What's this about?
A playful benchmark that makes top models "draw" images using colored pencils — by writing code that places strokes — instead of using an image generator, then compares how well each reconstructs targets like the Mona Lisa.
What exactly is it testing?
Spatial and visual planning without a built-in image model: the LLM has to reason about where colors and shapes go and express that as code. It probes whether a text/code model has an internal sense of composition — a very different skill from a diffusion model that paints pixels directly.
What's the community saying?
HN 143 points, 51 comments
- Enjoyment of it as a clever, low-stakes capability probe rather than a serious benchmark.
- Debate over how much is genuine visual reasoning versus the model recalling well-known images from training.
- Interest in the per-model differences as a window into how each reasons about geometry and color.