• Sources: primary, discussion
  • Summary: Honnibal tabulates the post's experiments and finds the raw floats in context setting only in three experiments on networks of 17K to 172K weights. Every experiment on a real language model used Python, and the Qwen3.5-0.8B MATH-500 gain from 7.9% to 48.1% also ran the model on up to 50 batches of practice problems, while the same method failed on Gemma 3 270M.
  • Why it matters: Engineers judging weight-space editing claims should treat the Qwen result as agent-written analysis code, not native weight reading.
  • Comments: Simon Willison separately reports Qwen 3.8 27B running locally can also interpret model weights, and Sean Goedecke argues LLMs are not superhuman at the task.