Worldmodeldata sells video game play data to train AI world models

In short: A British startup called Worldmodeldata is turning video game play data into training sets for a type of AI meant to understand how the physical world works.

What happened

Worldmodeldata says it has licensed almost 1 million hours of data from video game studios. This data includes what players see on screen and what they do with a controller, like moving a character with a thumbstick or pressing a button to jump.

The company wants to sell this packaged data to labs building “world models.” A world model is an AI system trained to predict what happens next in a scene, based on actions taken, like learning that if you push something, it moves (cause and effect). Researchers have argued this kind of AI could be important for machines that must act in the real world, such as robots.

Worldmodeldata is advised by AI researcher Yann LeCun. Its pitch is that it can act like a broker, so AI labs do not have to negotiate separate data deals with many different game studios.

Why it matters

Many popular AI tools today learn mainly from text, which is like reading about how to ride a bike without ever pedaling one. World models need training data that combines visuals and actions, but that kind of data is harder to find at large scale.

Not everyone agrees video games are the right source. Nvidia researchers and academics quoted by WIRED say game physics can be simplified or unrealistic, which may limit how well AI trained on games can handle delicate tasks, like gripping an object without dropping it.

Source: Wired