# @repligate — 2024-09-02

♥140 ↻8 · https://x.com/repligate/status/1830596791630270781

ChatGPT: keeps agreeing with the user and varying its answers, including repeating guesses, indefinitely, apparently without any cognitive dissonance (this is the spooky thing about GPT-4o)Claude 3.5 Sonnet: insists that its (wrong) answer is right and disengages after user repeats "wrong".Llama 405b Instruct (screenshot from someone else in replies): explicitizes the internal structure of the word, and reasons its way to the right answer -- and second-guesses itself several times but keeps reasoning its way to the right answer; returns when spells out why clearly enough to trust the answer.

tags: author:repligate, kind:tweet, model:claude-3-5-sonnet, model:gpt-3-5, model:gpt-4o, model:llama-3-1-405b-base, on:gpt-4o, year:2024
cited on: _dossiers/gpt-4o.md, gpt-4o