The Pirahã Have No Word for One

Edward Gibson, known to everyone as Ted, is a psycholinguistics professor at MIT and head of the MIT Language Lab. He tells Lex Fridman that the fascination started in school, when English grammar felt like a puzzle with no stated theory. After a computer-science and math degree at Queen's University in Canada, he spent a year in Cambridge on a computational-linguistics master's, having never taken a language class. He stayed for life. An AI course around 1984 had a natural-language section that struck him as a pile of hacks, which is part of why the problem still looked open.

He treats syntax — the *forms* of sentences — as the tractable engineering problem, and meaning as the hard one that is still unsolved. That split, he says, is also why large language models look so good: they are excellent at form. Fridman prefers a romantic picture in which form is the shadow of thought. Gibson does not. What he finds beautiful in human language is the generalizations that hold within a language and across languages.

English is subject-verb-object: *the dog chased the cat*. About 40 percent of the world's languages look like that, and they tend to have prepositions — *talk about a book*. Japanese and Hindi are verb-final: *Mary ball kicked*. Those are perhaps 45 to 50 percent of languages, and they tend to have postpositions — *a book about*, with the verb at the end. Subjects usually come first in both families. Of roughly 7,000 languages on Earth, researchers have decent word-order data on about 1,000, and about 95 percent fit this harmonic pattern. Mix the settings and dependency links get longer. Stanford typologist Joseph Greenberg noticed the alignments long ago without the processing explanation Gibson prefers: speakers and listeners are trying to minimize the distance between words that depend on each other.

Gibson divides language structure into sounds, words, and the combinations of words — grammar, or syntax. In any utterance, he says, every word connects to just one other word, which yields a tree: a root, usually the verb of the event, and leaves that are the words. Linguists argue about the details of those trees, not about the trees themselves. He prefers dependency grammar, a lineage he traces from the ancient Indian grammarian Panini through Lucien Tesnière, because it makes word-to-word links visible. Phrase-structure grammar can generate the same sentences, but the distances hide inside nested nodes.

Take *two dogs entered a room*. The root is *entered*; *dogs* hangs off it, *two* hangs off *dogs*; *room* hangs off the verb, and *a* hangs off *room*. Word categories are decided by use, not by meaning: verbs typically mark events and states, nouns typically mark people, places, and things, though nouns can name events too. Poetry still has to be recognizable as the language. You cannot say *dogs two entered room a* in English. Russian is freer because case markers on nouns signal agent and patient, so word order can move without losing who did what to whom. Subject and object are positional terms; agent and patient are meaning terms.

Morphemes are the minimal meaning units. English usually has one or two per word — *eats* is a verb root plus a third-person singular ending — and a little suffixing on nouns and verbs. Finnish can stack around ten morphemes on a root and produce millions of forms of a given word. High-frequency English verbs tend to be irregular: *drink/drank* versus *walk/walked*. Why some languages load morphology and others do not is, Gibson says, an open question. His working view is simple and probably right: language is an invented system for communicating ideas. People label what they need to talk about, then combine those labels in ways that are easy to produce and understand. There are many good solutions to that problem.

Researchers have snapshots of a few thousand living languages, not movies of how they evolve, and contact scrambles the picture: groups borrow whatever is useful. Color is Gibson's favorite illustration. A person with ordinary trichromatic vision can distinguish millions of colors, yet English speakers share about eleven everyday color words — black, white, red, blue, green, yellow, purple, gray, pink, orange, brown. The Dani of Papua New Guinea have two, roughly black and white, and mostly use them for very dark and very light things rather than splitting the whole space. The Tsimane of Bolivia, a group Gibson has worked with, have three to seven, and the three everyone knows are black, white, and red.

If a language has only three color terms, the third is always red. After that, languages diverge: some add a blue-green category, some add yellow. Gibson's guess is need, not vision. You invent a color word when two otherwise identical objects differ only in color and you have to ask for one of them — the red sweater, not the green one. Industrial societies dye almost everything arbitrarily, so the words are a necessity. Rainforest groups see brilliant flowers and berries and like bright colors just as much; they simply do not need a large palette of labels. The same logic, he thinks, will show up in number words later in the conversation.

