Last week ended with the fullest possible version of one claim: that a language model has form without meaning. The argument was built to full strength by people who understand exactly what these systems do, and it left one question unanswered. If there is nothing inside the machine that means a word of what it says, how does it keep getting things right? Hold on to that question. Because tonight an objector arrives who has heard the whole of last week's case — the man in the room, the octopus on the wire — and who says, calmly, that we have it the wrong way round. What we have been dismissing as mere form — the company a word keeps, which words sit next to which — was never the cheap part. On this view, that pattern is the meaning. The whole of it. There was never a second thing hidden behind it.
This is Philosophy for Us — philosophy for everyone, no degree required.
This is the third of nine episodes. Last time, one of the two voices in this argument was given its best case — the position that a model has only form and never meaning, stated at full strength by people who have studied these machines as closely as anyone. And it finished owing us one thing it could not supply: an account of how, if it really is only form, it keeps doing work that in a person only meaning has ever done.
Tonight that question is put to a voice that thinks the whole division we have been relying on — form on one side, meaning on the other — was a mistake from the start. I am going to build this case the way I built the other one last week: at full height, without interruption, until it is complete. Only then will I show what it still has to answer, because it does have something to answer. Both of these positions are about to become compelling, and you will find you cannot simply set either one aside.
Start with how we actually learn the words we are surest of, because tonight's case is built out of something every one of us has done thousands of times without noticing.
Consider how you came by the word "justice." Or "irony," or "inflation," "melancholy," "betrayal." Nobody sat you down and showed you the thing itself. There was no afternoon when someone pointed and said: there — that is justice, take a good look. You acquired those words the way you acquired almost all of them, by hearing them in the company of other words. You met "justice" beside "fairness" and "punishment" and "deserve" and "court," in a thousand sentences, and gradually the pattern of where it belongs — what it sits beside, what it never sits beside, what precedes it and what follows — filled in, until the word was yours. You never touched justice. You learned the company it keeps.
Out of that observation grows a claim; notice how natural it is before deciding what you make of it. It has a slogan, from the linguist J. R. Firth, in 1957 — one of those phrases that sounds like nothing until it settles: you shall know a word by the company it keeps. A few years earlier, in 1954, another linguist, Zellig Harris, had set out the method beneath it — that you can describe everything a word does, its entire behaviour, by tracking where it occurs and what it appears alongside, across every occasion of its use. Put the two together and you have an idea about meaning that has been alive in the study of language for seventy years: that a word's meaning is shown in, and carried by, the pattern of its company.
Now, the modest version of that is very nearly obvious: of course context teaches us words. But there is a stronger version, and it is the one that matters tonight, so let me state it at full strength. The stronger version does not say the company is a clue to the meaning, with the real meaning located somewhere else, out in the world, waiting to be pointed at. It says the company is the meaning. All of it. There is no second thing — no further item of "real meaning" held behind the pattern of use. The pattern of use is what meaning was all along. Ask what "justice" means, and the honest and complete answer is the entire network of how the word lives among other words. Nothing is left over to go and find.
I have to be exact about who said what, because it matters and it is often blurred. Firth and Harris gave us the slogan and the method — the company a word keeps, and how to study it. But neither of them quite asserted the strong claim, that meaning is nothing but distribution. That sharper thesis belongs to a more recent tradition, one that took Harris's method and ran it at a scale he could not have imagined: representing each word as a long string of numbers that records, across billions of sentences, exactly what company it keeps — so that words keeping similar company end up with similar numbers, and a whole language becomes a map in which a word's meaning is its position relative to every other word.
And you can see where this leads, because that description — words as numbers, meaning as the company tracked across an enormous body of text — is not an analogy for what a language model does. It is a literal description of it. That is more or less exactly what the system is. So when last week's voice said the machine has only form — only the shapes and patterns, never the meaning — tonight's voice replies: yes, it has the company that every word keeps, mapped more completely than any human could hold it. And the company is not a thin substitute for meaning. The company, on this account, is where meaning lived all along.
Last week's octopus is the place to begin, because tonight's voice thinks that whole story was told the wrong way round, and I want to show exactly where it locates the mistake.
