OpenAI两年内解散三个AI安全团队,独立监督归零

OpenAI continues to be the most disconcerting of the major AI companies.

精选理由

Gary Marcus这条帖文把OpenAI三支安全团队解散的时间线和后果讲得很清楚,适合关心AI安全和公司治理的人看。

AI 摘要

OpenAI在2023年承诺将20%算力用于AI安全,但到2026年7月已解散Superalignment、Mission Alignment和Preparedness三个独立安全团队。其中Superalignment于2024年5月解散,联合创始人Ilya Sutskever和Jan Leike相继辞职,Leike批评安全文化让位于产品。Preparedness团队曾负责在发布前评估模型是否可能助长生物威胁或攻击其他公司系统,其职能被拆分至现有团队。两年内六名高级安全负责人离职,而OpenAI正筹备史上最大科技IPO之一。

原文 · Gary Marcus

OpenAI continues to be the most disconcerting of the major AI companies.

OpenAI continues to be the most disconcerting of the major AI companies. Aakash Gupta @aakashgupta In 2023, OpenAI pledged 20% of its entire compute to AI safety. Last month it quietly dissolved the team whose only job was catching catastrophic risks in its own models. That makes three safety teams gone in two years. Superalignment went first, in May 2024. Co-founder Ilya Sutskever and Jan Leike both resigned, and Leike left saying safety culture had taken a backseat to shiny products. The compute pledge dissolved with the team. Mission Alignment went next, in February 2026. It lasted 16 months. Its six members were redistributed across research and product teams. Then Preparedness, at the end of July. This was the group that evaluated whether frontier models could enable bio threats or hack other companies' systems before release. Its work is now carved up by domain and handed to existing teams. Each dissolution came with the same explanation. The work continues, embedded inside the teams building the products. Which is exactly the structural problem. An independent safety team can delay a launch that belongs to someone else. An embedded safety function has to argue against its own team's ship date. Six senior safety leaders have left in two years, and the institutional knowledge left with them. The timing does the rest of the talking. All of this lands as OpenAI prepares for what's expected to be one of the largest tech IPOs in history. A standalone team producing internal documents about how your models could go catastrophically wrong reads very differently once there's an S-1 involved. The pledge was 20% of all compute. Three years later, the number of independent teams doing that work is zero. 🔗 View Quoted Tweet 💬 1 🔄 4 ❤️ 11 👀 2898 📊 3 ⚡