Qwen 3.8 27B on my laptop can interpret model weights, too
Simon Willison, writing on his weblog on November 3, 2026, tested a claim from a paper posted the day before. The paper, Sparks of Weight-Space Programming by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, said GPT-6 Astra can read a neural network's raw weights as text and infer what it was trained to do. Willison ran Qwen 3.8 27B on a laptop with only the weights of four small game-playing networks, and no code execution or tools.
The local model named three of the four games. It gave Pac-Man 72% confidence after 46 minutes and 43,506 reasoning tokens, Pong 75% after just over an hour, and Breakout 88% after 52 minutes. It missed Freeway, calling it a simplified Pac-Man-like grid game. Willison ran a model of about 17 GB in LM Studio, raising the context window above 8,192 tokens and extending the timeout.
The paper's most concrete editing result changed 0.016% of the weights in Qwen3.5-0.8B without gradients, lifting its MATH-500 score from 7.9% to 48.1%. Willison noted that the authors label their work a blog post, though it is paper-shaped, and cautioned that his test does not show models can improve themselves by editing weights.
Why it matters
For founders and operators, the test suggests that some model inspection work may run on open weights on ordinary hardware instead of a paid frontier API. A 27B model that takes close to an hour per answer is slow, though, so the cost case depends on how the task is built and how often it runs. Teams that ship on open-weight models can check whether the reading task holds for their own checkpoints before changing any tooling.
For investors, the evidence is thin: one laptop run per game, a single editor model in the source paper, and no independent replication yet. Diligence questions should cover which model and hardware were used, how many trials were run, whether results hold on tasks the authors did not pick, and whether any claim depends on frontier-only capability. Claims that a model can edit another model's weights also raise provenance and integrity questions for companies that distribute model files.
Key points
- 1Willison's run used Qwen 3.8 27B, an open-weight model released August 14, 2026.
- 2The model named Pac-Man, Pong and Breakout from raw weights, but missed Freeway.
- 3Confidence levels were 72% for Pac-Man, 75% for Pong and 88% for Breakout.
- 4Runs took 46 to 65 minutes on a laptop, with no tools or code execution allowed.
- 5The paper's editing test raised MATH-500 from 7.9% to 48.1% by changing 0.016% of weights.
