Qwen 3.8 27B on a laptop identifies games from raw RL policy weights
Overview
Simon Willison says Qwen 3.8 27B, running locally on his laptop in LM Studio, inferred which game a reinforcement learning policy was trained to play from its raw weights, as described in a recent paper.
The model identified a Pac-Man-style game with 72% confidence after 46 minutes and 43,506 reasoning tokens. It identified Pong in 1 hour 4 minutes and Breakout in 52 minutes, and failed on Freeway.
Written by AI from the articles below · updated Nov 3, 12:23 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- Simon WillisonBlogQwen 3.8 27B on a laptop identifies the games behind raw RL policy weights
AISimon Willison says Qwen 3.8 27B running on his laptop can infer which game a neural network was trained to play from raw weights, as described in a recent paper. In LM Studio, the model took 46 minutes and 43,506 reasoning tokens to identify the Pac-Man-style game and reported 72% confidence. It also identified Pong and Breakout in 1 hour 4 minutes and 52 minutes, respectively, but failed on Freeway.
Heat Trend
Not enough continuous observations to show a trend yet.