How will we do SFT on models with opaque reasoning? — LessWrong
lesswrong.com · saved by 1 readers
Current LLMs externalize lots of their reasoning in human interpretable language. This reasoning is sometimes unfaithful, sometimes strange and conce…