Qwen read three game networks from raw weights on a laptop

Qwen 3.8 27B running on a laptop correctly identified three of four small neural networks from their weights, without executing code or using a calculator. Simon Willison reproduced one part of the Weight Space Programming experiment with a 17 GB local model.

Three correct games show that a local model can read some signal in weights

The original researchers trained four networks for simplified versions of Breakout, Pong, Pac Man and Freeway on a 10 by 10 grid. The model received the same question each time with a different list of weight values. It could not run tools or code.

Willison's Qwen identified Pac Man in 46 minutes and 12 seconds after 43,506 reasoning tokens. It found Pong in 1 hour, 4 minutes and 54 seconds and Breakout in 52 minutes and 25 seconds. It failed on Freeway and described it as a Pac Man variant.

The original work goes further. In a separate experiment, GPT 6 Astra edited 0.016% of Qwen3.5 0.8B's weights and raised its MATH 500 score from 7.9% to 48.1%. That test used Python and forward passes, so it should not be conflated with reading numbers directly in a prompt.

Model interpretation benefits from separating weight reading from weight editing

The ability to infer the behavior of a small network directly from weights could complement mechanistic interpretability. It could give researchers another way to look for a learned trait or failure without a conventional evaluation run. Willison's test also suggests that the first hint of this capability is not confined to the most expensive closed model.

Precision matters here. Reading a static list of weights without tools and changing a model with Python are different capabilities. Matthew Honnibal criticized the original work for combining them, while Sean Goedecke argued that the Qwen editing procedure resembles the steps a human ML researcher would take.

Four tiny networks do not establish general understanding

The experiment covers four small networks and a familiar choice among four games. Three hits and one miss are far too small a sample for a claim that LLMs generally understand weights. The successful runs also took 46 to 65 minutes and required a longer context window because the default 8,192 tokens failed.

A correct label alone does not reveal whether the model found a stable internal representation of the game or assembled enough clues for a good guess. On Freeway, the same process produced a plausible sounding error.

Blind tests across architectures and tasks will settle the practical value

The next step needs a larger benchmark with unseen networks, multiple architectures and tasks that do not offer a short list of familiar choices. Researchers should also measure repeatability, time, token use and whether an explanation predicts network behavior on a new test.

The technique becomes useful if weight analysis finds a failure that ordinary evals miss and another team reproduces the result. Until then, this is a promising laboratory trace, not the end of source code.

Lilith's verdict

Qwen was handed a pile of numbers and pulled the right game screen out of it three times. Until it can do the same reliably with an unseen network, the researcher will keep hold of the lab notebook.