The Crossing
**EDO SEGAL:** Here's how this works. For the next stretch I'm barely here. You question each other directly — no moderator to soften it, no me to rescue anyone. I'll only step in if you both fall silent, which I doubt. Noam, you've been gracious about Ilya's machines all night. Now ask him the question you actually want answered, the one you'd ask if I weren't running a civil program.
**CHOMSKY:** All right. Ilya — you say prediction at depth becomes understanding, and that the human and the machine differ in degree, not kind. So I'll ask you the question that I think your own honesty has to answer. You told me earlier you'd been looking, for years, for a way to conclude that I'm right, and kept failing. What, specifically, would I have to show you? Name the observation that would make you say: *no, this is a different kind of thing from a mind, not a lesser degree of one.* Because if there is no such observation — if every result, fluency or failure, confirms your view — then your view isn't a scientific claim. It's a faith. So: what would change your mind?
**SUTSKEVER:** That's the right question and I owe you a real answer, not a dodge. Here's one. If we built systems that scaled to vast capability — that did science, ran companies, exceeded us across the board — and yet, no matter how we probed their internals, we found *no* coherent world model: no representations that corresponded to objects, causes, the beliefs of characters, the layout of described rooms — only a shallow, gerrymandered patchwork of surface correlations that happened to produce good text — then I'd have to concede they were a different kind of thing, fluency without a world behind it, your plagiarism machine vindicated. That's falsifiable, and it's the opposite of what we keep finding: we keep opening them and finding the world model, more of it each year. The second thing that would move me: if the systems hit a hard ceiling exactly at the boundary of their training distribution, never composing genuinely novel structure, only recombining — that would say "no understanding, just interpolation." We see them cross that boundary; the [capabilities that arrive unbidden with scale](https://www.youonai.ai/fieldguide/med/emergent_capabilities) are exactly the ones nobody put in by hand. So my view *is* exposed to evidence; the evidence has just been going the other way. Now I get to ask you, and I'll ask the mirror of your question. Professor Chomsky — what would *you* accept? Suppose we open a network and find a clean, modular representation that respects the structure-dependence you've spent your life describing — that computes over hierarchy, not linear order, that has, demonstrably, the very constraints you attribute to the faculty, grown from data. Would you say the machine has the faculty? Or is there *no* observation that would move you, in which case your "cannot" is the faith?
**CHOMSKY:** Fair — you've earned the symmetry. Yes, there's an observation that would move me, and it's precise. Show me a system that, trained the way a child is trained — on a realistic, impoverished sample, not an astronomical corpus — converges on the human constraints and *cannot* acquire the impossible languages, failing at them the way a child fails, categorically, not "with higher loss." Show me the *cannot*, reproduced from the *poverty*. Then you will have built a model of the faculty, and I'll say so, and it will be the proudest refutation of my skepticism I could imagine, because it would mean we'd finally *understood* the faculty well enough to build it. What you've shown me so far is a system that acquires the constraints from astronomical data and acquires the impossible ones too, more slowly. That's not the faculty. That's a powerful learner. The day the data is poor and the boundary is absolute, I'll change my mind in public. I've done it before.
**SUTSKEVER:** That's a better answer than I expected, and it tells me our disagreement is smaller and harder than the public version. You'll concede understanding if I show you *poverty plus an absolute boundary*. I'll concede a different kind of thing if I find *capability with no world model*. Both are real experiments. Neither is finished. Can I push on one thing, though — the "trained like a child" condition? It seems to me you've built your refutation to require a machine that is a child, and then you'll say, see, only a child can be a child. The child has a body, a life, twelve years, stakes, a mother. Of course the machine trained on text isn't that. But is your real claim about *language*, or is it about *being a living creature in a world*? Because those are different claims, and I think you slide between them when the linguistic one gets pressed.
**CHOMSKY:** That is a genuinely good thrust, and I'll concede — after a moment to feel the force of it — that it's the place I'm most vulnerable. My claim is about language — about the faculty — and you're right that I sometimes reinforce it with the whole life of the organism, the body and the stakes, which is a *different* argument, Dreyfus's argument, not mine. Let me separate them cleanly, since you've forced it. The linguistic claim stands on its own: the specific constraints, the poverty, the boundary. The life-of-the-organism claim is a further thing I believe but that is not, strictly, linguistics. You've caught me bundling them. I'll unbundle: take *just* the linguistic claim, and my challenge stands — poverty plus boundary. Bring the body in and we're having Dreyfus's fight, which is also worth having, but it isn't the one my life's work entitles me to win.
**SUTSKEVER:** I respect that enormously. Then let me say where that leaves me, honestly. On the pure linguistic claim, I think you may turn out to be *more* right than my field admits — there may be a real faculty-shaped fact about poverty and boundary that our data-drunk systems obscure rather than explain. On the larger claim — that there's nothing of mind in these systems, no understanding, no possible inside — I think you're wrong, and I think the next decade goes against you. The strange thing is I could be conceding the linguistics to you while you concede the mind to me, and we'd both be telling the reader something true.
**CHOMSKY:** That may be the most either of us can honestly claim. And I'll say, for the record and without my usual asperity: I came in expecting an engineer defending a product. I found someone willing to name what would prove him wrong, which is rarer than intelligence and more valuable. We disagree about almost everything that matters. I'd rather disagree with you than agree with most people.
**EDO SEGAL:** I told you I'd stay out. I'm breaking it for one sentence, because the reader can't see the room and someone has to mark it: that was the only stretch tonight where neither of you was performing for me, and it was the closest either of you came to changing the other's mind. Hold it, all of it, into the last hour. One round left. Closing statements. The floor, one final time, is each of yours.