Simon Willison gave Qwen 3.8 27B lists of weights, the learned numbers inside small neural networks, and asked what each network did. Running on his laptop, Qwen identified three training tasks without tools or code execution. It also got one wrong.
- The successes: Pac-Man in 46m 12s, Pong in 1h 4m 54s, and Breakout in 52m 25s. The Pac-Man run used 43,506 reasoning tokens, the model’s intermediate working text.
- The failure: Qwen called the Freeway network a Pac-Man-like game. Willison suggests the game’s obscurity might matter; that explanation remains untested.
- The setup: a roughly 17GB downloadable model through LM Studio. The default 8,192-token context setting failed, so he increased the space available for the prompt and response.
Willison separates this experiment from the original “Weight-Space Programming” post’s claims about editing other models with code. His linked prompts and transcripts make the narrower identification result inspectable.
Caveat: these four cases test recognition of small game-playing networks, not reliable interpretation or self-improvement of large language models.
The 137-comment thread on Hacker News adds questions for a stronger follow-up.
What the thread adds
- boznz — asks about damage to other capabilities when weights are edited: “has there been any study on What are these changes are breaking ?”
- NitpickLawyer — requests a different reproduction: “could you try reproducing Experiment 5 from the blog post, if you have a chance?”
- sagex — challenges the task’s difficulty: “Isn’t the network too shallow to be challenging to decode?” This is a criticism, not a demonstrated counter-result.
The recurring question
Two commenters, data-ottawa and tarruda, ask whether Qwen could modify another Qwen model or itself. Willison’s local experiment tests identification; that editing question remains open here.
HN handles are pseudonymous; HN publishes no per-comment scores. The sampled comments follow HN’s ranking, and this is a slice of the thread, not a consensus.