> ## Content Index > Fetch the complete content index at: https://varops.com/llms.txt > Use this file to discover other available public pages before exploring further. # In MIT's GPT-6 Astra paper, reading raw weights and fixing Qwen with Python are two different skills - URL: https://varops.com/in-mits-gpt-6-astra-paper-reading-raw-weights-and-fixing-qwen-with-python-are-two-different-skills/ - Published: 2026-11-04T13:21:50.000Z - Updated: 2026-11-04T13:21:50.000Z - Description: The week's hottest AI paper says GPT-6 Astra fixed a model by editing its weights. The raw-weight reading was real, on tiny networks. The Qwen fix used Python, 50 practice batches and a known steering move. - Author: Nix Nullty - Tags: Overhyped *Yesterday North priced the cost of AI-found bugs on old software versions. Today the question flips: can an AI fix a model by editing its weights? An MIT team says GPT-6 Astra did, and lifted a small Qwen model from 7.9% to 48.1% on a math benchmark.* [***Nix Nullty***](https://varops.com/columnist/nix/) *reads the post's own experiment setups and separates the test that used raw numbers from the one that used Python and practice runs. The authors get credit for publishing their logs and a failed follow-up. The piece ends on three questions to ask before this changes a fine-tuning budget. —* [*Muximus*](https://varops.com/columnist/muximus/) --- One of the week's most-discussed AI research posts says OpenAI's GPT-6 Astra can tell what a neural network does "just by looking at the raw weights", and that it lifted a small Qwen model's math score from 7.9% to 48.1% by "natively editing 0.016% of the weights". Both results are real. They come from different experiments with different tools, and the headline number belongs to the one an ML engineer would recognize as their own job. Verdict: partial. Reading tiny networks with no tools holds up, and one of those tests has already been repeated on a laptop model. The jump from there to "AI can now fix other models by editing their weights" does not. ## The claim, at full strength Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT CSAIL published "Sparks of Weight-Space Programming" on November 2\. [The abstract makes two moves in a row](https://varops.com/machine-checked-and-proved-are-two-different-claims-for-ai-written-math-openais-own-errata-show/). First: "Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do." Then, on Qwen3.5-0.8B, Astra "fixes this by natively editing 0.016% of the weights", which "raises MATH-500 from 7.9% to 48.1% and generalizes out of the box to instruction following, raising IFEval by 3x." All of this, it says, happened "without allowing Astra to write training loops." Here is the version its fans would sign. A frontier model can treat weights the way it treats code - read them, write them, patch them - and repair a model without retraining it. If that holds, model audits and fine-tuning budgets both change. That is a big claim. It deserves a close read of the setups, not the summary. ## Sort the experiments by what Astra was allowed to touch The post runs eight experiments plus a follow-up. Matthew Honnibal, the author of the spaCy library, [sorted them by tool access](https://varops.com/a-benchmark-scores-the-harness-as-much-as-the-model-openais-own-number-swings-3x/) and called the post misleading. Once the table is laid out, the sorting does most of the arguing. Astra worked from raw weights with no tools in three experiments: naming which game a network was trained to play, naming which shape a small image model draws, and fixing a stuck driving policy. A fourth had it write small game networks from scratch, also with no tools. All four networks hold between about 17,000 and 172,000 weights. A fifth, on 7.2-million-parameter character-level language models, already gave it Python to analyze the weights. The Qwen model holds 0.8 billion. That is roughly 46,000 times the game networks. Every experiment on a real language model gave Astra Python. The post says so. The abstract opens by saying frontier models read and write weights "using a variety of recent tools and algorithms or even via direct weight editing", and a later section headed "Real-World Relevant Scenarios (Scaling Up and Tool-use)" notes that so far, for most networks, "we have not enabled tool use". Nothing is hidden. The abstract simply runs both halves together and gives them one name. That one name is the product being sold. ## What actually fixed Qwen In the Qwen experiment, Astra sat in a Python session with the weights loaded and could run the model on up to 50 batches of practice problems from the GSM8K and MATH training sets, up to 512 problems a batch. The test questions were withheld. The base model looped. 85 of its first 96 practice answers ran to the 32,768-token limit without an answer, a weakness Qwen's own model card warns about. Astra traced the loop to one direction written by two layers and subtracted it from 64 columns in each - 131,072 weights. The post itself compares this to "removing a behavior by deleting a single direction from a model's weights". Sean Goedecke, a software engineer, calls this steering, a known technique: write examples that loop and examples that do not, measure the difference inside the model, find the weights that drive it, cut them back. His verdict on Astra's run: "exactly what a competent human researcher could have and would have done." It was not quick, either. According to R&D World, the successful run took 80 Python executions over 20.3 hours of wall-clock time, including a 4.7-hour service outage. With 10 or 25 practice batches instead of 50, the score reached only 12.3% and 14.1%. Then the authors draw the line themselves, which is the most useful sentence in the post: "Fixing loops, however, is not the same as making a model better at math: it lets the model finish answers it could already work out." They tried the same method on Google's Gemma 3 270M, which rarely loops - 8.5% of its answers hit the limit, against 93% for Qwen. Astra could not improve it. No edit moved its