• Sources: primary, discussion
  • Summary: Willison ran the published exp1_which_game prompts from the Sparks of Weight-Space Programming paper against Qwen 3.8 27B in LM Studio. Each run took 46 minutes to over an hour and tens of thousands of reasoning tokens, and the model correctly identified three of four games. He separates this no-tools result from the paper's tool-assisted weight edits, which Honnibal and Sean Goedecke criticized.
  • Why it matters: It shows the paper's weight-reading claim is reproducible with an open model on commodity hardware, while the weight-editing headline stays contested.