Gradient hacking: definitions and examples — AI Alignment Forum