level: technical
general intuition showed an ai agent playing a game for 100 hours straight while the same model controlled a quadruped robot walking around the office. the robot used only a single camera and eight minutes of real-world data to navigate, occasionally bumping into objects. the model learns spatial-temporal reasoning from millions of hours of gameplay clips from medal, a video clip sharing platform. the key data is not just video but records of exactly which buttons players pressed and when.
the company built a world model that generates frames on the fly instead of using a game engine. in demos, the agent understood walls, ladders, and shadows. this world model is a training environment, not the product. general intuition plans to sell the agentic model itself, arguing that action data helps the model separate self from environment and learn causality. the startup raised $320 million at a $2.3 billion valuation led by khosla ventures, with most funds going to compute and pre-training the next model.
general intuition wants to be a model provider for gaming, simulation, and robotics, not build end products like self-driving cars. it launched nerve, a jobs marketplace for gamers to earn money through data labeling and teleoperation. the company refuses military use cases but supports search and rescue. its proprietary data from medal is a key advantage, but scaling simulation-to-real transfer remains an open challenge.
why it matters: using gameplay data with action labels could lower the cost and time needed to train ai agents for real-world tasks like robotics and autonomous navigation.