“Reward hacking in the training of language models is a major blocker for the real-world deployment of more autonomous AI use cases.”