The lead
MIT CSAIL: GPT-6 Astra edited 0.016 percent of a small Qwen model's weights by reasoning, and its MATH-500 score rose from 7.9 to 48.1 percent
The authors call it Weight-Space Programming: the model reads what each part of a network computes and rewrites it, instead of running an optimiser like Adam. It is a blog post with experiments on networks of up to 0.8 billion parameters, and it took 2,300 points on Hacker News in a day.
Given only raw weights as text, Astra named which of four games a 17,253-weight network was trained on, fixed a failing control policy by editing 3.5 percent of its weights, and repaired a reasoning loop in Qwen3.5-0.8B with no gradient training. The same edit tripled the model's IFEval instruction-following score.
CONVOUse it as the week's AI talking point in interviews: models now debug other models' weights directly, so tuning work that needed an ML team is moving to a prompt. Say it is early results on small models, posted 2 November and not yet peer reviewed.