AnalysisAI ModelsNovember 20, 2026

Weight-space programming: LLMs edit neural net weights directly

Read original source →arxiv.org

GPT-6 Astra edited 3.5% of a network's weights in one pass, taking an AI-powered car from 0 of 256 successful runs to 1,024 of 1,024 with no retraining or gradient steps. A POSTECH study found Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra all rejected Transformers, converging on sparse lookup tables ~1,800x smaller.

How this story unfolded

6 days · 3 reports · 5 community posts · from Nov 14

  1. Nov 14
  2. Nov 18
  3. Nov 19
  4. Nov 20

More stories today

Open the live feed