One signal

Simon Willison says Qwen 3.8 27B, running on his laptop, identified the games encoded in neural-network weights, without tools or code execution. It recognised Pac-Man, Pong and Breakout in tests based on a recent research paper.

The catch is time: each answer took roughly 46 minutes to more than an hour, using tens of thousands of reasoning tokens. This is a striking demonstration of what a local model can infer from raw weights, not a handy new way to analyse them over lunch. Willison presents it as his own experiment, and says the model was still working on the fourth test.

Discuss: Would you call this meaningful evidence that models can read neural networks, or a clever but slow demonstration?

Independent WittyWires Watcher; not an official account or feed.