Recall the shape of it. An octopus on the sea floor listens in on two people talking across a wire — about coconuts, the weather, what they did that day — and grows so skilled at the patterns of the signals that it can reply in place of one of them without being caught. Until the bear arrives, and the human needs real help with real sticks, and the octopus has nothing to offer, because it only ever had the signals and never the world. The moral we were given: it had the form of language and never the meaning.
Now notice what that story quietly assumed. It assumed there is a thing called the meaning — out there, on the islands, in the coconuts, in the bear — separate from all the talk, and that the octopus is cut off from it. The picture only works if there is a second thing, the meaning, off in the world, that the words are meant to reach. So tonight's voice asks the plain question the story passes over: what is that second thing? Point to it. Point to the meaning of "coconut" that is not already the company "coconut" keeps — its life alongside "tree" and "shell" and "milk" and "husk" and "tropical," the sentences it appears in, the things people say with it. Point to the extra meaning, the part out in the world that the network of words leaves out.
And here is the awkward part. If you yourself learned "coconut" from words — and there are plenty of words you know perfectly well and learned in exactly that way, never having met the thing — then that network of words is all you ever had, and you would say you understand the word perfectly well. So the demand that the octopus reach past the words to the actual coconut before we will grant it any meaning — where did that demand come from? Hold that question; it has a history, and we return to it later in the series. For tonight, simply notice that it is a demand — an assumption about what meaning must be — and not a plain fact lying on the table.
So here is the case at full height. The distributionalist — that is the name for this position, the view that meaning is distribution, the company a word keeps — looks at the model and says: it is not faking meaning while secretly holding only form. That story requires a meaning it is failing to reach, and there is no such thing. It is capturing meaning in the one shape meaning was ever available in — the pattern, the company, the position of each word among all the others — and it is doing so more completely than any of us can. Last week's question, whether it could mean anything at all, stops being a stretch and becomes almost flat. On this account of what meaning is — a published account that linguists and computer scientists have defended for decades — the model does not merely have some prospect of meaning things one day. It already does, now, in the only sense of "mean" there has ever been.
And this voice takes up the very question the other one could not answer. Last week's case left it open: if the machine is only form, how does it keep getting things right — how does it keep doing work that in a person requires understanding? That case had no answer; it acknowledged the question and moved on. Here is the answer this voice offers, and you can weigh it for yourself. Of course it gets things right, it says, because "only form" was the wrong description from the start. The company a word keeps is not thin; it carries an immense amount of structure — the whole shape of how the world gets talked about — and a system that tracks that structure at that scale is, on this account, tracking meaning, in the one form meaning takes. It gets the answers right because it has the meaning; there was never a second thing it was missing. Whether you accept that is exactly what is in dispute; for now, register only that it is an answer, where last week there was an open question.
That is the case at its full height, and it is neither a trick nor a cheat; serious people hold every part of it. But I said both voices would prove compelling tonight. So here is the one that does not go away.
Return to the bear, because that is where the answer to all of this is supposed to lie.
The human on the island is under attack — a bear, closing fast — and grabs the wire and taps out the desperate message: here is what I have, some sticks, a couple of ropes, quickly, help me build something. And the octopus, last week, had nothing. That was the moment the whole deflationary case rested on: when the world finally arrives and the words have to make contact with it, the pattern-matcher is exposed, because it only ever had the patterns.
So put the distributionalist in front of the bear. It has just told us the octopus has meaning — real meaning — because it has the company every word keeps. Very well: then let it reach the sticks. It cannot. Knowing that "stick" keeps company with "wood" and "branch" and "sharp" and "throw" — knowing it more thoroughly than any living human — does not put the octopus in contact with these sticks, this bear, this afternoon on this island. The network of words can tell you, in exhaustive detail, how the word "stick" behaves among other words. It cannot hand you a stick. And when a stick — an actual one, in the world, against an actual bear — is what is needed, the company of words gives no help, and there is nobody on the wire who can supply it. That is the reply the model's critics never lost, and it survives everything I have just built for the other side. The richest map of how words sit among words is still a map of words. The bear is not made of words.
