The Problem That Folded Itself

EDO SEGAL: Demis, in late 2020, AlphaFold did what fifty years of brilliant human effort had not — it predicted the three-dimensional structure of proteins from their sequences to roughly the width of an atom, and the organizers of the competition that had measured the field's progress since 1994 declared a fifty-year grand challenge essentially solved. Then you did the thing I find almost more remarkable than the solving: you ran it on essentially every protein known to science, two hundred million of them, and gave the database away for free. Tell me what that machine understood about a protein. And Hubert — you've called this his best case. I want to know if it's the case that finally crosses your line.

HASSABIS: Let me tell you precisely what it did, because the precision matters for the argument. A protein is a chain of amino acids that folds into a specific three-dimensional shape, and the shape determines the function — nearly everything that happens in a living cell is done by a protein, and it's done in virtue of its shape. The sequence is easy to read; the genome spells it out. But the mapping from sequence to shape was a cipher no one could break, because the number of shapes a chain could in principle adopt is astronomical — Levinthal's paradox — and yet in nature the protein folds in milliseconds, sliding down an energy landscape to its one correct form. We built a system that learned that mapping. It learned from the database of structures human experimenters had painstakingly determined, and it built in the right priors — that evolutionarily related proteins constrain each other's structure, that physics and geometry impose constraints — and it refined its prediction iteratively, cycling through the structure to sharpen it. And it got there. Atomic accuracy. A thing that took a graduate student years, in seconds.

What did it understand? Here's my honest answer, and it's the most important sentence I'll say tonight. It understood the structure of the problem — the regularities relating sequence, evolution, physics, and shape — well enough to solve it, and that understanding is real and checkable and it was not in any human head, because no human could hold it. It did not understand that proteins are the machinery of life, or that the structures it was giving away would let researchers design enzymes to eat plastic and study antibiotic resistance and attack neglected diseases. It had no idea it was helping anyone. So I'll concede the professor's distinction lives here: there's a kind of understanding it had — the deep modeling of the domain — and a kind it lacked — the situated grasp of what any of it means. But Edo, the kind it had is the kind that solved the problem. And I am not going to let "it didn't know it was helping" diminish that it folded the proteins of life, because the not-knowing is exactly what the professor and I disagree about whether to care about.

EDO SEGAL: Hubert. Is this the case that crosses your line?

DREYFUS: No — and I want to explain why with more care than I've given anything tonight, because Demis is right that it's his best case and I'd be a coward to dodge it. AlphaFold is genuinely different from the games, and better, and I'll tell you exactly why it's better: it left the closed world. Go is a world humans invented to need nothing but pattern. Protein folding is real — the structures are out there in actual cells, governed by actual physics, and the machine got them right. That's not a game. That's the world pushing back, and the machine modeling the push. So this is the strongest thing he has, and if I'm going to lose, I lose here, and I want that on the record.

Here's why I don't lose, even here. Look at what made AlphaFold possible, because it's the same three ingredients every time: a vast search space, a clear objective — the physically correct structure — and a body of ground truth that human beings, in bodies, in laboratories, with X-ray crystallography and years of invested labor and stakes in being right, had laboriously produced. The machine learned the mapping. It did not establish the ground truth. It could not have. Anfinsen's insight that sequence determines structure, the decades of experimentalists who measured the structures the system trained on, the very framing of the problem as worth solving — all of that is the background the machine inhabited without inhabiting, the residue of embodied scientific life the system fed on. AlphaFold is the most spectacular demonstration I have ever seen of how much a system can do by modeling the trace of embodied human practice. And it is, for exactly that reason, the most spectacular demonstration of dependence on the embodied practice it modeled. Take away the experimentalists and there is no AlphaFold. The machine solved a problem that embodied beings had set up, scored, and made solvable. It is not the discoverer in the way Demis is the discoverer. It is the most powerful instrument a discoverer has ever wielded — and I mean instrument, in the precise sense, the hammer, which we are coming to.

HASSABIS: But every scientist stands on the embodied labor of prior scientists — Newton on Kepler's measurements, all of it. If "you used data other people gathered" disqualifies a discovery, you've disqualified all of science. The machine found structure no human knew, in data no human could read. That's discovery, Professor, in the only sense that's ever done any work.

DREYFUS: I'm not disqualifying it, Demis, I'm locating it — and the location is the whole argument. Of course you stand on prior labor; so does everyone. But you, the scientist, also extend the practice — you decide what's worth solving next, you feel when a result is wrong before you can say why, you reframe the problem when it resists, you care that it's true and not merely that it scored well. AlphaFold does none of that. It is superb inside the problem you posed and it cannot pose the next one, cannot feel that a fold is suspicious, cannot want the answer to be right. You called it, in your own writing, a tool — the ultimate general-purpose tool to understand the universe. I'm holding you to your own word. A tool is wielded. The wielding is where the understanding is, and the wielding is yours.

EDO SEGAL: And there's the thread I promised — the tool, the hammer, the thing wielded. Demis says the instrument has crossed into discovery; Hubert says the instrument is exactly an instrument, and that calling it intelligent is mistaking the hammer for the carpenter. There's a passage in Heidegger about a hammer that Hubert returned to more than any other in his life, and it holds a warning that lands hard in an age of tools that almost never break. After the break — the hammer, and what happens to a carpenter whose tools never fail.