Qwen 3.8 27B Reproduces Model-Weight Interpretation Demo
Simon Willison reproduced a “Sparks of Weight-Space Programming” example with Qwen 3.8 27B running on his laptop, identifying the game represented by raw neural-network weights. The paper’s authors report that GPT-6 Astra interpreted and edited weights, including a claimed MATH-500 increase from 7.9% to 48.1% after changing 0.016% of Qwen3.5-0.8B’s weights. Separately, a QwenBend project claims that a Qwen model type-checks and executes a Bend program inside its own forward pass using 53 million hand-set weights, without an external interpreter, compiler or tool call.
The laptop reproduction supports the paper’s central claim that a local LLM can reason about raw model weights, while QwenBend extends the demonstration to in-model program execution; however, the scope and tool-free nature of the reported results remain disputed.
Qwen 3.8 27B on my laptop can interpret model weights, too
Simon Willison’s Weblog
Sponsored by: doubleAI — doubleAI’s Maxwell makes software faster as it runs, no code changes needed. Now in early access
Qwen 3.8 27B on my laptop can interpret model weights, too
3rd November 2026