Perils of optimizing in social contexts — LessWrong