Aligning AI by optimizing for "wisdom" — LessWrong