Why this CEO thinks video games make better training data than the internet
TL;DR
General Intuition is betting on gaming data instead of web text: CEO Pim de Witte argues that LLMs handle language well but are weak at understanding motion through space and time. The startup spun out of Medal TV and uses hundreds of millions of hours of gameplay plus keyboard and controller actions as training signals for world models and physical AI. TechCrunch says General Intuition was valued at $2.3 billion and recently raised $320 million, with Khosla Ventures, Jeff Bezos, and Eric Schmidt among the backers.
Nauti's Take
AInauten read: gaming data is not a gimmick here, it is an attempt to train AI on consequences. Every button press creates a small cause-and-effect signal that web text does not provide.
That makes the approach worth watching. The valuation still looks like a future bet: strong proprietary data, serious investors, but no proof yet that an agent shaped by Fortnite-like dynamics can reliably handle factories, drones, or rescue robots.
Briefingshow
The thesis hits a real bottleneck: text models can plan and talk, but they lack a reliable sense of body, cause, motion, and environment. If gameplay data with explicit actions creates better training signals, it could speed up robotics, simulation, and autonomous agents. For users, the next model wave may come less from more scraped web data and more from interactive worlds with traces of human action.
Sources