The 'Astra can read and write weights to improve Qwen' post is misleading

What happened: A Hacker News post argues that the "Astra can read and write weights to improve Qwen" writeup is misleading, because its headline Qwen result does not come from reading raw weights alone. The author says the post mixes separate experiments that use different techniques. Key numbers: Raw-weight-only experiments covered networks of 17K to 172K weights. Why it matters: The author says the Qwen result shows Astra can be prompted to do tasks a human ML engineer could do, not that one technique both writes weights from scratch and understands a large existing network.

A Hacker News post argues that the "Astra can read and write weights to improve Qwen" writeup is misleading, because its headline Qwen result does not come from reading raw weights alone. The author says the post mixes separate experiments that use different techniques.

The author says the Qwen result shows Astra can be prompted to do tasks a human ML engineer could do, not that one technique both writes weights from scratch and understands a large existing network.