#Qwen

A Hacker News–discussed blog post contends that claims about Astra reading and writing model weights to improve Qwen are misleading, arguing that the demonstrated effect is narrower than the framing implies. The critique&\x27;s central point, as summarized in the discussion, is that the key result reduces reasoning loops—cases where the model assigns too little probability to the token that ends its reasoning—rather than making Qwen better at math. Commenters working in mechanistic interpretability say such loops can be addressed with simple steering vectors, while others note that the original post already described some of these limits and that the techniques resemble prior published work.

Simon Willison reports that Qwen 3.8 27B, running locally in LM Studio from a 17GB Q4\K\M build, identified Pac-Man from a prompt containing raw policy-network weights, answering after 46m12s and 43,506 reasoning tokens with 72% stated confidence; he also reports it identified Pong. The prompt comes from &\x27;Sparks of Weight-Space Programming,&\x27; a paper-shaped blog post by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab that claims frontier models, especially GPT-6 Astra, can interpret raw weights and, in one example, edit 0.016% of Qwen3.5-0.8B&\x27;s weights to raise MATH-500 from 7.9% to 48.1%. Those claims are contested: Matthew Honnibal called the paper misleading for conflating tool-free weight reading with code-assisted model updates, and Sean Goedecke argued the Qwen3.5-0.8B edit follows steps a human ML researcher would take. Willison&\x27;s laptop results are informal, individually reported runs of the game-identification experiment, not independent confirmation of the paper&\x27;s broader editing claims.