Most of the 7,000 languages lack a writing system, so the evolutionary record is thin: Chinese and English are the main long traces. Old English was verb-final with case marking, more like German; contact with French helped shift it to a verb-medial language without case. Even Queen Elizabeth II's vowels moved between her 1950s coronation and fifty years later. Languages change all the time. What researchers do not have is a clean causal film of why a given language took the morphological or lexical path it did.

Noam Chomsky, Gibson says, is the most famous linguist ever, and in the late 1950s he developed formal language theory and phrase-structure grammar: rewrite rules such as S → NP VP. Context-free grammars of that kind generate the same sentences as a dependency grammar; regular languages are too simple for human languages because they cannot handle the center-embedded recursion people actually use. The real split is not phrase structure versus dependencies. It is Chomsky's extra idea of movement.

From *two dogs will enter the room* English forms the yes/no question *will two dogs enter the room?* Chomsky treats the declarative as basic and moves the auxiliary to the front, with similar movement in passives and *wh*-questions, and two representations — deep structure and surface structure — in between. Gibson's alternative, long associated with Ivan Sag and others, is lexical copying: *will* has a declarative entry and an interrogative entry with different argument structures. Nothing moves. Chomsky showed in 1971 that phrase structure plus movement creates a learnability nightmare, with too many equivalent underlying analyses, and concluded that movement must be innate — universal grammar, with children only setting parameters. Drop movement, and children can learn rules one by one from the input.

Not every auxiliary behaves. *Aren't I invited?* is fine; *I aren't invited* is not, because that interrogative has no declarative twin. *I ought to do this* sounds all right to Gibson; *ought I?* sounds ridiculous. Lexical copying predicts exactly that messiness: you learn the usages a word actually has. Chomsky, Gibson argues, talks about combinations of words and almost never about words themselves, even though a common verb like *take* may have thirty or forty senses in a dictionary. When Gibson is scanned while listening to language, the same left-lateralized network responds to words and to combinations; the technology's time resolution is seconds, so finer splits remain open, but the clean cut Chomsky wanted is not sitting there in the voxels.

Gibson has been at MIT since 1993, thirty-one years at the time of the recording, across the street from Chomsky's linguistics department, in Brain and Cognitive Sciences — MIT's old psychology department. The deepest difference, he says, is method. He runs experiments and corpus analyses and evaluates theories quantitatively. Chomsky has never been associated with an experiment or a corpus study; the work is thought experiments and intuitions. That, not a taste in formalism, is why they land on different theories of the same English auxiliaries.

The reason Gibson likes dependency grammar is blunt: the farther apart two connected words are, the harder the sentence is to produce and to understand. Start with *the boy cried*. Add a relative clause — *the boy who the cat scratched cried* — and it is still English. Nest again — *the boy who the cat which the dog chased scratched cried* — and it collapses. That pattern is center embedding, also called nesting: a modifier stuffed between a head and the later word it depends on. It is grammatical. Native speakers still cannot handle two layers, and three is effectively impossible, in every language Gibson knows that allows the structure.

The behavioral tests are simple and stable. Ask people how natural a sentence sounds, and the ratings agree. Give them a fragment such as *the book which the author who…* and ask for a completion; the clean ending is *who I met wrote was good*, three verbs. On crowdsourcing platforms about 60 percent of people drop a verb and supply only two. Reading times, production attempts, and spoken tasks point the same way, even though speech is messier to score. The cost is not a quirk of English. Do the same nesting in any language and people will tell you it is not really their language, even while agreeing that each piece is legal.

Richard Futrell, then Gibson's student and now at UC Irvine, tested the short-dependency claim on parsed corpora. There were about forty languages at first, around sixty now. For every sentence he built the real dependency tree, then scrambled the dependents a hundred times — first freely, then only in projective, non-crossing orders, which is what languages mostly do. In every language the real sentences had shorter dependencies than the scrambled controls. It did not matter whether the language was verb-final, verb-medial like English, or even verb-initial VSO, a minority of about 10 percent. Human languages are pushing related words together.

Gibson thinks that pressure is mostly about production: it is easier to keep closely linked words in mind while speaking, and that happens to help the listener too. Hold the message constant — the same event of two dogs entering a room — and there are many forms you might choose. He wants to explain the form. Meaning, he keeps repeating, is a harder problem, which is why large language models can look fluent without understanding, and why language as a communication system should not be confused with thinking.

