A November 2 blog post from four MIT CSAIL researchers, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, claims that OpenAI's GPT-6 Astra can "natively read & write neural network weights." The most-shared result was an edit to the small open model Qwen3.5-0.8B: according to the authors, its MATH-500 score went from 7.9% to 48.1%. On November 3, a rebuttal calling the framing misleading was the top story on Hacker News.
Why it matters: if a frontier model can look at a network's weights and say what it does, that's a new tool for auditing models. That covers spotting a sycophantic fine-tune or a hidden behavior. But the post mixes two quite different setups, and the most dramatic number comes from the one that is least about "reading weights."
Two setups, not one
The post says it's describing early results. It isn't a peer-reviewed paper, and no arXiv version is linked. Going experiment by experiment, Astra had one of two kinds of access:
| Setup | Experiments | Network size | What Astra had |
|---|---|---|---|
| Raw weights as text, no tools | Identify which Atari-style game a policy plays; identify a diffusion model's shape; fix a failing car-on-ramp policy; write game policies from zero | 17K–172K weights | Weights pasted into the prompt; one reply; no code, no forward passes |
| Python sandbox over the weights | Language and copying-head analysis; Go network; spotting a sycophant; Qwen math edit | 7M to 0.8B weights | Code to inspect and rewrite tensors, and in some cases a budget of forward passes |
The no-tools results are striking. The authors report that Astra named the correct game for all four 17,253-weight policies. They also say a one-shot edit to a 171,664-weight car policy took its success rate from 0% to 100% over 1,024 episodes.
What the Qwen result actually was
For Qwen3.5-0.8B, Astra worked in a Python kernel over the weights. It got about 15,000 practice problems from the GSM8K and MATH training sets, and the authors say it never saw the benchmark problems. No gradients or training loops were allowed. Astra traced the model's tendency to loop endlessly to one direction written by the MLPs in layers 15 and 19, and subtracted it. That edit touched 131,072 weights, or 0.016% of the model.
How much it helped depended on the number of forward passes Astra was allowed:
| Qwen3.5-0.8B | MATH-500 | GSM8K | IFEval | Hit token limit |
|---|---|---|---|---|
| Base | 7.9% | 13.8% | 21.0% | 93.0% |
| Edited, 10 passes | 12.3% | 25.7% | 49.0% | 87.2% |
| Edited, 25 passes | 14.1% | 25.1% | 39.9% | 86.4% |
| Edited, 50 passes | 48.1% | 67.1% | 62.3% | 33.1% |
The authors read this as removing an obstacle, not adding a skill. The base model ran out of tokens on 93% of MATH-500 problems. With the loops cut, it could finish answers it already knew how to reach. As a control, the same approach on Gemma 3 270M, which rarely loops, barely moved its MATH-500 score (6.8% to at most 7.7%).
The pushback
The rebuttal on Hacker News came from Matthew Honnibal, creator of the spaCy library. His point, as Simon Willison summarizes it, is that the post conflates two topics. The raw-weight reading happens only on small networks. The Qwen fix is an LLM writing analysis code, which is closer to what a human ML engineer does. Willison also cites Sean Goedecke, who made the same comparison to a human researcher. Commenters split: some agreed the title oversells the work, while others said the body explains the setups clearly.
Willison then tried the no-tools game-identification test on his own laptop with the open Qwen 3.8 27B. It correctly named Pac-Man, Pong and Breakout, but each took 46 minutes to 65 minutes of reasoning. He called that impressive. He also argued that a patient human with pencil and paper could get there too.
What to take from it
- The interesting claim is the small one. Frontier and even mid-size open models can infer what a tiny network does from its weights alone. That's surprising, and Willison's partial replication suggests it isn't unique to Astra.
- The big number is a tool-use result. The Qwen gain came from code, practice data and 50 forward passes, and it fixes a known looping problem.
- It's unreviewed. The authors have published full experiment logs on GitHub and the edited Qwen model on Hugging Face. Independent replication is what to watch for, especially on bigger networks without tools.
GPT-6 Astra itself launched on September 3, according to Constellation Research. The post doesn't say how the researchers accessed it.
Sources
- Takehi, Teoh, Zhang, Khattab (MIT CSAIL) — Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights rikiyat.github.io
- Simon Willison — Qwen 3.8 27B on my laptop can interpret model weights, too simonwillison.net
- Hacker News — The 'Astra can read and write weights to improve Qwen' post is misleading (discussion) news.ycombinator.com
- SmartScope — GPT-6 Astra: research on directly reading and writing AI weights boosts Qwen math scores smartscope.blog
- Constellation Research — OpenAI launches GPT-6 Astra constellationr.com
