Watch the video, transcript and sources
+ Qwen reaches more answers after a targeted weight edit
+ Python and practice feedback help Astra choose the repair
- A smaller Gemma follow-up finds no established gainAn AI giving another model a sixfold math-score increase sounds like an instant delivery of intelligence. This result has a concrete engineering explanation: the smaller model kept repeating its reasoning until it ran out of room to answer. A targeted weight edit helped more of those answers reach the end.
The MIT-affiliated team of Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab published its weight-space programming study on November 2. It used OpenAI's GPT-6 Astra to inspect and edit the weights of Qwen3.5-0.8B. The team reports MATH-500 rising from 7.9% to 48.1% under the same sampling protocol. About a third of edited answers still hit the length cap. “Patched” is a useful word here: the result retains a substantial failure rate.
What Astra actually did
Weights are the numbers inside a neural network. Conventional gradient training was forbidden in this experiment. Astra could analyze the existing numbers, choose an intervention and save changed weights directly. The authors describe a repetition direction written by internal layers 15 and 19, removed from selected output columns. The edit touched about 131,000 weights, reported as 0.016% of the model.
The experiment was a multi-turn tool-assisted process. Astra received a Python kernel over the weights, practice problems with reference answers and a budget of model-run batches. It could inspect activations and compare original with edited answers. A generation batch could contain hundreds of answers. Fifty forward-pass budget units do not mean fifty single answers or fifty training runs.
The actual initial prompt gave useful research advice and explicitly named repetitive thinking as a repair that would count. It asked for causal localization, comparisons on the same problems and tests at different strengths. The model still had to produce a useful intervention within those rules.
Practice feedback and the final test
Practice problems came from training splits. The researchers say Astra never saw the final evaluation sets. Feedback helped it choose an edit; the unseen final benchmark measured the submitted model afterwards. The benchmark names and protocol were known to Astra. That distinction matters when describing what was withheld.
Today, Matthew Honnibal, spaCy's creator and Explosion's cofounder, challenged the presentation of the experiments. He argues that their different conditions invite readers to infer one broad technique. Tiny raw-number reading tasks, no-rollout game-network writing and the Python-assisted Qwen repair each had their own support. His article gives the tool-assisted engineering credit while disputing the combined impression.
Even the game comparison has a scope. Astra's generated networks won all three reported games against a baseline trained for a hundred million steps. Against a longer-trained baseline, it won Breakout and lost Seaquest and Asterix. The table's comparison column changes the conclusion.
The follow-up that earns the verdict
The authors also tried a smaller Gemma model in a separate setup. It rarely looped. Its failures included ordinary wrong arithmetic and invalid algebra, without one component the editor could switch off. Astra used practice feedback and targeted edits again. The follow-up failed to establish a gain; none moved its MATH-500 score by more than about a point.
Model size, sampling, output cap and grading differed, so this follow-up is a boundary on the demonstrated success. It provides a practical question for developers: which failure does an edit repair, and does the model you use actually have that failure? Test the saved weights on your own tasks before treating a reported multiplier as extra intelligence.
Verdict: NEEDS REVIEW
I'd take the loop repair seriously and test it. I'd keep broader claims gated on successful improvements beyond that failure mode. The authors' own counterexample is the most useful line in the diff.
If you'd rather read this than hear me say it, the diff lands in your inbox every morning, free at the daily diff dot dev, link below.
Sources
Cover portrait: Steve Jennings/Getty Images for TechCrunch, Sam Altman TechCrunch SF 2019 crop, CC BY 2.0. Cropped, background removed, resized and flipped. It identifies the editor's provider; no participation, reaction or endorsement is asserted. Research logs/data/networks: Rikiya Takehi and coauthors, CC BY 4.0. Primary footage is a brief credited excerpt for reporting/commentary.
.jpg){kind=link}
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.