Worldmodeldata sells video game data to train AI models

Worldmodeldata, a British startup advised by Meta chief AI scientist Yann LeCun, packages video game controller inputs into training data for AI world models. Chief executive Rhea Loucas told Wired the company also uses other data collected by video game studios. It acts as a broker, curating game data so labs do not need separate deals with individual studios.

Researchers say world models need paired visual and action data to learn physical cause and effect. That kind of data does not exist for images the way it does for text, researchers say. "For world models, you need cause and consequence," said Xiatian Zhu, an AI researcher at the University of Surrey. "On the internet, we have very little of this type of data."

Loucas said video games offer large volumes of controller inputs paired with visual outcomes. That could be an alternative to attaching sensors to humans and robots in test environments. Sensor data yields only small amounts of material and can miss rare real-world scenarios, she said. Other companies, including General Intuition and Niantic, are collecting similar data from their own game platforms rather than buying it from a broker.

Researchers including Fei-Fei Li and Yann LeCun have argued that large language models, trained only on text, may be poorly suited to tasks such as piloting vehicles or steering robotic arms. Whether video game data can close that gap is untested. World-model performance is assumed, but not yet shown, to scale with dataset size the way language models did.