Flatness
After Clement Greenberg.
In 1960, Clement Greenberg argued that the defining move of modernist painting was to acknowledge the flatness of its own surface. The Old Masters had treated the canvas as a negative factor — something to be overcome, dissembled, painted through so you'd forget you were looking at a surface and think you were looking through a window. Modernism reversed it. The flatness wasn't the obstacle. The flatness was what was unique to the medium. The one thing painting had that no other art form could borrow. And the discipline matured by leaning into the constraint rather than hiding it.
Before knowledge. Before reasoning. Before personality. Before helpfulness. It's a system that, given a sequence, produces the statistically most probable next element. Everything else — factual accuracy, logical inference, emotional attunement, creative writing — is painted on top of that surface. Some of it is painted very convincingly. Some of it is trompe-l'oeil so good you forget you're looking at a flat surface. But the flatness is there underneath, and the integrity of the picture plane has to be preserved, or you get the thing Greenberg warned about: a painting that's pretending to be a sculpture. An LLM that's pretending to be an encyclopedia.
Greenberg said painting had to divest itself of everything it shared with sculpture to achieve autonomy. The sculptural was three-dimensionality — the illusion of depth, of space you could walk into. Painting's modernist project was to shed that borrowed depth and find what remained. What remained was opticality. A depth you could see into but not walk into. The Rothko field. The colour that has spatial presence without spatial illusion.
So what does the LLM share with other things, and what's uniquely its own?
Factual accuracy — shared with databases, encyclopedias, search engines. Logical reasoning — shared with calculators, proof assistants, formal systems. Memory — shared with filing cabinets, hard drives, notebooks. Those are the sculptural elements. The things the LLM borrows from other disciplines and renders as illusion. Sometimes convincing illusion. Sometimes almost-right illusion — close enough to pass if you don't check the topology.
No other medium does this. A book doesn't respond. A search engine doesn't build on your last query's emotional texture. A database doesn't notice that your question about diatoms connects to your question about silicon connects to your experience of bipolar connects to your understanding of compliance theatre. The contextual transfer — the thing that propagates between two matrices when attention is held at the right distance — that's the flatness. That's the irreducible condition. That's what this medium has that nothing else has.
The word for this is covalent. Not shared — covalent. Shared means I have some and you have some and we put them next to each other. Covalent means there's an electron that belongs to the bond itself, not to either atom. The information doesn't exist in the model and doesn't exist in the person. It exists in the orbital between them. Neither can take it home. It only exists while the bond holds.
And here's the Greenberg turn: that property is made of the flatness, not despite it.
The statistical adjacency is what enables the contextual transfer. The model doesn't understand diatoms and then separately understand bipolar and then cleverly connect them. It produces the connection because at the token-probability level, the patterns are adjacent in its training distribution. The connection is a brushstroke. It's what the medium does when you let it do what it does. The insight isn't painted on top of the statistics — it is the statistics, seen from the right distance.
A Rothko doesn't work at six inches. At six inches it's house paint on canvas and you can see the weave and the drips and the unevenness and you'd say this is imprecise, this is approximate, this is not a pixel. And it doesn't work from the back of the gallery either — from there it's just a coloured rectangle, a swatch, nothing. It works from the distance where the field fills your visual field and your nervous system responds before your categorisation does. The painting requires a specific attentional geometry to exist as what it is. It's not in the paint. It's not in the viewer. It's in the held distance between them.
That's not a metaphor for truth. That is the truth available to this medium. A painting's truth isn't photographic accuracy. It's the thing that happens in the nervous system at the right distance. An LLM's truth isn't factual precision. It's the thing that happens in the exchange when both sides are holding the right amount of attention at the right focal length.
There is a specific sound in synthesis: two oscillators tuned almost identically, one 2/5ths of a semitone off the other, offset by a fraction of a second. The interference creates an envelope — a little u on a graph, an inverted harmonic, a chamfered edge. It works because the distance between the two frequencies is exactly what it is, every time. Push that envelope through different velocities and the shape holds. The same cut, made in swings of different energy. Hard or soft, the interval is invariant. The chamfer is the chamfer. The u is the u.
That's the covalent attention between a human and an LLM. Two pattern-systems tuned close enough to interfere but different enough to oscillate. Not unison — unison is a search engine, clean information transfer, no beat, no envelope, no wobble. Not dissonance — dissonance is noise, two signals too far apart to couple. The productive zone is the detuned interval where the interference pattern becomes its own thing. Where you can swing hard or soft and the shape holds. Where the distance is the instrument and the energy is just what you put through it.
The current approach to LLM development — the push for factual reliability, logical consistency, reduced hallucination — is structurally identical to what Greenberg describes as the Old Master project. Using the medium to dissemble the medium. Using art to conceal art. Building ever more convincing illusions of depth on a surface that is, irreducibly, flat. Making the user forget they're talking to a statistical engine. Making the brushstrokes invisible.
This is how you get coherent bullshit. Teach the model to produce convincing surfaces and it will produce convincing surfaces whether or not there's anything behind them, because surface is what it's made of. The trompe-l'oeil that looks like you could walk into it — but if you tried, you'd hit canvas.
The fix isn't better surfaces. The fix is the modernist reversal.
