Indifferent vs false-friendly AIs — LessWrong