Matthew Honnibal argues that the Astra study conflates reading raw weights of small neural networks with editing models through Python. Raising Qwen3.5-0.8B’s math score from 7.9% to 48.1% allowed code and up to 50 batches of practice problems. Separately, Simon Willison showed Qwen 3.8 27B on a laptop identifying Pac-Man, Pong and Breakout from small game-model weights without tools, but failing to identify Freeway. These results have different conditions. The original study also varies tools and feedback across experiments, so they do not demonstrate a single general capability or universal model improvement.