The AI industry is recognising that large language models, trained solely on text, are ill-equipped for physical-world tasks such as piloting autonomous vehicles or operating robots. Leading researchers, including Fei-Fei Li and Yann LeCun, are now focusing on "world models"—a different class of AI that requires training on visual and action data to understand real-world physics. However, unlike the vast text corpora available for language models, there is currently no comparable dataset for world models, creating a significant bottleneck to progress.

Worldmodeldata, a British startup advised by LeCun, aims to solve this problem by licensing video game data as training material. The company has secured approximately 1 million hours of gameplay footage from popular game studios, reasoning that games provide precisely what world models need: millions of diverse scenarios paired with player actions and consequences. Supporters argue that this abundance of diverse, varied data—including edge cases—makes video games ideal for training models on safety-critical tasks like autonomous vehicles and industrial robotics, though some, including Nvidia, remain sceptical of the approach.