Language, LLMs & Machine Understanding · Episode 2 · 19-min read

The most elaborate parrot ever built

Transcript · audio coming

Suppose you set out, this week, to defend the claim properly — to make the case out loud, from beginning to end. The claim is that the machine is a parrot: that it does not understand a word it produces. Many of us have said something like it, quickly, as a way of not being unsettled by the thing. But saying it and defending it are different tasks, and when you go to defend it you find you are not the first to arrive. The people who have thought hardest about this reached it before you. Some of them build these systems and can say, step by step, what happens inside one when it answers. Others have spent their careers studying what it is for words to mean anything at all. And many of them say the same thing you wanted to say: it does not understand. They are not being vague. They have an argument, a careful one, and it has been tested against the strongest objections for more than forty years.

That is worth pausing on, because it changes what kind of claim "it's just a parrot" is. Last week, in most mouths, it was a quick dismissal — a way of closing the question. Tonight it becomes something else: a position, held by serious people, that a machine trained on nothing but the shapes of words can never reach what those words mean. Form, they will say, is not meaning. The machine has all of the first and none of the second, and no quantity of the first ever amounts to the second.

This is Philosophy for Us — philosophy for everyone, no degree required.

This is the second of nine episodes. Last time, the argument you arrived with — parrot or genius, autocomplete or hidden understanding — turned out to be the wrong argument, not because either side was foolish but because neither had said what understanding is; and the moment you tried to say it yourself, you found you were holding an idea of meaning you had never examined.

Tonight, one of those two voices is presented in its strongest form: the parrot voice, the one that says there is no understanding in the machine at all. I am going to build it the way its most serious defenders build it — in two ways, through two well-known cases — and I am going to do it without interruption, because a case cannot be judged until it has been presented in full. One thing only I will promise: once it is built to full strength, I will show you the questions it has not answered.

Begin with something most of us have done, because the whole case tonight rests on a single distinction, and it is already familiar from ordinary experience.

Most of us have, at some point, sung along to a song in a language we do not speak — sung it accurately, the sounds right, the stresses in the right places, the syllables matched to the music closely enough that someone overhearing might think, for a moment, that we knew what we were saying. And we did not. We could have been singing a love song, or a threat, or an advertisement for laundry soap, and we could not have told the difference, because what we had was the sound of the words, the shape of them, and none of what they meant.

That gap — between having the shape of language and having its meaning — is the whole of tonight. Philosophers have a pair of terms for the two sides of it. The shape, the pattern, the rules for which symbol follows which: call that form. What the words are about, what they refer to in the world: call that meaning. When we sing along, we have the form and none of the meaning. And here is the claim the serious parrot view will build, in two different ways: the machine is exactly that, throughout. Form only. No meaning anywhere in it. And — this is the important part — one cannot get from form to meaning by accumulating more form. A million accurate performances of a song one does not understand do not, at any point, amount to understanding it. They only produce a more accomplished singer of sounds.

It matters who is making this claim. It is not only those who dismiss the machine out of hand — it's just autocomplete, next question. That is the shallow version, and last week we set it aside. The careful version is made by people who understand these machines as well as anyone alive — some who could describe the architecture exactly and say precisely what happens when you type a question and it answers — and by people who have thought as hard as anyone about what it is for a word to mean something. Many of them look at all that machinery and say: what it does, in the end, however vast the scale, is form. It has read an enormous quantity of text and become extremely good at one thing — given a stretch of words, continuing it the way the patterns say it should go. That is a fact about shapes: which words tend to follow which. It is the singing done at a level no human could match — and still, they say, only the singing. Sound without meaning. The most elaborate version there has ever been of what you did with that song.

Notice what this does to the question you brought in. Last time, the difficulty was that you could not say what the machine was missing; you reached for the sentence the machine does not understand because ___, and found nothing definite to complete it with. Here is a definite answer, the first we have met. The machine does not understand because it has only the form and never the meaning. That is not a dismissal. It is a claim, a precise one, and it has an honest case behind it.

