The startup General Intuition has raised $320 million in a funding round that valued it at $2.3 billion, betting that data on players’ actions inside video games can help artificial intelligence models develop something resembling human intuition, and then move from virtual environments to robots and the real world.
The round brings the company’s total disclosed funding to $454 million, following a $134 million round it raised at its launch last October. Khosla Ventures led the new round, with participation from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers from Google DeepMind and MIT.
From Video Games to Robots
General Intuition was founded after spinning out of Medal, a company that allows players to upload and share video game clips. The hundreds of millions of hours of clips collected by Medal provided the initial dataset for training the company’s model in spatiotemporal reasoning—that is, understanding how to move through space and time.
But the company believes the most important element is not the video itself, but the action labels embedded in it: records showing which buttons players pressed and when they pressed them. Co-founder and CEO Piem de Witte says competitors often try to infer actions from video alone, which he considers insufficient.
During a demonstration at the company’s New York office, an artificial intelligence agent appeared to play a Fortnite-like game, while the same model was operating a quadruped robot moving through the space using a single camera. Data analyst Josh Duplanty said the model needed only eight minutes of robotic data collected on the street to adapt it to the real world. The robot demonstrated an ability to explore, but it occasionally collided with chair legs and a trash can, reflecting the limits of its current performance.
The World Model as a Training Environment
General Intuition developed a world model that generates a simulated environment frame by frame, rather than relying on a traditional game engine. The company says this environment is not the final product, but rather its internal “arcade” used to train the agentic model. The model is intended to understand the relationship between the self, the environment, and causality, drawing on human movement and feedback data.
Nevertheless, the ability of this type of model to transfer from simulation to the physical world at scale remains an open question. Most competing approaches require huge amounts of real-world data collected slowly and at high cost, while General Intuition is betting that video games can provide a scalable shortcut.
Funding, Expansion, and Potential Uses
Most of the funding will go toward increasing computing capacity, under an agreement with CoreWeave, and focusing on pretraining the next version of the model. The company has also allocated part of the money to making its application programming interface more broadly available by the end of the summer.
The company currently has a limited number of customers in gaming, simulation, and robotics. The uses it cites include testing a robot inside a digital version of a factory floor, operating a humanoid-like robot inside a game studio, or sending a quadruped robot to explore hazardous environments. It has also tested the model with drones and other devices, as well as in driving games, and says it works with any device that can be controlled through a game controller or a keyboard and mouse.
The company has also launched Nerve, a labor marketplace that allows players to earn money using their existing equipment, starting with data labeling and extending to remotely operating robots and other tasks. De Witte says the company does not want to build end products such as self-driving cars, but rather provide a foundation model that makes building those products easier, while collecting new customer data to fuel a continuous development cycle.
General Intuition is also setting limits on the use of its technology. De Witte says its agents will not be used to harm humans, while expressing acceptance of using the models for search-and-rescue tasks. Investor Vinod Khosla believes that data on actions and human feedback in games could be the key to the emergence of “intuition” in world models, but the success of transferring these capabilities to the real world at scale has not yet been proven.