spaCy author Matthew Honnibal calls MIT's 'GPT-6 Astra edits weights' post misleading

Four MIT CSAIL researchers posted on Nov 2 that GPT-6 Astra can read and write neural network weights: it can tell what a network was trained for from its weights, and without training it edited about 0.016% of Qwen3.5-0.8B's weights to lift its MATH-500 score from 7.9% to 48.1%. Matthew Honnibal, author of the open-source spaCy library, wrote on Nov 3 that the post is misleading. Only three experiments use raw weights alone, without tools or running the network, all on networks of 17K to 172K weights. The language model experiments rely on Astra writing Python, and the Qwen one also ran the model on up to 50 batches of practice problems. He says Astra can do what a human ML engineer could, which models of a year or two ago would have failed at, but these are two methods, not one ability.

Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.