So we are going to present that case at full strength — not summarised, but constructed, the way its most serious defenders construct it. They have two ways of exposing the gap, two well-known thought experiments, approached from opposite sides. The first is a case of a person who has all the form, and whom we can watch, from outside, having none of the meaning. The second is a case of a creature that has nothing but form, and is caught the moment the world asks it for meaning.

Take the person first.

Here is the first case. It is almost fifty years old. It was devised by the philosopher John Searle, in 1980, and it is worth knowing that people have argued about it ever since without reaching agreement — which is precisely why it repays attention.

Imagine you are locked in a room. You speak no Chinese; to you, Chinese characters are simply shapes, intricate drawings with no more meaning than a doodle. In the room with you is an enormous rulebook, written in English, your own language. Slips of paper come in through a slot in the door, covered in Chinese characters. Your task is to look up the shapes you have been handed, follow the instructions — when you see this string of shapes, copy out that string of shapes — and push the result back out through the slot. That is all. You match shapes to shapes according to the book. You never learn what any of it means.

Now suppose the rulebook is extraordinarily good — so good that the slips coming in are questions, in Chinese, and the slips you send out are answers, in Chinese, and the answers are perfect: fluent, apt, sometimes funny. Outside the door, native Chinese speakers read your answers and are certain that someone in the room understands Chinese. To them it is a genuine conversation.

And inside the room you understand nothing — not one character. You are producing flawless Chinese with no idea what is being said, because all you ever had was the rulebook: the form, the shapes and the rules for arranging them, and never the meaning. It is the singing again, only now the case is airtight. Anyone outside would swear you understood; and we know, because we are in the room with you, that you do not.

Searle was writing about any computer running any program, decades before the machine now on your screen existed. Carried forward to a language model, his argument says this: the model is that room — not merely similar to it, but the same case. The machine has its rulebook, an enormous set of patterns for which symbols follow which, and it sends out characters that we, outside the door, read as understanding. But inside there is only the matching of shapes. Searle put it in a phrase, and it is the one the debate keeps returning to: syntax is not semantics. Syntax is the rules for arranging symbols — pure form. Semantics is what the symbols mean. One can have all the syntax there is, a rulebook of any size, and never cross into semantics, because arranging symbols by their shape is simply not the sort of thing meaning is made of. More rulebook never becomes understanding. It only becomes a larger room.

Some listeners met this room before, when this show asked a different question about the machine: whether it was conscious, whether there was any inner experience behind the screen. Set that aside tonight. You can grant, for the sake of argument, that there is no inner life at all — and the argument still holds, because the question now is not whether the machine feels anything. It is narrower and harder: whether its words mean anything. The room says no, and it does not need the machine to lack consciousness to say it. It needs only that the symbols are shapes.

That is the case at full strength. But I promised to build it as its most serious defenders build it, and that means putting the best replies to it into the room as well, because they have been there from the start; Searle has had to answer them, and the exchange is not finished.

The first reply is this: you are looking in the wrong place. The person in the room does not understand Chinese, true — but the person is only one part. The person, plus the rulebook, plus the whole apparatus, is the system; and perhaps the system understands, in the way that no single neuron in your head understands English although you do. This is called the Systems Reply. Searle's answer is direct: let the person memorise the entire rulebook, discard the books, do all of it in their head, and walk out of the room. Now there is no system left but the person — and the person still understands not a word of Chinese. So where did the understanding go? The defenders have answers to that in turn, and they are not foolish ones; I will not pretend the exchange ends where Searle wants it to.

