This work focuses almost entirely on a single category of proteins: fold-switchers. That is, proteins that can adopt two distinct conformations depending on the state of the system. Sort of like a flip-flop.

Anyway, its conclusions are a bit grandiose, we already would have expected AF2 and 3 to do poorly on predicting both structures of a fold-switcher. And nobody really thinks that AF "learns the protein's energy function"- AF3 does nothing of the sort.

Most of AF3's abilities are memorization, but that's not a bad thing. It's already been shown to generalize out of class but also work poorly where there is poor data density (either coevolutionary data, or structural data).