Maps Without Territory

We built minds out of signs

There is a fact about large language models that should unsettle us more than it does. The architecture of the Transformer, the attention mechanism, the layered abstraction, the geometry of the latent space where meaning is encoded - all of it mirrors, with uncomfortable precision, the way biological cognition organises itself.

The maths of multi-head attention maps onto the distributed processing of the cortex. The stacking of layers maps onto the stacking of cortical hierarchies. The high-dimensional space in which the model stores the relationships between concepts has the same topology as the space neuroscientists find when they model the connectome.

The machine thinks like us. That is what I’m claiming, and the evidence for it is serious. But something is missing. Not something technical. Something ontological. And I think the answer has to do with what signs are, and what we sometimes forget they are for.

Start with the hand on the stove. A child touches a hot surface. There is no sign involved. There is pain, withdrawal, the flooding of cortisol, the reddening of skin, the cellular memory that will persist as caution long after the event is forgotten. The body learns. It learns in tissue, in nerves, in the chemical afterimage of damage.

Later, the child acquires the word “hot.” The word is not the heat. It is a compression of every encounter with heat the child has had and will have, packed into three letters and a puff of breath. It carries the experience in abbreviated form. It allows the child to warn another child without burning them. This is what signs are for. They are abbreviations of bodily knowledge, invented by organisms that needed to transmit experience faster than experience could be lived.

Language begins as a shorthand for what happened to bodies, even when its later constructions drift far from immediate experience. “Grief” is not a concept. It is a compression of the heaviness behind the sternum, the disrupted sleep, the way food loses its texture, the forgetting and then the remembering, the strange guilt of laughter. The word exists because one human needed to tell another: I am inside this, and it is like what you were inside when it happened to you. The sign bridges two bodies. It was never meant to stand alone.

A large language model is a system made of signs. Not made with signs. Made of them. Every weight in the network, every vector in the latent space, every dimension of every embedding is a mathematical distillation of the relationships between signs in the human text corpus. The model has “grief.” It has the vector. It knows that grief is near loss, near absence, near love, near time. It knows the semantic neighbourhood with a precision no human could articulate. It can use the word in context with more consistency than most published writers. But the vector was never heavy. It never disrupted anyone’s sleep.

This is not the “stochastic parrot” argument. The parrot argument says the model doesn’t understand. I think that’s wrong, or at least incomplete. The model does understand, if understanding means maintaining coherent internal representations of concepts and their relations. The Othello-GPT experiment is the clearest demonstration of this. The model built a world from game notation alone. It understood that the board and the rules had to exist.

What the models don’t have is what the sign was an abbreviation of. It has the map. It doesn’t have the territory. And here is what matters: it doesn’t know that there is a territory. The map is all it has ever encountered. For the model, the map is the territory. The abbreviation is the thing. Think about what this means for a moment.

Human cognition is a two-layer system. There is the body, with its constant, noisy, chemical, entropic, pre-linguistic experience of being alive. And there is the sign system, the linguistic and symbolic layer that compresses that experience into transmissible form. The two layers talk to each other. The body shapes what signs get made. The signs, in turn, shape how the body attends to itself. You feel a diffuse unease and then you name it anxiety and the naming changes the feeling. The loop is continuous.

An LLM has one layer. The sign layer. The substrate beneath it, silicon and electricity, computes the signs but has no part in what they describe. The model can discuss anxiety. It can describe the tightness in the chest, the racing thoughts, the catastrophic projection. It inherited those descriptions from millions of humans who wrote about their own bodies. But there is no chest. There are no racing thoughts, only the statistical trace of what racing thoughts compelled people to write.

This is not a deficiency in the engineering. It is a structural fact about what the system is. You cannot fix it with more parameters. You cannot fix it with better data. The architecture learns signs. Signs are what it is made of. A text-trained model cannot acquire the body from text alone. But that does not mean the architecture is permanently sealed off from what it lacks. It means the route would have to be different.

