For years, learning English meant having words and sentences arrive before their meanings did. I could recognize kernel and shell as parts of an operating system, and switch, gateway, and bridge as pieces of a network. The same words belonged to ordinary life, but that second world did not come with them. A sentence could be grammatically sound and still feel like a building whose rooms I had never entered.

LLMs bring back that unease. They produce fluent language and handle technical terms with ease, and the fluency can seem to float free of the world the words name, as if the model had learned the architecture of the language without standing among the things it describes. This is not a claim that a model knows nothing, or that its limits match those of a person learning a second language. The resemblance is narrower: in both cases, the visible command of words can arrive before anyone can be confident that meaning is present.

I. What the Metaphor Leaves Out

Technical English makes the gap easy to see because its vocabulary is built from fossils. Kernel comes from the Old English cyrnel, a little grain of corn. Shell was coined by Louis Pouzin at the end of 1964 or the beginning of 1965, for the command language of Multics, and the two words together make a nut: soft center, hard outside. None of that helps you find a bug. A kernel is not a seed one can hold, and a computer shell is no animal’s discarded home.

Bridge is the instructive case, because it keeps more of the old meaning and then betrays it. A network bridge joins two segments, much as a stone bridge joins two banks. But a town is better off with two bridges over its river, and a network that connects two segments through two bridges can bring itself down: frames circle the loop and multiply until the traffic drowns everything else. Radia Perlman designed the Spanning Tree Protocol in 1985 to solve this, by shutting off the redundant paths until they are needed. No river teaches that. To learn the vocabulary is to learn which distinctions a community makes. To understand it is also to know what those distinctions do when something goes wrong.

II. The Company a Word Keeps

Much of what I knew about those words came from the other words around them. The bridge forwards the frame. The gateway routes the packet. Linguists had a theory of this long before anyone built a machine on it. Zellig Harris argued in 1954 that items with similar distributions tend to have similar meanings, and in 1957 J. R. Firth gave the idea its slogan: “You shall know a word by the company it keeps.” A language model is that slogan at industrial scale.

Wittgenstein would have asked what the company was doing. His builders in the Philosophical Investigations have four words, and when one calls “slab!” the other carries a slab. The word’s meaning sits inside the work; speaking is “part of an activity, or a form of life.” A manual gives you the word’s neighbors. The job gives you the slab.

Stevan Harnad turned that difference into a thought experiment. In “The Symbol Grounding Problem” (1990) he asks you to learn Chinese as a second language with only a Chinese/Chinese dictionary. That is hard, but you have a world to attach the words to. Now learn Chinese as a first language from the same dictionary. That, he thought, is impossible: a merry-go-round, “passing endlessly from one meaningless symbol or symbol-string” to another. Here the analogy with my English breaks. I was on the difficult version of the ride. I already knew, in another language, what a bridge and a seed were.

III. The Octopus and the Elm

The strongest modern statement of the skeptical case is Emily Bender and Alexander Koller’s octopus. Two people stranded on separate islands type to each other over a telegraph cable. A deep-sea octopus taps the line, learns the statistics, cuts the cable, and impersonates one of them. Small talk goes well. Then one islander, chased by a bear, asks how to make a weapon from two sticks, and the octopus has nothing, because it has never seen a stick. “Having only form available as training data, O did not learn meaning.”

The parable is good, and it is about an octopus who overheard two people. A large model has read something closer to a library, including what people who met bears wrote down, and a current one will give you usable advice. Its text was not produced by a random process. It was produced by people looking at things. That is why the easy conclusion, that a model’s fluency is empty, won’t hold.

Hilary Putnam offers a better description. In “The Meaning of ‘Meaning’” (1975) he admits he cannot tell an elm from a beech, yet when he says elm he refers to elms, because other speakers can tell them apart and his word leans on theirs. He called this the division of linguistic labor. For years my kernel ran on that credit: I used the word correctly and the people who had read the source code held the collateral. Whether a machine can borrow on the same terms is an open dispute, and I don’t want to settle it by analogy. But borrowing has a condition Putnam takes for granted: he knows there is a botanist, and he knows when to go and ask. Often the sign of a learner’s understanding is not an answer but the moment they say they would have to check.

IV. The Right to Say Green

The sharpest warning against simple verdicts comes from someone who was told she had no right to her vocabulary. Critics of Helen Keller assumed, in her words, that “we have no moral right to talk about beauty, the skies, mountains, the song of birds, and colours.” Educators later had a name for this kind of speech, verbalism: words used without the sensory experience behind them. Keller’s reply in The World I Live In (1908) is one of the best accounts of second-hand meaning ever written. The blind person, she says, “guesses, divines, puzzles out their meaning by analogies drawn from the senses he has.” And then: “I have talked so much and read so much about colours that through no will of my own I attach meanings to them.”

