What the Astra weight-reading demo actually showed

Last week the claim that an AI model called Astra could read and improve other neural networks just by inspecting their raw weights drew a lot of attention. Matthew Honnibal, who built the spaCy library, examines it closely and argues the headline oversells what happened. His point is not that the work is fake. It is that the original post bundles two very different kinds of experiment under one story, and the combination implies a general capability the evidence does not show.

Some of the experiments really do read raw weights with no code and without running the network. Astra named the game a reinforcement-learning policy had been trained on, four times out of four, identified the shape a small diffusion network drew, and fixed a stuck driving policy. But those networks were tiny, between 17,000 and 172,000 weights. The result that got the attention, raising Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing just 0.016% of its weights, was different: it leaned on Python code and on running the model over as many as 50 batches of practice problems. That is closer to what a human machine-learning engineer does with tooling than to "reading" weights directly.

Honnibal grants that the Qwen result is still interesting and would likely have failed with models from a year or two ago. His complaint is about framing. Putting the direct weight-reading tasks and the tool-assisted ones in the same writeup suggests one unified understanding of weights, when in fact the tasks used different methods and only some touched the weights at all.

Why it matters

If you read AI research by the abstract, this is a reminder to check which experiments actually support the headline. The gap between "the model read the weights" and "the model wrote code that analyzed the weights" changes what you should expect when you try it yourself.