GPT-6 edits just 0.016% of weights, boosting a 0.8B model's math 6x
SergioPaniego · x · 2026-11-03
An MIT paper, "Sparks of Weight-Space Programming," shows frontier LLMs can natively read and write neural network weights, with no gradients or training loops.
Key results:
- From raw weights alone, GPT-6 Astra can tell what a network was trained to do
- It writes small game-playing networks from scratch that outplay PPO networks trained for 100 million steps
- Asked to improve Qwen3.5-0.8B at math, Astra locates a weight issue causing reasoning loops and natively edits just 0.016% of weights, raising MATH-500 from 7.9% to 48.1% (6x) and tripling IFEval, with gains generalizing out of the box
- Astra is never allowed to write training loops; it reasons about what each part computes and changes weights accordingly
The authors argue this could matter for both optimization and interpretability. Code is open-sourced at rikiyat/lmunderstand-logs.
More from AGI Musings
- Every enterprise AI demo works — that's the problem; judge by month two — arieljalali · 2026-11-03
- Yoav Goldberg highlights InterpBench, a benchmark of Transformers with known circuits for interpretability — yoavgo · 2026-11-03
- Why robot inference companies must also deploy: moats differ from LLM serving — silver__tsuki · 2026-11-03
- A firm of heterogeneous AI individuals beats homogeneous agent clones, author argues — lfschiavo · 2026-11-03
- TechBio survey: Claude daily usage jumps 28 points as bullishness cools — nlarusstone · 2026-11-03
- Anthropic warns it is very close to RSI capability, aka self-building AI — TechCrunch AI · 2026-11-03