Ai's new frontier: training models with gaming data

A British startup, Worldmodeldata, is leveraging video game data to train world models, aiming to bridge the gap between AI and real-world physics. The company, advised by Yann LeCun, collects controller inputs and gameplay data from popular games, offering curated datasets to AI labs. CEO Rhea Loucas argues that video games, with their vast and diverse experiences, can teach AI to navigate complex environments. However, critics like Nvidia's Ming-Yu Liu caution that video game physics is often simplified, making it less suitable for tasks requiring fine motor control. Meanwhile, academic Xiatian Zhu notes that while games are simulators, they are inherently approximate. Despite skepticism, Loucas believes gaming data could be the key to a 'GPT moment' for world models, potentially revolutionizing their utility.