The Next Evolution of AI Is Learning From Your Dodgy Gaming Skills
wired ·
A British startup named Worldmodeldata is attempting to solve the data scarcity problem for AI world models by using video game controller inputs as training material. While large language models are limited by their inability to navigate the physical world, world models require vast amounts of visual and action data to understand cause and consequence in reality. The company argues that the millions of hours of gameplay data available from studios offer a diverse and abundant corpus that includes rare corner cases essential for robust AI performance. This approach aims to replicate real-world physics through the varied experiences found in video games, potentially leading to a breakthrough moment similar to the rise of GPT. However, experts like Nvidia’s Ming-Yu Liu remain skeptical, noting that video game physics are often simplified approximations unsuitable for tasks requiring fine-grained motor control. Critics argue that while games serve as coarse simulators, they lack the detailed physical interactions necessary for precise manipulation tasks. Consequently, the industry remains divided on whether gaming data will become the primary foundation for world models or if specialized simulation engines are required.
wired