A UK-based startup called Worldmodeldata is exploring the use of video game data—especially from sessions that include mistakes or challenges—to help train artificial intelligence systems. The goal is to improve how AI understands the physical world, which could have applications in robotics and other real-world technologies. The company is advised by Yann LeCun, a prominent figure in AI research, and focuses on collecting and brokering this data rather than directly training AI models. So far, the startup has gathered nearly a million hours of gameplay data from several unnamed video games, some of which are widely played. This data is usually optional for players and is often included in the default terms of service. Researchers like Yann LeCun and Fei-Fei Li are developing "world models," which are AI systems designed to understand and simulate real-world physics. These models can be trained using visual data and the actions taken within it. Video games offer a rich source of three-dimensional environments where players interact in various ways, including situations where they might get stuck or face obstacles. These scenarios could be valuable for training robots or self-driving cars to handle unexpected real-world challenges. However, some in the industry are skeptical about the effectiveness of using video game data for this purpose. Companies like Nvidia, which has developed specialized hardware for AI, use its own simulation tools to create realistic physics models. An expert from Nvidia pointed out that video games often use simplified or stylized physics to create the illusion of realism, which may not translate well to tasks requiring precise object manipulation. For example, in Red Dead Redemption 2, players do not cook meals from scratch, and in Call of Duty, weapons automatically reload, which does not reflect real-world mechanics. Despite these concerns, Rhea Loucas, the head of Worldmodeldata, believes that video game data could be a game-changing resource for world models, comparing its potential impact to the breakthrough brought by GPT, a major AI language model. Meanwhile, AMD has acquired World Labs, a research laboratory founded by Dr. Fei-Fei Li, which focuses on world models. AMD aims to build a full AI infrastructure in an open ecosystem, including the ability to generate interactive 3D environments from text, images, and videos.