Value learning: ultra-sophisticated Cake or Death — AI Alignment Forum