GPT-6 Astra Edited Another AI, Researchers Say

GPT-6 Astra edited a small Qwen model's weights with no training, four researchers who list MIT CSAIL say. In a blog post dated November 2, they say OpenAI's model changed 131,072 of about 800 million weights in Qwen 3.5 0.8B, and that model's MATH-500 score went from 7.9% to 48.1%. In the same run, answers hitting the length limit fell from 93% to 33.1%.