The second reply goes further, and it is one we cannot settle tonight; I will name it and set it aside. Of course the room does not understand — it is sealed. It has paper coming through a slot and nothing else. Give the symbols a way out into the world: put the system in a body, connect it to eyes and hands, let it see water and spill it, and perhaps then the word for water connects to the actual thing and begins to mean. This is called the Robot Reply. It is a genuine answer; it may be the most important one in the whole series. Searle thinks he can defeat it — he says that adding a camera merely feeds the room more shapes, more syntax, and brings it no closer to meaning. Whether he is right is a question we are nowhere near ready to take up. Note the point; we return to it, and when we do it will be one of the central questions of the whole nine.

So the room is not decisive. It is powerful — it holds — and the people trying to answer it are as serious as the man who devised it. Hold it as exactly that: the clearest single picture of "all form, no meaning." Now here is the same gap from the other side. This time it is not a person who lacks the meaning; it is a creature that has nothing but form, set loose in a world about to ask it for more.

The room shows a thing that has the form and lacks the meaning. The next case shows something stranger, and I think more pointed for our question: something that has only form — is made of nothing but form — and manages until the world asks it for something form cannot supply. It comes from two linguists, Emily Bender and Alexander Koller, in a paper published in 2020, and it runs as follows.

Two people are stranded on separate desert islands, far apart, connected by a telegraph wire running between them under the sea. So they talk, tapping out messages back and forth, year after year — coconuts, the weather, what they did that day. Ordinary conversation, turned into clicks on a wire.

Down in the deep water, a highly intelligent octopus finds the cable and begins listening. The octopus has never seen an island, a coconut, or the sky. All it has ever encountered is the traffic on the wire — the patterns of clicks, which signals tend to follow which, the rhythms of the exchange. And the octopus is remarkably able. In time it becomes so good at the patterns that it can predict what comes next with great accuracy — good enough that one day it cuts the cable, connects itself in, and begins answering in place of the person at the other end. It sends back clicks that fit. How was your day. On the far island the human reads the replies and notices nothing wrong; the conversation seems normal. The octopus is passing for a person.

And it works — for as long as the conversation stays within the shape of all the conversation that came before. Then one day the human on the island is in danger. A bear, say: they are being charged by a bear, and they seize the wire and tap out something urgent — here is what I have, some sticks, a couple of ropes; quickly, help me build something to defend myself. And the octopus has nothing. It has never seen a stick. It does not know what a bear is, or fear, or a weapon, or building. It has the form of ten thousand conversations and not one particle of what any of them was about. It can send something back — clicks that, by the patterns, resemble a reply — but there is no meaning beneath them, and when meaning is what is required, when the words must finally connect to the world, it fails. There was never anyone on the wire who understood. There was a very capable pattern-matcher that had only ever had the clicks.

Bender and Koller's claim — and Bender in particular has argued it forcefully and publicly — is that the octopus is the language model. The model was trained on text; on form; on an enormous quantity of which-words-follow-which, and never once on the world those words are about. It never saw a bear. The fluency is real, as the octopus's fluency was real, and it is fluency in form; and when we read meaning into it, the meaning is ours, not the machine's. We are the human on the island, hearing a conversation and supplying the understanding ourselves, then crediting it to whatever is at the other end of the wire. A year later, Bender and three colleagues — Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell — published a paper that gave the whole view its lasting name: the machine, they said, is a stochastic parrot. "Stochastic" means, roughly, according to the odds, at random within a pattern. A parrot, because it returns the shapes of words with no grasp of what they hold. It is, they wrote, a system for stitching together sequences of linguistic form it has observed, according to the statistics of how those forms combine — with no reference to meaning, no access to what the words point at, and no notion of anyone to say them to. And our error — the ordinary human error — is to hear all that fluent form and assume there must be meaning beneath it, because when we ourselves talk, that is how it works.

So now we have the case, both sides of it, at full strength. From one side: a person in a room with every rule and no understanding — all the form a thing could have, and still no meaning. From the other: an octopus on a wire with nothing but the patterns — only form to begin with, and exposed the instant the world demands meaning. Two cases, one claim, and it is a serious claim made by people who are not deceived by the machine's fluency: the machine has the shapes of our words and not the meaning of them, and no quantity of shapes ever becomes meaning. That is the strongest form of the view you arrived with. It is far more serious than the quick dismissal you began with, and it carries real weight.

