One of my favorite classes was a philosophy of science seminar on forms of explanation, where we were asked: what is the relationship between explanation, prediction, and understanding? The standard view being that understanding comes first, you learn the mechanism, and then you can explain it and predict what it will do. Working through the class, I came to think that this view is incomplete.

AlphaFold predicts the three-dimensional structure of a protein from its amino acid sequence. At the 2020 CASP14 competition, it achieved accuracy comparable to experimental determination, effectively resolving a problem that had been open for fifty years. But AlphaFold has no theory of how proteins fold that is legible to us. It learned patterns from millions of known structures in the Protein Data Bank, predicting new structures with extraordinary precision without being able to explain to us why a given sequence folds the way it does.

So does AlphaFold understand protein structure? I think the question is better asked differently. Ramsey had a useful principle: we judge habits of thought by whether they work. So rather than asking whether AlphaFold really, truly understands, the question is: what do we gain by calling what it does understanding, and what do we lose by not?

This is not Polanyi's tacit knowledge, where a skilled human knows more than they can tell through practice. AlphaFold has no experience and no body. It just learned from its training, and the regularities held. I call this pattern-predictive understanding.

Consider chicken sexers, whose job is to sort day-old chicks by sex shortly after hatching. Expert sexers do this with near-perfect accuracy and struggle to fully articulate what they perceive. Nobody watching would call it guessing, but nobody would call it understanding either. The knowledge is shallow, reliable for one classification and too unstructured to take anywhere new.

AlphaFold is different. Its input space is this effectively infinite set of all possible amino acid sequences, and to predict correctly on sequences it has never seen, the model must have captured something about how sequences determine structures, what chain properties make certain folds stable and what residue interactions recur. Those are regularities of the physical world, and they run deep enough that biochemists build on its outputs, designing experiments around predictions that no human understanding generated. That is the pragmatic test. If we call what AlphaFold does understanding, we treat its predictions as something to build on, and that label changes what science looks like.

There are real limits. AlphaFold's accuracy degrades on proteins too unlike its training data, on intrinsically disordered regions, and on interactions never represented. De Regt and Woodward would also note that AlphaFold cannot be interrogated or reason counterfactually. That narrows what this kind of understanding is, but it does not take understanding off the table.

The hardest objection to deflating understanding is that it seems to require something happening inside the understander, some internal representation that is more than just a prediction machine. But there is evidence that something is happening inside. Neel Nanda reverse-engineered a small neural network trained on modular addition and found it had independently discovered an algorithm using Fourier transforms to perform arithmetic as rotation on a circle. "This algorithm was purely learned by gradient descent... I did not predict or understand this algorithm in advance." The models are not empty. The question is whether we can learn to read what they have found.

What is the relationship between explanation, prediction, and understanding? I think prediction can be understanding when treating it as such leads to better science, and a lot of what we dismiss as prediction may already be knowledge.