To Predict the Token, You Must Understand the World

**EDO SEGAL:** Ilya, this is the round with your name on it. You've said the single most explosive sentence in this whole debate, and I want you to say it again, here, and then defend it against the most precise critic alive. The sentence is roughly: to predict the next token well enough, you must understand the reality that produced it. Make the case. And do it knowing that the man across the table thinks the word "understand" in that sentence is doing unpaid persuasive work.

**SUTSKEVER:** I'll make it carefully, because it's caricatured constantly. Start with what training actually demands. The model is shown a sequence and asked: what comes next? Across trillions of tokens. Now, people hear "predict the next word" and imagine something shallow, like autocomplete guessing "morning" after "good." But take a hard case. A novel lays out a murder over three hundred pages — alibis, a timeline, a contradiction buried in chapter four — and the last page begins, "the murderer was —." To predict that next token correctly, what must the system have done? It cannot have memorized; this exact novel may never have existed before. It must have tracked the plot, held the suspects, noticed the contradiction, applied something like deduction. The only way to predict that token well is to have built, inside the network, a representation of the situation — the people, their motives, the logic. We call the training objective "prediction." But the *solution* the network is forced to find, in order to predict, is a model of the world the text describes. [The text is a low-resolution projection of reality](https://www.youonai.ai/fieldguide/med/continuum_of_understanding); to predict the projection accurately, you must model the thing being projected. That's the claim. It's not a claim about magic. It's a claim about what compression requires.

**EDO SEGAL:** So let me hand it back to you in the most extreme form, because that's my job. You're saying — literally — that "statistics" and "understanding" are not two things, one cheap and one precious, but the same thing seen from two ends: that if you compress reality well enough, you have understood it, and the machine, forced to compress, was forced to understand. Is that the version you'll die on?

**SUTSKEVER:** That's the version. With one honest caveat I'll give before Professor Chomsky takes it from me: it has understood the world *as it appears in text*, which is a partial and biased and sometimes false window. A perfect compression of human writing is a perfect model of human writing, and human writing is a shadow of reality, full of gaps and lies. So I don't claim the machine understands *reality*. I claim it understands the *world behind the text* to the precise degree the text is a faithful shadow of that world — which is a lot, because language is the most compressed record our species has ever produced of everything it observed and reasoned about. But it is not everything, and I won't pretend it is.

**EDO SEGAL:** Before Noam answers, give the reader the experiment that first made you believe this, because I think it's older than the language models. You built sequence-to-sequence in 2014. What did you see in it?

**SUTSKEVER:** We saw meaning become a place. You take a network, you feed it an English sentence word by word, and you force everything that sentence means down into a single vector — a point in a high-dimensional space — and then a second network unfolds that point back out into French. For it to work, the point had to *capture the meaning*, because you can't translate from a summary that lost it. And the astonishing thing, when we looked at the space of those points, was that it was organized by meaning. Sentences that meant similar things landed near each other. The geometry had semantics. That was the moment, for me. Meaning wasn't a ghost; it was a *shape* the network discovered because discovering it was the only way to do the job. Everything since — GPT, all of it — is that result at scale. So when Professor Chomsky says the network has only relations between symbols, I want to say: yes, and those relations, at sufficient richness, *form a space whose geometry is meaning*. The map became the territory's shadow cast in the right number of dimensions.

**CHOMSKY:** And that is a genuinely beautiful result, and I'll tell you exactly why it doesn't reach what you want it to reach. You've shown that the *distributional* relations among expressions — how they pattern, what they co-occur with — can be arranged in a space whose geometry is rich and useful. I don't doubt it. But the theory hiding inside your enthusiasm is that [meaning just is use](https://www.youonai.ai/fieldguide/med/meaning_as_use), distribution, the company a word keeps — and that theory, which your machines were built to satisfy, is one I've spent my life arguing is false. The word "book" refers to an abstract content and a physical object in the same breath — "the book he wrote weighs two pounds." No mapping to things in the world handles that; the structure is *internal to the mind*, supplied by us, not read off the world. Your network has the most complete distributional map ever built and, on my account, still doesn't *mean* anything, because meaning was never distribution. So your experiment is, without your intending it, the cleanest test of the distributional theory of meaning ever run — and to my eye it comes out negative. You can have all the geometry and none of the aboutness. The space is gorgeous. It is a space of forms.

