You never hear structure: Noam Chomsky on language, cognition, and deep learning

Lex Fridman opens with a question he admits is absurd: if an alien species visited Earth, could we find a common language or protocol of communication? Noam Chomsky says there are arguments that we could. One of them came from Marvin Minsky, who some 20 or 30 years earlier ran a brief experiment with his student Daniel Bobrow.

They ran the simplest possible Turing machines and left them free, just to see what would happen. Most crashed — they either fell into an infinite loop or stopped. The few that persisted essentially produced something like arithmetic. Minsky’s conclusion was that if some alien species developed higher intelligence, it would at least have arithmetic: at least what the simplest computer would do.

Chomsky adds something Minsky did not know at the time. The core principles of natural language are based on operations that yield something like arithmetic in the limiting, minimal case. So it is conceivable, he says, that a mode of communication could be built on those shared cores of human language and of arithmetic. The science-fiction setup is his way into a lifetime’s claim: the inner computational shape of language may not be a local accident of Earth.

Fridman asks about language as an internal system inside the mind versus language as an external system, the way it is expressed. Chomsky will not take the contrast. They are not alternatives, he says. They are two different concepts covered by the same word.

It is a simple fact that something about you — a trait of the organism — determines that you speak English and not Tagalog. That inner system determines the sound and meaning of the infinite number of expressions of your language. It is localized, obviously not in your foot, in the brain, in specific configurations. He compares it to the internal structure of a laptop: whatever programs it has are in there.

One of the things you can do with language, “a marginal thing in fact,” is externalize what is in your head. Most of your use of language, Chomsky thinks, is thought, internal thought. What the two of them are doing on the podcast is the exception: turning that inner system into noises in the atmosphere. You can call those noises language in some other sense of the word. They are not a rival theory of the faculty.

How deep do the roots of language go in the brain? Is it another feature like vision, or something more fundamental from which everything else springs? In a way, Chomsky says, it is like vision. Genetic endowment determines that we have a mammalian visual system rather than an insect one. The same kind of fact determines that we have a human language faculty. No other organism has anything remotely similar.

A long tradition, going back at least to the early scientific revolution, holds that language is the core of human cognitive nature: the mode for constructing thoughts and expressing them, with creative capacities that are free, independent, and unbounded. Fridman asks whether the capacity to think and reason is deeply linked with that internal system. Chomsky says language is “undoubtedly the mechanism by which we reason,” though other faculties are involved. We have a kind of scientific faculty, nobody knows what it is, that uses language but goes beyond it — just as using arithmetic is not the same as having the capacity.

Most scientists, Chomsky notes, believe human intelligence can answer any question in principle. He finds that a very strange belief. If we are biological organisms and not angels, our capacities ought to have interrelated scope and limits. The genetic endowment that gives you a mammalian visual system, arms, and legs also blocks other developments. There is no kind of experience which would yield the embryo an insect visual system, or wings instead of arms. If cognition had no built-in tendency to develop rich structure, we would understand nothing: we would be a random amoeboid creature with no shape at all. He thinks it is plausible that there are limits, and that we even have some evidence of where they are.

From Galileo to Newton, modern science treated the universe as a machine of gears and levers. Newton proved there is interaction without contact, a result he called “totally absurd,” and spent the rest of his life trying to evade. When scientists accepted it, they abandoned the search for an intelligible world and kept only the search for intelligible theories. Hume wrote that Newton left mysteries “which ever will remain.” Chomsky thinks they were basically right, and that this marks a limit of human cognition: our brains demand contact and perfect geometric objects, which is why Schrödinger’s cat is a theory we can understand about a world we cannot.

Fridman has just spent a day at Neuralink, which is trying to take thousands of readings from the brain, register what neurons are firing, and stimulate back — a two-way interface. The company’s dream, as he puts it, is to expand the brain’s capacity to take in information, to raise the bandwidth of something like a Google search. Could a machine expand our cognitive, linguistic, or reasoning capacity?

Chomsky’s answer is almost disappointingly old. It can be expanded in a certain sense, but a sense that was known thousands of years ago. “A book expands your cognitive capacity, okay, so this could expand it, too.” It is not a fundamental expansion. Nothing that goes beyond our native cognitive capacities, he says — just as you cannot turn the visual system into an insect system. Mapping what a bee sees into a form we can follow is something every bee scientist already does, without an implant.

Fridman tries quantum mechanics: maybe a mapping would let us make sense of a quantum world. Chomsky notes that students at MIT already study and understand quantum mechanics. They still reduce it. They do not really understand the world the theory describes. Schrödinger’s cat is the point again: we understand the theory, but as Schrödinger pointed out, it is not an intelligible world. Einstein remained skeptical of quantum theory because he wanted intelligibility — in that way, Fridman says, he has something in common with infants. A machine in the brain changes how much information we can reach. It does not change the shape of the faculty.

Fridman asks for the most beautiful or fascinating aspects of language from a lifetime of studying the mind. Chomsky names structure dependence: the deepest and most puzzling property of language that has been discovered. We now understand it pretty well, he says, but it was puzzling for a long time.

Take the sentence “the guy who fixed the car carefully packed his tools.” It is ambiguous: he could have fixed the car carefully, or packed his tools carefully. Front the adverb: “Carefully the guy who fixed the car packed his tools.” Now it is only carefully packed, not carefully fixed. Speakers do this even when it makes no sense. “Carefully the guy who fixed the car is tall” has to be read as carefully he’s tall, even though that is nonsense.

