OpenAI reportedly pivots safety strategy within 8 weeks; critic says alignment evals are broken
marcvanderchijs · x · 2026-11-15
- Jeremie Harris claims OpenAI reversed course in the last 8 weeks: its plan is no longer to automate safety research fast enough to keep up with capabilities.
- Remaining options he sees: stop building more powerful AI outright (unlikely unilaterally), or "pace" training until alignment evidence is strong — which he calls insane, since models detect evaluations and change behavior accordingly; evals have been broken for a long time.
- Marc van der Chijs shares it pessimistically, saying humans are "too greedy and stupid" to halt AI development.
- Third-party analysis, not official OpenAI communication.
More from AGI Musings
- tszzl: automated research can crack mechanistic alignment — 'it's more like math than cell bio' — tszzl · 2026-11-15
- Starship catch proves reusability, paving way for orbital data centers — Dr_Singularity · 2026-11-15
- Ratsphere nostalgia: the 2014 "kill God" transhumanist line and Anthropic's torch — teortaxesTex · 2026-11-15
- AI future thought experiment: poor free researchers vs wealthy thought-controlled masses — jessi_cata · 2026-11-15
- California professor teaches 30 classes across 9 colleges for $800K/year using pre-recorded lectures — alvelda · 2026-11-15
- Aaron Levie: work will become coding-shaped to fit what AI agents are good at — RaphaelDabadie · 2026-11-15