Kyle O’Brien comments on Maybe we should pretrain on synthetic data about good-but-reward-hacking AIs