The next big breakthrough will be AIs learning on the job

Summary

Dwarkesh Patel makes the case that the next major AI scaling frontier is not just larger models or longer context windows, but AIs that learn efficiently during deployment and then compress that experience back into their parameters. He argues that verifiable tasks help, but many important domains are not merely verifiable—they must also be “grindable,” meaning they support deterministic, replayable rollouts, which is why coding and math have progressed faster than open-ended computer use. He describes RLVR as a plausible route to training agents that generalize from containerized environments to more ambiguous real-world work, while noting that this remains an empirical question. He also argues current systems waste a large share of inference compute because deployment produces the most valuable learning signal, yet that signal rarely updates the base model. The episode closes with speculative but concrete mechanisms for continual learning, including on-policy self-distillation, learned “dreaming” simulations, and a 2027–2028 workflow where agents do real work, get reviewed, and their experience is distilled into future models.

Notes

• The central bet is that AGI-like agents can be trained on millions of verifiable tasks across thousands of RL environments.
• The desired capability is long-horizon problem solving under ambiguity, errors, and setbacks.
• Model usefulness depends on in-session intelligence, generality, and sample efficiency.
• RL training is presented as improving longer-horizon task completion.
• The host says computer use lags because a domain must be both verifiable and grindable.
• “Grindable” means deterministic, replayable simulators with parallel rollouts from the same starting state.
• Coding and math improve faster because their training environments are more replayable than real-world tasks.
• Open-ended domains like business, courts, trading, and elections cannot be recreated inside a data center.
• That makes real-world competence harder to train with brute-force simulation.
• The episode frames RLVR as a generalization bet from containerized tasks to broader agency.
• Whether RLVR actually generalizes is explicitly unresolved.
• Inference now consumes a large share of lab compute.
• Deployment reveals the most valuable learning signals.
• Current systems do not feed inference-time learning back into weights.
• Continual learning requires returning information to the weights.
• A bigger and bigger KV cache is not a scalable memory strategy.
• The comparison to humans is that learning is consolidation, not raw transcript storage.
• Cursor’s tab model is cited as a narrow example of online learning.
• That objective is accepted user edits across more than 400 million requests per day.
• The host says current online learning has not yet specialized by user.
• Sparse attention is mentioned as one possible architectural path.
• KV cache compaction is another suggested optimization for continual learning.
• Architecture may not be the fundamental bottleneck.
• OPSD is proposed as a way to distill session learning into weights.
• OPSD avoids needing an outer-loop verifiable reward.
• OPSD only needs the model to learn correctly within the context window.
• Naive supervised fine-tuning would reward memorizing the whole session transcript.
• The right target is job-relevant learning, not perfect recall.
• The host introduces “dreaming” as a potential fourth scaling axis.
• Dreaming would use learned simulations of reality for rehearsal.
• The point of dreaming is orders-of-magnitude more simulated experience per unit time.
• EfficientZero is used as an example of extreme data efficiency.
• The model could beat a novice human after two hours on an unseen Atari simulator.
• The model’s efficiency partly comes from internal lookahead.
• One real step can correspond to dozens of simulated games.
• The speculative deployment loop is work, review, distill, repeat.
• By 2027 or 2028, agents may handle unfamiliar projects for a week before review.
• A thumbs-up review would trigger distillation into the base model.