Rikiya Takehi wanted Qwen3.5-0.8B to do better math, so he skipped the usual training run. Instead, a bigger model read Qwen's weights and changed them directly1. That model was OpenAI's GPT-6 Astra.
Qwen's score on a standard math test jumped from 7.9 percent to 48.1 percent1. The test is called MATH-500. That is six times higher.
Takehi changed just 0.016 percent of the model. That got him there.
What Astra Actually Changed
Weights are the numbers inside a model that decide what it says next. Training usually sets them, through thousands of practice rounds. Takehi's edit touched just 131,072 of them, inside two layers, out of 800,000,0001.
Astra is OpenAI's current flagship reasoning model. Takehi is a first-year PhD student at MIT's Computer Science and Artificial Intelligence Laboratory3. He worked with three co-authors: Jayden Teoh, Alex Zhang and Omar Khattab. Together, they gave Astra a direct channel into Qwen's weights.
They asked Astra to find whatever was holding Qwen's math back, instead of retraining the model from scratch. Astra traced the problem to a repetition loop. The loop lived in two of Qwen's layers, 15 and 191.
Astra subtracted that pattern from 64 output columns. It worked by hand.
The fix worked on three tests at once. GSM8K, a set of grade-school word problems, rose from 13.8 percent to 67.1 percent1. IFEval grades how well a model follows instructions, and it rose from 21.0 percent to 62.3 percent.
Answers that once ran until they hit Qwen's token limit changed too. That rate fell from 93 percent to 33 percent, as the model stopped repeating itself1.
The Fix Removed a Bug, Skipped the Lesson
Researchers tried the same trick on a second model1. Its name was Gemma 3 270M. Gemma rarely loops.
Its math score barely moved. It stayed close to 7 percent across five attempts.
Editing weights could clear a blockage, in the researchers' own words1. It left any missing skill exactly where it started, since a model that already reasoned cleanly gave Astra little to unblock.
One Edit Lifts Qwen's Math Score From 7.9 to 48.1 Percent
Qwen3.5-0.8B's scores on three benchmarks, before and after GPT-6 Astra's weight edit, Nov. 2, 2026
The blue dot is Qwen's score before the edit, the magenta dot is its score after, and the rod is the gap the edit opened. Higher is better on every test.
Source: Rikiya Takehi's blog [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Before the edit | After the edit |
|---|---|---|
| MATH-500 | 7.9% | 48.1% |
| GSM8K | 13.8% | 67.1% |
| IFEval | 21% | 62.3% |
Puneet Nanda writes on the DEV Community platform4. He compared the fix to a familiar kind of software repair. "Removing an obstacle that hid an ability does not install an ability the model lacked," he wrote. He added a sharper test of his own: "Deleting while (true) is not a math curriculum. It is a favor to whoever pays for the electricity".
The Twenty Hours Behind One Edit
Finding that one fix took real work. The strongest attempt ran 158 conversational turns and 80 separate Python runs2. It lasted 20.3 hours in total, including 4.7 hours stuck waiting through a service outage. A person watched every one of those turns and decided when to stop.
Astra Touched 0.016 Percent of Qwen's 800 Million Weights
Weights GPT-6 Astra edited directly, out of Qwen3.5-0.8B's full weight count
Source: Rikiya Takehi's blog [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Weights edited | 131,072 |
| Whole | 800,000,000 |
Takehi arrived at MIT after a degree from Waseda University and internships at NVIDIA Research and Mixedbread3. Omar Khattab, his adviser at MIT CSAIL, oversaw the project. The team named their method Weight-Space Programming.
Weight-Space Programming means a model reads and writes another model's weights directly, skipping the gradient and the loss function of normal training1. Runtime Wire called the near-term use diagnosis and patching of small models3.
Rikiya Takehi@rikiyatakehiWe asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming:… pic.twitter.com/DwTwzXzSbE
A Warning Anthropic Filed the Same Day
Takehi's post landed the same week Anthropic filed its own warning with securities regulators5. The filing updates the Responsible Scaling Policy, the company's own rulebook for when a model turns dangerous. It said Anthropic may imminently meet its definition of dramatically accelerated AI progress.
