Minimax as an approach to reduced-impact AI — LessWrong