Moderator's Dilemma: The Risks of Partial Intervention — LessWrong