Taking a tight corner, intercepting a ball, or jumping at just the right moment: gamers are constantly practicing how to observe, predict, and act. Dutch entrepreneur Pim de Witte is building an AI company around those skills. General Intuition was valued at $6.2 billion in its latest funding round.
The company raised another $220 million, De Witte announced. Behind that eye-catching valuation is a compelling idea: the billions of clips gamers record could help machines learn how to move through an environment and understand what their actions will do.
A game clip with an extra layer
De Witte previously founded Medal, a platform that lets gamers record and share moments from their games. For AI, the valuable part is the combination of footage and the actions behind it. What did a player see, which button did they press, and what happened next? Together, those details provide examples of decisions made in a changing world.
According to De Witte, Medal is on track for 3 billion uploads a year. General Intuition trains on game footage labeled with player actions and plans to make its models available to partners in gaming, simulation, and robotics. A platform for spectacular kills and perfectly timed turns gets a second life as training material.
How is this different from ChatGPT?
A language model like ChatGPT is good at processing and generating language, and it can also analyze images. A robot needs to go a step further and perform a physical action at the right moment. A world model tries to predict how an environment will change; an action model selects what the system does next.
Think of a racing game. An explanation of braking is useful, but making the turn also requires a system to judge speed, distance, and movement. It can anticipate possible actions: steer harder, brake earlier, or wait. World models can provide a practice environment and a predictive building block for AI that performs tasks itself.
Try it yourself: soccer inside an AI world model
There is already more to see than a funding announcement. With MIRA, General Intuition and Kyutai, working with Epic Games, have introduced a playable world model built around Rocket League—the game where cars chase a giant soccer ball.
The model learned from 10,000 hours of gameplay collected using publicly available bots. It generates a game world with four players and, according to its creators, responds to their inputs at 20 frames per second. The researchers describe their approach in a technical paper. It is a concrete demonstration of a model that handles interaction rather than simply producing a video.
Google and Nvidia see the same opportunity
De Witte is entering a busy research field. Google explains how world models can generate interactive environments through Project Genie. Nvidia is exploring world action models that let robots predict both their next movement and the camera image they expect to see afterward.
For a robot arm, that could mean anticipating what happens when it picks up a cup or moves a cloth. The potential applications extend beyond games: the same research questions matter in warehouses, vehicles, and factories. Making that leap still requires additional training on real machines and real conditions.
A Dutch founder with international ambitions
General Intuition has an international team and, according to De Witte, is hiring in New York and Europe. The company says on its website that it has raised more than $650 million over the past year. Its new valuation reflects what investors are willing to pay for a stake; it is not a revenue figure.
The most striking part of this story is still the raw material. What started as a way to share fun gaming moments could become a resource for machines that learn to anticipate what comes next. If the approach takes off, gamers will have trained more than their own reflexes—they will have helped train part of the next generation of AI.
