Learning human preferences: optimistic and pessimistic scenarios — AI Alignment Forum