MIT researchers say GPT-6 Astra can read and edit neural network weights directly; a critic says the headline result needed code
The original post has the experiments and full logs; Matthew Honnibal's rebuttal tables which ones really worked from raw weights alone, and Simon Willison tried one on a laptop model.
In a November 2 blog post, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT CSAIL report that OpenAI's GPT-6 Astra can "natively" read and write neural network weights, which they call Weight-Space Programming. Given only the raw numbers of small networks, Astra named which of four grid games each was trained on and what shape each of three diffusion models draws. It also wrote small game-playing networks from scratch. Asked to make the Qwen3.5-0.8B language model better at math without training, it found a cause of the model's reasoning loops and edited 0.016% of its weights; the authors say MATH-500 rose from 7.9% to 48.1% and IFEval from 21.0% to 62.3%. A similar attempt on Gemma 3 270M did not help. The logs are public on GitHub.
spaCy author Matthew Honnibal called the post misleading: by his count only three experiments, on networks of 17,000 to 172,000 weights, worked from raw weights with no tools, while the Qwen result used Python and up to 50 test runs of the model, much as a human ML engineer would. Simon Willison found that Qwen 3.8 27B on his laptop also named three of the four games from the published prompts, and wrote that some wider commentary on the paper did not feel justified.
-
Mechanistic interpretability
After exploring Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights, for the field this work builds on, read about Mechanistic interpretability.
The research field that tries to understand what neural networks compute by analyzing their internal structures and circuits, much like reverse engineering conventional software.
Wikipedia · Wade · 5 min
0