Reward Hacking from a Causal Perspective — LessWrong