When reward hacking gets in the way of AI progress

Just updated my homepage for the first time since “What (if any) are the limits of large deep learning models?” was an interesting question