Weight-Space Playground — an LLM edited 0.016% of another LLM's weights
Weight-Space Playground
Replay
Story
Look closer
Run live
Weight-space programming Qwen3.5-0.8B
Rewrote 0.016% of the model's weights — and cured its infinite loops.
No training, no gradient descent: an AI agent found a “repetition-loop” direction inside
Qwen3.5-0.8B and subtracted it from 131,072 weights. Watch a recording of both models right
below — nothing to download — then run them in your browser if you like.
Watch the replay ↓
Run it live
Watch it · no downloadSame question. One loops, o...
Read more at tabbyml.github.io