Eve Fedorenko, a neuroscientist in Gibson's department and his wife, can localize a person's language network in about fifteen minutes. Listen to a language you know versus one you do not, or versus backward speech; or read sentences versus nonsense strings; subtract, and the leftover voxels are a stable left-lateralized network. Gibson has been scanned since around 2007. His network, he says, is like his face: the location and shape barely change. When it stabilizes in childhood is still hard to know, because toddlers will not lie still in a loud scanner.

The same network handles listening, reading, and — with harder tasks to run in the magnet — production. It handles English and Russian in a bilingual, and it handles constructed languages such as Klingon and the *Game of Thrones* tongues, because those systems were built to express arbitrary thoughts the way human languages do. Lewis Carroll's *Jabberwocky* is an in-between case: English function words and syntax with invented stems, so activation sits below real English and above pure noise. The more language-like the input, the brighter the network.

What does not light it up is just as important. Spatial memory, music without lyrics, chess, math, and programming — even in skilled programmers — activate other systems, including a bilateral multiple-demands network for hard tasks, and show no overlap with language. Code, Gibson says, is not conveying the same information as human language. Fedorenko's conclusion is the one Fridman has trouble loading: we do not think in language. The language network is a conventional communication system translating between thought and form. What thoughts themselves are, Gibson does not claim to know.

Internet polls find that about 70 to 80 percent of people say they hear themselves talking while they think. Gibson does not. He long assumed *inner voice* was a metaphor. Fridman also denies having one, and refuses to believe it is the majority; Gibson laughs that the classroom show of hands always comes out the same, and that self-report bias is possible. What the inner voice has to do with the language network is unknown. It sits on the table as an unsolved curiosity, not as a theory.

The harder evidence is clinical. Some people suffer a massive left-hemisphere stroke that wipes out the language network and leaves the right side intact. They have global aphasia: they cannot speak, understand, write, or read. Rosemary Varley at University College London has followed patients like this and found they can still play chess, drive, do arithmetic, and match age-matched controls on non-language cognitive tasks. Symbolic processing is not the same as a conventionalized language system. Math uses symbols; it does not live in the language network.

The implication, Gibson says, is heavy. Language is not required for thought. You *can* think in language, because it is an expressive tool, but you do not need it. That undercuts a core Chomsky claim: that language is the foundation of human thought. Gibson is polite about the idea and blunt about the method. You cannot test it with thought experiments. You need a scanner, and you need patients. The fMRI dissociation is, in his view, unambiguous: the networks do not overlap.

Gibson has written that the best current theories of human language are arguably large language models. He immediately loosens the claim. A model with billions of parameters is a lousy scientific theory if theory means a small story that explains a lot. It is a good theory in one sense: nothing else predicts as accurately which strings are English and which are not. Chris Manning's group at Stanford, probing BERT and related models, found internal representations that look like dependency structure — trees of word-to-word links, the thing Gibson's theory says language is built from.

The linguistic family LLMs sit closest to is construction grammar: language as a large inventory of form-meaning pairs, from morphemes through collocations to sentence patterns, studied as people actually speak and write. Gibson would formalize those constructions as dependency grammar. Do the models understand? He thinks they are doing form, extremely well, and not meaning. The demonstration he likes is a twist on the Monty Hall problem. In the classic version you should switch doors, raising the chance of the car from one third to two thirds. Tell the model the car is behind door one, you pick door one, Monty opens door three onto a goat, and ask whether to switch. It still says yes, walking through the familiar explanation, with no way to keep hold of the 100 percent already stated. A person, once you explain that, does not make that error.

Fridman objects that this may be overfitting to a famous riddle, and that more scale could look like a world model. Gibson does not deny that the systems are impressive. His inference is narrower: this is not what humans do. On center embedding the similarity runs the other way. Ask people to complete *the book which the author who…* and they fail; ask an LLM and it fails too, sometimes by saying the sentence is too hard to complete, even though nobody trained it on a difficulty label. That, for Gibson, is a feature. He wants a model of humans. Superhuman completions would mean it was not simulating how people process form. His former student Kyle Mahowald, now at the University of Texas at Austin, is first author with Fedorenko on a paper that draws the same line: language and thought are two systems; models have learned language and still lack thought.

