A British startup called Worldmodeldata is trying to solve a bottleneck in AI development by turning video game telemetry into training data. The company, advised by Yann LeCun, curates controller inputs and visual footage from game studios and packages them as datasets for so-called world models—AI systems designed to understand and navigate physical environments. CEO Rhea Loucas argues that games offer abundant, varied examples of cause and consequence that text-based models lack.
The pitch comes as researchers including Fei-Fei Li and LeCun push world models as a path beyond large language models, which struggle with tasks like piloting robots or vehicles. But suitable training data is hard to come by: manual sensor collection is slow and misses edge cases. Worldmodeldata claims to have licensed almost 1 million hours of game data, though it declined to name the studios involved.
Not everyone is convinced. Ming-Yu Liu, who leads world model work at Nvidia, says game physics often takes shortcuts that make manipulation look realistic without capturing real forces, such as finger pressure. Xiatian Zhu of the University of Surrey agrees that game simulations are only coarse approximations of physical reality. Both suggest game data may be better suited for generating virtual environments than for precise robotic control.
The article presents the startup's approach as one of several competing strategies, with no consensus yet. As investor Nicole Fraenkel puts it, "There are many paths to the promised land," and the jury is still out on which will work best.