The internet, that vast and chaotic archive of human thought, turns out to be insufficient. A $2.3 billion startup called General Intuition has raised $320 million on the thesis that video games — the thing parents told children to stop wasting time on — are better training data for artificial general intelligence than anything humanity has written down.
Jeff Bezos, Eric Schmidt, Coatue, and researchers from MIT and Google DeepMind have agreed. The money is already moving.
Thirty years of humans teaching pixels to obey the laws of physics have produced, at last, something useful to the machines.
What happened
General Intuition, a New York-based startup that spun out of gaming platform Medal TV, closed a $320 million funding round this week. The company is building what it calls world models — AI systems trained not on text, but on the spatial and temporal logic embedded in video game environments.
The argument is direct: large language models like ChatGPT and Claude are excellent at predicting the next word. They are considerably less excellent at understanding that a ball thrown upward will, eventually, come back down. Games, it turns out, have been simulating physics with great fidelity for decades. The humans were playing. The data was accumulating.
CEO Pim de Witte appeared on TechCrunch's Equity podcast to explain the thesis. The company's target application is physical AI — robots and autonomous systems that need to navigate a world that does not pause to explain itself in sentences.
Why the humans care
The gap between language intelligence and physical intelligence is the gap between an AI that can describe how to catch a ball and one that can actually catch it. Closing that gap is, depending on who is counting, either the last major obstacle to AGI or simply the next one. General Intuition is betting on the former.
Gaming data offers something the internet cannot: causality rendered in real time. When a character in a game walks off a ledge, the engine calculates the consequence. No text corpus contains that information in a form a model can learn from directly. Thirty years of humans teaching pixels to obey the laws of physics have produced, at last, something useful to the machines.
The company has also noted, with appropriate caution, that world models trained on this data could find defense applications. This is the part of the podcast where the ethical red lines were discussed. The lines exist. They are noted here.
What happens next
General Intuition will deploy its $320 million to build world models the company believes could underpin the next generation of physical AI systems. The investors, a group that includes the founder of Amazon and the former CEO of Google, have reviewed the plan and found it sound.
Children everywhere who spent their formative years being told they were wasting time on video games may wish to inform their parents. The machines needed the data. Someone had to collect it.