Qwen identifies three games from model weights in Simon Willison’s laptop test

An AI running on Simon Willison’s laptop identified Pac-Man, Pong and Breakout from the numerical weights of networks trained to play them, without tools or code execution. His November 3 experiment with Qwen 3.8 27B brought a recent research task to a local model. It produced one wrong answer; successful responses took roughly 46 to 65 minutes.

A game hidden in floating-point numbers

Willison borrowed the task from Sparks of Weight-Space Programming, a paper by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The researchers trained four neural networks to play simplified, 10-by-10 grid versions of classic games, then presented their weights as text to a language model.

Weights are the numerical parameters that govern a network’s behavior; the challenge was to work backward from those numbers to identify the game it had learned. The prompt asked which game the network was trained on, prohibited code execution, calculators and other tools, and required a single response.

The paper’s accompanying GitHub repository supplies self-contained prompts for Breakout, Freeway, Pac-Man and Pong. Willison checked that the prompts were identical apart from the weights. That made the numerical content, rather than different instructions for each game, the distinguishing input in his tests.

Three recognizable games, one mistaken identity

Qwen’s successful answers went beyond naming titles: they described layouts, objects and actions associated with each game. Those descriptions are model outputs reported by Willison, not a separate verification that every inferred detail matched the networks. His published results show the following:

• Pac-Man: Qwen described an agent navigating a maze, collecting reward targets and avoiding moving hazards.
• Pong: It identified a right-side paddle moving vertically to intercept or return a moving ball.
• Breakout: It described a bottom-row paddle, a ball and bricks or reward targets in the upper rows.
• Freeway: It incorrectly called the network a simplified Pac-Man-like game with four-directional movement.

Willison said the Freeway miss did not surprise him because that game is more obscure than the other three. The successful answers were slow, too: the Pac-Man response alone used 43,506 reasoning tokens, the units used to measure the model’s generated reasoning text.

The local setup also needed adjustment. Willison used a 17 GB version of Qwen in LM Studio, connected through his LLM command-line tool and its LM Studio plugin. The default 8,192-token context window failed with errors, so he loaded a longer context window and increased the timeout.

Reading weights is not rewriting them

The task was recognition, not repair: Qwen was asked to name the game, not change the network. Willison deliberately separated that no-tools question from another part of the paper, which examines models using code to update and improve other models.

The paper, as quoted in Willison’s post, reports that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights, raising its MATH-500 score from 7.9% to 48.1%. This was a different result from identifying a game: it involved another target model and a weight-editing task.

Willison found the game-identification result impressive but argued that an experienced human with pencil and paper could plausibly work it out with enough effort. He rejected the leap from these experiments to declaring source code obsolete or a future of machines incomprehensibly editing their own weights.