When David Robinson joined OpenAI in May 2023, the day after Sam Altman first testified before the Senate, the company’s idea of shipping a new frontier model generally meant training one from scratch. “We were going to bake a fresh cake with a new pretraining run, do the whole thing from scratch,” Robinson told Ezra Klein on the latest episode of “The Ezra Klein Show,” describing a process that took months and limited major releases to a few each year.

“We were going to bake a fresh cake with a new pretraining run, do the whole thing from scratch,”

Robinson, who led the writing of the safety reports that accompany OpenAI’s frontier models, announced his resignation on Oct. 3 in an essay for The Atlantic, and OpenAI responded with a statement from CEO Sam Altman posting on X, saying it is working to keep its models from outpacing its safety measures.

In his first interview since leaving, Robinson described a release cycle in which new capabilities no longer have to wait for a new pretraining run. Reasoning training layered onto existing base models, new tool integrations, and coding agents that accelerate OpenAI’s own research can all change what a system can do between major model releases.

Back then, a new generation started with another massive pretraining run. Only after that came post-training and safety testing. Robinson recalled that OpenAI even made a point of talking about the “month or months” of safety work that went into GPT-4 after the model itself was finished.

Source link