AGI safety from first principles: Superintelligence — AI Alignment Forum