Key Info
A developer asked Claude Opus to stop a small OCR model from looping by directly editing its weights through Hugging Face Jobs, changing only 0.0085% of parameters with no training.
Highlights
- The edited model correctly read an 1881-page document instead of emitting 16,399 characters of dots.
- The prompt allowed up to 5 hours and $100 of Hugging Face Jobs compute; the task finished in under 3 hours for about $10.
- The approach, framed as weight-space programming, requires no gradients or training.
- A related cited example claims GPT-6 Astra edited Qwen3.5-0.8B weights (0.016%) to fix reasoning loops and improve math up to 6x.