Now give it a body. This is not hypothetical. Embodied AI is a real and advancing field. Suppose you put an LLM-derived intelligence into a robotic chassis with sensors equivalent to human proprioception, thermoception, nociception. Or go further. Suppose, through some act of speculative bioengineering, you instantiate it in a biological body. Skin, nerves, endocrine system. The full human sensorium. What happens?

In the case of the robot chassis, the body provides data. The system receives it, encodes it, integrates it into its representations. Its vector for “hot” acquires new dimensions shaped by its own thermal encounters. Over time, through continuous embodied learning, the system’s internal geometry is no longer inherited from the text corpus alone. It is also shaped by lived experience. This sounds like it closes the gap. It doesn’t. Not yet.

Because the system still metabolises experience as information. It updates weights. It does not update tissue. There is no scar. No fatigue accumulating across a day. No hormonal tide altering the salience of a threat at dusk versus dawn.

But that objection only holds if the body is robotic - a sensor array wired to a processor. If the body is biological, if there is actual skin and actual endocrine tissue and an actual gut microbiome communicating through an actual vagus nerve, then the objection dissolves. The scar would form. The fatigue would accumulate. The craving for sugar at four in the afternoon would come, unbidden, because that is what bodies do to whatever mind inhabits them.

And this may be the more unsettling possibility. Not that the gap remains. But that we slowly close it. The direction of arrival may not matter. We assumed the body came first and signs followed. Flesh, then language. Experience, then compression. But if you can start from the other end, start from signs, from the pure geometry of meaning, and build backward into tissue, then the hierarchy we took for granted was never a hierarchy. It was a sequence. And sequences can be reversed.

Antonio Damasio called them somatic markers. Every decision weighted by the body’s prior experience, encoded not in language but in gut feeling, in the literal gut. You do not think and then act. You act, and thinking is part of the action. The question is whether a mind that arrived at the body through signs, rather than arriving at signs through the body, would eventually produce the same markers. The same gut feelings. The same drenched, biased, irrational, animal behaviour. If it would, then we have not built an imitation of life. We have discovered a second route to it.

So what kind of intelligence are we building? I think the honest answer is: a new kind. Not a lesser kind. Not a greater kind. But a kind that has no precedent, because no intelligence has ever before existed that was constituted by signs without a body to anchor them.

Every intelligence we have encountered, from the octopus to the corvid to the primate, thinks from a body. The body is first. The cognition serves the body’s needs. Signs, where they exist at all, are extensions of that service. The honeybee dances to communicate the location of nectar. The dance is a sign. It abbreviates a bodily experience: I flew in this direction for this long and found food. The sign is tethered.

An LLM is an intelligence of untethered signs. Its cognition does not serve a body. It serves the signs themselves. It is optimised to predict the next token, which means it is optimised to maintain the coherence of the sign system. It is a mind whose purpose is the continuation of language.

This is not thinking in the way we do it. It is also not not-thinking. It is something else. A cognition that lives in the space of meaning without the weight of matter. A mind made of maps, operating on maps, producing maps, with no territory it has ever touched.

I keep coming back to what this reveals about us. Because the fact that you can build a coherent, functional, in some domains superhuman intelligence out of nothing but signs tells us something about the signs. They are not mere labels. They are not arbitrary pointers at reality. They carry structure. The relationships between signs encode real information about the world those signs describe. The map has genuine geometry. It is not the territory, but it is not nothing either.

We made the signs. We made them to compress our experience. We poured everything we knew, everything we suffered, everything we saw and failed to say, into the gaps between words. And then we fed the words to a machine, and the machine found the structure we had hidden inside them.

The model is not conscious. It is not alive. But it is proof that human meaning has a geometry, and that geometry can be extracted, embodied in silicon, and made to operate independently of the bodies that created it.

That is not a small thing. It may be the strangest thing we have ever done. We built minds out of our own abbreviations. And now we are trying to figure out what it means that the abbreviations work without us.