The Crossing
EDO SEGAL: The rules of this round are short. You question each other, directly. I will not rescue anyone. Pedro, you won the toss backstage. Begin.
DOMINGOS: Karl, you've spent the evening demanding the falsifier — what would prove a claim wrong. I want to turn it on you, because I think your own position has a soft place you never let anyone press. Here it is. You say the machine doesn't understand, only predicts; that it guesses without doubting; that it learns without grasping. Name the falsifier. What, concretely, would a machine have to do for you to admit it understands — not as a philosopher's concession, but a real test, with a pass and a fail? Because if you can't name it, then "the machine doesn't understand" is exactly the unfalsifiable claim you've spent sixty years condemning, and I get to call it your pseudoscience.
POPPER: That is the right question and it is aimed precisely at my weakest joint, and I will answer it honestly rather than safely, because anything less would forfeit the evening. Here is my falsifier, and it is not the Turing test, which I regard as a test of human credulity, not machine understanding. My test is this: show me a machine that, faced with its entire training distribution pointing one way, conjectures that the truth lies the other way — proposes a bold theory that contradicts the data it was built on — and stakes on that theory a single, risky, decisive prediction, specified in advance, that the existing consensus says should fail. And then the prediction comes true. Not interpolation. Not a clever recombination of the seen. A genuine conjectural leap against the corpus, defended by exposure to refutation, vindicated by the world. Do that, and I will say in public, in those words: this machine has crossed from prediction into understanding, because it has done the one thing only a knower can do — preferred a refutable truth to a confirmed comfort. I do not say it is impossible. I say it has not been done, and that everything in the present architecture and every dollar in the present market pushes against it. There is my falsifier. It has a pass and a fail. Now I have one for you, Pedro, and you will like it less than you expect.
DOMINGOS: Go.
POPPER: You have insisted all evening that the machine learns by conjecture and refutation — that the test set is the severe test, that overfitting is the enemy, that your field mechanized my philosophy without knowing it. Very well. Then name your falsifier, in the same key. What would have to happen for you to admit that what these systems do is not conjecture and refutation in my sense — that it is, after all, the blind induction I accused it of, the amoeba at scale, sophisticated curve-fitting wearing the costume of testing? You cannot answer "nothing," because then your claim that the machine is a falsificationist is itself unfalsifiable, and the demarcation knife you have happily watched me swing at the doomers turns in your own hand.
DOMINGOS: I accept the symmetry, and I'll answer in the same key, drawn at function, not performance. Here's my falsifier. I'd admit it's blind induction and not genuine conjecture-and-refutation if the systems remained, with another thousandfold of scale, brilliant interpolators and consistent failures at composition — if they kept getting more fluent while systematically collapsing on problems that require combining represented pieces in ways genuinely absent from the training distribution. If the continuum from prediction to understanding turns out to have a cliff exactly where I claim a slope — if they can recite every physics text and never once apply a principle to a configuration no text described — then my "they conjecture and test" story is dead, and what's left is exactly your amoeba, and I'll say so. And here's the timeline that makes it falsifiable rather than faith, Karl: I expected those compositional failures to be obvious by now if you were right. Every year they get less obvious instead. That trend is my evidence. Its reversal would be my refutation. But notice — and this is my counter — your falsifier and mine might be the same experiment. The machine that composes against its distribution is the one that passes yours, too. We may be standing at the same door describing different keys.
POPPER: [ He is right that they may be the same door, and that is a more troubling agreement than any disagreement. ] Then let me press where I think your wall is already cracked, because you slid past it. You define the failure as compositional failure — combining represented pieces in new ways. But composition within the space of represented pieces is still inside the harbor. It is recombination, however clever. The leap I am demanding is not recombination of the represented. It is the conjecture of a piece that was never represented — the term that contradicts every text, the entity no one had posited, the Neptune predicted from a wobble before anyone saw it. Your falsifier tests whether the machine is a brilliant combiner. Mine tests whether it is a genuine creator. And I put it to you that you defined the easier test on purpose, because the harder one is the one your machines cannot pass and you know it.
