How useful is mechanistic interpretability? — AI Alignment Forum