Gradient hacking: definitions and examples — LessWrong