Worldmodeldata, a UK startup, collects game control data to create datasets for training AI that better understands the physical world.
Worldmodeldata, a British startup, is gathering player control data from game studios—including joystick movements, button presses, and changing 3D scene images related to those actions—to compile datasets for training World Model AI designed to understand and operate in the physical world.
Worldmodeldata states it has secured rights to nearly one million hours of data. However, some researchers, including Nvidia’s lead development team, question whether game physics are detailed enough for real-world applications.
Worldmodeldata views player control data already collected by game studios as a massive byproduct resource that can help solve the shortage of training data for World Models. Some companies like General Intuition and Niantic already build models using their own platform data, but Worldmodeldata positions itself as a middleman organizing and curating data so AI labs need not negotiate individually with numerous game studios.
Rea Lucas, CEO of Worldmodeldata, told WIRED that millions of quality games exist today, increasingly close to reality. She questioned why such abundant, diverse video game experiences are not used to train AI and said the company plans to eventually enable individual players to earn compensation for their gameplay data.
Lucas believes that ultimately, video game data will become the primary raw material for training World Models, which can then be fine-tuned with environment- or task-specific real-world data.
Not everyone shares this positive view. For example, Nvidia, which released a World Model tailored to run on its chips using a proprietary physics engine to simulate real-world physics as AI’s core, has skepticism. Ming-Yu Liu, Nvidia’s World Model lead, thinks models trained on game control data likely won’t perform well in tasks requiring precise motion control, such as carefully grasping objects, because game physics often use shortcuts and approximations. For instance, a character reaching for an apple on a table may not have code specifying individual finger pressures needed to hold the apple securely.
Similarly, Xietian Zhu, an AI associate professor at the University of Surrey, expressed concerns that video games are essentially simulators based on rough physics approximations rather than detailed, accurate models.
This concept arises from a growing belief among some AI researchers that Large Language Models (LLMs), which learn only from text, have fundamental limitations and may be ill-suited for tasks demanding precision and subtlety, such as autonomous driving.
/source:Wired