AnalysisAI ModelsNovember 2, 2026

MIT CSAIL: GPT-6 Astra edits Qwen weights, MATH-500 7.9% to 48.1%

Read original source →venturebeat.com

MIT CSAIL's "Sparks of Weight-Space Programming" reports GPT-6 Astra natively edited 0.016% of Qwen3.5-0.8B's weights, lifting MATH-500 from 7.9% to 48.1% with no gradients or training. IFEval rose 3x; Astra also wrote small game networks from scratch that beat PPO-trained ones.

How this story unfolded

1 day · 2 reports · 9 community posts · 11 of 14 shown

  1. Nov 2
  2. Nov 3

More stories today

Open the live feed