ryan_greenblatt comments on How useful is mechanistic interpretability?