What makes you think that a machine any less complex than the entire organism is capable of "grokking what is actually going on"?

I agree with you that humans may never be able to understand most of what's going on in biology. But I'm even less enthusiastic about gradient descent. The interesting results in machine learning for biology all seem to amount to better ways of indexing the data we already have (like crystallographically-discovered protein shapes) and interpolating between the existing matches nearest to a query term.

I don't claim to know what "understanding" really truly is, but I do know what it isn't.