General Intuition’s $2.3B bet that video games can train AI agents for the real world

A startup is using hundreds of millions of hours of video game footage to train AI agents that can understand how to move through and interact with physical spaces. The company's approach relies on action labels embedded in gameplay clips that show exactly what buttons players pressed and when, rather than trying to infer actions from video alone. This matters because the startup believes gameplay data is a scalable shortcut to training AI models that work in the real world, something competitors typically achieve only through expensive collection of real-world data. The company has secured substantial investment backing, including from major venture firms and notable investors, based on the belief that this approach could lead to general-purpose AI agents.