Post

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

Matthew Honnibal argues the striking Qwen result combined coding and model runs, not direct inspection of raw weights in the model’s context. The genuinely raw-weight experiments involved much smaller networks; the Qwen experiment used Python and up to 50 batches of practice problems.

He still regards the coding achievement as meaningful, but says the experiments should not be conflated. HN readers dispute whether the original post’s details made the distinction sufficiently clear; some find the tool-assisted result impressive regardless.