source: techcrunch ai: general intuition’s $2.3b bet that video games can train ai agents for the real world
level: technical
general intuition showed an ai agent playing a fortnite-like game for 100 hours straight while the same model powered a quadruped robot wandering its office. the robot used a single camera and defaulted to exploration mode, occasionally bumping into chairs like a toddler. fine-tuning took only eight minutes of real-world robotics data collected on a street, not in the office.
the startup's key data comes from medal, a platform for sharing game clips. hundreds of millions of hours of gameplay include action labels—records of exactly which buttons players pressed and when. this lets the model learn causality and distinguish itself from the environment, unlike competitors that infer actions from video alone. a world model generates frames on the fly, learning that walls are solid and shadows shift with the sun.
the $320 million round, led by khosla ventures, values the company at $2.3 billion. most funds will scale compute via coreweave and pre-train the next model version. the api will broaden by summer's end. the startup aims to be a model provider for gaming, simulation, and robotics, not a self-driving car company. it also launched nerve, a jobs marketplace where gamers earn money labeling data or teleoperating robots.
why it matters: using game button-press data could make training real-world robots faster and cheaper than collecting physical data alone.
source: techcrunch ai: general intuition’s $2.3b bet that video games can train ai agents for the real world