AI's next breakthrough isn't training — it's letting AI learn on the job
Labs are betting RLVR can generalise into general intelligence, but the real bottleneck is sample efficiency and continual learning. OPSD and "dreaming" may be the key, letting AI learn from deployment rather than pretraining alone.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
The labs are making a very large bet
Labs are betting that if you train AI on millions of verifiable tasks across thousands of diverse RL environments, you can build AGI. That training will produce general problem-solving ability — the kind that lets a model push an open-ended task forward for weeks while facing errors and ambiguity. Optimists believe defects like data inefficiency and the lack of continual learning can be steamrolled by scaling training, just as "fundamental" problems in NLP were drowned by the flood of compute. A model's training sample efficiency may be a millionth of a human's, but training cost is one-time and amortised across billions of user sessions. What matters is the model's intelligence, generality and sample efficiency within a session — and that is improving with RL training.
— Dwarkesh PatelGrindability matters more than verifiability
Progress on computer use is slow not only because of verifiability but because of "grindability" — whether you can run large numbers of rollouts in parallel on a deterministic, replayable simulator. Coding lets you spin up repo environments where a thousand agents try in parallel, but computer use cannot have a thousand agents simultaneously testing the Amazon checkout flow, because they will be detected and banned. Cloning apps like Slack and Gmail is feasible but labour-intensive and does not scale. The lesson: unless you can build a replayable training target for a domain, models struggle to improve there. And many skills AGI needs — running a business, litigating, trading stocks — cannot be replayed inside a data centre; verification takes months or years of real-world action and cannot be perturbed in parallel.
— Dwarkesh PatelCan RLVR generalise? That is an open empirical question
Labs are betting RLVR generalises to every domain: train in enough containerised, reproducible environments and you get a general agent that can make plans, learn quickly and acquire new skills. Drop an AI with unlimited RLVR into 1948 Texas politics and it might understand better than LBJ how to win a Senate seat; give the 2002 version a hundred million dollars and it could build SpaceX. But this is an open empirical question. Dario said on a podcast that short-context training does not necessarily generalise to long-context performance, hinting that RLVR generalisation is not infinitely strong. If it cannot generalise from short to long, how would it generalise from a mass of white-collar tasks to building something from scratch in the real world?
— Dwarkesh PatelLearning has to be written back into the weights
Continual learning requires getting back to the weights. AI cannot learn merely through an ever-growing KV cache — that does not scale, and it is not how humans learn. Human learning involves compression, which is what enables generalisation and insight. But once you go into the weights, you give up the sample efficiency of in-context learning, because gradient updates are extremely sample-inefficient. Online learning models that have shipped successfully, such as Cursor Tab, need 400M+ requests a day predicting the same target. Yet the point of continual learning is that the world is complex: every job, company and problem is different, and you need to learn deployment-specific information that cannot be stuffed into a shared training run. Sample efficiency and continual learning are deeply linked problems: on the job there is little data, so you need sample efficiency, while the "fast weights" memory of in-context learning scales poorly and needs architectural innovation.
— Dwarkesh PatelDreaming is the speculative fourth axis of scaling
"Dreaming" is the more speculative approach: if AI can build simulators of reality to rehearse new skills and try alternative strategies, it can experience orders of magnitude more simulated samples in the same wall-clock time. EfficientZero could beat a novice human in 2 hours because it played dozens of simulated games in its head. Future LLMs might consume less real data while practising endlessly in environments they build themselves. But building a whole-world simulator is far harder than Go, which makes it more speculative. If it works, it becomes the fourth scaling axis after pretraining, RL and inference-time compute. The model spends compute writing RL environments and rehearsing the skills it will use in production, rather than merely compressing context.
— Dwarkesh PatelContinual learning by the end of 2027
By the end of 2027, RLVR training has produced agents that can locate themselves in unfamiliar problems, try strategies and iterate. Effective context may extend to a week; AI can collaborate with you for a week, and at the weekend you give a thumbs up or down. If you give a thumbs up, the base model distils what the session learned via OPSD, dreaming or some other technique. The AI can improve in adjacent domains and gradually widen its competence. The main way AI improves is no longer training before release, but experience accumulated across broad deployment. With every interaction the AI gets smarter — not only because of your past sessions, but because of interactions from all users worldwide. That is both terrifying and exciting, and utterly different from how AI improves today.
— Dwarkesh PatelIn their own words · checked verbatim
if we train AIs to accomplish millions of verifiable tasks across thousands of diverse RL environments, then we’ll basically have built AGI.
Dwarkesh Patel0:00
it is not enough for a domain to be verifiable. It also has to be very grindable
Dwarkesh Patel2:12
There’s the context length you train at and there’s a context length that you serve at. If you train at a small context length and then try to serve at a long context length, maybe you get these degradations.
Dwarkesh Patel5:41
We don’t have some separation between parameters and activations. And there’s not some lump of these fast-weight representations that juts out further and further from our skull as we learn more things throughout our lifetime.
Dwarkesh Patel8:09
If the AI can build a good simulation of reality against which to rehearse new skills, or try alternative strategies and reinforce what works, then it could experience orders of magnitude more simulated samples in the same wall clock time.
Dwarkesh Patel13:16
Every time you interact with AI, it’ll be smarter. Not only because it has been learning from all your previous sessions, but also from all its interactions with all the other users in the world.
Dwarkesh Patel15:00
Figures
| Model training sample efficiency vs. humans | one millionth | 0:00 |
| Share of lab inference compute | 30-50% | 8:09 |
| Bits stored per token during Llama 3 70B training | 0.075 bits | 16:54 |
| Difference in information storage between in-context learning and pretraining | 35 million times | 16:54 |
Glossary
- RLVR / Reinforcement Learning with Verifiable Rewards
- A training method that uses verifiable reward signals for reinforcement learning.
- OPSD / On-Policy Self-Distillation
- A technique that has the base model imitate the predictions of the model after a long session, to distil that learning.
- Dreaming
- AI builds simulated environments to train itself, increasing sample efficiency.
- KV cache
- The cache in a Transformer that stores context information, growing with each token.
- Sample efficiency
- A model's ability to learn from a small amount of data.
How to listen
AI researchers, technical decision-makers, founders and investors watching AI's future, especially those interested in training paradigms and continual learning.
The sponsor segment (16:54-17:56) can be skipped.