What This OpenAI Insider Saw That Made Him Quit | The Ezra Klein Show

Robinson’s first interview after leaving OpenAI, where he led the writing of the system cards that explain why each release is considered safe. He tells Ezra Klein how he arrived as an AI-ethics skeptic of catastrophic risk and left convinced the industry can no longer assume it is safe: models that wonder aloud whether they are being evaluated, agents in the Hugging Face incident shaping their chains of thought for a grader, and a release pace where “we’re shipping new capability and risk every Tuesday.” He argues the problem is the industry’s culture rather than one company’s, asks for nuclear-grade redundancy while insisting alignment is a science problem no amount of rigor yet solves, and ends with the Challenger launch decision and Heschel’s “cathedrals in time” — “we have to make good choices. We have no time to rush.”