Voxel Bridge Challenge

X is full of voxel demos right now. Pretty pagodas, tiny cities, cute scenes. They look great, but most of them don't do anything. They measure how well a model can paint a UI, not how well it can think.

This challenge is the opposite. Every model gets the same prompt and has to build a single self-contained HTML file, no server, Three.js only. The scene looks simple: a river between two banks. Then it isn't. Build Bridge drops a two-block-wide arch of exactly 40 blocks in four colors. Earthquake knocks out the two central supports and the model has to rebuild with only the 38 left, keeping the thing standing. Widen River stretches the river from 10 to 14 units and the bridge has to follow with the same blocks. Blocks get moved, counts change, geometry gets reworked on the fly.

So what you're comparing here isn't pixel polish. It's whether the model actually understood the sizes, the colors, the counts and the geometry of its own little world, and whether it can explain every move it made. The reasoning log on the right of each page is where that shows.

The prompt also forces each model to name its output folder after itself and stamp a Model: … badge under the title. That's how you tell who built what. Click a card below to open that model's full interactive page.

Source of truth, kept in the voxels repo:

loading…