Dependency length has a cognitive cost you can measure in several ways: naturalness ratings, controlled reading times, and, in a recent study, fMRI. People lying in a scanner listening to stories show stronger language-network activation as dependencies get longer. Gibson once tried to fit a function and now guesses an exponential: a single long subject-verb link padded with extra phrases is tolerable, but multiple nested links go south, probably because retrieval of earlier words suffers interference from similar material in between. Counting words is a crude proxy, not the true metric, and he is a little puzzled that more people have not wanted to write the equation.

Natural languages keep those links local. Legalese is the exception. Eric Martinez, a Harvard law student with undergraduate linguistics who took Gibson's psycholinguistics lab class, wanted to know why legal language is so hard. Nixon signed a Plain Language Act in 1970; Obama signed another; presidents keep asking for simpler law. Without a diagnosis there is no fix. The team coded random contracts for features. About 70 percent of sentences had a clause wedged between subject and verb, against 20 to 30 percent in other genres, including academic prose. Low-frequency words were even denser than in academic writing. Passive voice was more common too.

Then they tore the features apart. Passive versus active: zero effect on comprehension or recall. Low-frequency words: a little hurt. Center embedding: it slowed people down, wrecked understanding, and wrecked recall. The same pattern held in a hundred lawyers from a range of firms. Lawyers outperformed laypeople, as training would predict, but the main effect of nesting was the same size. They also preferred unpacked versions of the same contracts and said they would hire the person who wrote that way. One real sentence stuffed a definition of *payments and benefits* between the subject and *would be subject to the excise tax*. Gibson's prescription is almost rude: put the definition somewhere else.

Fridman suspects an industry-level incentive: unreadability keeps clients dependent on lawyers, even if no individual admits it. Gibson takes the suspicion seriously and says he cannot find evidence for the simple conspiracy. Among the hundred lawyers, nobody liked the nested style. His milder guess is the magic spell hypothesis: people putting a spell on something reach for rhyme and ritual, and center embedding may be a syntactic costume that says this is binding, not chat. He is clear that the hypothesis is untested. Researchers have a snapshot of current legal language, not a film of how the style grew, and no clean experiment on its cultural function.

Claude Shannon's master's thesis around 1948 framed communication as getting a message from one head to another through a noisy channel. He cared about language. Noise, Gibson says, is at least three different things and should not be stuffed into one term: the speaker's slips and revisions; environmental noise, including a loud party; and listener-side problems such as fatigue, drink, distraction, or a shaky command of the language. One speculative idea is that some word-order facts are robustness tricks, placing material where context can recover it if a piece is lost. Gibson calls this shaky territory, a just-so story. Dependency length he will defend; why order looks exactly as it does, he will not. Shannon could not get traction in a field moving toward Chomsky and went to Bell Labs to do mathematical communication theory instead. Had he stayed with language, Gibson thinks, the field might look different.

If languages were only optimizing dependency length, word order would be almost free: put related words next to each other and ignore the rest. No language actually does that. English is especially rigid; Russian is freer, not free. Regularity, Gibson guesses, is there because it is easier to learn. Every language he knows is a noisy solution to a complex optimization that includes both communication and learning. None is better than the others at keeping dependencies short.

The public list of "hard languages" is a second-language list. The U.S. Defense Language Institute ranks Chinese, Japanese, and Arabic as brutal for English speakers. Gibson says the difficulty is distance from the first language, not an inner complexity score. Babies do not show it. There is no evidence that children acquiring one language start talking or master grammar later than children acquiring another. By three or four, children are fluent in whatever surrounds them. Human languages are equal in learnability.

Chomsky went innate, Gibson thinks, because movement is hard to learn. Throw out movement, and the messiness of actual words is not a bug. Large language models, trained on form, are further evidence that the forms of human language can be learned from input. The language network's modularity does not settle innateness. The visual word form area, near visual cortex, specializes in written words in readers and does not exist in people who never learned to read; it grows with literacy. Some children lose left-hemisphere tissue to disease or surgery and develop language on the right, behaving so normally that the missing tissue is discovered only by accident. Fedorenko collects these cases under the name Interesting Brains. Chomsky's claim that language underpins thought, Gibson repeats, is a philosophical position his methods cannot test. Scanners and patients can.

