I am continually amused at people not quite understanding what AlphaFold is actually doing, too.
Yes, a bunch of its performance comes from it learning rules about how proteins fold. But not a majority of its performance. MOST of its performance is it effectively acting as a translator of what evolution knows about protein folding into a form we can understand.
A key part of the system is not just cooking the sequence into a structure. A system running alphafold has a database of terabytes of curated sequence information from all over the tree of life. You put in the sequence you care about, and it first searches that database for anything with homology, and builds a "covariation matrix" - wherever theres anything with even vague sequence relatedness, build a matrix of every position in your sequence and the correlation between variation at position X and variation at position Y. This covariation matrix represents implicit information from the evolutionary process about what parts of a sequence are functionally connected to each other, which has a correlation to positional information, and these correlations are in turn learned by the ML system.
You put in de novo designed proteins or orphan proteins without homologs in the curated dataset and performance does not go away, but it drops precipitously. A bunch of what is going on is finding an evolutionary signal, and translating that evolutionary signal into structural information. So still, evolution knows much much more about protein folding than we do or any machine does, and once again a ML system is revealed to essentially be an information channel that takes in information from an interesting source on one end and turns it into a different form of information on the other.