Funny how there was a lot of concerns then about reward hacking, something I never hear anyone talk about with current AI
I think it just got folded under the umbrella concept of model alignment. And it moved from theoretical discussions to practical daily struggles with LLMs deleting failing unit tests