Apollo Research's CEO on OpenAI Audits: Why Embedded Evaluations Still Fall Short
Apollo Research CEO Marius Hobbhahn doesn't mince words. An "AI takeover" scenario by 2027? It's now plausible. He calls it a "speed problem," if advanced AI outpaces humanity's capacity to intervene. Hobbhahn points th
Apollo Research CEO Marius Hobbhahn doesn't mince words. An "AI takeover" scenario by 2027? It's now plausible. He calls it a "speed problem," if advanced AI outpaces humanity's capacity to intervene.
Hobbhahn points the finger at OpenAI's audit rules. He deems them insufficient for "complete embedded evaluations," even with Apollo integrated into the pilot program. Restrictions on interviewing OpenAI staff, via a whitelist system, present a clear limitation.
Apollo Research investigates whether OpenAI's models can subtly subvert training, evade evaluations, exhibit misalignment in deployment, or conceal their reasoning. A crucial verification, when the stakes are development speed and transparency. Sources: BigGo Finance and Apollo Research.
https://il-quotidiano.web.app/edizioni/report-2026-11-22.html
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.