November 2 saw an AI-circles sensation: the paper "Sparks of Weight-Space Programming" claimed GPT-6 Astra can tell what a neural network does just by looking at its raw weights, and can hand-edit 0.016% of Qwen3.5-0.8B's weights to lift its MATH-500 score from 7.9% to 48.1%. Within 24 hours, two veteran developers delivered audits pointing in opposite directions: one shows a laptop 27B can do the slow version of the trick, the other argues the viral framing is seriously misleading.
The laptop replication: a 27B reads weights too
Simon Willison skipped the frontier APIs and loaded a 17GB quantized build of Qwen 3.8 27B into LM Studio, then ran the public prompts straight from the paper's GitHub repo: look only at the floating-point weights and guess whether the policy network was trained to play Breakout, Pong, Pac-Man or Freeway.
Results: Pac-Man identified in 46m 12s at 72% confidence; Breakout in 52m 25s at 88%; Pong in 1h 4m 54s at 75%; Freeway wrong. Kimi K3.1 Pro via OpenRouter nailed Breakout in 9m 41s with 19,493 reasoning tokens.
Willison stays measured. The reasoning traces are "fun to skim, and almost incomprehensible", but "a sufficiently experienced human being armed with a notepad and a pencil could get there with enough effort". He is "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 also argues the no-tools question is the more interesting of the paper's two topics.
The dissection: ten experiments, three techniques
The same day, spaCy creator Matthew Honnibal published a rebuttal bluntly titled "The 'Astra can read and write weights to improve Qwen' post is massively misleading".
He sorted the post's 10 experiments into a table by method. Only experiments 1, 2 and 4 truly satisfy "just looking at the raw weights" — no Python, no running the network — and those networks hold just 17K-172K weights. Every experiment involving a real language model uses Python, and the headline Qwen math fix lets the model run itself on up to 50 batches of practice problems first. The from-scratch game networks beat PPO-500M only on Breakout (517 vs 25), losing on Seaquest and Asterix; the same weight-repair recipe failed outright on Gemma 3.
His analogy is sharp: a stage magician performs the same trick with three different techniques, so each performance seems to debunk your theory of how it works. Yet he explicitly credits the Qwen result as "work a human ML engineer could do" — and something models of a year or two ago would have failed at. His objection is not that the results are fake, but that three distinct techniques were packaged into one narrative of a deep capability.
So what
Put together, the two posts tell the complete story: the capability is real, and so are its limits. A local 27B can do the slow version of weight reading, so the trick is not reserved for flagship APIs; but the unified "natively understands weights" story is, so far, only evidenced on networks of a hundred-odd thousand weights and one dissected headline result.
Extraordinary claims are priced by replication and dissection. This time the community settled the account within 24 hours. Next time you see "models natively rewrite weights" in a headline, ask three questions: how big is the network, did it use Python, and did it run practice problems first?
References: Willison's replication; Honnibal's dissection; the original post.