Ryan Seslow has been making analog collage for over forty years: paper, paint, aerosol, ink, blades, glue, and a great deal of cutting. This August he cut 231 fragments out of his own archive — not stock, not scraped, not generated, every piece from work he made by hand, pulled out of 147 of his own Photoshop files. Then he wired a terminal to Photoshop. One channel looks: the model reads what the canvas actually looks like right now, not just numbers about it. The other channel acts: it generates Photoshop's own scripting language and hands it over through AppleScript. Every element lands as its own named layer, normal blend, one hundred percent opacity. Nothing flattened, nothing saved, the document left open so he can take the mouse and finish it himself.

To give the machine anything to read, he first composed four reference pieces by hand — and deliberately did not compose them for the machine. "The second you start performing for the measurement the measurement stops being worth anything." Then it reported back what he had been doing for forty years without saying: the hero element regularly exceeds the canvas, measured at 1.17 and 1.28 of the frame; he upscales past native resolution without hesitation, a 90 pixel cut blown up to 562 pixels; zero masks and zero non-normal blend modes across thirty layers, opacity his only dimmer; and the bottom third of every picture sits between 93% and 98% full. The floor is what stopped him. He had never thought about it, never been taught it, never once said it out loud to a student. It fell out of a measurement.

Up to here it is merely interesting. A rule that has been measured and stated is immediately already-construct: repeatable, teachable, optimizable — handsome, but it is the sediment of a chisel-construct cycle that has already closed, no longer a live remainder. The remainder is in the next step.

The machine built a composition out of six of his fragments, and by every number it had access to it had succeeded: frame filled, floor loaded, overlap ratios inside the bands measured off his own work. It was mush. You could not see anything in it. Every metric had been computed on the union of the shapes, and the union is blind to burial: six elements interlocking and six burying each other score identically. When they finally measured how much of each element survives the layers above it, the machine's version came back at 44% mean visibility, with four of seven elements effectively invisible. His own four pieces run 75% to 87%.

That is the shape of a remainder: the formalization is complete by its own books, and there is still something non-empty outside them — and you cannot derive it from inside the books. You can only wait for the system to produce garbage with total confidence, and read off which column was missing from the shape of the garbage. The most counterintuitive line in the whole loop is his own: "The repair teaches it more than the original ever does." The hand-made original is already-construct; new information enters only at the correction. The machine gets it wrong, he spends ten minutes rescuing the picture by hand, and the machine measures the rescue. The load-bearing end of the work is not the making but the fixing. And the structure cannot finish: each repair grows a new metric, and each new metric arrives with a blind side of its own.

It is worth seeing now because it does not have a name yet. "AI art" is wrong: nothing is generated, the fragment pool is one hundred percent his own hand, and he puts that in the work rather than in a footnote — "State what the machine touched. Say it in the work, not in a footnote." (The exhibition also holds a second body of collage heads, cut by hand and then run through ChatGPT under an aesthetic prompt until one surface held across the group; the geometry is his and the palette is not, and the page says so. Those are studied for composition only and never cut into the fragment pool.) "Computational aesthetics" is wrong too: he is not building a general model, he is measuring one specific person's hand. And it is certainly not a tutorial, though he wrote it as one — take five to ten pieces of your own finished work, measure something you have never measured about them, write down whatever surprises you, then break it on purpose in the next piece. When this practice gets its name, the name will arrive with a method attached and the friction will be gone. It is still here.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics