Why this CEO thinks video games make better training data than the internet
TL;DR
General Intuition CEO Pim de Witte argues that LLMs are strong at text but weak at space, time, motion, and causality. His startup uses gameplay data from Medal: not just video clips, but also button and mouse actions showing what players did and when. TechCrunch says General Intuition was valued at $2.3 billion and recently raised $320 million, with backers including Khosla Ventures, Jeff Bezos, and Eric Schmidt.
Nauti's Take
This is one of the more credible counters to the idea that more web text automatically creates more general intelligence. Gameplay data has something many AI datasets lack: intent, reaction, and timing inside a controllable environment.
Still, the valuation is not the breakthrough. General Intuition is selling a strong hypothesis with strong investors, not proof that robots trained through game data can reliably operate in factories or streets.
Briefingshow
The important part is not gaming as a fun data source, but labeled action: the model sees a scene and knows which action changed it. For robotics, simulation, and autonomous systems, that could be more useful than internet text alone. The hard question is whether game worlds contain enough physical reality to train reliable real-world machines.
Sources