ShaderLM

An experiment in using large language models to generate WebGPU shaders from natural language descriptions. Describe a visual effect in plain English, and the system produces working WGSL shader code that renders it in real time.

Shader code is an interesting test case for LLM generation because the output is immediately visual — you can see whether the model understood your intent without reading a line of code. The project explores the boundary between what LLMs can reliably produce in a highly constrained domain (strict type systems, parallel execution models) and where they break down.