Safety via selection for obedience — LessWrong