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

1000 points · 204 comments on HN · read original →

Points and comments are a snapshot, not live.

The original blog post misrepresents how Astra improved Qwen by conflating multiple techniques into one claim.

The article examines experiments from the 'Astra can read and write weights' post, showing only three of eight experiments (on networks of 17K-172K weights) involve reading raw weights directly. The Qwen 3.5-0.8B math improvement (7.9%→48.1% on MATH-500) used Python tools and up to 50 model runs, not raw weight reading. The author argues these diverse techniques do not constitute a single method, and conflating them overstates the capability claimed.

What commenters are saying

Commenters split: many argue the original post explained its methods clearly and this critique adds nothing new, while others agree the abstract and title mislead readers who skim. Several note that writing code to analyze and modify weights is a known capability, citing prior work like Tracr and 'Thinking Like Transformers'. A few call the weight-reading experiments impressive but not unexpected, given linear relationships in small networks.