AnalysisAI ModelsNovember 2, 2026

MIT CSAIL paper: GPT-6 Astra reads and edits model weights directly

Read original source →venturebeat.com

In "Sparks of Weight-Space Programming," four MIT CSAIL researchers report GPT-6 Astra natively edited 0.016% of Qwen3.5-0.8B's weights, lifting MATH-500 from 7.9% to 48.1% and tripling IFEval, with no gradients or training. Astra also wrote small game-playing networks from scratch that beat PPO-trained ones.

How this story unfolded

5 days · 4 reports · 10 community posts · 14 of 16 shown

  1. Nov 2
  2. Nov 3
  3. Nov 4
  4. Nov 7

More stories today

Open the live feed