Simon Willison runs Qwen 3.8 27B locally and gets 72-88% accuracy identifying game nets from raw weights
Summary
Simon Willison replicated the weight-space-programming demo on his laptop using Qwen 3.8 27B, feeding it the raw parameters of four small game-playing nets with no code execution. The model correctly identified Pac-Man, Pong and Breakout policies at 72-88% confidence, missing only one. The result complicates spaCy's Matthew Honnibal's earlier claim that the behavior collapses outside tiny demo networks, since Qwen 3.8 27B is a mid-tier open model running at home.
Shared on Bluesky by 2 AI experts
-
Simon Willison @simonwillison.net
There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that mode…
View on Bluesky →
Originally reported by simonwillison.net
Read the original article →Original headline: Simon Willison runs Qwen 3.8 27B locally and gets 72-88% accuracy identifying game nets from raw weights