Optimizing Human Collective Intelligence to Align AI — LessWrong