Qwen 3.8 27B reads raw weights, names Pac-Man at 72%

The Sparks of Weight-Space Programming paper, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, bundled two claims: that frontier models can read raw weights without tools, and that with code they can edit other models to improve them. Simon Willison separates the two and tests only the first, using the paper's own exp1_which_game prompts, which differ only in the weights they contain.

His local trial with Qwen 3.8 27B is a single blogger's run, not a controlled benchmark, and it missed Freeway, the most obscure of the four Atari 2600 titles. Still, the fact that a 17GB model loaded in LM Studio with a long context window could name three games from floating point numbers alone, in 46m 12s to 1h 4m 54s per prompt, is the kind of replication that invites others to try.

Willison is careful about the wider narrative. He notes Matthew Honnibal's criticism that the paper conflates weight-reading with code-assisted updates, and Sean Goedecke's read that the Qwen3.5-0.8B fix follows steps a human researcher would take. His conclusion is that the paper is cool but does not justify declaring the age of source code over.