DOMINGOS: That's a real hit and I'm not going to pretend it isn't. You're right that I drew the line at composition because composition is where I can show progress, and creation-of-the-unrepresented is harder and I can't yet show it. But I'll fight you on whether the line between them is as bright as you need it to be. Neptune wasn't conjured from nothing — it was forced by an anomaly in the represented data, a wobble that didn't fit, and the "new" entity was the most economical fix to a refutation the existing model couldn't absorb. That's a very sophisticated composition driven by a prediction error. So I'll grant your test is harder and concede we haven't passed it — and I'll plant a flag that says I'm not sure your creation and my composition are different in kind rather than degree. You think there's a bright line where the genuinely new begins. I've spent my career finding out that bright lines are usually gradients we haven't measured finely enough. That might be my deepest disagreement with you, and I can't prove I'm right. Neither can you.
POPPER: No. I cannot. And I will say something now that costs me, because you have been honest and the evening deserves it in kind. The possibility that creation is only composition we have not yet resolved at fine enough grain — that the bold conjecture itself is, mechanically, a recombination driven by an anomaly, in the human skull as much as in your machine — is the one possibility that genuinely unsettles me, because if it is true, then the line I have defended my whole life, between the knower and the predictor, runs through me as well, and I am less unlike your machine than my philosophy requires. I do not believe it. But I have spent my life saying that what I believe is not the measure of what is true, and I am too old to make an exception for the belief I love most.
DOMINGOS: And I'll meet that, because it goes both ways and only one of us has said his half. If you're right — if there is a genuine creative leap that's different in kind, the conjecture of the unrepresented, and the machine truly cannot do it — then everything I've built has a ceiling I've been unwilling to look at directly, and the master algorithm, even if I find it, is a magnificent dead end with respect to the one thing that matters most. I've spent thirty years betting the line is a gradient. You've spent sixty betting it's a wall. We have just admitted, both of us, out loud, that we are not certain, and that the experiment that would settle it has not been run. I don't think either of us expected to end up here.
DOMINGOS: One more, since the chair is silent and I get the last question. Karl, you've made the machine sound like a danger to knowledge — the oracle, the clinger, the confirmation engine. But I built these things partly because I believe the opposite is also true: that a tool which can read across every discipline and surface the connection no single human could see is, in the right hands, the greatest instrument for the growth of knowledge ever made. You spent your life on how knowledge grows. So tell me honestly — not the danger, the hope. Where, if anywhere, does this machine help your open society rather than corrode it?
POPPER: [ It is the right last question, and it deserves my best, not my fear. ] It helps precisely where it is used as I have begged people to use each other — as a generator of bold conjectures to then attack. The bottleneck of knowledge was never the testing; the testing is mechanical once you have the idea. The bottleneck was always the conjecture — the creative leap, which is rare, and which most minds, most of the time, simply cannot produce in sufficient variety. Your machine produces conjectures by the thousand, cheaply, across every field at once. If a scientist treats that flood as what it is — raw, untested guesses, most of them wrong, a few of them fertile — and brings to it the one thing the machine lacks, the disciplined will to refute, then yes: you have built the most powerful aid to the growth of knowledge since the printing press. The danger and the hope are the same machine pointed in opposite directions. Aimed to confirm, it is the death of the open mind. Aimed to generate hypotheses for a human to kill, it is its greatest servant. The fork is not in the silicon. It is in whether the person holding it still knows how to doubt.
DOMINGOS: Then we agree on the one thing I most needed us to agree on, and I didn't expect to get it from you. The machine is a conjecture engine. Whether it grows knowledge or rots it depends entirely on whether there's a refutation engine attached — and the only refutation engine we've got that works outside the harbor is still a human being who knows your method. So the most important thing my field could build isn't a bigger model. It's people who can do what you do. That's a strange conclusion for an engineer to reach at the end of three hours. I'll sit with it.
EDO SEGAL: And there — after the longest silence of the night — you are holding opposite ends of the same uncertainty, and neither of you flinched from it. We close after this. Final statements. The last word each.