MIT: GPT-6 edited a rival AI's weights by hand, lifting its math score to 48%, and Legora's CTO says Claude hand-wrote a better legal classifier
What's the deal? MIT CSAILDealroom has a profile for this one. Try Dealroom → researchers report that OpenAI's GPT-6 Astra directly edited the weights of Alibaba's small Qwen3.5-0.8B model, with no retraining, and raised its MATH-500 score from 7.9% to 48.1%. Astra changed just 0.016% of the model's weights, according to the paper published November 2, 2026.
How it works: Rather than running training loops or gradients, Astra read the raw weights and reasoned about what each part should compute. The researchers call this "Weight-Space Programming."
Why it matters: The edit targeted a flaw Qwen's own model card already flagged: repetitive reasoning loops. So this is better understood as a precise repair of a known failure than an AI independently learning new mathematics.
The upside: The fix generalised. The same weight correction improved instruction-following, tripling the model's IFEval score, the researchers say. Astra also identified which game a separate 17,253-weight network was trained on (Breakout) and wrote small game-playing networks from scratch.
Already in the wild: A day later, Jacob Lauritzen, CTO of legal AI company Legora, said he got Anthropic's Claude Fable to hand-write the weights of a roughly 100,000-parameter network that tells real legal citations from fake ones, and that the one-shot result beat training it. In the results he shared, the hand-written network saw zero labelled examples, uses 3,793 non-zero parameters and scored 97.8% test accuracy; one revision after seeing 5,062 labels took it to 98.9%. Conventionally trained versions of the 107,745-parameter network reached 76.5% with 512 labels, 92.9% with 5,062 and 94.6% with 8,192. These are his own figures and have not been independently checked. Separately, Simon Willison reported that Qwen 3.8 27B running on his laptop could also interpret the weights in one of the paper's examples.
What could go wrong? The result is preliminary. It comes from one small model, relies on repeated test-and-edit attempts, and has not been independently replicated or peer reviewed. Lower-budget attempts reportedly produced much smaller gains, leaving open whether the approach scales. Critics such as engineer Sean Goedecke argue the Qwen repair is what a competent human researcher would have done, and that training normally remains the best way to update a model's weights.
The signal: For years, neural network weights were shaped only by optimisers like Adam or by training code. The prospect here is different: capable language models that diagnose, patch and even write other models by reasoning directly about their weights. If it holds up, it points to a new path for both optimisation and interpretability, though that remains a big if.
Read more: MIT CSAIL paper, VentureBeat, Jacob Lauritzen (X), Matt Stockton (X), Simon Willison, Sean Goedecke
Table: Jacob Lauritzen on X
