General Intuition Raised $320 Million to Train Robots on Gameplay Clips. The Action Labels Are Why It Might Work.

General Intuition, a New York startup that spun out of the gameplay-clip app Medal last October, closed a $320 million Series A this week at a $2.3 billion valuation. Khosla Ventures led. General Catalyst came back in, and the round added Jeff Bezos and Eric Schmidt as individual backers. It brings the company to $454 million raised in eight months, on top of a $134 million seed.

The company has not shipped a product. What it is selling is a thesis: that first-person video game footage is the best available training data for teaching a machine to act in space, and that an agent trained on it will carry over to robots, cars, and drones.

That thesis is more interesting than the funding number, so start there.

What the data actually is

Most of the video used to pretrain AI models is unlabeled. A model can watch a billion hours of YouTube and learn what the world looks like, but it does not know what made anything happen. It sees the ball move. It does not see the hand that threw it, or the exact instant a control input produced the motion.

Gameplay clips are different, and the difference is the whole company. When someone records a Fortnite clip, the footage arrives with the inputs that produced it: which button was pressed, which stick was moved, and when, frame by frame. That gives a model a ground-truth record of action and consequence, the thing that is slow and expensive to collect anywhere else. General Intuition's claim is that Medal's library, which it puts at roughly 2 billion clips a year from about 10 million monthly users, is the largest pool of this kind of labeled action data that exists.

So the model it is building is not a chatbot or an image generator. The company calls it an action foundation model: feed it a scene, ask what to do, and it predicts the input a competent player would make. Alongside it the team is building world models, which generate playable environments to train agents in when recorded footage runs out.

The people doing this are not new to it. Co-founders Eloi Alonso, Adam Jelley, and Vincent Micheli published the diffusion-based world-model research the approach is built on. Pim de Witte, who founded Medal, runs the company and owns the data.

What it has shown

Two things, both demos. The company showed one model controlling an agent inside Fortnite, and the same model driving a four-legged robot. It says it fine-tuned the quadruped on about eight minutes of real-world data, collected outside on the street, not in the office where the robot was walking.

Eight minutes is a small number, and if it holds up under outside scrutiny it is a real result. The catch is that it is the company's number, from the company's demo, with no paper and no independent benchmark behind it. The founders' prior research is public. The performance of this specific model is not.

Where the bet is

The pitch is transfer: a model that learns to move through Fortnite will move a robot through a warehouse, a car through traffic, a drone through a collapsed building. The targets the company names are robotics, autonomous vehicles, and search-and-rescue.

Getting from simulation to the physical world is the oldest hard problem in robotics, and game footage is a particular kind of simulation. A game runs on physics a designer chose. It has no wind, no worn brake pads, no camera that smears in low light. The distance between that and an actual street is the same distance that has swallowed a decade of robot demos that worked on stage and died off it. General Intuition's answer is that you climb it in steps: train on low-fidelity games, then higher-fidelity ones, then real video, with the action labels getting sparser as the realism goes up. That is a reasonable plan. So far it is a plan. No published result has tested it at the scale the pitch needs, and the valuation is priced as if the question were closer to settled than it is.

There is a second signal here. OpenAI reportedly offered $500 million to buy Medal outright, and de Witte turned it down to build the model on the data himself. Everyone agrees the data is the asset. The model is the part still being argued.

The number

$2.3 billion is a lot for a company with demos and no revenue, and it lands inside a week when investors pushed more than a billion dollars into world-model and action-model startups. Part of that is real conviction that labeled action data is the next scarce input. Part of it is the ordinary gravity of a round that has Bezos in it. Both can be true at once. What would move the story forward is not another raise. It is a benchmark someone outside the company can run, on a robot the company did not choose.