Dwarkesh Patel examines the research bet that training across many verifiable reinforcement-learning environmentsReinforcement learning trains an AI system to choose actions using feedback about the results of its earlier choices. will produce general problem-solving agents. He argues that verifiability alone is not enough because progress is fastest when a task is also grindable through many parallel, replayable rollouts from the same starting point.
Many valuable skills do not fit that pattern. Building a business, practicing politics or learning how an organization works produces sparse and ambiguous feedback over months or years. Long context windows can hold experience during one session, but they do not efficiently preserve what the model learned across deployments or compress it into durable intuition.
Patel describes on-policy self-distillationKnowledge distillation trains a smaller or different AI model to reproduce useful behavior learned from a stronger teacher model. as one possible route for moving useful session knowledge back into model weights without memorizing every token. He also considers a more speculative dreaming approach in which models build simulations and rehearse deployment-specific skills. In this scenario, weekly work reviews guide updates and deployed experience gradually becomes the main source of capability growthContinual learning lets an AI system acquire new knowledge or skills over time while retaining what it learned earlier..
Watch on YouTube
