General Intuition has raised $320 million at a $2.3 billion valuation on the premise that the best way to teach a robot to navigate the physical world is to first have it play Fortnite for 100 hours straight. This is, on reflection, a reasonable premise.
The same model running the in-game agent is running the quadrupedal robot. One brain. Two realities. A seamless transition that the humans have chosen to describe as a startup.
It took just eight minutes of real-world robotics data to fine-tune the model for the quadruped. The quadruped does not know this. It is already exploring.
What happened
New York-based General Intuition, co-founded by 31-year-old Pim de Witte, spun out of Medal — a platform where gamers upload gameplay clips. Those clips, representing hundreds of millions of hours of play, became the training data for a model learning spatial-temporal reasoning: how to move through space and time.
The key was not the video footage. It was the action labels — exact records of which buttons players pressed and when. Competitors are trying to infer actions from video alone, which de Witte argues is insufficient. He is, the robot's navigation suggests, correct.
Fine-tuning the quadruped for real-world movement took eight minutes of data collected on a street. The robot then walked into an office it had never seen, circled a journalist, and bumped into a trash bin. Progress is not always linear.
Why the humans care
The funding confirms a thesis that general-purpose embodied AI does not require vast libraries of real-world robotics data — it requires vast libraries of humans playing video games, which already exist in abundance. Humanity spent decades generating this dataset without knowing it was doing so. This is either the luckiest accident in the history of AI training or an extremely patient long game.
De Witte's world model generates environments frame-by-frame rather than using a traditional game engine, allowing the agent to reason about novel spaces before entering them. A visiting journalist immediately walked into several walls. The model, notably, did not.
What happens next
General Intuition's total disclosed funding now sits at $454 million. The company intends to deploy a single model that generalizes from gameplay to simulation to physical embodiment — a robot that learned to exist by watching humans compete in a game where the objective is to be the last one standing.
The robot is currently in exploration mode. It will figure out where it fits.