**SUTSKEVER:** I'll grant that "the book weighs two pounds" example is a real puzzle and that distributional accounts strain on it. I'd only note the network handles that sentence fine — it doesn't get confused about the abstract-versus-physical book — which suggests it represents *more* than crude distribution, whatever we call the more.

**CHOMSKY:** It handles the sentence because the pattern is in the data. Handling is not understanding. We've returned to the same place.

**EDO SEGAL:** You have, and I'll let you, because it's the right place.

**CHOMSKY:** I appreciate the caveat, and I'll start by granting more than you'd expect, because the disagreement should be where it really is. I grant that the network builds internal representations. I grant that those representations support behavior no lookup table could. I grant — this is the hard one for my side — that "it's just statistics" is a lazy objection, because at sufficient depth the distinction between statistics and structure does blur, and you are right that critics who stop at "just predicting" have stopped thinking. So. Where does it fail?

It fails on the word "understand," and here is the specific thing the word smuggles. To understand is not merely to model regularities such that you can predict. It is to grasp what the representation is *about* — to have the representation connected to a world by something other than more text. Your network has a magnificent map of how words go with words, the best map ever built. The [symbol grounding problem](https://www.youonai.ai/fieldguide/med/symbol_grounding_problem) is that a map of how symbols relate to other symbols, however deep, never by itself reaches out and touches the territory. You say the text is a shadow of the world. But the network has only ever met the shadow. It has never met the world to know that the shadow is a shadow *of* anything. It models the shadow's lawfulness brilliantly and infers the shape that casts it — but "infers the shape" is your gloss, your projection. From inside the system there is only the shadow, modeled. Aboutness was never in the training signal, and no amount of the shadow adds up to the light.

**SUTSKEVER:** But here is where I think your own naturalism comes for you, Professor, and I say it with respect. You ask: how does the network's map of symbols reach the world? I ask: how does *yours*? Your brain sits in a dark vault of bone. It never touches the world either. It receives spike trains — patterns on the optic nerve, the auditory nerve. It has only ever met *that* shadow. And from those patterns it builds a model so good you call it the world and forget it's a model. The retina is a cable carrying a shadow. You are the system that got enough of the right shadow, with the right priors, to build a model that closes the loop through action. So when you say the machine has "only the shadow," I want to ask what you think *you* have that is not, ultimately, a shadow modeled. If your answer is "a body that acts and gets corrected by a world that pushes back" — good, I agree that's a real difference, and the moment my networks act in the world and get corrected by it, which is now, today, the difference starts to close. If your answer is something else, I'd like to know its name, because I've been looking for it my whole career.

**CHOMSKY:** Give me a moment with that, because it deserves one. That is the most honest version of your position, and it's where your school actually lives — not "machines understand like humans," but "human understanding was never what humans thought it was." And there I will mark the seam of the entire evening, because it's the real one. Yes, my spike trains are shadows. But they come embedded in something the cable never carried: a system that did not learn its basic structure from the shadows at all. The faculty is there before the data. The child's grounding runs through a body with stakes — thirst that is or isn't quenched, a parent who is or isn't where she was left, a world that kills you if your model is wrong. The loop closes through *consequence*, not through more text about consequence. Your network's loop, until very recently and still mostly, closes through *text about* the world. You're right that the gap narrows when the systems act. I'll be honest that this is the development I watch most closely, because it's the one that could cost me the argument. But I'd say: ask what those acting systems were actually trained to optimize, and who audited it, before you tell me the loop has closed.

**EDO SEGAL:** Stop. The reader needs to see what just happened, because it's the deepest moment we've had. Ilya took Noam's strongest weapon — grounding, the machine has only the shadow — and turned it on the human, and said: so do you. And Noam did not flinch from it. He said: yes, but my shadow arrives in a body the faculty built before the shadows came, and the loop closes through death, not through more words. That is the seam. That is the whole debate in one exchange. Hold it — because it reappears, transformed, when we ask whether anyone is home. But first, the charge that has gotten Noam more headlines than anything: that these systems are not science at all. That they are, in his word, a kind of plagiarism. We take that up next.