Critic says the GPT-6 Astra weight-editing post overstates its headline

A post on honnibal.dev argues that the write-up on GPT-6 Astra reading and editing model weights mixes different techniques under one claim. Its breakdown of the experiments finds that the pure setting, raw weights in context with no tools, appears only on three small networks of 17K to 172K weights. The headline Qwen3.5-0.8B maths result (MATH-500 from 7.9% to 48.1%) used Python and let Astra run the model on up to 50 batches of practice problems, and the same approach failed on Gemma 3. The author still says models from a year or two ago would have failed these tasks. On Hacker News (879 points), the biggest reply thread says the original post already explained its methods, and another says the critique adds nothing new. A commenter who says they work in interpretability research agreed the original oversells, while others say editing weights through Python is still impressive. No consensus. Sorted take: when a lab result sounds like magic, read the methods before planning around it.