Understanding mesa-optimization using toy models — AI Alignment Forum