Qwen 3.8 27B: reading model weights on a laptop

Simon Willison writes that Qwen 3.8 27B, running on his laptop, could interpret raw model weights. He was working from an example in a paper on what the authors call Weight-Space Programming. The paper's first question is whether a language model can work out what a network does just by reading its weights. The authors trained four small networks to play simplified 10x10 grid versions of Breakout, Pong, Pac-Man and Freeway. They then gave the weights, as floating point numbers, to a model and asked which game each network had been trained on. The prompt told the model it had no tools: no code execution, no calculator, a single response.

The paper comes with a GitHub repository, RikiyaT/lm_understand-logs. The exp1_which_game folder holds the prompts used, with a separate file for each game: breakout, freeway, pacman and pong. Willison checked the files and found that, apart from the weights, every prompt is identical. That matters if you want to try it yourself: the difference between one prompt and another is only the numbers describing the network. Willison ran them through his LLM tool, which has a -f/--fragment option that takes the URL of a text file and uses it as a prompt. He passed the raw GitHub URL of each prompt file directly.

The paper's abstract describes larger claims for GPT-6 Astra. According to the authors, it can tell what a network was trained to do from the raw weights. It can also write small game-playing networks from scratch that, they say, outplay networks trained with PPO for 100 million steps, without environment feedback. In another example, asked to improve Qwen3.5-0.8B at maths without gradients, Astra found a problem in the weights that caused reasoning loops and edited 0.016% of them. The authors report MATH-500 rising from 7.9% to 48.1% under the same sampling settings, and IFEval tripling.

Not everyone read the results the same way. Willison notes that the paper mixes two subjects he thinks should be kept apart. The first is what a model can do with raw weights, with no tools and no code. The second is how well a model, once it can run code, can carry out operations to update and improve other models. Matthew Honnibal, author of spaCy, called the paper misleading for merging the two, Willison writes. Sean Goedecke went through the Qwen3.5-0.8B example and described it as taking the same steps a human machine learning researcher would take. Willison says he focuses on the no-tools question, which he finds the more interesting of the two.

The four prompt files are plain text on GitHub, and each one carries the weights of a network trained on one of the grid games. If you have a local model and a tool that can read a prompt from a URL, you can send the same question the paper used and see what comes back. Willison's own result is that a 27B model on a laptop answered it. The paper's bigger numbers for GPT-6 Astra come from the authors, and the two readings of the Qwen3.5-0.8B example remain disputed.