@Turn_Trout @Michael05156007 In practice, the situation is more complicated; the RL process itself *does* upweight/upvote the circuits that make “I should optimize the evaluator score” a salient…

Posters: davidad. davidad

2 likes recorded by the source corpus.

Topics: AI & LLMs, AI safety

View Post: https://x.com/davidad/status/2001661824383520792