And now, as promised, the questions the case has been carrying the whole time and has not answered.

Two things. The case for the parrot is the most complete one we have built so far, and it arrived holding two pieces of unfinished business, neither of which it has resolved.

The first you have already heard, and I asked you to set it aside. The best replies to the room were never defeated; they were answered, and then they answered back. Perhaps the system understands although the person does not. Perhaps a sealed room was always the wrong test, and a thing with a body in a world would cross over into meaning. Searle has responses; the defenders have responses to his responses. That exchange is open — I would be misleading you to close it tonight — and the second reply, embodiment, we deliberately hold for later, because it is too important to rush.

The second piece is the one that ought to keep the parrot view honest, and it is the very voice you arrived with last week: the one that asked, then how is it getting things right? Consider the octopus again. The reason the octopus is caught is the whole point of the story: the bear demands meaning, the octopus had only form, and so it fails. But the machine on your screen frequently does not fail the bear. Give it a problem that looks genuinely new — not in the shape of anything you would expect it to have been trained on — and a surprising amount of the time it does what the octopus could not: it answers, it adapts, it builds you the defence out of the materials you actually have. This does not prove that it understands. Perhaps the quantity of form it absorbed is so vast that almost nothing is truly new to it. But look how much the parrot view now has to explain. It has told us, with real force, that form can never do the work of meaning — and then the machine goes on doing work that, in a person, only meaning ever did. The case is not wrong. It is unfinished. It owes us an account of how pure form keeps passing the test that was meant to expose it.

So here is where we now are, and it is further along than an hour ago. You arrived with a reflex — it's a parrot — something said to steady yourself. You leave with the developed version of that reflex: form is not meaning, syntax is not semantics, and you can now say it precisely, with two hard cases behind it, the way its most serious defenders say it. That is the gain: a quick dismissal has become a considered position. But the same hour has shown where that position is weakest — the replies it has not closed, and the successes it cannot yet explain — so you also leave unable, honestly, to declare it the winner. You have the most developed version of the parrot view there is. You cannot quite call it the last word.

This was the second of nine. Last week we found the empty place in the argument; tonight we filled it with the first serious answer, and saw where that answer is under strain.

Next week the other voice receives what this one has just received. Not the careless version — it obviously understands, look at it — but its serious champion. For there is someone who hears the octopus story and says you have told it the wrong way round. Exactly how — what they take the whole parrot case to get wrong — is next week's subject, and it will have the floor without interruption from me, precisely as the parrot view did tonight. For now it is enough to know that it is coming, and that it is not the credulous reflex but a serious and formidable theory.

Before then, consider running the distinction once yourself, on a case I have not touched.

Consider a student who did very well on an examination. Top marks — but you happen to know how it was done: the student never understood the subject at all. They had memorised which answer-shape goes with which question-shape, drilled the patterns until they were flawless, and reproduced them on the day without any grasp of what the material meant. Ask the question: did the student understand the subject? Plainly not — and yet every answer was correct. There it is, in a human being, with no machine involved: all the form, none of the meaning, and from outside no way to tell the difference. The distinction is real, and now it is yours to use.

And then, since nothing here is to be settled cheaply, take it one step further. The student could leave the examination, go and do the thing, connect the symbols to the world, and finally understand. That route is open to the student. Is it open to the machine? Is the machine like the student — form now, with a way through to the rest later — or is it the sealed room, with no way out at all? I am not going to answer that; I cannot yet. Hold the two side by side, the student and the machine, and notice that you are not sure they are the same case. That uncertainty is exactly where you should be, going into next week.

Thank you for listening. I will see you next time.

← Back to Language, LLMs & Machine Understanding Back to top ↑