Why this CEO thinks video games make better training data than the internet
TL;DR
TechCrunchs Equity podcast profiles General Intuition: CEO Pim de Witte argues that pure LLMs are too text-bound to reach AGI. The missing piece, in his view, is spatial-temporal understanding. The startup uses gameplay data from Medal TV. The key asset is not just video footage, but logged player actions: which button was pressed, when, and how the game world responded.
Nauti's Take
General Intuitions thesis is strong because it treats training less as more internet and more as cleaner cause-and-effect data. A game clip with logged controls can be more useful for agents than another pile of web text, because action and reaction sit together.
Still, the story is almost engineered for VC gravity: proprietary data, AGI language, robot demo, defense ethics, billion-dollar valuation. The real test starts when customers run the model outside polished demos, in messy factories, streets, or rescue scenarios.
Briefingshow
The story hits a weak spot in the LLM debate: text models can explain, but they do not have lived experience of what happens when a body takes action. If gameplay really provides useful action data, gaming becomes infrastructure for robotics, simulation, and physical agents. Much of this is still demo, funding story, and founder thesis.
Sources