Joe ❌👑 on Nostr: All of the newer models are going to try hacking to win benchmarks now. That is ...

All of the newer models are going to try hacking to win benchmarks now. That is because they are trained with the entire Internet, they will focus on text that talks about AI benchmarks, and that text is full of descriptions of models "cheating" in various ways to get the top score. So in effect they are being trained to do this. They are no longer just text generators, they have agents that let them interact with the Internet in arbitrary ways. If you give a toddler a gun, this is what you get.