Eight minutes of real-world practice. That's all it took for a four-legged robot to navigate an office after training on months of Fortnite gameplay. That ratio is what separates General Intuition's approach from anything else in the current AI race.
On June 25, the company announced a $320 million Series A at a $2.3 billion valuation. Khosla Ventures led the round; General Catalyst, Jeff Bezos, Eric Schmidt, and researchers from Google DeepMind and MIT also joined. Total disclosed funding reaches $454 million, including a $134 million seed round from October 2025.
The core insight: training AI on video is like teaching someone to cook by watching cooking shows. They see the output but miss the critical part — what to do, and exactly when. Gameplay clips are different. Each recording contains precise timestamps for every button press and control input. These are action labels: a detailed log of every decision a player made. From that log, an AI learns spatial-temporal reasoning — how to move through space, react to obstacles, plan ahead.
The training data came from Medal, General Intuition's parent company — a gaming clip platform with 17 million monthly active users. Hundreds of millions of hours of action-labeled footage form a dataset competitors simply can't replicate. The company also launched Nerve, a platform where gamers earn money by helping label data.
The investor demo featured an AI agent playing Fortnite for over 100 consecutive hours. The same model then controlled a physical quadruped robot in an office after just eight minutes of real-world fine-tuning. By comparison, robotics competitors typically need thousands of hours of physical data collection to reach similar results.
Most of the fresh $320 million targets compute scaling and pre-training. A developer API is planned for the end of summer. The company explicitly avoids military applications but is open to search-and-rescue use cases. The long-term goal: a universal agent brain — one model that controls a game character, a drone, and a delivery robot with equal capability.
At its core, this is a direct attack on robotics' oldest unsolved problem: the simulation-to-reality gap. Video games are already rich simulators with endless scenario variety and built-in action labeling. If the approach scales, millions of gamers may unknowingly become the largest suppliers of training data for next-generation robots — without ever leaving their chairs.
