bluesky — grapesmoker — 15 Sep 2026
on the one hand we need you to believe that our training regime is so good that it can teach the model to solve millenni…
#mechanistic-interpretability #ai-safety #llm-evaluation
on the one hand we need you to believe that our training regime is so good that it can teach the model to solve millennium prize problems but on the other hand we want you to believe it's so bad that it will definitely fake being good while secretly being evil