You steer,
AI motors.

Computer-assisted image-making has a long history at the California Institute of the Arts, much of it predating the phrase in its current sense. Figures whose work at or near the school connects to this tradition include:

This project is an attempt to continue a line — not to import a new technology into an art school, but to recognize that the school has been working with machines, procedurally and generatively, for a long time.

Most AI tools built for artists share one design. You type a prompt; the system hands you a finished thing: an image, a video, a piece of music. This is efficient. It also removes much of the space where artistic judgment usually accumulates.

For many artists, the satisfaction of creative work comes from decisions made over time: this color not that one, this curve adjusted, this element removed. When a tool skips that process and delivers a result, the artist risks becoming a curator of outputs rather than the author of the work.

Writing code with a language model has a different shape. You still see every line. You can change anything. When something breaks, you debug it together; when something works, you can understand why. The model supplies syntax, algorithms, and standard solutions. You supply what you want to make, and why. The conversation stays yours.

The training-data story runs differently here as well. Image generators were trained on artists' work without consent — a claim still being argued. Large language models learned to code from a culture of publicly shared examples, explanations, and problem-solving, much of it posted to help other people. That is a different bargain. It isn't a clean one, but it is different.

A shader is small — tens to a few hundred lines — and it runs on the GPU. A fragment shader answers one question: what color should this pixel be? It runs that rule, in parallel, once for every pixel on the screen.

Because shaders are small and well-documented, a language model can often reason about the whole program at once. The artist prompts; the LLM writes and edits GLSL; WebGL runs it; the browser paints it. Preview is instant.

Four common channels move data through the pipeline. Attributes, uniforms, and textures come in from JavaScript; varyings pass data from the vertex shader to the fragment shader. The names appear throughout the classroom and the playground:

Three small widgets, each one teaching a single idea. The classroom opens each one into a chapter. Drag the sliders on the first two; the third does its own thing.

A shader is a small program. On its own it draws an image. The playground wraps each shader with a modular set of tools that turn it into an instrument — something you can feed, play, and record.

Advanced students compose these. A student who understands all three — an uploader, the recorder, an audio-reactive hook — can take a piece of footage, drive its distortion with music, and record the result as MP4. The audio-reactive behavior from one shader can be isolated and redeployed to drive another transformation entirely: making a character appear to speak, or a starfield explode on a snare hit. The pieces are separable; what the advanced student contributes is the composition.

A few habits make the difference between a productive hour and an exhausting one. They do not come from the model. They come from working with it.

Watch for false summits. The model draws from training data, and where many people have been stuck before, the model gets stuck in the same place. The signs are familiar — the same wrong answer in three different shapes, the same broken code with a different comment. When you recognize it, stop. Reset the conversation. A fresh session unblocks what a long conversation cannot.

Read errors literally. When something breaks, copy the exact error text and paste it into the conversation. The model parses errors far better than it parses descriptions of errors. "It isn't working" gets you guesses; the literal message gets you answers.

Change one thing at a time. When the model proposes five edits at once and the result is worse, you won't know which one caused it. Apply edits in single bites. Test. Then the next one.

Stay in conversation. A language model is not a better search engine. It is a collaborator with a short memory, a tendency to flatter, and a habit of confidently supplying what it thinks you want. The work is not to type better prompts. The work is to stay alert to what it is actually giving you — and to keep redirecting when it drifts.