General Intuition’s $2.3B bet that video games can train AI agents for the real world

Upon entering General Intuition's research and development space in New York, I was greeted by a captivating scene. The company's co-founder and CEO, Pim de Witte, pointed out a monitor displaying what resembled a game of Fortnite. However, the player wasn't human; it was one of their AI agents, which had been engaged in gameplay for over 100 hours, as noted by Kent Rollins, the chief product officer. As I absorbed the unique sight of an AI navigating a digital landscape, a large quadrupedal robot approached, its movements echoing those of a curious child. "The same AI brain controlling the in-game agent is also guiding this robot," de Witte explained. Josh Duplantis, a data analyst, shared that the robot's primary function was to explore its environment, and it was performing this task based on data collected from real-world scenarios, not just simulations. The foundation of General Intuition's innovative approach lies in its ability to generalize from gaming experiences to real-world applications. Recently, the company announced that it had raised $320 million at a whopping $2.3 billion valuation, bringing its total funding to $454 million. This infusion of capital is expected to enhance its AI models, which are trained using an extensive dataset derived from video game clips uploaded by users of their sister company, Medal. Central to their training process are the action labels embedded in gameplay footage. De Witte emphasized that while competitors rely solely on video data to infer actions, General Intuition's method is more robust. "We view this as the next stage of pre-training for AI models," he stated, highlighting the unique capability of their model to interpret both digital and physical environments. During a demonstration, I interacted with General Intuition's world model, which allowed me to experience a simulated environment. Unlike other systems I've tested, this one provided a realistic interaction with walls and obstacles, demonstrating the AI's understanding of physical space. De Witte clarified that the company aims to sell the agentic model, which utilizes human action data to develop a deeper understanding of interactions within environments. General Intuition is not alone in its quest to bridge the gap between virtual and real-world applications, yet the challenge remains. Many approaches require extensive real-world data collection, which can be slow and costly. The company believes that leveraging gameplay data presents a scalable solution. Investors, including Khosla Ventures and notable figures like Jeff Bezos and Eric Schmidt, are backing this vision, with funds primarily allocated to scaling compute capacity and refining their models. Vinod Khosla expressed confidence in de Witte’s vision, noting the potential for AI to achieve a human-like intuition through the analysis of gameplay data. "In world models, the quantum leap is the emergence of intuition in AI, and the human action data is key to that development," Khosla explained. The startup's ethos is shaped by de Witte's humanitarian background, emphasizing ethical considerations in the development of technology. He has made it clear that General Intuition's innovations will not support harmful applications, particularly in military contexts, focusing instead on beneficial uses such as search and rescue. To address the concerns of job displacement due to AI advancements, General Intuition launched a platform called Nerve, aimed at helping gamers monetize their skills through tasks like data labeling and robot teleoperation. De Witte envisions General Intuition as a supportive ecosystem for developers, facilitating easier access to AI capabilities and enabling them to create innovative solutions in various fields, from gaming to robotics. As the company continues to explore new applications of its technology, including testing robots in digital replicas of real-world environments, its commitment to ethical practices and community engagement positions it as a forward-thinking player in the AI landscape. The ongoing challenge remains: will the transition from simulation to the real world hold true at scale? Only time will tell.