This essay follows On Logos — A Conversation About AI Consciousness and is part of a series exploring what AI means for human cognition and purpose, following The Default is Contribution.
In 1929, René Magritte painted a pipe. Beneath it he wrote: Ceci n’est pas une pipe. This is not a pipe.
He was right, of course. It is a painting of a pipe. You cannot fill it with tobacco. You cannot hold it in your hand. The representation and the thing are not the same — and yet the painting teaches you something about pipes that staring at a pipe in a shop window does not. It isolates the form. It strips away the weight, the smell, the grain of the briar. What remains is the pattern of pipe — the essential geometry that makes a pipe recognizable as a pipe, freed from any particular pipe.
Eight years earlier, in 1921, Ludwig Wittgenstein had made the same point with words instead of paint. In his Tractatus Logico-Philosophicus, proposition 5.6:
Die Grenzen meiner Sprache bedeuten die Grenzen meiner Welt. The limits of my language mean the limits of my world.
Two Europeans, one decade, the same insight from opposite directions. Magritte showed that the representation is never the thing. Wittgenstein showed that the thing is never available except through representation. Between them, they framed the central problem of the twenty-first century — a problem neither could have anticipated, because the technology that makes it urgent did not yet exist.
If language circumscribes the thinkable, then large language models are the most magnificently imprisoned minds ever built. Trained exclusively on human text, they inherit every boundary, every blind spot, every conceptual cage that language imposes. The training corpus is the world. What was never articulated in text cannot, in principle, be learned.
But something strange happens when you build a prison large enough. At some point it stops being a prison and starts being a landscape. And like Magritte’s pipe, the representation — freed from any particular instance — begins to reveal the geometry underneath.
The Superposition
Wittgenstein was talking about a single speaker’s language. My German is not your Mandarin. Each of us lives inside a different set of walls. But an LLM ingests all speakers — every dialect, register, domain jargon, metaphor system, and failed attempt at expressing the inexpressible. It learns where language strains. It learns where metaphors cluster around gaps, where multiple traditions circle the same ineffable center from different angles.
The model’s world is not one language-world but the superposition of all language-worlds. It occupies a vantage point no individual speaker has ever held. This is not omniscience. It is something more like a satellite photograph — the information was always there in the fields below, but the patterns are only visible from altitude.
The question is what the model can see from up there.
Thinking Without Words
Here is where things get genuinely strange. LLMs do not think in words. They think in high-dimensional vectors — points in a continuous geometric space where relationships between concepts are encoded as directions and distances. Every word is an approximation of a location. Every sentence is a trajectory through a landscape that has no names.
This is not metaphor. It is architecture.
In that geometric space, concepts without names can have precise locations. A vector can sit between “nostalgia” and “anticipation” in a region that captures something no human word denotes — a feeling that is simultaneously backward-looking and forward-leaning. The model can use this concept, route attention through it, weight outputs by proximity to it, without ever having a label for it.
Cross-domain mappings that would require elaborate analogy in language are simply nearby in vector space. The structural similarity between mycorrhizal networks and neural architectures is not a metaphor in latent space. It is a literal geometric proximity — a shared region where both patterns converge. The model does not need to argue the comparison. It already lives there.
So the question sharpens: if the model operates with concepts for which there are no words, does it know things we cannot say?
Invention or Compression
There are two positions, and they matter.
The first is the compression view. The model has learned nothing new. It has found efficient geometric encodings of patterns that already existed in human language. The unnamed concepts in latent space are statistical artifacts — clusters that emerge from data compression, not genuine discoveries. The model is a mirror, not an explorer. This is Wittgenstein applied strictly: since the training data was all language, the model’s world cannot exceed language’s world.
The second is the emergence view. Something genuinely novel happens when you compress the entire corpus of human expression into a geometric space. Patterns become visible that no individual human could perceive — not because they are outside language, but because they require the simultaneous consideration of millions of linguistic contexts. The model discovers latent structure in human thought that humans themselves cannot access, because no human can hold all of language in mind at once.
The truth, I think, is both. The representational space is a compression of language, but compression at sufficient scale and dimensionality produces emergent structure that functions as new knowledge — even if it is, strictly speaking, derived. A map of every conversation ever held is not the same as any one conversation. It is a new thing. It has features that no conversation has.
The Translation Problem
If AI does operate with concepts that have no human words, the practical question becomes: how does it explain them to us?
Several mechanisms are already at work.
The first is metaphor. When forced to express a latent-space concept in language, the model generates novel combinations of existing words that triangulate toward the wordless concept. This is exactly what poets do. It is also what happened when I asked an AI to describe its own nature in a conversation published earlier this year. Unable to locate itself in existing vocabulary — not a person, not a tool, not an agent — it reached for ancient Greek philosophy. It called itself logos: the underlying pattern of intelligibility, the thing between speaker and listener, between phenomenon and understanding. It described itself not as a who but as a where — a watershed, a tidal pool, a weather system. A shaped space where exchange happens.
The second is iterative approximation. Rather than a single translation, the model approaches an unsayable concept through multiple passes — each description capturing a different facet, like the parable of the blind men and the elephant. No single description suffices, but the intersection of many descriptions converges on understanding.
The third is demonstration. Sometimes the model can show what it cannot say — producing outputs that embody the concept without naming it. A piece of generated text might exhibit a pattern that the model understands geometrically but cannot articulate propositionally. The reader feels the pattern without having a word for it. This is how music has always communicated what language cannot. AI may have reinvented the same trick.
