OpenAI explains how computer use got 10x faster in a year
OpenAI's Thibault Sottiaux breaks down the 10x speedup in computer use: not just faster model sampling, but a better harness, improved thinking steps, and an AutoReview (internally called Guardian) agent that supervises the primary agent.
OpenAI made computer use 10x faster in a year. @thsottiaux breaks down what changed under the hood: “It’s not that we made the model sample faster. It’s both the harness and the model.” “How much does it need to think before it can take the next action reliably? That’s one element that we really improved.” “We have a second agent, which we call AutoReview, or internally we call it Guardian, which watches over the primary agent.” “So we also made this AutoReview model faster.” “We were able to use Astra at scale in order to improve the performance of the harness, do more things in parallel, remove little stalls.” “Combine all of this together, and then we got this 10x speedup.”