Qwen 3.8 27B interprets model weights on laptop
The gist
Blogger Simon Willison reports Qwen 3.8 27B on his laptop identified a Pac-Man-style game from raw weights in 46m 12s.
The test suggests small local models can inspect other networks' weights without tools.
How it unfolded
- Nov 2, 2026The paper "Sparks of Weight-Space Programming" is highlighted, reporting that frontier models, in particular GPT-6 Astra, can interpret and modify raw model weights.
- Nov 3, 2026Simon Willison reports Qwen 3.8 27B on his laptop identified a simplified Pac-Man-style 8x8 gridworld from raw weights after 46m 12s and 43,506 reasoning tokens, at 72% confidence.
- Nov 3, 2026In the same run, the model identified Pong in 1h 4m 54s; Kimi K3.1 Pro guessed Breakout in 9m 41s and 19,493 reasoning tokens.