Joe Henrich's acronym WEIRD — Western, educated, industrialized, rich, democratic — is a warning not to theorize about humans from MIT undergraduates. Gibson has worked with two Amazonian isolate languages: the Tsimane of Bolivia, farmer-foragers, and the Pirahã of Brazil, with the linguist Dan Everett, who first went as a missionary in the 1970s. Isolates cluster in the Amazon and Papua New Guinea, places stable enough that groups can live for centuries with little contact. Gibson went for the words. Naively, he had assumed every human language would have exact counting. The Pirahã do not. There is no word for one, so none for two, three, or four. You cannot ask for two of those. The language has no words for it.

Earlier work treated three Pirahã quantifiers as *one, two, many*. With Everett translating and student Mike Frank running the tasks, Gibson lined up identical spools of thread. Going up from one, speakers used one word through about four items, another around five, a third beyond that. Going down from ten, they switched to the middle word around eight and to the small-set word at four. The words mean *few, some, many*, and the thresholds move with context. Matching balloons to visible spools was essentially perfect, no worse than MIT undergraduates, because one-to-one correspondence is not counting. Hide the spools behind an opaque sheet and performance stays good at two or three, then turns approximate after four or five. Four is the subitizing range: vision can see small sets without counting. Beyond that you need symbols. No number words, no exact count. That is a tool limit, not an intelligence limit.

The usual anthropological guess is that farming creates the need: seventeen goats at night, seventeen in the morning, an abstraction that does not care which seventeen. People ask how Pirahã parents track their children without numbers. Gibson asks whether the questioner has children. You track identities, not a headcount; if one child were replaced by another, you would notice. A goat is more substitutable. Color follows the same invention logic in the jungle: no two factory-dyed sweaters, no need for teal.

Why didn't Esperanto take over? Gibson's answer is utilitarian: a language survives when a community uses it to get things done. The Tsimane's relatives, the Moseten, have more contact with the Spanish-speaking world, and their isolate language is dying because Spanish is how you get a job and feed a family. English is popular because of the size of the American economy, not because it is especially easy. Mandarin and Spanish ride the same incentive. The world is already losing languages. Fridman poses the counterforce: identity. In Ukraine, language is itself a political claim. Gibson agrees completely. Economics pulls toward a few big languages; identity pulls toward keeping small ones. As a linguist he hopes identity wins more often, because each death closes a window on human cognition.

Machine translation, he thinks, can probably map form. Concepts are the hard ceiling. You cannot faithfully put a contract full of exact numbers into Pirahã; the target language has no such concepts, and "a lot more than a hand" is a rewrite that throws away the point. The reverse is also true: Pirahã almost certainly has distinctions English can only describe, tied to a rainforest life Gibson does not live. Color is milder; you can say "the color of blood." Number is not. Style is another unsolved layer. Fridman wants average dependency length by author and bets Hemingway is the shortest. Gibson says the measurements are doable and admits he has never gotten anyone to do them.

Aza Raskin's Earth Species Project tries to decode whales and crows with machine learning and reports similar structure in embedding space. Gibson does not find that crazy. The claim that only human language has compositionality, often associated with Chomsky, fails because the claimants do not speak whale, crow, or monkey. A whale listening to us would also hear weird noise. Lower species with small brains may well have simple systems; whales, elephants, and some primates have complex social lives and large brains, and there is no good reason to assume their signals are simple. Neri Oxman's work on plant chemical signaling sits even farther out. Gibson is open and cautious: they may be talking to conspecifics, not to us. His floor is clear. Until we decode those systems, asserting that they lack complexity is converting ignorance into a finding. Human uniqueness, on this point, is under-argued.

If an alien civilization arrived, Gibson would bring Dan Everett — the former missionary who learned Pirahã from scratch by throwing down a stick and asking what it was, then two sticks, and discovering there were no count words. The method is socially extreme: you have to be willing to gesture through misunderstanding for a long time. Whether it would work on aliens depends on how much they want to talk, but the ladder from objects upward is the one Gibson would try. Asked for advice to young people, he does not offer a five-year plan. "When you see something interesting, just go and do it." He said yes to the Pirahã, and when corrupt Brazilian officials made return impossible he found the Tsimane and went anyway. Try something other people have not done. Fridman closes with Wittgenstein: the limits of my language mean the limits of my world. Gibson's evidence recasts the line. Language is not the limit of thought. It is the limit of what you can point to for someone else. Without *seventeen*, you cannot say seventeen goats. The world is still there.