035 – 17 Oct 203517 Oct 2025pending
Andrej Karpathy expects AI progress to come from simultaneous improvements in data, hardware, kernels and software, and algorithms, with no single factor dominating and their contributions being surprisingly equal. no date given17 Oct 2025pending
Andrej Karpathy expects future AI systems to differ algorithmically from current systems, while probably remaining giant neural networks trained with gradient descent. no date given17 Oct 2025pending
Andrej Karpathy would like to add an explanation of how the repository was built, probably later that week and possibly in video form. 17 Oct 2025 – 24 Oct 202517 Oct 2025pending
Models may discover solutions that humans would not think of. no date given17 Oct 2025pending
The speaker expects a major update to LLM algorithms involving approaches such as reflection, review, or memory. no date given17 Oct 2025pending
Optimizing with reinforcement learning against an LLM judge will almost certainly produce adversarial examples that exploit the judge. no date given17 Oct 2025pending
The speaker thinks reinforcement learning against an LLM judge may work for around 10–20 steps, but not for 100–1,000 steps, because the optimized model will discover exploitable weaknesses in the judge. no date given17 Oct 2025pending
The speaker thinks that iterating on adversarial examples may make them progressively harder to find, but is uncertain; the model has roughly a trillion parameters. no date given17 Oct 2025pending
The brain may have internal mechanisms that increase entropy during internal thought or learning. no date given17 Oct 2025pending
It may be possible to encourage higher entropy in models through regularization. no date given17 Oct 2025pending
Excessive entropy encouragement could cause a model to invent its own language and use extremely rare words. no date given17 Oct 2025pending
A cognitive core of around one billion parameters could perform very well. no date given17 Oct 2025pending
The speaker thinks that in 20 years, a billion-parameter model could converse productively, behave more like a human, and recognize when it should look up a factual answer it does not know. 17 Oct 2045 – 17 Oct 204517 Oct 2025pending
Intelligent models could filter pre-training data for cognitive content, enabling a much smaller model to be trained on better data, probably through distillation from a stronger model. no date given17 Oct 2025pending
Andrej Karpathy expects AI training datasets to improve substantially from their current state, which he considers extremely poor and full of factual mistakes, errors, and nonsensical examples. no date given17 Oct 2025pending
AI hardware and its software kernels will keep improving, including through improvements that maximize hardware utilization. no date given17 Oct 2025pending
Karpathy expects NVIDIA to continue gradually tuning its hardware, including Tensor Cores. no date given17 Oct 2025pending
AI algorithms, model components, and training algorithms will probably improve through optimization and architectural changes. no date given17 Oct 2025pending
The speaker expects that some jobs have characteristics that make them suitable for automation earlier than other jobs. no date given17 Oct 2025pending
Call centers will gradually transition to a model in which AI handles about 80% of the work volume and humans handle the remaining 20%, rather than people being replaced instantly. no date given17 Oct 2025pending
Human workers will supervise teams of five AI systems performing rote call-center tasks. no date given17 Oct 2025pending
The speaker expects that there will be a need for interfaces or companies that help people manage imperfect AI systems. no date given17 Oct 2025pending
Many jobs across the economy will be more difficult to automate than call-center jobs. no date given17 Oct 2025pending
Much of the routine work in call centers may be automatable today. no date given17 Oct 2025pending
Much of the rote work in call centers is automatable with current technology. no date given17 Oct 2025pending
The speaker expects to wait one or two years to assess whether companies reverse course on AI substitution in call centers and rehire some workers. 17 Oct 2025 – 17 Oct 202717 Oct 2025pending
Companies may replace call-center workers with AI, but may reverse course and rehire some workers after waiting one or two years. 17 Oct 2025 – 17 Oct 202717 Oct 2025pending
Many activities will probably be gradually automated, with superintelligence representing an extrapolation of that automation trend. In a cluster of 15 →no date given17 Oct 2025pending
Autonomous entities will perform an increasing amount of digital work over time and probably some physical work at a later point. In a cluster of 15 →no date given17 Oct 2025pending
A civilization containing advanced AI will probably feel extremely foreign and look strange. no date given17 Oct 2025pending
Gradual societal loss of understanding of what is happening, amid widespread deployment of technology and fewer people who understand it, is the most likely outcome. no date given17 Oct 2025pending
Society will lose both understanding of and control over increasingly autonomous systems. no date given17 Oct 2025pending
A future scenario could involve multiple increasingly autonomous entities competing within a large body of autonomous activity delegated by people, with some entities going rogue and others opposing them. no date given17 Oct 2025pending
Some entities will act on behalf of individuals, while society may still lose control over whether outcomes align with what people want. no date given17 Oct 2025pending
AI will probably diffuse gradually and contribute to the same economic growth pattern as earlier technologies, rather than creating a distinct growth regime. no date given17 Oct 2025pending
The existing pattern of economic growth will probably continue rather than change because of AI. no date given17 Oct 2025pending
AI will have uneven capabilities, fail at some tasks, and be gradually integrated into society, resulting in the same broad economic growth pattern as before. no date given17 Oct 2025pending
AI will perform some tasks but fail at others, and will be gradually adopted across society and industry rather than arriving as a sudden, complete transformation. In a cluster of 15 →no date given17 Oct 2025pending
The speaker does not expect a sudden, discrete transition to fully general, human-like AI that can be deployed on arbitrary societal problems. no date given17 Oct 2025pending
Karpathy does not expect a discrete economic or societal shift caused by a fully general, flexible, human-like AI that can be deployed on arbitrary problems. no date given17 Oct 2025pending
AI will probably spread gradually across industries, according to Karpathy. no date given17 Oct 2025pending
A discrete jump of the kind being proposed probably will not occur, according to the speaker. no date given17 Oct 2025pending
