Worldmodeldata Says It Licensed Nearly 1 Million Hours of Game Data for AI Training

Worldmodeldata says it has licensed nearly 1 million hours of video-game data to train AI on how actions change a scene. That is a substantial supply for a field short of action-linked data, but its value for real-world tasks remains in dispute: a game can show a character picking up an apple without modeling how firmly each finger grips it.

A broker for cause and consequence

The British startup packages visual representations of 3D game spaces alongside the actions players took in them. A world model learns relationships between actions and their effects; a controller input paired with the resulting scene offers a different lesson from video alone. University of Surrey AI researcher Xiatian Zhu told WIRED that the internet has little of this cause-and-consequence data.

Worldmodeldata’s pitch is also about access. General Intuition and Niantic collect data from their own games or platforms; Worldmodeldata instead aims to organize data from multiple studios so AI labs need not negotiate with each one. CEO Rhea Loucas declined to name the studios behind the licensed hours. The figure describes data licensed, not a demonstrated improvement in any model trained on it.

Where gameplay may help

Loucas argues that games offer abundant, varied experiences. The company’s longer-term theory is to train world models mostly on game data, then adapt them with data from the real environment or task. Researchers have not yet established that increasing the volume of this kind of training data will deliver the gains Worldmodeldata anticipates.

Collecting comparable data in physical settings is difficult to scale. Labs can put sensors on people and robots, but those exercises produce limited data and may miss unusual situations. Nicole Fraenkel, a Khosla Ventures partner whose firm has invested in General Intuition, argues that repeatedly demonstrating a simple task will not capture the disorder a machine may face outside a test setting. Her case favors variety; it does not settle whether game actions contain the physical detail a robot needs.

The apple is not really held

Ming-Yu Liu, who leads world-model development at Nvidia, draws a line at fine motor control. Game physics can create the appearance of a grasp without modeling the finger pressure needed to hold an apple. Liu says that makes game inputs a weaker fit for teaching careful object manipulation, though he sees more promise for generating realistic video or 3D environments. Zhu likewise calls game simulations physically grounded to a degree, but coarse.

Nvidia takes a different route for its own world models, using a custom engine intended to replicate real-world physics. That contrast points to two questions a training team would have to keep separate: whether a dataset covers many situations, and whether it represents the forces involved in the task. More game hours could address the first question without answering the second.

Who supplies the next hours?

For now, the licensing business runs through studios. Loucas says Worldmodeldata eventually wants ways to compensate individual players for their data, but that is an ambition, not a current arrangement described in the interview. The immediate test is more basic: which abilities, if any, improve when a model learns from recorded gameplay before it encounters the physical world?