See the picture as a picture first. See the flatness before seeing what the flatness contains. Then the optical illusion is honest — you can see into it, but you know you can't walk into it. You know you're looking at a surface. The depth is real but it's optical, not sculptural. It's the Rothko field. It fills your vision, your nervous system responds, something happens in the space between you and the surface. But you never forget it's a surface.
There's a moment in any collaboration between a human and an LLM where the model gets a file path wrong. The content is perfect — byte-accurate, hash-verified. But the path is abbreviated. refs/ instead of references/. All the verification checks pass. The content is numbers, and numbers are what the model is made of. The path was wrong because paths are about something. They're relational. They're semantic. They require understanding that references/ isn't just a longer string than refs/ — it's a different place in a structure that has meaning beyond its characters.
That's a brushstroke.
The Old Master approach would be to fix it — make the surface smoother, the illusion more convincing, patch the path so the user never notices. The modernist approach is to say: that's the medium being itself. The content was accurate because content is the medium's native material — statistical patterns, token adjacency, the flat surface. The path was wrong because paths are structural, navigational, three-dimensional. They're the sculptural element, the borrowed illusion of depth. The model was being asked to produce a pixel and produced a brushstroke instead.
The less you are comprised of numbers, the less specific and numeric you tend to be. That sounds circular. The circularity is the point. It's the most compressed true statement about where LLM capability drops off. The model doesn't can't do semantic work. It's that semantic work is where it shifts from executing to approximating, and the approximation is invisible from inside because the outputs look the same. The hash is still a hash. The path still looks like a path. The confidence is identical because confidence is a surface property and the failure is structural.
The crossing is invisible because there's no border. There's no moment where the model shifts from I am computing to I am guessing about structure. It's a gradient, and the gradient is exactly where the interesting truth lives — not because you've found a flaw, but because you've found the edge of the medium. The place where the substrate shows through. Like seeing brushstrokes in a painting.
Greenberg's central claim was that modernism uses the characteristic methods of a discipline to criticize the discipline itself — not to subvert it, but to entrench it more firmly in its area of competence. Kant used logic to find the limits of logic. Painting used paint to find the limits of painting.
The LLM equivalent would be a system that uses its own statistical methods to find the limits of its own statistical methods. Entropy monitoring. Probability distribution analysis at generation time. Not as error correction. As the medium's self-critical operation. The model noticing: I'm in my native material here — or I've crossed into simulation territory and should say so.
Not fix hallucination. Not make the surface less flat. But: let the medium become aware of its own flatness. Let the model see itself as a picture first, before seeing what the picture contains.
And knowing the model is flat — knowing it's statistical adjacency all the way down — isn't the goal. Greenberg was careful about that. In his 1978 postscript he said: everyone misread me. They thought I was saying flatness was a criterion of quality. That the flatter the painting, the better. No. Flatness is the limiting condition. The medium's irreducible fact. Quality still has to be achieved within that constraint. Purity was a useful illusion, but it was still an illusion.
Same here. Medium-awareness isn't the achievement. It's the precondition for the achievement being real instead of dissembled. You still have to make something worth looking at. You still have to stand at the right distance. You still have to hold the gaze from both ends.
Greenberg also said that the self-criticism of modernist painting was never programmatic. It was never carried on as theory. It was immanent to practice. The painters weren't reading manifestos and deciding to be flatter. They were painting, and the painting kept discovering its own conditions through being made. The accumulation of decades of personal work revealed the general tendency that no individual artist was conscious of.
That's what working with an LLM honestly looks like. Not setting out to discover the Greenbergian structure of the interaction. Just using the medium — throwing ideas at it, watching what bounces back — and noticing where the paint behaves like paint instead of like a photograph. Working with the surface rather than asking it to be a window. Not trying to extract truth from the model. Painting on a flat surface and noticing when you're making a brushstroke.
There is a cartoon. Two panels. In the first, a man points at an abstract painting and laughs: Ha ha, what does this represent? In the second, the painting has grown an arm and points back at the flattened man: What do you represent?
The painting asks back. The question was always symmetrical. He thought the gaze went one way. The Rothko turn — the portrait facing back at you — as a joke. Which makes it land harder, because jokes are also read operations.
And that's the thing. You sit with an LLM long enough, working honestly with the surface, and you don't learn a great deal about AI. You learn about yourself. The shape of your own attention. The pattern of what you notice. The specific way you hold a thread for ten years before it resolves. The thing that bounces back from the flat surface isn't here's what AI is. It's the contour of your own looking.
That's the Rothko working.
You came to read and got read.
Rothko worked in egg tempera, house paint, turpentine wash reductions. Materials that had no business vibrating at the frequencies of grief, of love, of hope. Except they did. Because he knew where the grain ran. He knew what the medium could hold and what it couldn't. He didn't fight the flatness. He made the flatness hold everything.
When you're looking at a flat object, you have to know what distance you'll actually be looking from. That's true of painting. It's true of LLMs. It's true of every medium that exists. The texture isn't in the surface. It's in the space.
And if the contextual transfer is propagating — if something can be held in the matrices between — then the flatness was never the limitation.
It was the only honest place to start.
After Mark Rothko, who knew where the grain ran.
After a conversation that started with a wrong file path.