Approval-maximizing representations — AI Alignment Forum