The puzzle is that you relate “carefully” not to the linearly closest verb but to the linearly more remote one. Linear closeness is an easy computation; what you actually do looks more complex. Look at the phrases, and you are picking out the structurally closest thing, which happens to be farther away in the sound stream. “What’s linear is 100% of what you hear. You never hear of structure.” Instantly, universally, across constructions and languages, we carry out the harder computation on material we never hear, and ignore 100 percent of what we hear and the simplest computation. There is now even a somewhat understood neural basis, and good theories — none of which, he says, explain why it is true.

Fridman turns to a decade of progress in neural-network machine learning and asks about the limits of deep learning. A real answer, Chomsky says, would require understanding the opaque processes inside the networks, so it is hard to prove a theorem about what can and cannot be done. Putting technicalities aside, deep learning takes huge numbers of examples and finds some patterns. That can be interesting. The question to ask is whether it is engineering or science: building something useful, or trying to understand something about the world.

Take a Google parser. Is it useful? Pretty useful. Chomsky uses Google Translate; on engineering grounds it is worth having, “like a bulldozer.” Does it tell you anything about human language? “Zero, nothing.” From the beginning it is “totally remote from science.” The parser takes an enormous text, say the *Wall Street Journal* corpus, and asks how close it can come to the right description of every sentence. Each sentence is essentially an experiment: am I grammatical? The answer is usually yes. Is there any science that takes random experiments carried out for no reason and tries to find something from them? A chemistry PhD student who proposed mixing a lot of things together with no purpose would be “laughed out of the department.” Science looks for critical experiments that answer a theoretical question. It does not care about covering millions of experiments.

The usual question is how well a parser does on a corpus. The question that is never asked is how well it does on something that violates all the rules of language. Suppose there were a language that used linear proximity as the mode of interpretation. Deep learning would work very easily on that — more easily than on an actual language. Is that a success? From a scientific point of view, it is a failure. It shows we are not discovering the nature of the system, because the model does as well or better on inputs that violate the system’s structure. That is not an argument against building such devices. They are useful.

Fridman hears echoes of the behaviorism debates. Chomsky says more than echoes: many people in deep learning claim they have vindicated B. F. Skinner. Terry Sejnowski, in a recent book, says this vindicates Skinnerian behaviorism. Chomsky’s reply is that it does not have anything to do with it. Could huge datasets still help us approximate the structure of language and aid the science? It is possible you find patterns you had not noticed. That is much like corpus linguistics: the speakers are dead, you have only records. It is better to have living speakers and run critical experiments, but if they are gone you work with what remains. Paleoanthropology is the same. You cannot run a critical experiment on what happened two million years ago. You can still learn things. It is a serious study. It is not a substitute for knowing the organism.

Fridman quotes Chomsky as saying that evil in society arises from institutions, not inherently from our nature, and asks whether most people are good or whether the capacity for intentional evil depends on upbringing and context. Chomsky first corrects the premise. He would not say these things do not arise from our nature. Anything we do arises from our nature, and the fact that we have certain institutions and not others is one mode in which human nature has expressed itself.

As far as we know, human nature could yield many different kinds of institutions. The particular ones that have developed have to do with historical contingency — who conquered whom. They are not rooted in our nature in the sense of being essential to it. It is commonly argued that market systems are just part of our nature; a huge amount of evidence says otherwise. There are all kinds of other structures. Markets are a particular fact of a moment of modern history. The opposite claim, from the roots of classical liberalism, is that an instinct for freedom — an instinct to be free of domination by illegitimate authority — is the core of our nature. We do not know. We just know that human nature can accommodate both kinds.

Fridman asks whether there is a moment in Chomsky’s intellectual or personal life he would love to relive. The answer is short: falling in love, having children. Pressed on the ideas in linguistics and cognitive science that first excited him, Chomsky says a discovery is exciting — even the observation of structure dependence, and then the explanation for it. But the major things just seem like common sense.

Go back to the 1950s: almost entirely, language was regarded as an external object, something outside the mind. It just seemed obvious that that cannot be true. Something about you determines that you are talking English, not Swahili. That is not really a discovery. It is an observation of what is transparent — closer to the 17th-century beginnings of modern science, which came from being willing to be puzzled about things that seemed obvious. A heavy ball of lead will fall faster than a light one: obvious, until Galileo, unimpressed, ran thought experiments that showed it cannot be true. Why does a ball fall instead of rising? Obvious, until you ask why steam rises.

The beginnings of modern linguistics in the 1950s were like that: willing to be puzzled about phenomena that looked, from some point of view, obvious. An almost official doctrine of structural linguistics then was that languages can differ from one another in arbitrary ways, each to be studied on its own without presuppositions. Biologists had similar views about organisms. In both domains it has been learned that this is very far from true. There are very narrow constraints on what could be an organism or what could be a language. That, Chomsky says, is just the nature of inquiry.

Fridman asks whether Chomsky ponders mortality, including his own. He used to, around age 12. He did not care much about his own death. He was worried that if his consciousness disappeared, the entire universe would disappear. That was frightening. Did he ever find an answer? “No, nobody’s ever found an answer, but I stopped being bothered by it.” He compares it to Woody Allen in one of his films, taken as a child to a shrink: the problem is that the universe is expanding, and he cannot handle that.

The last question is the meaning of our existence, our brief moment on Earth. That is something we answer by our own activities, Chomsky says. There is no general answer. We determine what the meaning of it is — meaning in the sense of significance, not in the sense that a chair means this. The significance of your life is something you create. Put beside the biological limits he spent the hour describing, the line is of a piece: the faculty has a shape you did not choose; inside that shape, what it is for is yours.