A British startup, Worldmodeldata, is utilizing video game inputs to create training datasets for artificial intelligence models, particularly those designed to navigate the physical world. This approach aims to address the current limitations of large language models, which struggle with real-world applications due to a lack of suitable training data.
By packaging controller inputs and other data from video games, Worldmodeldata hopes to provide a rich resource for developing world models. These models require diverse training data to improve their performance, and the startup positions itself as a broker to streamline data acquisition for research labs, potentially transforming AI training methodologies.
Watch for Worldmodeldata's partnerships with game studios to expand their dataset offerings. As they refine their approach, the implications for robotics and autonomous systems could be significant, potentially leading to breakthroughs in real-world applications of AI.