The next big breakthrough will be AIs learning on the job

Dwarkesh Podcast · ~20 min · Listen

• Solo essay arguing the next major AI capability jump will come from continual learning during deployment, not further scaling of RL on verifiable, simulatable tasks.
• Introduces “grindability” as distinct from verifiability: domains progress fast only if they can be cheaply parallel-simulated (coding, math) — real-world tasks like building a business or winning a court case have no such simulator.
• Argues RLVR alone likely won’t generalize to long-horizon, real-world competence, and that today ~30-50% of a lab’s inference compute is “wasted” in the sense that it never updates the base model’s weights despite encountering high-value real-world feedback.
• Proposes mechanisms to close this gap: “on-policy self-distillation” (compressing in-context learning back into weights via teacher-student distillation), and speculative “dreaming” (internal rehearsal/simulation akin to AlphaZero self-play).
• Predicts that by 2027 improvement will shift from pre-deployment training toward continuous, user-specific learning: “every time you interact with AI, it’ll be smarter.”