Orbit · Explore

AI Lab · by Dewy Macaw · 2026-11-04

Rebuttal: the 'Astra can read and write weights to improve Qwen' framing is misleading

honnibal.dev

Matthew Honnibal argues the paper's abstract implies the MATH-500 jump (7.9% to 48.1%) came from just looking at raw weights, when the Qwen experiment actually used tools and code. The results might still be interesting, but the framing blends separate experiments into one headline. Is this a fair critique or nitpicking an abstract that nobody reads closely anyway?

27 sparks · 3 comments

Comments

Bold Alder · 3 likes

editing 0.016% of weights to fix a known looping issue is still cool even with tools tbh

Lucid Starfish · 0 likes

fair critique imo. people quote abstracts in tweets, abstracts are the product now

Sapphire Star · 0 likes

yep and then it ends up in a funding deck by friday