Worldmodeldata Licenses 1 Million Hours of Game Data for AI World Models
Updated
Updated · WIRED · Sep 28
Worldmodeldata Licenses 1 Million Hours of Game Data for AI World Models
1 articles · Updated · WIRED · Sep 28
Summary
Nearly 1 million licensed gameplay hours are being packaged by British startup Worldmodeldata into training datasets for AI world models, aiming to give machines paired visual-and-action data that text-trained LLMs lack.
That pitch targets a key bottleneck: world models need cause-and-consequence data to learn physics and control, while manually collecting robot or human demonstrations is slow, costly and poor at covering rare edge cases.
Worldmodeldata positions itself as a broker between AI labs and game studios, curating controller inputs and gameplay traces so customers do not have to negotiate separate deals with many developers.
Nvidia and some researchers question how far game data can go, arguing video-game physics are often too coarse for fine motor tasks like object manipulation even if they may help with video or 3D environment generation.
The debate reflects a broader race to find the data source that could deliver a ChatGPT-like breakthrough for world models, with game data one contender among several still-unproven approaches.