Claude fixes OCR doom loop by editing 0.0085% of weights, no training, $10
huggingface · x · 2026-11-04
A developer asked Claude Opus to fix an OCR model's repetition doom loop by editing its weights, with a prompt allowing up to 5 hours and $100 of Hugging Face Jobs compute. It changed just 0.0085% of weights and, with no training, fixed a model that was spewing 16,399 dots on an 1881 contents page — in under 3 hours and about $10.
A related experiment dubbed "Weight-Space Programming" had GPT-6 Astra edit Qwen3.5-0.8B weights: changing only 0.016% fixed reasoning loops and improved math up to 6x and instruction following 3x, with no gradients or training — suggesting frontier LLMs can natively read and write neural network weights.
More from coding & agent
- Google Cloud ships Antigravity plugin with guardrails and live docs grounding — rseroter · 2026-11-04
- Celesto CLI Lets AI Agents Inspect CI Jobs and Logs — aniketmaurya · 2026-11-04
- DeepMind adds Cheminformatics Skills to Antigravity for retrosynthesis and enzyme design — jueseph · 2026-11-04
- Vercel Launches Security Review Beta: Auto-Scans Vulnerabilities on Every Push, One-Click Fixes — cramforce · 2026-11-04
- Codex security plugin scans GitHub repos with OpenAI Daybreak Blue, fewer cyber guardrails — drdanielbender · 2026-11-04
- A wishlist for the ideal personal AI assistant: parallel agents, scheduling, autonomy boundaries — koltregaskes · 2026-11-04