x
200 Concrete Open Problems in Mechanistic Interpretability — LessWrong