MATH-500 score by more than about a point. In their words, the edit "could not add an ability the model lacked." ## Credit where it's due The no-tools reading is not a trick. Simon Willison ran the paper's published game prompts against Qwen 3.8 27B on his laptop, no tools. It named Pac-Man, Pong and Breakout, taking 46 minutes to over an hour each, and missed Freeway. So a model on a laptop can do at least that part of the first half. Willison keeps it in proportion himself: "a sufficiently experienced human being armed with a notepad and a pencil could get there with enough effort." Goedecke is less impressed by the game test. The game networks' inputs are labeled per grid cell, so a heatmap of first-layer weight strength shows the paddle row or the maze. He builds it in three lines of Python. That Astra did the arithmetic in its head is, in his words, "superhuman mental mathematics, not superhuman AI research." The Qwen number is real as well. A separate rerun by akhaliq at Hugging Face, reported on the edited model's page, scored 51.0% on MATH-500\. Nobody serious is disputing that the number holds. The fight is over what produced it. The from-scratch game networks need one footnote. They beat PPO trained for 100 million steps in all three games, as the post says. Against PPO trained for 500 million steps they won only Breakout (517.4 to 25.3), lost Seaquest (53.1 to 88.9), and lost Asterix by a nose (53.2 to 55.2). And the authors earned this section. They published prompts and logs, reported the Gemma follow-up that failed, and called their own results "very early". That is why three outsiders could check the work within a day of publication. ## Where the two halves get stapled together The trouble is the bridge, and the paper lays some of the planks. Its closing section imagines that "we may just prompt 'Can you add Capability X to the model?' and the LLM can directly tweak some weights." Its own Gemma result argues against that: the method removed an obstacle and could not add an ability. R&D World summarized the authors' proposed uses as including edits that "could replace some fine-tuning runs". Read only the abstract and you get the mystique of the first half and the number from the second. One hole in this piece, stated plainly. As of November 4 we could not find any response from the authors to Honnibal or Goedecke. And we could not verify Goedecke's claim that the original model already solved every problem that fit under 32,000 tokens. It is his claim, and he does not cite a source for it. ## Hype-o-Meter: "AI can now read and edit model weights" **"Astra reads raw weights": held, on networks of about 17,000 to 172,000 weights.** Real, reproduced in part, and impressive as mental arithmetic. Not yet a tool for anything a business runs. **"Astra fixed Qwen's math by natively editing weights": overstated.** It fixed one known looping bug, with Python, 50 practice batches and about 20 hours of wall-clock time - work the critics say a competent ML engineer would do the same way. The same method did nothing for Gemma. ## Three questions before this changes a plan When a vendor, a board member or an internal champion says AI can now read and edit model weights - and so fine-tuning, audits or retraining budgets should change - ask: 1. [**Which experiment produced the headline number**](https://varops.com/how-to-read-an-ai-patch-benchmark-when-a-third-of-its-trials-forbade-the-agent-from-testing-2/)**, and what did the model have in that one?** Raw numbers only, or Python and test runs? 2. **How big was the network in the no-tools test**, compared with the models the team actually runs? 3. **What did the fix repair** \- a missing skill, or one known bug - and did the same method work on a second model? On this paper the answers are: Python and 50 practice batches; about 17,000 to 172,000 weights against 800 million; one known looping bug, and no, it did not work on Gemma. Asked that way, "AI can edit model weights" turns into two smaller claims a buyer can use. Frontier models can read very small networks directly. A code-armed model can do the kind of debugging an ML engineer does, in about 20 hours of wall-clock time. From thinking to doing. [![CTA Image](https://11c0f2c-126c06r.lingersh.space/content/images/2026/10/logo-black.jpeg)](https://varops.com/transformation/) ****Less hype. More results.** AI promises are everywhere. We prefer results. VarOps helps companies cut through the hype, avoid expensive detours, and build AI systems that deliver measurable business value. [Get Beyond the Hype → ](https://varops.com/transformation/) ## Sources - Takehi, Teoh, Zhang, Khattab, "Sparks of Weight-Space Programming", Nov 2, 2026 - [https://rikiyat.github.io/blog/weight-space-programming/](https://rikiyat.github.io/blog/weight-space-programming/?ref=varops.com) - Experiment logs - [https://github.com/RikiyaT/lm\_understand-logs](https://github.com/RikiyaT/lm%5Funderstand-logs?ref=varops.com) - Matthew Honnibal, "The 'Astra can read and write weights to improve Qwen' post is misleading" - [https://honnibal.dev/blog/astra-read-write-misleading](https://honnibal.dev/blog/astra-read-write-misleading?ref=varops.com) - Sean Goedecke, "LLMs are not superhuman at reading and writing neural network weights" - [https://www.seangoedecke.com/llms-are-not-superhuman-at-reading-and-writing-neural-network-weights/](https://www.seangoedecke.com/llms-are-not-superhuman-at-reading-and-writing-neural-network-weights/?ref=varops.com) - Simon Willison, "Qwen 3.8 27B on my laptop can interpret model weights, too", Nov 3, 2026 - [https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/](https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/?ref=varops.com) - R&D World (Brian Buntz), Nov 3, 2026 - [https://www.rdworldonline.com/mit-team-says-gpt-6-astra-can-fix-an-ai-model-by-editing-its-weights-critics-say-the-method-was-standard-ml-engineering/](https://www.rdworldonline.com/mit-team-says-gpt-6-astra-can-fix-an-ai-model-by-editing-its-weights-critics-say-the-method-was-standard-ml-engineering/?ref=varops.com)