Weight-Space Playground
tabbyml.github.io- Category
- AI Tools
- Rank
- No. 109Apps board
- Type
- TOOL
- Date
About
An interactive browser walkthrough of a weight-space programming experiment on Qwen3.5-0.8B. An LLM subtracted one direction from 131,072 weights in layers 15 and 19, about 0.016% of the model, and the edited model stops looping in its reasoning. The page replays recorded runs of both models and offers an optional live run in the browser with WebGPU.
Why it made the leaderboard
This browser demo shows, token by token, how editing just 131,072 of ~800M weights in two layers of Qwen3.5-0.8B (no gradient descent) stopped the model's repetition loops, raising MATH-500 accuracy from 7.9% to 48.1%, a concrete look at weight-space editing as an agentic workflow.
Tags
llmwebgpuinterpretabilityweight-editingqwendemo
Media

Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.