← November 3, 2026 briefing · added in the 04:00 KST update
Weight-Space Programming follow-up: critics say the Qwen improvement relied on Python execution; partly reproduced with Qwen 3.8 27B on a laptop (community reports)
Announced November 3, 2026
Verified (fact)
A rebuttal has emerged to the paper "Sparks of Weight-Space Programming" (authors: Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab). According to a Reddit post and a GeekNews summary, the post that introduced the paper presented two different kinds of experiments as a single capability: one in which the model read raw weight values directly, and another in which it used Python to analyze and modify the model. Only three experiments put weight numbers alone into the context without tools, and they targeted small networks of 17,000 to 172,000 weights. The widely discussed improvement in Qwen3.5-0.8B's math performance used Python, under conditions that let the model run up to 50 batches of practice problems. An attempt to improve Gemma 3's math performance reportedly failed. Separately, Simon Willison said he reproduced one of the paper's examples with Qwen 3.8 27B running on a laptop. Another Reddit user rebuilt the modified model from the published weight diff and shared an analysis of its repeated "Wait" segments using Anthropic's Jacobian lens.
Why it matters (interpretation)
This gives grounds to re-examine how far the claim of improving performance by editing weights directly, without training, actually holds. Whether the model could run tools and get feedback substantially changes how the results should be read, so anyone citing the original paper's numbers should state the experimental conditions.
Confidence medium · official source pending
Sources
- Community The 'Astra can read and write weights to improve Qwen' post is misleading Reddit
- Community ‘Astra가 가중치를 읽고 써서 Qwen을 개선한다’는 게시물은 오해를 부름 GeekNews
- Community Qwen 3.8 27B on my laptop can interpret model weights, too Simon Willison
- Community What Qwen is thinking when it gets stuck saying "Wait" Reddit
Story thread
- 2026-11-03 GPT-6 Astra reportedly edits just 0.016% of Qwen weights without training, lifting MATH-500 from 7.9% to 48.1% (community report)
- 2026-11-03 Weight-Space Programming follow-up: critics say the Qwen improvement relied on Python execution; partly reproduced with Qwen 3.8 27B on a laptop (community reports)
More about MIT
-
GPT-6 Astra reportedly edits just 0.016% of Qwen weights without training, lifting MATH-500 from 7.9% to 48.1% (community report)
A research blog post, 'Sparks of Weight-Space Programming,' published early experimental results with GPT-6 Astra, saying it can...
-
MIT and MIT-IBM researchers' 'Swarm Telepathy (SwaT)' lets models from different companies share KV caches directly instead of text, claiming 40% higher accuracy and 4x speed (report)
Researchers from MIT CSAIL and the MIT-IBM lab posted a paper titled 'Swarm Telepathy: Prefill-Free KV Cache Transfer Across Model...
-
NVIDIA and MIT release weights for 'Open-H3-Regenerate-2K', a video enhancement model built on MiniMax H3 that regenerates up to 4K (report)
According to QbitAI (量子位), NVIDIA and MIT researchers uploaded the model weights for the video enhancement project...
-
Follow-up to virtual-island Claude agent experiment: MIT paper 'Is this machine playing?' finds a throwing technique unknown to the developers (reported)
MIT researchers posted a paper on arXiv, 'Is this machine playing?', analyzing how AI agents released on a virtual island without any...
-
MIT researchers develop a multimodal model that understands 3D atomic arrangements together with natural language, applied to solid-state battery electrolyte discovery (report)
According to AI Shinmun, MIT researchers have built a multimodal model that learns by linking crystalline materials' atomic arrangements...