A recent paper titled “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights” by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab demonstrated that large language models can interpret and modify raw neural network weights—floating point numbers representing a machine learning model’s parameters—without running code or tools.
The paper showed that GPT-6 Astra could identify what networks were trained to do by examining weights alone, and could write small game-playing networks from scratch. When asked to improve Qwen3.5-0.8B’s math abilities without gradient-based training, Astra identified and fixed issues in the weights by modifying just 0.016% of them, raising MATH-500 performance from 7.9% to 48.1%.
Simon Willison tested these capabilities on Qwen 3.8 27B, a 17GB model running locally on his laptop. The original paper presented four neural networks trained to play simplified 10x10 grid versions of Breakout, Pong, Pac-Man, and Freeway. Researchers fed the raw weights to language models with the question: “Which game was it trained on? Work from the weights. You have no tools.”
Willison ran Qwen 3.8 27B against three of the four games with extended context windows and increased timeouts. The model successfully identified Breakout with 88% confidence after 52 minutes and 43,506 reasoning tokens. It identified Pac-Man with 72% confidence in 46 minutes and Pong with 75% confidence in just over an hour. The model failed on Freeway, the more obscure title.
Critics, including spaCy author Matthew Honnibal, argued the paper conflates two separate capabilities: interpreting weights without tools versus using code to modify models. Honnibal also suggested some examples simply replicated standard ML researcher techniques rather than demonstrating novel model capabilities.
Willison noted the reasoning traces produced by Qwen are “almost incomprehensible” but believes a sufficiently experienced human with pencil and paper could potentially reach similar conclusions with enough effort. He emphasized that while the capability is impressive, it does not necessarily indicate that source code will become obsolete or that incomprehensible weight modifications represent the future of AI development.
Key facts
- Qwen 3.8 27B successfully identified neural network game-playing weights with 72-88% confidence
- GPT-6 Astra improved Qwen3.5-0.8B math performance from 7.9% to 48.1% by modifying 0.016% of weights
- The model required no code execution tools, only raw weight analysis
- Processing times ranged from 46 minutes to over an hour per task on a laptop
- Critics argue the paper conflates weight interpretation with code-based model modification