25launches

The first launch of Weight-Space Playground.

Weight-Space Playground page

Weight-Space Playground

Launching today

See how an LLM fixed another LLM by editing its weights

Hunted by IlyasHunter
Visit tabbyml.github.io (opens in a new tab)

An interactive walkthrough of one weight-space programming result: an LLM, with no training, found a 'repetition loop' direction in Qwen3.5-0.8B and removed it from 131,072 weights (0.016%). Replay both models on the same prompts, see which neurons changed, then run the base and edited models side by side in your browser via llama.cpp on WebAssembly and WebGPU (about 1 GB download, nothing sent to a server). The edit itself was made offline. Built by the TabbyML team. Free.

Meet the team

Hunted by Ilyas. The makers are not on 25launches yet.

H (Hunter)
  • Weight-Space Playground screenshot 1
  • Weight-Space Playground screenshot 2
  • Weight-Space Playground screenshot 3
  • Weight-Space Playground screenshot 4

Discussion 1

Ask questions, give feedback. Comments don't count in the ranking.

Sign in to join the discussion.

  • IlyasHunter Pinned

    Weight-space programming, where one model reads and edits another's weights directly, is hard to picture from a paper. This page makes one case concrete: small Qwen hits its token limit on 93% of math problems, and after a 0.016% edit that falls to 33% and accuracy goes from 7.9% to 48.1% (figures from the original experiment). The edit was computed offline; the page shows the result and lets you run both models in a browser tab.