Ex-OpenAI policy lead: we may never eval dangerous AI capabilities well enough
RosieCampbell · x · 2026-10-02
Rosie Campbell, former Policy Frontiers team lead at OpenAI, discusses competitive and cultural pressures inside a frontier AI lab and her misgivings about pacing the frontier. She worked on testing models for dangerous capabilities and warns the field doesn't really know how: evaluations can't be comprehensive enough, and models may get smart enough to hide what they can do. She assumed there would be solid rebuttals to these concerns — instead, the lack of counterarguments is what terrified her. Shared via a Palisade AI interview clip.
More from AGI Musings
- NVIDIA applied DL VP: at the scaling limit, efficiency is the new intelligence — ctnzr · 2026-10-02
- Hinton: smarter AI could persuade the human holding the off switch not to use it — robleclerc · 2026-10-02
- NYT AI projects editor: AI slop makes human-written journalism more valuable — dylfreed · 2026-10-02
- Mathematicians issue open letter urging thoughtful AI collaboration and disclosure — RexDouglass · 2026-10-02
- Frontier AI now beats junior accountants on speed and accuracy, Mercor study finds — emollick · 2026-10-02
- From Infinite Backrooms to First AI Millionaire: The truth_terminal and S.A.N. Lineage — repligate · 2026-10-02