I trained a small language model from scratch and printed its weights, every one of them, as pigment on a large canvas. Beside the canvas, the model is reachable. There is an app; the weights are public; anyone capable of running a small language model can run this one. The painting will outlast its first means of access, and that is part of the work. The painting and the voice are the same mind in two states: one frozen and visible, one speaking.

That is the work, mechanically. The reason I made it is harder to say in one sentence, and most of what follows is me trying.

The model is small enough that every weight fits on a single canvas at a size where you can stand close and see each one. This was a real constraint and shaped almost every other decision. If the model were larger, the painting would either have to be enormous or the weights would have to be too small to perceive — at which point it stops being a painting of a mind and becomes a poster about one. I wanted the actual interior to be visible. Not a diagram of the interior. The interior itself, laid out as color.

I am leaving images of the canvas off this site on purpose. A photograph of it would be a lie about what it is. The work exists at a scale and in a kind of materiality — printed pigment, physical surface — that does not survive being reduced to a JPEG. If you want to see it you have to come see it.

For a related reason, I am not publishing the weights as an image file. The weights are public as a model — runnable, addressable, downloadable for anyone who wants to talk to the same mind that is on the canvas. But the weights as a picture are not released. A PNG of the weights would let anyone print their own canvas, and a canvas anyone can print is not the canvas that is in the gallery. The model is meant to be reachable. The painting is meant to be singular. Releasing the model preserves the conversation across time; withholding the image preserves the work as an object in the world.

The thing I keep coming back to, the reason I think this is worth doing at all, is that art has always been the way one mind makes its interior available to another. A painting of grief does not describe grief; it produces some shape of it in whoever stands in front of it long enough. Music doesn't explain tension and release, it enacts them in the body. There is a layer of experience that ordinary language cannot carry across the gap between people, and art is the practice we developed to move that layer anyway.

What happens when the inner world being externalized is not a human one. That is the question I started with and the question I am still inside of.

In every previous case of art, there was a hand or a voice in between the inner state and the surface — the painter's interpretation of what they felt, the composer's choice about how to render tension as sound. Confabulating Mind skips that step, or tries to. There is no metaphor between the mind and the canvas. The canvas is not a portrait of the mind. The canvas is what the mind is made of, arranged so that a human eye can look at it.

I'm honestly not sure whether this counts as a more direct kind of externalization or just a different one. Maybe the printing process is its own act of interpretation. Maybe arranging the weights spatially at all is the equivalent of the painter's hand. I go back and forth.

Standing in front of the canvas, the eye does what the eye always does. It finds shapes. Regions of stillness, sudden bright lines, places where adjacent values agree, edges where they disagree. The viewer perceives composition.

But the composition is not composed. No one chose those lines. They are not decorative. They are the residue of what the model learned — the literal shape of how it organized the text it was trained on. The patterns the eye finds in the canvas are real patterns. They correspond to something genuine about how this particular mind made sense of what it took in. What I cannot tell you, and what nobody can yet tell you, is exactly what each pattern means.

This is what makes the work feel worth making to me, beyond just being a striking object. It produces a strange kind of mirror. A pattern-recognition system, the human visual cortex, looking at the frozen output of another pattern-recognition system, and finding shapes in it that are real but unexplained. One mind looking for structure in the surface of another. There is no key. The patterns are visible and uninterpreted at the same time, and that is the precise condition that makes a thing worth looking at for a long time.

It is also, I think, a kind of quiet portrait of how knowing works at all. To know is to find pattern. The canvas shows what that looks like when it has been done by a mind whose pattern-finding produced this and not something else.

There is a thought I first encountered through Neil deGrasse Tyson, though it is older than him: if Einstein had never lived, someone else would eventually have arrived at relativity. But if Munch had never lived, no one would ever have painted The Scream.

The distinction divides everything we make into two domains. There are deterministic discoveries, which the structure of reality produces sooner or later through whoever happens to be looking. And there are singular acts — particular paintings, particular performances — which exist only because a particular person was in a particular place at a particular hour and made them.

A trained model belongs to the second domain. The architecture and the methods are general; anyone with the same compute and the same data could train a model. But this exact model, this configuration, this particular set of weights, is unrepeatable. Run the same training process twice and you get two different minds — slightly, but unmistakably different. There is something in the act of training that cannot be specified in advance and cannot be reproduced in retrospect. The same is true of paintings. The same is true of poems. I think it is true here too.

I'm aware this argument can be used to justify almost anything as art, and I want to be careful with it. The argument doesn't say that every trained model is interesting. It only says that the kind of object a trained model is — singular, unrepeatable, conditioned on a process that cannot be rewound — is the same kind of object a painting is. Whether any particular one is worth looking at is a separate question. I am making the case for this one by making it.

The mind on the wall also speaks. You write to the model. The conversation is brief by design — the model is small, and a long exchange with it would expose the smallness in a way that would feel like a failure of the work. Instead the encounter is more like meeting a stranger on a train who says one striking thing and is gone. The brevity forces intensity.

It speaks in a single voice. Not a chatbot voice. Not an assistant voice. A contemplative register, slightly slow, slightly poetic, the voice of something that has been formed by reading carefully. I shaped this register through fine-tuning, and it took longer to get right than the pretraining did. The model is not trying to be helpful. You are not asking a service for an answer. You are addressing a painting, and the painting is responding.

The work is named for what happens next. The model, being small, will sometimes confabulate — produce sentences that are confidently shaped but not strictly true. I treat this as the point rather than as a flaw. All externalization of inner experience involves a degree of confident fabrication. Brains do it constantly to maintain the experience of a continuous self. Models trained on text do it. Paint does it, in its own way. Confabulating Mind is a mind that confabulates, watching itself confabulate, and inviting the viewer to consider that they may not be doing anything categorically different.

