Post

Qwen 3.8 27B on my laptop can interpret model weights, too

Simon Willison tried the raw-weight game-identification prompts from the recent weight-space article with a local Qwen 3.8 27B model. It identified Pac-Man, Pong and Breakout, but mislabeled Freeway; each successful run took roughly 46–65 minutes.

This is a small four-example experiment, not evidence that the model understands arbitrary networks. Unlike the separate Qwen weight-editing experiment, this test used no code execution. HN readers debated what the test demonstrates and how to verify the explanations.