Worldmodeldata licenses almost 1 million hours of gaming data

Large language models are trained on oceans of text, but world models—AI systems designed to understand and navigate physical environments—need something different: paired visual and action data. That type of data barely exists on the open internet, creating a bottleneck for researchers like Fei-Fei Li and Yann LeCun who are betting on world models as the next frontier of AI.

Worldmodeldata's solution is to treat video game play as an exhaust product available in massive quantities. The startup, advised by LeCun, licenses controller inputs and gameplay data from game studios and packages it into curated training datasets, positioning itself as a broker that saves labs from negotiating individual deals. Some companies, like General Intuition and Niantic, already collect their own video game data, but Worldmodeldata's pitch is aggregation and organization.

The idea is not universally accepted. Nvidia's Ming-Yu Liu argues that video game physics is often eccentric and lacks the fine-grained detail needed for tasks like careful object manipulation, suggesting the data may be better suited for generating hyperrealistic video or 3D environments. Nicole Fraenkel of Khosla Ventures, which backs General Intuition, counters that video games offer the disorder and corner cases that matter most. The debate is likely to hinge on whether world models actually scale with dataset size the way language models do—and whether video game physics can ever be detailed enough for real-world dexterity.