Honnibal Says GPT-6 Astra Qwen Math Run Used Python and Practice Batches

Matthew Honnibal says a post claiming GPT-6 Astra reads and writes raw weights to improve Qwen3.5-0.8B is misleading. Honnibal notes the raw-weights method only applied to networks of 17K to 172K weights, while the Qwen maths run used Python and up to 50 batches of practice problems. That run lifted MATH-500 from 7.9% to 48.1%, he writes.

• Astra writes MinAtar game networks that beat PPO-100M only on Breakout, 517 to 25, Honnibal notes.
• The Python-assisted maths method fails on Gemma 3 270M, Honnibal reports from the same experiment table.
• Astra lifts a stuck Kinetix car policy from 0% to 100% by changing 3.5% of weights, Honnibal says.