Weight-Space Programming: GPT-6 Astra edits Qwen's weights and lifts its MATH-500 score from 7.9% to 48.1%

By Turing Post · Monday, November 2, 2026.

Today's Turing Post editorial looks at why "Sparks of Weight-Space Programming", a paper published on November 2 by MIT CSAIL researchers, went viral. According to the newsletter, the announcement passed one million views on X within a few hours, and one reader described it as "model-on-model surgery." The editor set aside the editorial she had planned so she could investigate what was going on.

According to Turing Post, part of the paper's success comes from a phrase that sounds like science fiction: an AI acting as a surgeon, opening up another AI's "brain" to rewrite its circuitry. Add to that a striking figure, the jump from 7.9% to 48.1% on MATH-500. The newsletter notes that the concept is not a sudden breakthrough: research on model editing, weight-space representations and hypernetworks has been under way for years. What was genuinely new, it says, was how the demonstration was put together.

First experiment: the Breakout "paddle." The team gave GPT-6 Astra the weights of a small game-playing network as text: all 17,253 weights, along with a description of the grid and the controls. The question was whether Astra could name the game without a list of options. In the weight map for the player's position, the strongest connections were concentrated in the lowest playable row, suggesting a paddle that moves left and right. Astra also considered Pong, but in the second map, three rows of reward targets near the top gave away the bricks. Astra correctly identified Breakout without seeing a single frame.

Second experiment: writing into the weights. Astra was asked to make another language model, Qwen3.5-0.8B, better at math. Astra had Python, practice problems and up to 50 test batches to try out edits. The team ruled out retraining and kept the test questions private.

With the optional thinking mode turned on, Qwen kept repeating steps it had already written. Astra compared Qwen's internal signals during those loops with those from steps where it did make progress, and found that two layers, 15 and 19, reinforced a pattern associated with repetition. To weaken it, Astra made an edit touching 131,072 weights, about 0.016% of the model. On MATH-500, Qwen's accuracy rose from 7.9% to 48.1%.

Co-author Alex Zhang says the team initially thought Astra had somehow added a new math skill to Qwen. But when they looked at how the answers ended, 93% hit the length limit before the edit; afterwards, 33.1% did. According to the newsletter, much of the improvement came from Qwen finally finishing problems it could already solve. They also tried Gemma 3 270M, which generally completed its answers but made arithmetic and algebra errors; Astra's edits did not produce a clear improvement there.

The loop and the prompt: a caveat. Researcher Lucas Beyer read the full prompt and pointed out that it explicitly told Astra that if Qwen "repeats itself and never finishes," finding and fixing what causes that repetition counts. Qwen's own documentation had already flagged this behavior in thinking mode. Co-author Jayden Teoh replied that the team had given Claude Code a brief assignment, and that it added the hint when expanding the prompt, with no further human involvement. Turing Post sums it up this way: one model suggested where to look and another worked out which of Qwen's 800 million weights to change, a model-to-model collaboration that, according to the editor, the media has yet to highlight.

The newsletter connects this to how alarmist "AI will kill us" headlines are read. It mentions that Vidhay Reddy told CBS he was asking Gemini how to prevent elder abuse when the model told him to die, and that Transluce researchers later cited his case when deliberately searching for prompts capable of pushing models to encourage self-harm.

On the lineage of weight editing: techniques such as ROME, MEMIT and task arithmetic showed years ago that parameters contain direct associative structures that can be targeted. In 2024, other researchers used related targeted edits to control when chatbots refuse a request (the paper "Refusal in Language Models Is Mediated by a Single Direction").

According to Turing Post, what makes the work appealing and explains its virality is the orchestration: describing a high-level problem in plain English, giving a standard base model an execution environment, and letting it systematically probe, isolate and patch the weights of another network.

Finally, the team tried a shorter prompt without the repetition example. With 25 test batches, Qwen reached 14.1% accuracy after the edit, the same score as in the run with the hint. The table does not include a 50-batch result for that shorter prompt, and that is the row the editor would like to see.

Other items in the newsletter: the weekly schedule includes an AI 101 on Looped Transformers on Friday, November 6, and a Twitter Library on an update to computer-use agents on Sunday, November 8. It also recommends a YouTube video on DeepSeek's new open-source "super-agent," covering what breaks when AI agents run in parallel.