When should we worry about AI power-seeking? — AI Alignment Forum