Only 3 of 10 Weight-Space Experiments Read Raw Weights: Critics Split Astra's Qwen Result as Qwen 3.8 27B Names 3 of 4 Games on a Laptop
The “Sparks of Weight-Space Programming” write-up, which reported GPT-6 Astra lifting Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 by editing 0.016% of its weights, now has two public rebuttals and one local replication within two days of release.
The dispute is over which experiments actually count as a model reading raw weights.
The Honnibal breakdown
Matthew Honnibal, the author of spaCy, tabulated the paper’s ten experiments by four properties: whether the model reads raw weights, whether it writes weights, whether it uses Python, and whether it runs the network.
- The “just by looking at the raw weights” setting, with raw floats in context, no tools and no running of the network, appears only in experiments 1, 2 and 4.
- Those three use networks of 17K to 172K weights.
- Every experiment involving a real language model uses Python.
- The Qwen result, the headline one, also runs the model on up to 50 batches of practice problems.
His objection is the abstract, which says Astra can tell what a network was trained to do “just by looking at the raw weights” and, in the same breath, fixes Qwen. A single technique that did both would imply a deep capability, he argues. Here the technique is not single. What the Qwen result shows, in his reading, is that Astra can be prompted to devise and perform tasks a human ML engineer could do, which “is not nothing,” but is a different claim.
His table also carries caveats the abstract skips. The from-scratch MinAtar networks beat PPO at 500M steps only on Breakout (517 vs 25). They lost on Seaquest (53 vs 89) and were roughly level on Asterix (53 vs 55). The Gemma 3 270M variant of the Qwen edit failed. Spotting a sycophantic Gemma from raw weights failed.
Sean Goedecke separately argued that the Qwen3.5-0.8B update follows the same steps a human ML researcher would take.
The laptop test
Simon Willison took the paper’s no-tools experiment, which feeds a policy network’s weights as text and asks which game it was trained on, and ran it on Qwen 3.8 27B. The model is a 17GB Q4_K_M quantisation served from LM Studio on his laptop, and the prompts are the ones published in the authors’ repository. The default 8,192-token context failed, and the timeout had to be raised to 900 seconds.
| Game | Qwen 3.8 27B (laptop) | Time | Stated confidence |
|---|---|---|---|
| Pac-Man | Correct | 46m 12s (43,506 reasoning tokens) | 72% |
| Pong | Correct | 1h 4m 54s | 75% |
| Breakout | Correct | 52m 25s | 88% |
| Freeway | Wrong (answered Pac-Man-like) | not stated | not stated |
For comparison, Kimi K3.1 Pro via OpenRouter named Breakout in 9m 41s using 19,493 reasoning tokens.
Willison read the result as impressive but bounded. The reasoning traces are, in his words, “almost incomprehensible,” lines of channel indices and float values, and he expects “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.
What it changes
Taken together, the three accounts narrow the claim. Reading small policy networks from raw weights, with no tools, works on a four-way multiple choice at the 17K-weight scale, and it is not exclusive to Astra: a 27B open-weights model gets three of four. Editing a 0.8B language model works with code execution and forward passes, and the Qwen edit was a fix for a documented looping failure, not new capability.
The authors’ own table already showed the failures. The argument now is over what the headline should have said.