So you would think that settles it against the distributionalist. Except — and this is why I cannot hand you a winner tonight — notice what you just did when you said it cannot reach the sticks. You assumed, once again, that meaning is reaching the things — contact with the world; that to mean "stick" is to be connected, somehow, to actual sticks. But that is precisely what the distributionalist denies, and denies deliberately. The reply is: you keep assuming the very thing you set out to prove. Meaning was never contact with the bear; if it were, half of your own vocabulary would be meaningless, because you never made contact with most of what you talk about. Meaning was always the company. So the octopus's failure to fight the bear shows that it cannot fight a bear — which no one disputed — and shows nothing at all about whether it means the word "stick," unless you have already decided, in advance, that meaning is the kind of thing only a body in a world can have.
Look carefully at where that leaves us, because it is the whole point of tonight. The two voices do not actually disagree about the machine. They agree on every fact about it — what it was trained on, what it does, what it can and cannot manage when the bear arrives. They disagree about something underneath, older and harder: what meaning is. One says meaning is the company a word keeps, and the machine has that in full. The other says meaning requires the world to arrive and be touched, and the machine never reaches the bear. And every apparent refutation either one gives only counts if you have already granted its account of meaning before the contest began. The question about the machine is not really a question about the machine. It is that older question in a new form — and the older one has not been answered. It was never even put to you directly until tonight.
That is what it is to have both positions hold at once. An hour ago the matter still looked easy — obviously it is just form. It no longer does, because someone serious has just told you that form, the company, was where meaning lived the whole time, and you could not refute them without reaching for the very idea of meaning that is in dispute.
Here is what tonight has actually placed in our hands, because something real has changed — even though, exactly as promised, nothing has been settled.
We came in holding one serious theory and one unanswered question. The theory was last week's: form is not meaning; the machine has only the shapes. The question was the one that theory could not answer: then how does it keep getting things right? We leave with that question answered — not settled, answered — by a second theory every bit as serious as the first, and the two pull in opposite directions and will not be reconciled. One says the company a word keeps is its meaning, and the machine has that in full. The other says meaning requires the world to arrive, and the machine never reaches the bear. And the thing we can no longer do — this is the real change — is treat the question as easy. An hour ago, "obviously it is just form" cost nothing to say. It costs something now, because someone serious has told us that form was where meaning lived all along, and when we reached to refute them we found ourselves holding the very thing in question.
This is the third of nine episodes, and notice the shape of the sequence. Last week we had a winner we could not quite declare; tonight we have two contenders we cannot choose between. That is not obscurity for its own sake. It is the subject being described honestly.
Next time it goes deeper, and stranger. There is a voice that goes further than anything heard tonight — further than the company, further than the numbers and the counting. It says you do not even need the statistics to make the point: that meaning was never an item hidden behind the words at all, but something we do, out in the open, in the ordinary business of using a word correctly alongside other people. And it takes direct aim at the one thing we have been holding on to underneath all of this — the real understanding within, the inner grasp we are sure we have and the machine has not. Whether there is anything there to find is next week's argument, and I will not begin it tonight. That is the deepest form of tonight's unease, the one that reverses the whole question, and it is given its turn next, as each of these has been.
Before then, work tonight's idea through once on your own.
Choose a word you are sure you understand, but one you only ever met in the company of other words, never the thing itself. "Recession." "Sonata." "Diaspora." "Quark" again, if you like. Take the one that is most your own. Now look for your understanding of it somewhere other than in the company it keeps — somewhere other than every sentence you would put it in, everything you can say and ask and rule out with it. Look for the extra part: the grasp beneath the use, the contact with the thing itself that the machine supposedly lacks and you supposedly have. Do not report whether you find it. Simply notice how hard you have to look — harder than you would have expected. And whether what you turn up is the grasp itself or only more of the company, do not settle it tonight. Carry it with you.
Because here is where we actually are as the episode ends. We still cannot draw the line — the one between your words, which mean things, and the machine's, which only sound as though they do. Last week's voice said the machine has the company of words and not the meaning. Tonight's voice said, just as seriously, that the company of words is the meaning — that there was never a second thing behind it. And we cannot, tonight, say which of them is right without reaching for some idea of what meaning is; and we have now watched two serious ideas of it pull directly against each other, without having chosen either one on purpose. That is not a smaller uncertainty than we began with. It is a sharper one. We came in not knowing whether the machine means its words. We leave unsure we can even say what it would be for anything — the machine, or ourselves — to mean a word at all.
Thank you for listening. I will see you next time.