An idea for creating safe AI — AI Alignment Forum