Qwen 3.8 27B Identifies Video Games Based on Raw Weights

A Qwen 3.8 27B model launched on a laptop identified Pac-Man, Pong, and Breakout from raw weights, with significant computation times and token volumes. These trials partially replicate an idea popularized by researchers, the "programming in the weight space," attributed to the capabilities of GPT-6 Astra. The local results come with failures and technical constraints that the experiment details.

Successes, a Failure, and Measured Delays

On a Pac-Man prompt, Qwen 3.8 27B produced a response after 46 minutes and 12 seconds, using 43,506 tokens of reasoning. The description evoked a grid world in the style of Pac-Man with an 8x8 maze, central and corner walls, moving objects or ghosts, and reward targets. The model also identified Pong in 1 hour, 4 minutes, and 54 seconds, with a formulation where the agent controls a vertical paddle on the right moving up and down to intercept or return a ball. Breakout was recognized in 52 minutes and 25 seconds as a paddle, ball, and bricks game similar to Arkanoid. In contrast, the Freeway prompt led to a poor hypothesis, equating the game to a Pac-Man style grid with movements in four directions. The generated reasoning traces, dense with references to channels and grid positions, are described as amusing but almost incomprehensible.

Local Technical Constraints and Configuration Used

The Qwen 3.8 27B model was run locally on a laptop; the version used weighs 17 GB. Loading required a longer context window, as the default of 8,192 failed with errors. To drive inference via LM Studio, the llm-lmstudio extension was installed. Running the command also required the environment variable LMSTUDIO_TIMEOUT=900.

Method: Autonomous Prompts and Comparison with Kimi

The evaluation prompts are autonomous and available on GitHub. The llm tool accepts a remote file as a prompt using the -f/--fragment option. One example involved running the Breakout prompt against Kimi K3.1 Pro via OpenRouter, with the required llm-openrouter plugin. In this context, Kimi K3.1 Pro guessed Breakout after 9 minutes and 41 seconds, using 19,493 tokens of reasoning.

What Research Attributes to GPT-6 Astra

Researchers describe the ability of certain LLMs to read and write the weights of neural networks. They cite GPT-6 Astra as capable of interpreting and modifying raw weights, which are considered a textual description filled with floating-point values. According to this work, Astra can determine a network's task by observing its weights and write small game networks from scratch without environmental feedback, achieving better performance than that obtained after 100 million training steps by PPO. Astra detected reasoning loops in Qwen3.5-0.8B during an attempt to improve mathematics without gradients and fixed the issue by modifying 0.016% of the weights. With the same sampling parameters, MATH-500 would have improved from 7.9% to 48.1%, and the effect would generalize to instruction execution, with an IFEval tripled. The whole is referred to as "programming in the weight space," and it would occur without writing training loops, relying on the model's reasoning to determine what each component of the network computes.

Reservations on the Scope of Conclusions

A local trial concludes that Qwen 3.8 27B can interpret weights to infer games, but the author expresses reservations about the generalist interpretations that arise from this. He finds the experiments reported in the literature interesting while considering that some ambitious comments are not substantiated. He neither claims the end of the source code era nor the advent of a future where machines modify their weights in incomprehensible ways. He also believes that an experienced human could, with enough effort, achieve similar identifications.