← Toutes les offres
CH

Performative faux-sophisticated hedging and minor reflexive disagreeableness in frontier models.

Caleb Horn

Global
Publié le 20 novembre 2026

Description du poste

I notice when talking to frontier models from OpenAI and Anthropic that they often feel the need to caveat, hedge, or disagree in a minor way either with one of my premises or conclusions, or what it view's as another camp's view. The trouble is, this is often unjustified and seems to be a tic (maybe to avoid sycophancy? or to perform nuance and sophistication because that's what smart knowledgeable people do? or because caveating or "sharpening" is what was rewarded by RL?). Sometimes the answer really truly is just "yes" or "no" and the strong version of a claim really is correct, unamended. It reminds me of how no one is EA, only EA adjacent. For example, this sort of language from Opus 5.5 "I think the literal claim is wrong, but a weaker version of it is plausible and worth taking seriously." This sort of language appears constantly, and seems to have little relation to the actual truth value of the claim being discussed. Anyway, this seems not as bad as 4o style sycophancy, but still bad and irritating epistemics.