Teddy Warner built a wall-mounted pen plotter called GPenT — Generative Pen-trained Transformer. The name is a pun: it sounds like GPT, but it's literally a pen-trained transformer. The pun is the least interesting thing about it. The core logic is this: you give the machine a text seed, and it translates that seed into a combination of generators and parameters, which are then physically drawn on paper by stepper motors, belts, and a pen gondola. The entire pipeline — from text to generator selection to G-code to physical ink — is open-source on GitHub.

The most structurally interesting piece is a generator called dcode. Most text-to-image pipelines follow a predictable path: text → image → (optionally) vectorize → machine instructions. dcode skips the image step entirely. It is a diffusion model trained to convert text directly into G-code — raw machine instructions. The image is never the goal. The image is what happens after the machine executes its instructions. The image is the remainder of the instruction.

This has a precise location within the SAE framework. The typical chisel-construct cycle works like this: the artist begins with a visual intention (construct), uses tools to chisel toward that intention, encounters deviations and accidents along the way (remainder), and then absorbs the remainder into a new construct. GPenT's dcode interrupts this cycle at its origin. There is no "visual intention" step. The text is not a description of a desired image — it is a seed for machine behavior. The lines the machine produces are not approximations of any picture. They are the physical traces of executed instructions. The image is a leftover, something nobody explicitly asked for. This is the definition of remainder.

It goes deeper. GPenT does not rely on dcode alone. It includes procedural pattern generators, image-upload converters, and other machine learning modules. The crucial design decision is that the system uses Gemini to randomly shuffle and select combinations of these generators. You provide a text seed; the system decides which generators to use, with what parameters, in what stacking order. Each run produces an accidental combination. This means the remainder doesn't only occur at the dcode layer — at the level of "skipping the image" — it also occurs at the level of generator composition. The entire system is a multi-layered remainder-production machine.

Seeing this work now matters more than seeing it later, because it has not yet been named. It is not "AI art" — it deliberately misuses AI. It is not "plotter art" — the plotter community generally pursues precise physical reproduction of generated images, and GPenT has no "source image" to reproduce. It is not "conceptual art" — it involves extensive concrete engineering and physical materiality. It sits in the gaps between all these categories, belonging to none of them. This is precisely where the remainder lives: in the cracks between existing categories, where names have not yet hardened.

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