Corrigibility doesn't always have a good action to take — AI Alignment Forum