Simon Willison tested whether Qwen 3.8 27B could identify a game by examining the raw weights of a small neural network, without running code or using other tools. Running the 17GB model on his laptop, he says it correctly identified Pac-Man, Pong and Breakout, but mistook Freeway for a Pac-Man-like game.
The successful answers took 46 minutes to more than an hour each. Willison’s experiment tests interpreting weights; it does not show Qwen editing or improving another model, a separate claim explored in the paper he was following.
Why it matters
For people running large models locally, the test suggests they may be able to infer what a small network does from its weights, though it was slow and failed on one of four games.