Anthropic went further: "Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress"5.
The filing said its current models stay short of that line5. Maxime Fournes read Takehi's thread and saw the exact risk Anthropic described in its own filing. He is a former deep learning researcher and the chief executive of Pause AI2.
Pause AI wants frontier AI development halted. It says development should wait until proven safe2.
"Gosh, may this be hype. Because if real, that could massively accelerate progress which is very bad news and brings us much closer to the global catastrophe. What a terribly reckless research direction," Fournes wrote2.
The researchers' own blog names that exact worry. Weight-Space Programming offers a more direct path to recursive self-improvement than a model writing code to launch training jobs, they wrote1. Fournes was reading the result of one edit, though.
That one edit took 158 turns and 80 code runs, most of a day, guided by a person at every step2. The gap between one slow, supervised patch and a model rewriting itself alone is real. Anthropic's filing says that gap still has to close5.
What to Do
Read a "model edited itself" headline as two separate claims. Ask what changed, in plain, checkable numbers. Then ask whether a model that loops rarely still improved under the same edit1.
Check Hugging Face's download count before you trust a benchmark jump. Qwen3.5-0.8B has already drawn 3.69 million downloads. A fix to it reaches real deployments fast7.
Run the same control, if your own work tests a self-editing claim. Try the identical edit on a model that loops rarely. Watch whether its score moves at all.
By the Numbers
- 0.016 percent: the share of Qwen3.5-0.8B's weights GPT-6 Astra changed directly1.
- 131,072 weights: the exact count the edit touched, inside two layers1.
- MATH-500 rose from 7.9 percent to 48.1 percent after the edit1.
- GSM8K rose from 13.8 percent to 67.1 percent over the same run1.
- 20.3 hours: how long the strongest edit attempt took, a service outage included2.
- 3.69 million: Qwen3.5-0.8B's download count on Hugging Face7.
- Nov. 2, 2026: the date Takehi posted the finding on X, the same day MIT published the full write-up6.
What to Watch
Watch whether anyone reproduces Weight-Space Programming on a fresh model, outside MIT's own team. Anthropic's next scaling policy update deserves a look too. The company said it continues assessing whether any model has crossed its self-improvement line. A reproduction on a larger model, run by an independent team, would settle how far this finding reaches.
Sources
- Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, "Weight-Space Programming," Rikiya Takehi's blog, Nov. 2, 2026, https://rikiyat.github.io/blog/weight-space-programming/
- Carl Franzen, "'Like gene editing for neural networks': MIT researchers used GPT-6 Astra to edit Qwen's open weights directly, boosting math performance," VentureBeat, Nov. 2, 2026, https://venturebeat.com/technology/like-gene-editing-for-neural-networks-mit-researchers-used-gpt-6-astra-to-edit-qwens-open-weights-directly-boosting-math-performance
- Ryan Merket, "GPT-6 Astra edited 0.016% of Qwen's weights and its math score went up 6x," Runtime Wire, Nov. 2, 2026, https://runtimewire.com/article/gpt-6-astra-edits-qwen-weights-weight-space-programming
- Puneet Nanda, "131,072 Weights, and None of Them Were a Math Lesson," DEV Community, Nov. 2, 2026, https://dev.to/puneet_nanda_dc25d8b489ad/131072-weights-and-none-of-them-were-a-math-lesson-1b65
- Julie Bort, "Anthropic warns that it's very close to RSI capability, aka self-building AI," TechCrunch, Nov. 2, 2026, https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/
- Rikiya Takehi, "We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math," X, Nov. 2, 2026, https://x.com/rikiyatakehi/status/2117313742891327994
- Alibaba, "Qwen/Qwen3.5-0.8B," Hugging Face, accessed Nov. 2, 2026, https://huggingface.co/Qwen/Qwen3.5-0.8B
