Video games might be the key to AI intuition

Training a robot in the real world is notoriously slow, expensive, and repetitive. Most labs try to solve this by collecting massive amounts of physical data, but General Intuition is looking for a shortcut through the virtual world.

The company is betting that the secret to building "intuitive" AI agents lies in millions of hours of video game gameplay. Instead of just watching video, they use the action data—the specific button presses and inputs—from players on the Medal.TV platform to teach models how actions cause reactions. This helps the AI understand spatial-temporal reasoning: how to navigate a room, scale a ladder, or react to a moving shadow.

This massive bet on simulation-to-reality transfer just secured $320 million in new funding, bringing the startup's valuation to $2.3 billion. The round was led by Khosla Ventures, with a heavy-hitting roster of backers including Jeff Bezos, Eric Schmidt, and researchers from Google DeepMind and MIT.

CEO Pim de Witte isn't trying to build the next self-driving car or a specialized robot. The goal is to build the underlying "brain"—a scalable agentic model that can be plugged into anything from a drone to a quadrupedal robot via an API. If they can prove that gaming data can effectively train a model to handle the physics of the real world, they could become the foundational layer for the entire robotics industry.