量子位· Jay·· 3 hours agoSelectedAI score69
Matthew Honnibal Questions MIT's Astra Research: Editing Qwen Weights Is Not Direct "Mind Reading"
MIT神作被锤：Astra爆改Qwen权重？就吹吧！
AI guide
spaCy creator Matthew Honnibal published a critique arguing that MIT's Astra research is seriously misleading, pointing out that its method for modifying Qwen's weights is not the "mind-reading" feat of solely inspecting raw weights as advertised.
Why it matters
The original article compares two experimental setups—inspecting weights alone versus executing code to test them—breaking down the actual boundaries of models directly rewriting weights and dissecting how the paper packaged the claims.
Source:量子位 · qbitai.com