Qwen reads game-network weights in a laptop experiment

Simon Willison tried Qwen 3.8 27B on his laptop with prompts containing a game-playing network’s numerical parameters, or weights. He reports that it identified Pac-Man, Pong and Breakout, but mistook Freeway for a Pac-Man-like game. Before prompting, he explicitly loaded the model with a longer context after default-setting errors and increased LMSTUDIO_TIMEOUT. These are reported experiments, not a general guarantee of interpreting arbitrary networks.

The research behind the prompts deliberately gives each game network the same dimensions, input channels and available actions. The authors supply textual weights and describe the grid without naming the games. That design makes the task about interpreting the parameters rather than recognizing a different network shape. Their broader findings also include failures: Astra could not distinguish their sycophantic language models from weights alone; observing attention patterns changed that result. Tool access differs between experiments.

For readers following the local setup, the plugin documentation supplies concrete prerequisites. Plugin version 0.3.0 onward needs LLM 0.32 or newer, and LM Studio must be at least 0.4.0. Install LLM before the plugin, start the LM Studio server, then list models and use the actual returned identifier. In general plugin use, a selected downloaded model need not already be loaded: the plugin loads it automatically, and the loading endpoint returns after completion.