Qwen 3.8 27B Identified Three Games From Neural Network Weights on an Ordinary Laptop

Neural network weights usually look like a gigantic table of decimal numbers to a human. To understand what a model has learned, an engineer runs it on test data, plots activation graphs, or examines its architecture with specialized tools. A new experiment points to another path: simply show a language model the weights of a small network and ask it to explain what the network does.

Simon Willison tested this scenario with Qwen 3.8 27B running locally on a laptop. The model received the parameters of four small game-playing agents and the same question each time: which game was this network trained for? No Python, calculator, or model execution—just one long prompt filled with numbers.

The result was unexpectedly strong. Qwen correctly identified Pac-Man, Pong, and Breakout. It got Freeway wrong, mistaking the game for a Pac-Man variant. In other words, the local model solved three out of four tasks, although each analysis took between 46 minutes and more than an hour.

What exactly the model did

Imagine a small network that controls the paddle in Breakout. Some inputs describe cells on the game board, while other layers connect the ball's position to the commands “left” and “right.” After training, these relationships remain embedded in the weight matrices.

Qwen attempted to reconstruct them manually: it searched for large coefficients, matched them to cell coordinates, and checked which objects might affect the agent's actions. For Breakout, it arrived at the correct picture: the paddle sits at the bottom, moves horizontally, and bounces the ball toward the blocks. Its confidence in the answer was 88%.

For Pong, the model described a vertical paddle on the right and a ball moving across the field, reporting 75% confidence. For Pac-Man, it recognized a maze, targets, and moving threats with 72% confidence. This was not merely guessing a title from the prompt format: all four tasks differed only in their sets of weights.

Why this matters to developers

The practical value is not in replacing a debugger with a chatbot. A much more useful capability is quickly forming an initial hypothesis about an unknown or damaged model. For example, an engineer could ask an LLM to find unusually strong connections, explain the purpose of a small layer, or compare two sets of weights after fine-tuning.

This approach could potentially help audit small models, investigate regressions, and analyze neural network components whose documentation has been lost. A language model can connect numbers to understandable engineering concepts and suggest where to look next. Its conclusions would still need to be checked by running the model and using conventional diagnostic tools.

There is also a bolder scenario: editing weights without full-scale training. The authors of the original work reported that GPT-6 Astra changed 0.016% of Qwen3.5-0.8B's parameters and raised its MATH-500 score from 7.9% to 48.1%. Critics note, however, that some such experiments involve code and operations familiar to conventional ML researchers. Reading weights without tools and automatically repairing a model should therefore be treated as two different tasks.

Where the boundary lies

The experiment cannot yet be called a reliable reverse-engineering method. It tested compact game-playing networks with understandable inputs and a small number of possible answers. Even under these conditions, Qwen got one of the four games wrong. Analyzing each example required tens of thousands of tokens and nearly an hour of computation, while parts of the reasoning are difficult to verify.

The task is far harder for large language models: billions of parameters cannot fit into a prompt in full, and an isolated fragment of weights rarely has an unambiguous meaning. Moreover, confident model output does not guarantee a correct interpretation.

The practical takeaway is simple: treat an LLM as an additional analytical tool, not as a replacement for experimentation. Start with small networks, known control examples, and specific questions. If the model consistently finds patterns, its conclusions can guide the next test. Reports of the end of source code are still premature, but a new class of neural network auditing tools is already beginning to emerge.