Willison: Qwen 3.8 27B names 3 of 4 game nets on laptop

Simon Willison has replicated a widely-circulated weight-space-reading demo on his own laptop, running Qwen 3.8 27B (the Q4_K_M.gguf quantization pulled from Hugging Face, under LM Studio), and the model correctly named three of four small game-playing policy networks from the raw parameters alone. Writing on his blog, Willison reports Pac-Man identified at 72% confidence after 46 minutes and 43,506 reasoning tokens, Pong at 75% after 1 hour 4 minutes, and Breakout at 88% after 52 minutes. Freeway was the miss, which Qwen guessed as "A simplified Pac-Man-like grid game."

The experiment replicates "Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights" by Rikiya Takehi and colleagues, whose headline claim is that "Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do." That paper has drawn pushback from Matthew Honnibal, who Willison notes "called the paper misleading for conflating the two": analysis without tool use versus analysis with code execution.

A sample of Qwen's internal reasoning hints at what the model is actually doing: "What about agent channel0? We see some large channel0 rows near row8 col7/8 and maybe bottom?"

Willison declines to read his replication as vindication of the paper's bigger claims. "It's a cool paper, but some of the wider commentary I've seen around it does not feel justified. I'm not ready to declare that the age of source code is over and our future involves incomprehensible machines editing their own model weights in incomprehensible ways," he writes.