British startup Worldmodeldata is building a training-data business around video games, pairing recordings of simulated 3D environments with the controller actions taken inside them. The company says it has licensed almost 1 million hours of data from game studios, although it has not identified the studios or titles involved.
A Different Kind of Training Corpus
World models need examples that connect observations with actions and consequences. Such data could eventually support systems that navigate spaces or control physical equipment, but the internet does not offer an equivalent to the enormous text collections used for large language models. Some developers instead collect demonstrations from instrumented people or robots, an approach that can produce relatively limited and repetitive datasets.
Games as a Data Supply Chain
Worldmodeldata, which is advised by Yann LeCun, proposes using the large volume of interaction data already produced by games. It aims to act as a broker that licenses, organizes and packages material from multiple studios, reducing the need for AI laboratories to negotiate separately with each developer. Other companies, including General Intuition and Niantic, collect game data from their own platforms for model development.
CEO Rhea Loucas expects game data to provide broad initial training before models are refined with information from a particular real-world task or environment. The startup also wants to create a future mechanism through which individual players could be compensated, but the source gives no launch date or operating details for that plan.
Scale Does Not Guarantee Realism
The central hypothesis remains unproven: larger collections of visual and action data may improve world models, but game environments simplify or alter physical behavior. A controller input can show that an action produced an outcome without capturing the forces, grip pressure or precise movements required to reproduce that outcome with a machine in the physical world.
Useful for Some Tasks, Limited for Others
Nvidia world-model lead Ming-Yu Liu told WIRED that game-derived data is unlikely to transfer well to tasks requiring fine motor control, such as carefully manipulating objects. Nvidia instead uses a custom engine designed to reproduce real-world physics. Liu said gaming data may be better suited to models that generate realistic video or 3D environments, while University of Surrey researcher Xiatian Zhu similarly characterized game simulations as physically grounded but coarse approximations.
AI-generated editorial illustration; not a documentary photograph.
Source: The Next Evolution of AI Is Learning From Your Dodgy Gaming Skills
SIGNAL TITANS
Leave a Reply