Dwarkesh Patel on How AI Agents Could Learn on the Job

Dwarkesh argues that a domain must be not only verifiable but also grindable: it needs replayable, parallel training targets. Coding is grindable; real-world domains like politics, business, law, markets, and operations are much harder.

Dwarkesh argues that deployment contains the most valuable learning data, but in-context learning does not scale as durable memory and gradient updates remain sample-inefficient. He presents on-policy self-distillation as one candidate way to compress session learning back into weights.

Dwarkesh says labs are betting RLVR-trained agents will generalize from containerized tasks to open-ended real-world work, but he treats that as open rather than settled.

Dwarkesh sketches a scenario where RLVR produces agents competent enough for real-world work, longer contexts let them accumulate week-long experience, and successful sessions are distilled back into future models.

Dwarkesh describes a speculative training axis where a model builds simulations of a real-world task, rehearses inside them, and uses extra compute to turn scarce real-world experience into many simulated samples.

Dwarkesh says the frontier-lab bet is that millions of verifiable RL tasks across diverse environments can create general problem-solving agents, while skeptics may still care about sample efficiency and continual learning.