Some real examples of gradient hacking — LessWrong