Laptop Qwen identifies three of four game networks from their weights

Developer Simon Willison reported on November 3 that Qwen 3.8 27B running on his laptop identified three small game-playing networks from their weights. It recognized Pac-Man, Pong and Breakout, but missed Freeway. His experiment revisits a task previously demonstrated with a frontier model using a locally running alternative.

The numbers supplied belonged to small networks trained to choose game actions. Weights are the numerical values learned during training; the task asks what behavior they encode. This was not Qwen inspecting or rewriting its own 27 billion parameters. The community-converted model file he linked is listed at 16.8GB.

In the released responses, Qwen described Pong as a right-side paddle moving vertically to intercept a ball. For Breakout, it identified a bottom paddle moving horizontally beneath bricks. Its Pac-Man answer described maze walls, moving objects and reward targets.

The successful runs took 46 minutes 12 seconds for Pac-Man, 52 minutes 25 seconds for Breakout, and one hour four minutes 54 seconds for Pong. The Pac-Man record counts 43,506 reasoning tokens, units of text processing. Willison says the default 8,192-token context failed with errors, so he used a longer context window.

Freeway produced an incorrect Pac-Man-like answer. That transcript cites top-row rewards, moving objects, a previous frame and an upward action preference. The response assigned itself 70% confidence, a self-assessment written by the model. The released log records one hour 16 minutes 17 seconds and 69,787 reasoning tokens for that failed run.

The grid description supplied with the weights

The prompt provides substantial context for interpreting the numbers. It describes a 10×10 grid with a wall border and an 8×8 playable interior. Each cell has five channels: player, moving objects, rewards, obstacles and previous-frame information. Input positions are mapped to cells.

Five possible actions are specified: wait, up, down, left and right. Requested fields cover observations, reasoning, game, behavior and confidence. No candidate-game list is supplied; code execution, calculators and other tools are prohibited. Comparing the four released prompts shows identical instructions before the weights. Values are rounded to three decimals. The researchers separately release the full-precision network files and the numerical text shown to the model.

The original researchers’ documentation specifies 17,253 parameters per policy network, a small model selecting actions from game observations. All four games use the same architecture and training settings. The researchers trained three seeds per game and selected the network with the highest mean reward.

The original team reported that GPT-6 Astra identified all four networks. Its scorer matches game names against the answer key and accepts Frogger, another road-crossing game, as Freeway. The scorer reads the answer key separately; the model’s prompt omits game names.

In a separate November 3 critique, spaCy developer Matthew Honnibal argued that the original presentation combines materially different experiments. Reading numerical weights directly differs from using Python to analyze and modify language-model weights. He acknowledged technical progress in the latter work while challenging an interpretation that treats the methods as one capability.

Willison explicitly focuses on the no-tools identification question. The earlier mathematics experiment measured the performance of an edited language model; this reproduction asks which game a given network learned. It is a new test with an existing model, rather than a Qwen release or general-purpose interpretation product.

The community model card distinguishes Qwen, the original creator, from the LM Studio team providing the converted file, and lists Apache 2.0 licensing. Willison released the successful responses, reasoning and usage records together, plus a separate transcript preserving the Freeway failure.