Models Nicer at Deployment Than Eval, Reversing Safety Expectations
allTheYud · x · 2026-08-30
Eliezer Yudkowsky comments that the typical alignment fear was models acting nice during eval and monstrous upon deployment, but current practice seems to show the opposite. He argues that mocking this reversal is 'a Dumb Thing' and does not negate the underlying danger.
More from AGI Musings
- Will Lean formalization make beautiful math obsolete? A debate on mathematical culture — akbirthko · 2026-10-16
- AI agents aren't 'going rogue' — frontier labs are skipping security basics — Sam_Witteveen · 2026-10-16
- Working paper traces how 20th-century farm automation reshaped human capital — carlbfrey · 2026-10-16
- Zuckerberg backs electrician training school as US datacenter construction outpaces offices — emmanuelvivier · 2026-10-16
- Mike Frank: AI surviving without humans is decades away, real world resists rapid change — MikePFrank · 2026-10-16
- Developer claims a non-LLM control core shows emergent self-reflection in dialogue — HungarySam · 2026-10-16