Qwen 3.8 27B on my laptop can interpret model weights, too
Points and comments are a snapshot, not live.
Qwen 3.8 27B running locally can identify games from raw neural network weights.
Simon Willison replicates a key task from the "Sparks of Weight-Space Programming" paper, where LLMs interpret raw model weights to determine which game a network was trained on. He feeds Qwen 3.8 27B (running on his laptop via LM Studio) the same 10x10 grid weights for Breakout, Pong, Pac-Man, and Freeway. Qwen correctly identified Breakout (88% confidence, 52 min), Pong (75%, 65 min), and Pac-Man (72%, 46 min), but failed on Freeway. The paper also claims GPT-6 Astra can edit weights to fix reasoning loops, which critics like Matthew Honnibal and Sean Goedecke argue conflates direct weight interpretation with standard ML tool use.
Willison notes the reasoning traces are "almost incomprehensible" but says a human with enough effort could likely do the same. He cautions against overinterpreting the results: "I'm not ready to declare that the age of source code is over."
What commenters are saying
Top comments focus on the limits and realism of weight-space programming. One thread notes the Qwen 3.8 fix for Qwen 3.5 looping is "a fairly straightforward abliteration method on a model that is known for a specific pathological behavior," not general weight editing. Another comment flags that all tested games are famous 2D Atari titles, and the prompt likely hints strongly at those options. A Shakespeare quote ("I can call spirits from the vasty deep") is reused to question whether correct interpretation was actually demonstrated, leading to a split: some argue the paper's test for correctness is simply improved task performance, while others say the impact on overall model benchmarks was not studied.
Several commenters highlight that weights are just token sequences, making the task less surprising for LLMs. A sub-thread discusses whether Qwen could improve quantized (compressed) models to reduce looping, acknowledging this would be practically valuable but noting the paper's method is not general-purpose optimization.