Chomsky, Noam

linguist, cognitive scientist, philosopher; Institute Professor Emeritus at MIT

Noam Chomsky: Language, Cognition, and Deep Learning

Lex Fridman Podcast · Nov 29, 2019

• Noam Chomsky
• universal grammar
• structure dependence
• internal vs external language
• limits of human cognition
• philosophy of science
• deep learning critique
• Skinnerian behaviorism
• quantum mechanics intelligibility
• human nature and institutions

Core ideas

Chomsky opens by distinguishing two entirely different senses of "language": an internal system (a specific configuration in the brain, part of an individual's biological endowment, that determines the sound-meaning pairings of an infinite number of possible expressions in "your" language) and the external act of using that system to communicate — noises produced in the atmosphere. He stresses these are not competing definitions but different concepts entirely, and that most actual use of the internal language faculty is for thought itself, with externalized communication being comparatively marginal. He argues language is not simply one cognitive module analogous to vision but is plausibly the generative core of human thought — the mechanism by which humans construct and reason with ideas — while acknowledging other faculties (a "scientific faculty" of unknown nature that lets us judge what makes sense and pursue certain lines of inquiry) also contribute to reasoning and extend beyond language itself, just as having arithmetic capacity is distinct from actively using it.

On the limits of human cognition, Chomsky argues — against what he says is the prevailing scientific consensus that human intelligence can in principle answer any question — that because humans are biological organisms rather than "angels," cognitive capacities should have built-in scope and limits just as physical traits do (a human embryo's genetic endowment that enables a rich, complex body plan simultaneously forecloses other possible body plans, like insect vision or wings). He offers a striking historical case for this: the shift from Galileo/Newton's mechanical philosophy — the founding assumption of the early scientific revolution, that all natural phenomena could ultimately be explained via contact-based mechanisms — to Newton's own discovery of gravitational action at a distance, which he and contemporaries first regarded as an absurd return to mysticism. Chomsky notes that Newton spent the rest of his life trying to explain away his own finding, and that philosophers like Hume and Locke recognized this moment as science permanently abandoning the goal of a fully "intelligible" world in favor of merely predictive, mathematically consistent theories — evidence, in Chomsky's reading, that human cognition has a built-in preference for (and possibly a hard limit around) mechanical, contact-based causal intuitions, further illustrated by infants perceiving distorted drawn triangles as "true" triangles and inferring invisible connecting causes between separately-moving objects. He extends this to quantum mechanics: physicists can learn to correctly use quantum theory's mathematics, but he argues (citing Einstein's and Schrödinger's own discomfort, via the Schrödinger's cat thought experiment) that the world quantum theory describes may be permanently unintelligible to human cognition in the deeper, mechanistic sense that satisfied Newton's contemporaries. On brain-computer interfaces (prompted by a Neuralink visit), Chomsky is skeptical that such technology could extend cognition beyond native human cognitive capacities in any fundamentally new way, comparing it to how a book already "expands" cognitive capacity — useful, but not a qualitative leap past the boundaries of what human cognitive architecture can natively represent, just as a device can map what a bee perceives for human study without granting a human bee-like perception itself.

Chomsky identifies "structure dependence" as the deepest and most puzzling property of language he's encountered: in a sentence like "the guy who fixed the car carefully packed his tools" (ambiguous between "fixed carefully" and "packed carefully"), moving "carefully" to the front of the sentence forces it to modify the more distant verb ("packed") rather than the linearly closer one ("fixed") — and speakers do this reflexively even in sentences where it produces nonsense, indicating the grammar operates on hierarchical syntactic structure rather than linear word order, despite linear order being the only thing actually present in the acoustic signal. He calls this computation over unheard structure, rather than the simpler linear computation over what's actually heard, a universal feature of human language with a partially understood neural basis but no accepted explanation for why it must be true.

Turning to deep learning, Chomsky draws a sharp distinction between engineering (building something useful) and science (explaining something true about the world), arguing large language-model-style systems like a Google parser are almost entirely the former: they achieve broad coverage over a corpus of "grammatical" sentences without any regard for whether the patterns found reflect the actual generative structure of human language, calling this scientifically comparable to a chemistry graduate student randomly mixing chemicals with no guiding hypothesis, whereas real science seeks targeted, structure-revealing "critical experiments." He argues the crucial, rarely-asked test of such a system is not how well it performs on real corpora but how well it performs on a hypothetical language that violates the actual rules of human syntax (e.g., one based on pure linear proximity rather than structural dependence) — since a system that would learn such a language equally well or better reveals it isn't capturing anything specific to how human language actually works, only pattern-matching whatever data it's given. He directly rejects the claim (which he attributes to figures like Terry Sejnowski) that deep learning's success "vindicates" Skinnerian behaviorism, while conceding some value to using such models the way corpus linguistics or anthropology use fixed historical records when direct experimentation is impossible — a legitimate but scientifically weaker substitute for controlled inquiry.

Asked whether human evil arises from institutions or human nature, Chomsky rejects the framing that these are separable: any institution humans build is itself an expression of human nature, but our nature is broad enough to accommodate radically different institutional forms (market systems being a contingent historical outcome, not an inevitable expression of human nature, just as classical liberal thought has argued for an opposing "instinct for freedom" from illegitimate authority) — meaning existing institutions reflect historical contingency rather than a fixed human essence, and we simply don't yet know which of these tendencies, if either, is more fundamental. Reflecting on his own intellectual history, he frames his foundational contributions less as sudden discoveries than as a willingness to be puzzled by things that seemed obvious (comparing this to Galileo questioning the "obvious" fact that heavier objects fall faster), citing 1950s structural linguistics' assumption that languages could differ arbitrarily from one another as a comparable false "obvious" belief later overturned by finding narrow universal constraints on possible human languages. He closes reflecting that he stopped being troubled by mortality after childhood worries about whether the universe itself would disappear with his consciousness, and states that life's meaning or significance is not something answered in general but something each person determines through their own actions.

Excerpts