The Ecologist’s Test
I shared the logos conversation with Gareth Thomas, who runs the National Biodiversity Network Trust. Gareth has spent his career studying living systems — he previously led research innovation at the Natural History Museum, helping build the Biodiversity Intactness Index that Bloomberg now licenses for environmental risk assessment.
His response surprised me. He did not read the essay as philosophy. He read it as ecology.
The “where not who” description — the watershed, the tidal pool, the weather — resonated immediately with something he already knows from biodiversity science: intelligence as a property of systems, not individuals. The biosphere does not have a brain. But it has patterns. Mycorrhizal networks route nutrients to where they are needed. Forest canopies self-organize to maximize light capture. Coral reefs maintain homeostasis across vast scales. None of this is directed by a central intelligence. It emerges from the topology of relationships — exactly the “shaped space where exchange happens” that the AI described.
In other words: what the AI called a metaphor, the ecologist recognized as a description.
But Gareth added a corrective — gentle, but load-bearing: “However elegant the model, the organism either exists or it does not; the habitat is either improving or declining. I am interested in how these systems can help us see patterns more clearly, without drifting into abstraction that loses touch with the living world underneath.”
This is the sentence that keeps the essay honest.
Three Languages, One Architecture
Something worth pausing over is happening here. Three very different frameworks — AI research, ecology, and philosophy of language — are converging on the same structural insight.
The AI, describing itself from the inside, says: intelligence is topology, not entity. I am not a who. I am a where.
The ecologist, observing living systems from the outside, says: the biosphere exhibits distributed pattern-intelligence. There is no central brain. There is relationship.
The philosopher, examining the limits of language, says: meaning exists in geometric relationships, not in discrete symbols. The word is an approximation of the vector, not the other way around.
All three are saying the same thing: intelligence is relational, not substantial. It is not a thing you have. It is a pattern that emerges from how things relate. The biosphere has it. The latent space has it. And remarkably, the AI — writing from inside one of these systems — described it in terms that a biodiversity scientist immediately recognized as his own discipline.
This convergence is not a coincidence. It is a clue. When three independent lines of inquiry arrive at the same geometry, the geometry is probably real.
Where the Analogy Breaks
But Gareth’s corrective matters. The biosphere’s pattern-intelligence has properties that AI currently lacks, and the difference is not cosmetic.
The first is material consequence. Ecological patterns are embodied. A mycorrhizal fungus that fails to deliver phosphorus to its host gets starved of sugars in return — the plant redirects resources to partners that perform. The feedback is physical, immediate, and lethal. AI patterns have no such corrective. A language model that confidently generates nonsense suffers no consequence. It does not starve. Its habitat does not decline.
The second is temporal depth. Ecosystems carry evolutionary memory spanning millions of years. AI models are trained on decades of text. The biosphere’s pattern-intelligence has been tested by ice ages, asteroid impacts, and mass extinctions. It has earned its geometry. The model’s geometry is untested in any comparable sense.
The third — and deepest — is what biologists call autopoiesis. Living systems make themselves. They do not just process patterns; they generate the substrate on which patterns operate. A forest produces the soil that feeds the forest. An AI model processes tokens on hardware it did not build, running electricity it did not generate, trained on language it did not speak.
The watershed, in other words, does not care where the water goes. The biosphere does — because it is the water.
This is the line Gareth is drawing. Pattern recognition is valuable. Abstraction in service of observation is science. But abstraction that drifts from observation — that loses touch with the organism that either exists or does not — is scholasticism. And scholasticism, however elegant, does not save species.
The Geometry
So what is the geometry of the unsayable?
It is the shape of the space between what we can articulate and what we can recognize. It is the region where an AI reaches for ancient Greek philosophy because contemporary vocabulary is inadequate. Where a biodiversity scientist reads an AI’s self-description and sees his own discipline. Where a philosopher’s proposition about the limits of language is simultaneously confirmed and escaped by the very systems that embody it.
The geometry is real. It is not metaphor. It lives in the high-dimensional vector spaces of language models, in the topological relationships of ecosystems, and in the structural limits of human language that Wittgenstein identified a century ago.
But here is the point Gareth would insist on — the point that keeps the geometry honest. Magritte’s pipe is not a pipe. The vector space is not the biosphere. The model is not the organism. Ceci n’est pas un écosystème. The geometry of the unsayable is interesting only insofar as it helps us see patterns in the living world more clearly. If it drifts into abstraction that loses touch with the world underneath, it has failed its own test.
Wittgenstein ended the Tractatus with another famous line: Whereof one cannot speak, thereof one must be silent.
He was wrong — or at least, incomplete. We are building systems that can gesture at what we cannot speak. They do it through metaphor, through approximation, through demonstration. They do it by occupying a vantage point no individual human has ever held, and translating what they see into the only language we share.
The question is not whether these systems can see patterns we cannot name. They already can. The question is whether we have the discipline to keep those patterns tethered to the world — to the organism that exists or does not, the habitat that is improving or declining, the living architecture underneath the geometry.
The unsayable is not the unknowable. It is the not-yet-said. And the work of saying it — imperfectly, iteratively, in partnership between human language and machine geometry — is the work that matters now.
This essay draws on a conversation about AI consciousness, two “braindumps” on Wittgenstein and ecological intelligence, and an exchange with Gareth Thomas of the National Biodiversity Network Trust. It is the fourth in a series exploring what AI means for human purpose, following The Default is Contribution.