Barbara Landau and Lila Gleitman found the same capacity in a child. Kelli, blind from birth, understood look as an instruction to explore with her hands. She was not copying a sighted meaning, and she was not repeating noise. The word fit the world she had.

So language travels a long way past perception, and the critics were wrong. But notice where Keller’s journey starts. In The Story of My Life her teacher holds her hand under a spout and spells a word into the other, and she knows that “w-a-t-e-r” means “the wonderful cool something that was flowing over my hand.” Her analogies are drawn “from the senses he has.” A model’s analogies are drawn from text. That difference should not be casually mapped onto human experience, in either direction: it neither proves the model’s words are empty nor grants them what Keller had at the pump.

V. Words That Fail

Understands turns out to be several exams. Knowing a word’s definition, using its syntax, applying it correctly, having met the thing, acting on it when the situation changes: a person or a machine can pass some and fail others, and fluency only shows the first three.

J. L. Austin noticed in How to Do Things with Words that some sentences act rather than describe. “I name this ship” names the ship. When such a sentence goes wrong it is not false but, in his term, unhappy. Technical English is full of these. kill kills a process. mount mounts a disk. A shell is the place where a sentence meets a machine that does not grade on plausibility, and that gives the fifth exam an odd twist. The register that felt emptiest to me, the one with no river in it, is where a model’s words can most easily be tested. Connected to a terminal, it can run the command, read the error, and try again. It gets something distribution alone never gave the octopus: a reply from outside its own text.

That is not a pump. The reply arrives as more text, a passing test proves less than it seems, and the Chinese Room is still in the building, and someone has to read what the machine did. But the useful question about a fluent sentence has changed. It is no longer whether anything lies behind the words, which nobody can inspect. It is what happens if the sentence is wrong, and who finds out.

Two bridges in parallel, with no spanning tree, will take a network down fast enough to get the person on call out of bed, and from then on that person knows what the word means. Fluency can’t teach that. It can only describe it, sometimes very well, which is how I learned it the first time.

VI. The Language Nobody Teaches

There is a harder version of this story, and it needs no machine. Suppose you already speak the language. You passed the exam, you read the manuals, you do your job in it. Then you move to the country where it lives, and you don’t go to school there, you don’t make close friends, and you don’t marry anyone. You can explain a routing failure to a full room and still not know what to say to the pharmacist, why the table laughed, or whether a neighbor’s we should have you over sometime is an invitation.

Jim Cummins described the usual order among schoolchildren: conversational language in about two years, academic language in five to seven. The adult I am describing has it backward, and may never finish. Technical language can be learned alone, because it is written down and it holds still. The daily language is stored mostly in other people.

The research has a bleak word for what comes next. Larry Selinker called it fossilization in 1972: the learner’s language stops developing short of the target, however long the stay. John Schumann’s Alberto, a Costa Rican immigrant in the United States, barely advanced in English while the other learners in the study moved on, and Schumann looked for the cause in social distance, not in ability. Bonny Norton followed immigrant women in Canada and found the same thing from the other side. Eva, the only worker at her fast-food restaurant without English as a first language, was surrounded by it, and her colleagues rarely spoke to her. Exposure is not access. A language can fill the street and still not be addressed to you.

Think about what a school, a friend, and a marriage have in common. Each obliges someone to keep talking to you after the business is finished. They tease you, finish your sentence wrongly, correct you, and stay in the room when you fail. The teller and the pharmacist do none of this. They understood you well enough, the transaction closed, and the mistake goes home with you intact, to be repeated for thirty years. This is Wittgenstein’s building site seen from the fence: you can hear “slab!” every day and never be handed one.

Agota Kristof fled Hungary in 1956, at twenty-one, and landed in a factory in Neuchâtel with a husband and a baby, surrounded by a French that belonged to other people. After five years she left the factory, went to study the language, and became a celebrated novelist in it. Even that was not enough. In L’Analphabète, three decades on, she wrote: “I have been speaking French for more than thirty years, I have been writing it for twenty, but I still do not know it.” She called it an enemy language.

A model can read what was said in every kitchen. A person can live forty years a few meters from one. Neither arrangement gets anyone inside, because the rooms of a language open from the inside, and someone there has to want you in.

Further reading