What happened
Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT CSAIL published "Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights" on November 2, 2026. It is the authors' research blog post, not an OpenAI announcement, and the results have not been independently replicated in full.
What the authors report
The authors write: "Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do." In one example it reads a small game-playing network of 17,253 weights without tools and names Breakout. In another it edits 3.5% of a stuck car-driving policy, taking it from 0% to 100% success.
For Qwen3.5-0.8B, they say Astra found a weight issue causing the reasoning loops that Qwen's model card warns about and fixed it by editing 0.016% of the weights, which "raises MATH-500 from 7.9% to 48.1%" and triples the IFEval score. They stress this works "without allowing Astra to write training loops" and call the approach Weight-Space Programming.
The criticism
Matthew Honnibal, author of the spaCy library, calls the framing misleading in a November 3 post. In his reading, the no-tools "just by looking at the raw weights" setting appears only in three experiments, on networks of 17,000 to 172,000 weights, while the Qwen result uses Python and runs the model on up to 50 batches of practice problems. He sees it as work a human ML engineer could do, which he says is still notable, but not the same capability.
Simon Willison, also on November 3, separated the same two questions and repeated the no-tools game test locally: Qwen 3.8 27B on his laptop identified Pac-Man, Pong and Breakout from the weights, taking 46 minutes to over an hour per game, but failed on Freeway.
What it means for you
For GPT-6 Astra users nothing changes in the product. The post is evidence of what the model can be prompted to do in research settings, with results reported by the authors. If you cite the MATH-500 figure, say that it came from a tool-assisted experiment, not from reading weights alone.
Go to the source
Sources & further reading
Based on the official documentation below. Features and access can change; check the provider before making a decision.
Was this page useful?
Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.
Go a little further