Matthew Honnibal breaks down a claim that Astra can read and rewrite neural-network weights, the numerical parameters a model learns during training. The experiments combine several techniques, making their shared headline suggest a broader capability than any one experiment demonstrates.
- Direct reading happens on small networks. The raw-numbers-in-the-prompt experiments cover networks with 17,000–172,000 parameters.
- The Qwen improvement uses tools. According to Honnibal, Astra writes Python and can run Qwen on up to 50 batches of practice problems. The reported maths-test score rises from 7.9% to 48.1%.
- Transfer has limits. A corresponding Gemma maths experiment fails. Constructed game networks also lose two of three comparisons against a stronger training baseline.
Honnibal credits the engineering achievement. His objection is that reading small networks’ raw parameters and improving a language model through generated programs are different demonstrations. The Qwen score is an original experiment’s reported result, not an independent replication by this critic.
The 44-comment thread on Hacker News adds earlier work and disagreement about the framing.
What the thread adds
- micoti — connects the result to Arditi et al. (2024), arguing that the gain “mostly comes from unblocking the looping.” That mechanism is the commenter’s interpretation.
- GavinAnderegg — still finds the capability surprising: “the examples provided in the original piece were eye-opening.” They accept that tooling exists while valuing Astra’s apparent ease with it.
- phoghed — disputes the need for the critique: “Wasn’t all that obvious to anyone who read the original post when they got to the section that explicitly explained what the author of this post is now summarizing?” kenjackson makes the same disclosure objection.
Handles are pseudonymous; HN publishes no per-comment scores, and its ranking determines the fetched order. This is a slice of the thread, not a consensus.