Reward hacking is swamping model intelligence gains · HackerLangs