Towards Deconfusing Gradient Hacking — LessWrong