Rikiya Takehi posted a bold claim on X on Nov. 2. GPT-6 Astra had edited Qwen3.5-0.8B's weights directly, he wrote, and lifted its math score sixfold3. The post reached 52,608 likes within a day. Readers argued back fast.
Matthew Honnibal read the same post and called it misleading2. He co-founded Explosion, the company behind the spaCy code library. His complaint centers on how Astra found the fix. The fact that Qwen's weights changed stays undisputed.
What the Viral Post Claimed
Takehi is a first-year PhD student at MIT's Computer Science and Artificial Intelligence Laboratory1. He worked with three co-authors. They are Jayden Teoh, Alex Zhang and Omar Khattab, his adviser. Together they call their method Weight-Space Programming.
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
Their claim is specific. GPT-6 Astra, OpenAI's current flagship reasoning model, looked at Qwen3.5-0.8B's weights. It found a loop that made Qwen repeat itself. Then it changed 131,072 of those weights by hand, just 0.016 percent of the model13.
Qwen's score on a standard math test, MATH-500, jumped from 7.9 percent to 48.1 percent1. That is six times higher. The edit alone did it, Takehi wrote, "without gradients or training"3.
Press coverage picked up the comparison fast, and VentureBeat's write-up called the fix "like gene editing for neural networks" the same day5. The gene-editing framing spread before Honnibal's correction did.
Honnibal Says the Test Mostly Ran Code
A rebuttal followed the next morning, Nov. 3, from Honnibal himself2. He is a computational linguist, and Explosion builds AI tools for other coders. His read of the method differs sharply from the viral one.
Two of the post's eight experiments do read raw weights alone, Honnibal found2. The Qwen fix sits outside that pair. "Every experiment involving a real language model uses Python, and the Qwen result, which is the headline one, also gets to run the model on up to 50 batches of practice problems," Honnibal wrote.
Astra wrote Python code to test ideas before it touched a single weight, in Honnibal's account2. "What the Qwen result actually shows is that Astra can be prompted to devise and perform tasks that a human ML engineer could do," he wrote. That is a real skill. It differs from the claim Takehi's headline seemed to make: a model editing another by inspection alone.
Qureshi Calls the Fix Ordinary Engineering
Haseeb Qureshi posted his own thread about four hours after Takehi's, the same day4. He writes about AI investing and development. His post framed the story differently too.
"The story is not 'Astra used its LLM X-ray vision to surgically alter the weights of its brethren.' It's more like 'Astra found a published interpretability technique that can solve this particular kind of failure modes, and implemented it on its own,'" Qureshi wrote4.
Qureshi added his own verdict. "It's a cool anecdote, but well within our understanding of LLM capabilities. It's not RSI, and it's definitely not LLM telepathy or whatever people on HN seem to think," he wrote4. His post had drawn 486 likes by Nov. 3, next to 52,608 on Takehi's original3.
Takehi's Viral Claim Draws 108 Times the Likes of Qureshi's Correction
Likes on X, both posts from Nov. 2, 2026, measured Nov. 3, 2026
Each bar is one post's likes on X. The tall bar is Takehi's original claim, and the short bar is Qureshi's reply, posted about four hours later.
Sources: Rikiya Takehi on X [3]; Haseeb Qureshi on X [4]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Takehi's claim | 52,608 likes |
| Qureshi's reply | 486 likes |
Takehi's claim has drawn roughly 108 times the likes of Qureshi's correction so far34. A thread on Hacker News built around Honnibal's post climbed to 706 points over the same stretch6.
The One Number Every Side Accepts
Every account of this story agrees on one fact. Each side takes it as given. Astra changed 131,072 of Qwen's weights, a slice just 0.016 percent wide13. The figure came from a control test the researchers ran on a second, smaller model, Gemma 3 270M.
Qwen is Alibaba's open-weight model, free for anyone to download and test1. Qwen's openness gave Takehi's team a real target to test on. Gemma rarely loops the way Qwen does. Only 8.5 percent of its answers hit the token limit that trips up 93 percent of Qwen's.
Researchers tried the identical edit on Gemma anyway. Its math score moved from 6.8 percent to a high of 7.7 percent, a gain the blog itself called "not established"1.
Astra's Edit Lifts Qwen's Score 40 Points, Leaves Gemma's Flat
MATH-500 percent correct, before and after Astra's identical edit method; higher is better
Each row is one model. The first dot is its score before Astra's edit, the second dot is after, and the rod shows the gap the edit opened.
Source: Rikiya Takehi's blog [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Before the edit | After the edit |
|---|---|---|
| Qwen3.5-0.8B | 7.9% | 48.1% |
| Gemma 3 270M | 6.8% | 7.7% |
The gap is the whole argument. Qwen's loop gave Astra something concrete to remove. Gemma's math skill stayed flat, leaving little for code or a direct edit to fix. The Qwen-Gemma gap is what Honnibal's critique rests on.
What to Do
Read a "model edited another model's weights" claim in two layers. First ask what Astra actually changed, in a number anyone can check. Then ask how much code ran before that number existed12.
Check whether a control test ran on a model that behaves differently from the one in the headline. Gemma barely moved under the identical edit, the clearest sign the fix targeted one bug, separate from any new math skill1.
Read the correction before the original headline convinces you. Honnibal's post reached Hacker News within a day and gathered more engagement there than many outlets gave the first claim26. The same caution fits the next big claim that skips how much code sat behind it.
By the Numbers
- Takehi's original post on X drew 52,608 likes, counted Nov. 3, 20263.
- That same post gathered 2,205 replies and quote posts3.
- Qureshi's clarifying post drew 486 likes, posted roughly four hours later4.
- Hacker News gave Honnibal's critique 706 points6.
- Astra changed just 0.016 percent of Qwen3.5-0.8B's weights directly1.
- Qwen's MATH-500 score rose from 7.9 to 48.1 percent after the edit1.
- Gemma 3 270M's MATH-500 score moved from 6.8 to only 7.7 percent under the same method1.
- Fifty batches of practice problems ran before any weight changed, in Honnibal's count2.
What to Watch
Watch whether Takehi or his co-authors answer Honnibal in public. A reply would settle how much of Weight-Space Programming survives outside one viral thread12. Watch too for a team outside MIT running the identical test on a third model. One more control run would show how far the fix 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/
- Matthew Honnibal, "The 'Astra can read and write weights to improve Qwen' post is massively misleading," honnibal.dev, Nov. 3, 2026, https://honnibal.dev/blog/astra-read-write-misleading
- 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
- Haseeb Qureshi, "Re: the Astra modifying Qwen 3.5-0.8B paper today, important clarification," X, Nov. 2, 2026, https://x.com/hosseeb/status/2117380109355352166
- 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
- syllogism, "The 'Astra can read and write weights to improve Qwen' post is misleading," Hacker News, Nov. 3, 2026, https://news.ycombinator.com/item?id=50351111
