AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

OpenAI's Identity Crisis with Ex-Researcher Steven Adler

· The Cognitive Revolution

Former OpenAI researcher Steven Adler discusses the evolution of OpenAI's culture, safety practices, and mission over his ~4 year tenure. He walks through four chapters of his work: product safety (including calibrating content filters), leading GPT-4 deployment, dangerous capability evaluations, and AGI readiness research on personhood credentials. Adler reflects on how the Anthropic split in late 2020/early 2021 was a formative crisis that leadership navigated by reaffirming the nonprofit charter, and how the company has since shifted from research-first to product-first as its workforce swelled with commercially-oriented talent. Adler is one of 12 ex-OpenAI employees who filed an amicus brief in the Musk v. OpenAI lawsuit, arguing the nonprofit mission is central to the company. He expresses concern about OpenAI's attempted conversion away from nonprofit control, arguing the fiduciary question—whether governance is accountable to humanity or shareholders—is fundamental regardless of dollar amounts paid. He notes the same week's announcement that the nonprofit would retain control of a new public-benefit corporation appeared to be a partial win, though trust between former staff and leadership has eroded significantly. On AI race dynamics, Adler argues the industry is trapped in a defection equilibrium where cautious safety testing becomes a competitive disadvantage. He advocates for a minimum testing period floor, praises Anthropic's practice of publishing explicit commitments, and urges companies to be more honest publicly about the risks they are taking. He doubts OpenAI has rigorous internal analysis on recursive self-improvement timelines, and warns that increasingly siloed information controls may leave even internal safety staff behind the curve on what's being built.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.