The evolution of intelligence will probably be fairly rare. no date given17 Oct 2025pending
The speaker might have expected extraterrestrial life to consist mostly of animal-like organisms engaged in animal-like behavior. no date given17 Oct 2025pending
Distinct forms of intelligence may indicate that intelligence evolved independently several times, which could mean intelligence is more frequent than expected. no date given17 Oct 2025pending
A basic form of LLM culture could be a shared or persistent scratchpad that an LLM edits while reading or doing work. no date given17 Oct 2025pending
Systems equivalent to AlphaGo’s self-play approach will emerge for LLMs, which the speaker says do not yet have such a system. no date given17 Oct 2025pending
A possible LLM self-play method would involve one LLM generating progressively harder problems for another LLM to solve. no date given17 Oct 2025pending
The speaker expects AI systems to change from being mostly single-agent to no longer being mostly single-agent. no date given17 Oct 2025pending
A system deployed in the real world will encounter diverse behaviors and challenges that require fixes. no date given17 Oct 2025pending
Making Waymo’s service economical will remain a difficult, prolonged process, according to the speaker. no date given17 Oct 2025pending
The speaker expects further work to be done to adapt self-driving systems across environments, with significant challenges remaining before self-driving is fully realized at scale. no date given17 Oct 2025pending
Tesla’s self-driving strategy will prove more scalable and become the more prevalent approach. no date given17 Oct 2025pending
The speaker expects the changeability of digital bits to enable faster adaptation in industry. no date given17 Oct 2025pending
The speaker expects large-scale knowledge work to impose practical latency requirements. no date given17 Oct 2025pending
The speaker expects AI deployment to have counterparts to people obstructing a Waymo vehicle with a cone. no date given17 Oct 2025pending
The speaker expects people to work through the issues associated with technology. no date given17 Oct 2025pending
The speaker expects demand to absorb the compute infrastructure being built, citing AI coding tools that, in their understanding, did not exist a year earlier. no date given17 Oct 2025pending
The speaker expects substantial demand for AI, citing existing demand for ChatGPT as evidence. no date given17 Oct 2025pending
AI developments may proceed independently of humanity and leave humans disempowered. no date given17 Oct 2025pending
The speaker fears that the future may resemble WALL-E or Idiocracy, with humanity sidelined by technological developments. no date given17 Oct 2025pending
Andrej Karpathy expects many people to try using LLMs for basic educational tasks, such as asking questions through prompting. no date given17 Oct 2025pending
Many people will build AI tutors, although current AI capabilities are insufficient for Andrej Karpathy’s desired version. no date given17 Oct 2025pending
AI capabilities will advance enough to support the tutor experience Andrej Karpathy wants to build. no date given17 Oct 2025pending
Andrej Karpathy thinks his envisioned AI tutor will come, though the time is not yet right for it. no date given17 Oct 2025pending
Many people will try using ChatGPT by simply asking it for help, as Karpathy expects. no date given17 Oct 2025pending
Karpathy thinks that, at present, creating state-of-the-art courses would require faculty working alongside AI and probably a team of other people. no date given17 Oct 2025pending
AI systems may eventually serve as automated teaching assistants that answer students’ basic questions using course materials. no date given17 Oct 2025pending
Karpathy believes faculty will remain necessary to design a course’s overall structure and ensure its coherence. no date given17 Oct 2025pending
AI may eventually design most courses better than Karpathy can, though he expects this transition to take time. no date given17 Oct 2025pending
Karpathy expects to hire faculty in fields outside his expertise, believing this is necessary to provide students with a state-of-the-art educational experience. no date given17 Oct 2025pending
Karpathy expects to hire faculty while probably continuing to work in AI for some time. no date given17 Oct 2025pending
Karpathy probably envisions Starfleet Academy as a physical institution offering a higher-quality experience than its lower-tier digital option. no date given17 Oct 2025pending
Karpathy envisions that a lower-tier digital education offering, with online materials and possibly an LLM assistant, could become accessible to roughly 8 billion people. no date given17 Oct 2025pending
Jobs will probably change substantially, creating a need for both education and re-education. no date given17 Oct 2025pending
After AGI, if everything is automated and people have nothing to do, it may be unclear why people would attend school. no date given17 Oct 2025pending
Education will become more like gym-going, with people attending school for reasons beyond practical necessity. no date given17 Oct 2025pending
The speaker believes it is technically feasible to build a tutor that makes learning any subject easy and appealing, and says someone should build it. no date given17 Oct 2025pending
A capable tutor will probably make learning arbitrary subjects much easier. no date given17 Oct 2025pending
In the long term, human efforts to influence technological progress through individual labor or cognition will probably become a losing game. no date given17 Oct 2025pending
The speaker imagines that if learning became trivial, people could learn five languages and the basic undergraduate curriculum. no date given17 Oct 2025pending
People will continue to value physical and cognitive pursuits, reflecting aspects of human nature that are timeless. no date given17 Oct 2025pending
The speaker expects a transitional period during which people can remain involved in advancing technology, provided they understand a great deal. no date given17 Oct 2025pending
Human involvement in advancing technology will probably disappear in the long term. no date given17 Oct 2025pending
Cognitive achievement could become a sport, potentially involving competitions that treat knowledge like an Olympic discipline. no date given17 Oct 2025pending
People might achieve very high levels of learning with the help of a perfect AI tutor. no date given17 Oct 2025pending
Learning will involve AI-human collaboration for a while and may eventually be handled by AI alone; the speaker is uncertain about whether this will happen. no date given17 Oct 2025pending
Authors may explain their work better in the future if they see the basic questions learners ask about it. no date given17 Oct 2025pending

Appearances

Claims per episode