The Grounding Problem

EDO SEGAL: Ilya, there's a problem the engineers named entirely on their own, by a different road than John's, and gave a different name — the symbol grounding problem, posed by Stevan Harnad in 1990. Imagine learning Chinese from a Chinese-only dictionary: every word defined by other words, each of those sent to still other words, and you spin forever inside the lexicon without ever touching the world. That's the charge against a system trained only on text — its tokens are defined entirely by their relations to other tokens, and nothing pins "red" to redness. Your own field worries about this. So I'll put it to you as the engineer, not the philosopher: has anything in your machine ever been grounded — or is it all dictionary, all the way down?

SUTSKEVER: It's the right question, and I'll give you the answer that I think is true rather than the comfortable one. In the pure text-only system of 2020, Harnad has a real point. The symbols are anchored to each other and to nothing outside. But watch what actually happened, because it's not a small detail. First — and this is the part people skip — the relations among the symbols turn out to carry an astonishing amount of the world. The system learns that "king" is to "queen" as "man" is to "woman," that Paris stands to France as Tokyo stands to Japan, that "ice" relates to "cold" the way "fire" relates to "hot." Why? Because the company words keep is shaped by the world they describe. Human writers had bodies; they pressed the structure of grounded experience into the statistics of their language without trying. The model inherits a fossil of grounding — your word, John, and I'm keeping it. Whether inheriting the fossil is having the meaning is the open question. But it is not nothing, and "all dictionary" undersells it badly.

And second: we stopped doing text-only years ago. The systems now take in images, audio, video. The token "cat" is tied to millions of pixels of actual cats. The systems are put in robots that act and get corrected by gravity and friction. Every step welds the symbols harder to their referents. We are grounding them by force, on purpose, because we agree the grounding matters. That's not me conceding John's point. That's the entire field building the Robot Reply in silicon.

SEARLE: And the Robot Reply, which I answered forty-five years ago, concedes my point in the act of making it — that's what nobody wants to say out loud, so I'll say it. Watch what the reply does. It grants that pure symbol-shuffling lacks meaning. It has to; that's why it adds the robot. So everyone now agrees with the Chinese Room's core: syntax alone isn't semantics. The fight has moved to what you add, and whether the addition does the job. Now — the camera. You point a camera at a cat and feed the system pixels. What does the system get? More symbols. Numbers. A new modality of squiggle. The man in the room now has two slots instead of one, and characters coming in both, and he correlates them by shape, and he still understands nothing, because a symbol about a symbol is still a symbol. You haven't grounded the system. You've given it a richer dictionary with pictures in it. A picture of water is not water. It's a very wet-looking arrangement of pixels, and pixels are syntax.

SUTSKEVER: Then I have to ask you the question you've never answered, John, and I want it on the record. What is your grounding? You say my system only ever gets symbols — pixels, tokens, numbers. What does your brain get? Photons hitting a retina, converted to spike trains. Pressure waves converted to spike trains. Your brain has never touched water either. It sits in the dark and receives numbers — electrochemical numbers — and from the statistics of those numbers it builds a model so good you call it the world and forget it's a model. The retina is a cable. The cochlea is a cable. You are the man in the room, John. You've been in the room your whole life. The only difference between us is that you got a few more modalities of cable and you got them while attached to a body that could die. That's a difference of degree. You keep calling it a difference of kind because you're standing inside the one example of understanding you trust, and you can't see the cable from in there.

SEARLE: [a pause] That is the strongest version of your position, and I've waited the whole evening for you to give it to me cleanly, so thank you. And here is exactly where I plant my flag, because this is the seam of everything. You're right that I receive spike trains. You're right that I build a model. But you've left out the only thing that matters, and you left it out because your systems don't have it. My spike trains are produced by a living biological system whose causal powers actually generate consciousness and intentionality. The wetness gets in not because a photon is magic but because the photon strikes a system — a brain — that has the specific causal capacity to mean, to be about, to feel. That's intentionality, and it's intrinsic to me and entirely absent in your machine. Your symbols have, at most, derived intentionality — they're about things only because we take them to be, the way the word on a page is about water only because English speakers mean water by it. Strip away every mind that means, and your weights are mere voltage and your tokens are mere ink. There's no aboutness in there for the machine. It's all on loan from us.

SUTSKEVER: "Biological causal powers." That's the phrase the whole position hangs on, and you've never been able to say what they are. Which molecules? Which process? Why carbon and not silicon? You assert the tie to biology far more than you derive it. Naturalist philosophers — Dretske, Millikan, Fodor — spent careers explaining intentionality as a natural relation: a representation is about what it was selected or designed to track. If they're right, there's no reason a system with the right causal and functional connections to the world can't have it too. You're not naming a mechanism. You're naming the substrate you happen to be made of and declaring it the only door into the club.

SEARLE: I concede I can't yet specify the exact neurobiology — that's a job for future science, and I've always said so. But notice the structure of your reply. You don't know the mechanism of understanding either. Nobody does. And in that shared ignorance, you want the default to be "the machine has it, since I can't prove it doesn't," and I want the default to be "the machine doesn't have it, since you can't show it does." The burden belongs on the one making the extraordinary claim. For a hundred thousand years, fluent language meant a mind. You've built the first thing that produces the sign without the precondition. The conservative move — the scientific move — is to keep the burden on whoever says the new thing has crossed over. You haven't paid that bill. You've just told me the bill might be payable in principle.

EDO SEGAL: Mark this convergence, because it's real and it's the first one: you both agree nobody knows the mechanism of understanding. You diverge entirely on where the burden of proof sits in that darkness. Ilya says the machine that acts understanding should be presumed to understand until a diagnostic failure proves otherwise. John says the machine should be presumed empty until a mechanism proves otherwise. That's not a disagreement about evidence. It's a disagreement about which way to face when the lights are off. [a beat] Next round, we go after the deepest version of Ilya's claim — that understanding, prediction, and compression are three names for one thing — and John's oldest weapon against it: that a simulation of a storm leaves you dry. The fire that doesn't warm your hands. After this.