Qwen 3.8 27B on my laptop can interpret model weights, too
Qwen identified three of four games from neural-network weights alone, without tools or code execution.
Willison’s laptop running Qwen 3.8 27B correctly identified Pac-Man, Pong and Breakout, taking roughly 46–65 minutes per response. It failed on Freeway, calling it a Pac-Man-like game. His experiment tested whether a model could infer behavior from raw weights, separate from the paper’s experiments with modifying weights. Willison calls the results impressive but rejects claims that they signal the end of source code.
Simon Willison's note
Willison’s laptop running Qwen 3.8 27B correctly identified Pac-Man, Pong and Breakout, taking roughly 46–65 minutes per response. It failed on Freeway, calling it a Pac-Man-like game. His experiment tested whether a model could infer behavior from raw weights, separate from the paper’s experiments with modifying weights. Willison calls the results impressive but rejects claims that they signal the end of source code.
Simon Willison's note
score 5