Honnibal Challenges MIT's Astra Weight-Editing Claims
Matthew Honnibal argues that MIT's Astra research misleadingly groups small-network weight inspection with the repair of Qwen3.5-0.8B. According to Honnibal, the Qwen experiment used Python and up to 50 forward tests rather than directly reading raw weights, unlike the restricted small-network experiments. MIT claims Astra changed 0.016% of the model's weights and raised its math score from 7.9% to 48.1%, though Honnibal says the underlying tool-use progress remains significant.
The critique materially narrows the scope of a prominent result claiming that a model can inspect and edit another model's weights, separating small-network raw-weight experiments from an iterative, tool-assisted Qwen repair experiment.
