Friendlier AI through politics — LessWrong