Researchers say GPT-6 Astra can edit another model's weights

Four researchers at MIT CSAIL, including Omar Khattab, published a blog post reporting that OpenAI’s GPT-6 Astra can read and write the weights of neural networks directly, without gradient descent. The post, by Rikiya Takehi and colleagues, calls this “Weight-Space Programming”. It is a blog post, not a peer-reviewed paper, and the authors describe the results as “very early”.

The headline result concerns Alibaba’s small Qwen3.5-0.8B model. The authors said Astra identified a looping failure and, by editing 131,072 weights (0.016%), raised its MATH-500 score from 7.9% to 48.1% with the same sampling settings. They reported smaller gains on GSM8K and IFEval. Other experiments had Astra name the game a small reinforcement-learning network was trained on, repair a failing robot-control policy, and write small game-playing networks from scratch that, they said, beat networks trained with PPO for 100 million steps. The authors also reported failures: Astra could not pick out a sycophantic model from its weights alone, and the same editing procedure did nothing for Gemma 3 270M.

The framing drew a rebuttal the next day. Matthew Honnibal, author of the spaCy library, argued that the abstract’s “just by looking at the raw weights” holds only for three experiments on networks of 17,000 to 172,000 weights. The Qwen result, he wrote, used Python and up to 50 batches of practice problems, so it shows the model can be directed to do what a human machine-learning engineer might do, not a single deep skill. He credited the work as non-trivial. The post’s own caveats say the Qwen edit used forward passes on practice problems, with evaluation problems held out.

Independent preprints followed within days, with mixed results. One paper, by Jusheng Zhang, Keze Wang and Liang Lin, reported that Astra edited 65 token-embedding rows of a robot policy and raised success on held-out paraphrases from 86.3% to 92.0% in the best of three sessions, and concluded the evidence supports precise lexical repair rather than general acquisition of missing capabilities. Another ran 16 editing trajectories on Qwen3.5 models and found results varied widely, with six trajectories keeping the original weights. A single-author paper tested Claude Opus 5.5 on small image classifiers and concluded its weight reading was better described as local arithmetic plus dataset priors. All are small, unreviewed preprints by independent groups, and none of the sources read here reports a replication by a major lab.