Nothing Inside the Word: What Language Models Reveal About Meaning

Say your own name aloud, slowly, ten times. Somewhere around the seventh repetition it comes loose. The sound stays, the shape of the mouth stays, and the meaning drains out of it like water from a cracked cup. For a moment the most familiar word you own becomes a bare noise in the throat. The feeling is faintly vertiginous, and it is worth trusting, because it is showing you something true. You went looking for the meaning inside the word, and found the word empty.

This is an old habit, and a deep one. For four hundred years the West has pictured meaning as a thing that must sit somewhere. The Cartesian inheritance is not only that mind and world fall apart. It is that meaning is imagined as belonging somewhere, either inside the subject as an inner representation, or out in the object as a determinate property waiting to be grasped. Descartes crystallised it, and Christian soul-body dualism, early modern science and the long habit of treating the mind as a container facing an external world all fed it. Once meaning is placed on one side of the split, the whole difficulty becomes how the other side reaches across to it.

Then a machine started using words well without ever opening one to look inside. Do language models actually understand meaning? No. And the reason they do not is the interesting part.

The machine that never looks inside

A large language model has no inner theatre where the essence of a word waits, understood, before it speaks. It has relations. Self-attention, the operation at the centre of every transformer, does not read off the essence of a token. It composes a token's operational profile through its relations to every other token in reach. This is distributional semantics taken to scale, the old linguistic insight that a word is known by the company it keeps. The model does not grasp the word "father." It places "father" in a vast field of contrast, proximity and use, and the field does the work.

Wittgenstein saw the shape of this long before the machines arrived. To learn what a word means you never look inside it. You look at everything it stands in relation to. The words beside it, the world it points at, the life it is used in. A child learns "father" from a face at a table, a voice, a run of ordinary mornings, a thousand small corrections. The machine does the same, mapping tokens, contexts, inferences and repairs across a corpus. When we are honest about ourselves, so do we. The competence was never the sign of a hidden essence, because there is no hidden essence. There is only the web, and the web is enough to translate, summarise, analogise and reply.

Nagarjuna said this two and a half thousand years ago, in the Madhyamaka, and he said it of everything, not only words. Things carry no self-standing nature. They arise in dependence on one another. This is pratityasamutpada, dependent origination, and its consequence is emptiness, the absence of any intrinsic core sealed within a thing. The word "father" has no father-essence inside it. It arises in relation to a child, to a particular man, to absence, to grief, to the morning table. Remove the relations and nothing remains to point at. The machine stages this in technical form. Its embeddings do not store meaning as a thing. They stabilise patterns of proximity, contrast and use. I have come to call this the Empty Representation Hypothesis. The model does not climb towards a Platonic ideal of meaning standing behind the words. It converges on the relational web itself, which is no-thing. Here is the line I would keep. Language models do not understand meaning as a private possession. They expose how much of what we call meaning is a relational field arising through use.

Where relation runs out

Now the guardrail, and I will state it twice because it is easy to lose. First, relational structure can produce genuine semantic efficacy with no intrinsic meaning to ground it. The public behaviour of meaning does not require a private inner grasp. Second, and this matters more, relation without embodiment, care, accountability and lived stakes has limits, and the machine shows those limits as plainly as it shows the field.

The machine maps the field. It does not stand in it. It has never lost a father. Nothing is at stake for it when it uses the word. It carries the traces that human practice, embodied life and social correction have left in its training, inheriting meaning rather than originating it. To announce that meaning is only relations in a vector space would repeat the Cartesian error in fresh dress, moving the essence out of the soul and into the geometry. The field is real. The field is also lived, and the machine does not live. So this is no triumph for the machine. The interesting result runs the other way. The machine does not tell us that it understands. It tells us something about us. We had assumed meaning must be owned, held inside, privately grasped before it could ever be shared. A system with no inner life has now reproduced much of the public behaviour of meaning by tracking relations alone. The assumption was never necessary, and the machine is the proof.

Which is oddly good news, and close to home. The word you wore thin with repetition comes back the instant you set it down among the things it belongs to, the person who gave it to you, the room, the years. Meaning was never a kernel to be prised open and never a possession that could be lost for good. It is the field a word arises from, and the field is always there to walk back into. When the word in your mouth goes dead, you are not empty. You have simply stepped out of the relations that make it warm, and you can step back in. The machine has shown us this from the far side, without once feeling a thing, and the warmth it only maps is ours to return to.