Alignment through atomic agents — LessWrong