Why this CEO thinks video games make better training data than the internet

July 9, 2026 - 03:34

General Intuition CEO Pim de Witt recently sat down for an interview, making the case that video game data is far superior to internet scrapings when it comes to training artificial intelligence. His startup just closed a massive $320 million funding round, and de Witt argues that the interactive, physics-rich environments of games offer something the open web cannot: structured, cause-and-effect feedback.

According to de Witt, the internet is a chaotic mess of text, images, and videos that lack real-world constraints. A model trained on internet data might learn what a cup looks like, but it won't understand what happens when you push it off a table. Video games, by contrast, simulate gravity, collision, and spatial logic. Every action triggers a predictable reaction. This makes gaming data ideal for teaching AI about physical interaction, which is exactly what robotics needs.

The CEO explained that his company is not just building chatbots or image generators. They are focused on "physical AI" -- systems that can navigate, manipulate objects, and operate in the real world. To get there, they need training data that mirrors reality without the noise of the internet. Games provide millions of hours of simulated physics, all labeled by the game engine itself.

De Witt also addressed the scale of the challenge. While $320 million sounds like a lot, building the next generation of robotics AI requires enormous compute and data infrastructure. He believes that the gaming industry, often dismissed as entertainment, holds the key to unlocking machines that can actually move and work alongside humans. The pitch is simple: if you want robots that don't bump into walls or drop things, train them in a world where physics actually matters.