The next big breakthrough will be AIs learning on the job

Patel begins with a detour that turns out to be the spine of the argument: why has progress on computer use been so slow, given how obviously verifiable it is? Ordering the Etsy item, booking the venue, filing the taxes — success or failure is unambiguous, and yet the domain has advanced much more slowly than coding or math, which are equally verifiable.

He lists the standard explanations — less high-quality multimodal data during pretraining, video eating the context window faster — and then names the one he thinks is underrated: verifiability is not sufficient. A domain must also be grindable, in the specific sense that you can run many parallel rollouts against a deterministic, replayable simulator. Coding qualifies because you can hand a thousand agents identical containers with a repo and a missing feature.

Computer use does not, at least not trivially. A thousand agents cannot all try the same Amazon checkout flow, because the bots get detected and shut down. The workaround is cloning Slack, Gmail and the rest, which Patel calls labor-intensive and unscalable today — though he expects it to become tractable once models can rebuild complex applications faithfully, a procedure that doubles as an excellent RL objective for coding.

The generalization he draws is the important part: unless a domain admits a very replayable training target, models struggle to progress in it. The underlying reason is sample inefficiency during training, which is where the previous monologue left off.