Matthew Honnibal · 2026-11-03 · notable
Matthew Honnibal — the 'Astra reads and writes weights' post is misleading
Matthew Honnibal, creator of spaCy, argues the viral post on GPT-6 Astra editing network weights overstates it. Only 3 of 10 experiments read raw weights, on networks of 17K to 172K weights; the Qwen result used Python and test runs.
Matthew Honnibal says the GPT-6 Astra weight-editing results are real but come from different techniques, not one deep skill.
What is it?
Matthew Honnibal, co-founder of Explosion and creator of spaCy, has written a rebuttal to the post "Astra can read and write weights to improve Qwen". He argues the post's abstract suggests GPT-6 Astra works just by looking at raw weights, and that the experiments do not show this.
How does it work?
Honnibal sorts the original post's 10 experiments by method. Only experiments 1, 2 and 4 read raw weights directly, and those networks have 17K to 172K weights. Every experiment with a real language model uses Python, and the Qwen result also runs the model on up to 50 batches of practice problems. The game networks built from zero beat PPO only on Breakout (517 vs. 25) and lost on Seaquest and Asterix.
Why does it matter?
The weight-editing claim spread fast as a sign of a new model ability. Honnibal still credits the Qwen result — MATH-500 rising from 7.9% to 48.1% — as work a human ML engineer could do and that models of a year or two ago would have failed. His point is that mixing these results with raw-weight reading suggests one unified capability that was not shown, so the two should be reported apart.
Who is it for?
ML researchers and anyone judging frontier-model capability claims