Efficient prestige hypothesis — AI Alignment Forum