Should we publish mechanistic interpretability research? — LessWrong