笔记本上的 Qwen 3.8 27B 也能解读模型权重
原标题：Qwen 3.8 27B on my laptop can interpret model weights, too
AI 摘要
博主 Simon Willison 复现了论文《Sparks of Weight-Space Programming》中的实验：该论文称 GPT-6 Astra 等前沿模型能直接读取和修改神经网络原始权重，例如识别网络训练的游戏、在无梯度条件下修改 Qwen3.5-0.8B 的 0.016% 权重，将 MATH-500 从 7.9% 提升到 48.1%。Willison 用笔记本上的 Qwen 3.8 27B 运行论文提供的提示词，成功识别出 Pac-Man、Pong 和 Breakout 三个游戏，但耗时 46 分钟到 1 小时以上。spaCy 作者 Matthew Honnibal 等人批评论文混淆了「无工具读权重」与「借助代码改权重」两类能力。
正文节选
Qwen 3.8 27B on my laptop can interpret model weights, too 3rd November 2026 Yesterday’s hot paper was Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. The paper found that frontier models, in particular GPT-6 Astra, can both interpret and modify the raw weights of a machine learning model—literally a text description of the model containing a bunch of floating point numbers. I’ve been playing w