Downstream applications as validation of interpretability progress — LessWrong