Current themes in mechanistic interpretability research — LessWrong