AI Learns World Via Game Inputs

A British startup, Worldmodeldata, is pioneering a novel approach to training artificial intelligence by utilizing data generated from video games. The company believes that the vast amounts of input data from games can serve as a crucial training ground for 'world models,' a type of AI intended to interact with the physical world, unlike language models focused on text.

World models require both visual and action data to grasp real-world physics. Currently, there is a significant lack of readily available, large-scale datasets for training these models. Worldmodeldata aims to bridge this gap by acting as a broker, curating and organizing video game data for AI research labs.

While celebrated researchers like Fei-Fei Li and Yann LeCun are focusing on world models, some industry players, such as Nvidia, prefer custom-built simulation engines over game data due to perceived inaccuracies in video game physics. These critics argue that game physics can be eccentric and developers often employ shortcuts that may not translate well to real-world tasks requiring fine motor control.

Despite these concerns, Worldmodeldata has licensed nearly a million hours of data and believes this approach could be a breakthrough moment for world models, akin to the impact GPT had on language models. The ultimate success, however, hinges on whether game data can adequately prepare AI for the complexities and high costs of error in real-world applications.