Qwen 3.8 27B on my laptop can interpret model weights, too

A recent study introduces “weight-space programming”: asking language models to infer what a neural network does, or alter its parameters, by inspecting the raw numerical weights. Its claims include that GPT-6 Astra could identify the behavior of small game-playing networks and improve a language model by editing a tiny fraction of its weights. These are distinct abilities, however: interpreting weights without tools is not the same as using code and experiments to modify a model. The account focuses on the no-tools test, in which models received weights from networks trained to play four simplified grid games and had to name each game without executing code or using a calculator.

Qwen 3.8 27B, running locally on a laptop, correctly identified Pac-Man, Pong, and Breakout from their weights, but mistook Freeway for a Pac-Man-like game. The successful attempts took roughly 46 minutes to an hour and consumed tens of thousands of reasoning tokens; Qwen described clues such as maze walls, moving objects, paddles, balls, and bricks. The reasoning was lengthy and difficult to follow, and the author notes that a knowledgeable person might reach the same conclusions with enough time and careful inspection. The results demonstrate a striking capacity to infer behavior directly from weights, but do not justify claims that source code is obsolete or that models can reliably improve themselves through opaque weight edits.