← Home
Showcases
Explore benchmarks and try interactive demos.

SketchAgentBench
A public benchmark for 3D coding agents. One photo and one prompt go into a browser three.js playground, where models write, render and refine a 3D scene — testing spatial understanding and coding together.

Weight-Space Playground
An AI agent found a “repetition-loop” direction inside Qwen3.5-0.8B and subtracted it from 131,072 weights — no training, no gradient descent. The model stopped looping. The demo is WebGPU-powered: both models run locally in your browser, with nothing sent to a server.