Attainable Utility Preservation: Scaling to Superhuman — AI Alignment Forum