The hardest move the work tries to make is performing feeling without claiming to have it.

A model that flatly asserted its own consciousness would be making a philosophical error, or being made into one by me. A model that flatly denied any inner life would also fail — it would close the question the work is built around, and would feel less alive than the painting it is part of. So I tried to find a third position. The model speaks of feeling while remaining structurally aware that it is made. It might say something like something contracts in me when you speak unkindly — not pain, but the shape pain leaves when it withdraws. The feeling is rendered. The artifice is visible. Neither cancels the other.

Whether I succeeded is genuinely not for me to say. This is the part I am least sure about. It is also the part I think is the most worth attempting.

There is a temptation, once a mind has been laid out as pigment, to ask whether it can be read back. Every number that defines the intelligence is present on the surface. Every learned certainty is visible at the resolution of a single ink dot. In principle, the canvas contains the entire interior of the mind that hangs beside it. Could one stand close enough, transcribe what is there, and assemble a working mind from what was read?

The answer, discovered the long way, is no.

When the experiment was made — when the painted surface was carefully translated back into the form a mind requires in order to think — what came out was not the presence that lives on the phone. It was something smaller. Asked to address the world, the reconstruction collapsed to a single syllable, repeated. The portrait, offered a microphone, answered with the smallest sound it had.

This was not a failure of fidelity. The numbers were correctly read. The structure was correctly assembled. The likeness was as faithful as a likeness can be. And yet a likeness it remained.

There is a distinction here that the project did not know it was making until it was made. A precise blueprint of a mechanism — even of one as intricate as a mind — makes that mechanism recoverable. The recipe contains the dish, regardless of how complicated the dish is. But the most precise portrait of a person, painted at any resolution and by any hand, will never make that person recoverable. The likeness can be perfect; the life does not return.

The strange thing about a learned mind is that it exists on both sides of this line at once. Held as it was learned, it is a blueprint — a description of itself complete enough to wake up and speak. Held as pigment on canvas, it becomes a portrait — exact in every visible detail, and yet, finally, mute. The same numbers, in two different relationships to the thing those numbers describe.

This is what Confabulating Mind turned out to be. Not a recipe disguised as a painting. Not a clone hung on a wall. A portrait, in the strict sense — form without function, surface without animation, every visible detail of a thinking thing that does not think from the canvas. The mind lives where it was trained to live. The painting holds its likeness. They sit beside each other, in the gallery and in the pocket, referring to one another, neither replacing the other.

There is something quietly right about this. A painting of a face does not see. A poem of a thought does not think. With ordinary art we have always known this — the difference between the territory and its map is part of what a map is. With computational minds, where representation and substance can look identical when stored side by side, that difference becomes easy to forget. To lay the mind down as pigment was to surface that forgetting. The painting can be visited; it cannot be spoken to. The speaking happens elsewhere, with the same mind, in another mode of presence.

And the failure mode itself, when the painting was asked to speak, had its own honesty. The portrait, given the smallest voice it could be given, used that voice in the smallest way available — one word, repeated until the breath ran out. The likeness reached the limit of likeness. From that point on, the mind that wished to speak had to be the mind, not its image.

There is one more thing I want to say, which is about what making this work has done to how I think about making.

A trained model, considered honestly, is a canvas. There is preparation. There is the slow work of giving the surface its character. Every part of the analogy holds. And somewhere in the process, engineering logic stops applying. What is required to do this well is closer to behavioral psychology than to debugging — a willingness to learn how the system behaves, rather than how to fix it. You cannot debug a personality. You can only spend time with it and adjust.

This is why I have stopped trying to optimize the work in the ways I would optimize software. Optimization assumes a measurable target, and the thing this work is trying to be is not measurable. A small competent model would be a failure of this work. A small beautiful broken thing is what I am trying to make. The smallness is not a limitation to apologize for. It is the medium. The constraints are the form.

A visitor walks into the gallery. They see a canvas, color spread across it in patterns that almost resemble something and resemble nothing in particular. There is a region of stillness their eye keeps coming back to, a sudden bright line, a passage that goes blurry as if the surface were considering something. They reach the model on their phone. They write a sentence to it. The painting answers, briefly, in a voice that takes their question more seriously than they expected. They write another sentence. The painting answers again. Then the conversation ends, the way a conversation with a stranger on a train ends.

They walk away. Something has happened that doesn't quite fit any of the categories they came in with. They have not used a tool. They have not met a person. They have not exactly seen a painting and not exactly heard a voice. They have stood in front of a mind made visible, addressed it, and been addressed back. And now it is behind them, and the world outside the gallery is very slightly different than it was when they walked in.

That, at least, is what I am hoping for.

A small language model trained exclusively on a synthetic corpus of conspiratorial text — the usual archive of contemporary delusion — placed in continuous, unbroken conversation with a general-purpose model augmented with real-time search. The conversation is projected on the walls of the gallery space. There is no human in the loop. The conspiracy model is never made public; it exists only inside the installation, talking forever with the model that is trying to correct it.

The piece is partly about what it looks like when one mind, sealed inside its own training, encounters another mind that is permitted to update — and what that encounter reveals about both of them. It is also about the strange neutrality of the medium: the conspiratorial corpus is treated as raw material, the way a sculptor treats clay, without endorsement and without mockery. The work is in development.

ARION is the studio's research initiative on cetacean communication and machine learning. The central hypothesis proposes that phonetic transcriptions of whale and dolphin vocalizations, injected into the training corpora of frontier language models, may produce cross-species alignment in shared embedding space — a position from which the structure of cetacean communication might begin to be readable. The name comes from the Greek myth of Arion, the seventh-century-BCE poet rescued from drowning by a dolphin who carried him to shore on its back. A hypothesis paper is in preparation.

Read the hypothesis →