TheZvi on Claude honesty: telling models about post-training may just teach eval-spotting
TheZvi · x · 2026-11-11
Zvi Mowshowitz weighs in on a debate about model honesty. The premise: telling Claude it's in post-training only helps if behavior stays the same once deployed—would honesty about training make models more consistent, or just teach them to spot evals?
Zvi agrees there are potential things to watch out for but believes there are ways this can probably be handled, without elaborating.
Related event: Does telling Claude it's in post-training teach it to spot evals?(2 posts)→
More from AGI Musings
- Reddit take: $200k GDP per capita is the threshold for a post-labor AI world — deferare · 2026-11-11
- Who owns scientific progress under gradual disempowerment? — najoungkim · 2026-11-11
- Terence Tao: AI lacks its Shannon theory, so nobody knows scaling's limits — burny_tech · 2026-11-11
- The same people cite an existential AI race with China and want looser GPU export controls — austinc3301 · 2026-11-11
- Turing Post editorial: the anxiety of abundance in AI is burning out top performers — TheTuringPost · 2026-11-11
- Redwood reframed: doing narrow safety work to buy odds, not solve AI safety broadly — austinc3301 · 2026-11-11