Hard problem of corrigibility — LessWrong