A 1.3 million parameter model beats GPT-4o-mini at real-time game control — and costs almost nothing to run

Researchers built a tiny neural network that plays a classic video game vastly better than language models 92,000 times its size, using a fraction of the computing power. The finding: when you train a small model on the specific task it needs to solve, instead of training a giant general-purpose model, you get better performance at lower cost — which matters for anything that needs to make fast decisions on regular hardware.

This is a direct measurement of something the AI industry has been betting against for three years: that bigger is always better. The paper shows that a model trained on 31,000 human gameplay demonstrations outperformed models trained on trillions of tokens of internet text, at real-time speeds on consumer hardware. What matters is not the game itself — it is the evidence that task-specific training on relevant data beats raw scale for control problems. This suggests the current trajectory (make models bigger, hope they work for everything) is not the only path, and may not be the cheapest one.