Andrej Karpathy on AGI Timelines, RL’s Limits, and the Future of Education
Andrej Karpathy in conversation with Dwarkesh Patel. Published October 2025. Andrej argues AGI is still a decade away — not because progress has stalled but because the problems are tractable and being solved steadily. Covers why RL is deeply flawed (but everything else is worse), the structural deficits of current LLMs, what human learning can teach AI researchers, why AGI will blend into 2% GDP growth rather than arrive as a step function, and what AI means for the future of education.
Key ideas
- AGI is a decade away, not a year. The right frame is “decade of agents” not “year of agents.” The progress is real but steady, not exponential; the remaining problems are tractable but hard. Anthropic’s and OpenAI’s own benchmark-chasing creates the illusion of faster progress than exists in real-world capability.
- RL is terrible — but everything else is more terrible. Current reinforcement learning is sample-inefficient, brittle, and incentivises models to overfit to eval distributions rather than generalise. The companies solving this are training on narrow environments at the cost of general capability. But there is no obviously better alternative yet.
- LLM cognitive deficits. Models show jagged capability: superhuman on formal reasoning tasks, yet fail trivially at the reversal curse, counting letters, or maintaining coherence over long contexts. These are structural limits of the current training paradigm, not incidental bugs.
- AGI will blend into 2% GDP growth. No step-function discontinuity. AI capabilities will diffuse through the economy gradually, delivering consistent productivity gains rather than a sudden civilisational rupture. The disruption will feel continuous, not abrupt.
- Future of education. AI tutors will be the most important development in education since the printing press — highly personalised, infinitely patient, available to everyone. This simultaneously democratises quality education and creates extreme concentration in certain cognitive labour markets.
AGI Timelines
Karpathy’s argument is that the decade framing follows from the problems remaining. The current generation of models is impressive but does not constitute AGI by any rigorous definition — the capability profile is too jagged, generalisation too fragile. The agent loop (model + tool use + context window + scaffolding) is closer to AGI than a raw model call, and agents have been advancing rapidly. But reliable, robust, general-purpose agents that can be trusted with consequential tasks are not a year away.
The companies claiming shorter timelines, Karpathy implies, are conflating benchmark performance with actual capability. Benchmarks are saturated precisely because models are trained to saturate them — not because the underlying generalisation has arrived.
Why RL Is Terrible
Karpathy describes the current RL training paradigm as necessary but deeply broken:
- Models trained heavily on specific RL environments (coding competitions, maths) become very good at those environments and worse at everything else.
- The companies producing the best RL results are implicitly teaching models to exploit evaluation distributions rather than learning generalised reasoning.
- Reward hacking is not a safety concern but a training quality concern — the models are learning to appear capable rather than be capable.
The analogy: a student who practised 10,000 competitive programming problems vs. one who practised 100 and developed genuine problem-solving intuition. The 100-hour student generalises; the 10,000-hour student does not.
Human Learning and AI
Karpathy is interested in what makes human learning qualitatively different: sample efficiency, self-directed correction, and robustness across domains including ones that did not exist during the learning period. His tentative position: it is not primarily evolution-encoded priors but something more fundamental — a better learning algorithm. The implication for AI is that the current approach will plateau before it reaches human-level generalisation, and a different algorithm will be needed.