Weight-Space Playground - Visualize LLM weight editing with WebGPU

Show HN: Watch weight-space programming in action with WebGPU

Weight-Space Playground - Visualize LLM weight editing with WebGPU

Weight-Space Playground lets you watch an AI agent edit just 0.016% of Qwen3.5-0.8B's weights to cure its infinite loops—no training or gradient descent required. It subtracts a 'repetition-loop' direction from 131,072 weights, cutting token-limit hits from 93% to 33.1% and boosting MATH-500 accuracy from 7.9% to 48.1%. The site offers a replay of both models on the same prompts and lets you run them locally in your browser via llama.cpp, WebAssembly, and WebGPU—nothing sent to a server. It's a hands-on demo of weight-space programming and a promising fit for agent-driven model editing.

Rewrote 0.016% of the model's weights — and cured its infinite loops. No training, no gradient descent.

More from this day

2026-11-19