Gaming Revolution: General Intuition's $2.3 Billion Bet to Train AI for Reality

General Intuition, a startup co-founded by 31-year-old CEO Pim de Witte, is pioneering a novel approach to artificial intelligence with its agentic model capable of generalizing from virtual gameplay to real-world embodiment. The company recently announced a substantial funding round, raising $320 million at a $2.3 billion valuation, confirming earlier reports and bringing its total disclosed funding to $454 million. This significant investment underscores the industry's belief in General Intuition's unique methodology.

Upon entering the company's New York R&D floor, the immediate impression of General Intuition's capabilities was palpable. A monitor displayed an AI agent autonomously playing a game akin to Fortnite, having accumulated over 100 hours of continuous gameplay. Simultaneously, a large quadrupedal robot, powered by the very same AI brain, navigated the office environment. This robot, with its singular camera acting as its "eye," demonstrated its default "exploration" mode by moving around, occasionally bumping into objects like a learning toddler. Remarkably, it took only eight minutes of real-world robotics data, collected outdoors, to fine-tune the AI model for this quadruped, showcasing the model's impressive generalization abilities.

General Intuition's foundational innovation lies in its ability to leverage vast datasets derived from video game clips. The startup was spun out of de Witte's other company, Medal, a platform where gamers upload and share video content. While hundreds of millions of hours of gameplay footage provided the raw material, the critical element for training General Intuition's model in spatial-temporal reasoning — understanding movement through space and time — was the embedded action labels. These labels record the precise buttons a player pressed and when, offering a richer, more causal understanding of actions compared to competitors who primarily infer actions from video alone.

The company employs a sophisticated "world model" internally referred to as "the gym." This simulated environment, generated frame-by-frame without a traditional game engine, serves as a crucial training ground. Through millions of hours of gameplay, the model learns fundamental environmental physics and causality, discerning, for instance, that walls are solid, ladders are for climbing, and shadows respond to light sources. This world model is not the end product but a vital tool for enabling the agentic model to distinguish between the "self" and the "environment," fostering a deeper understanding of cause and effect.

The recent $320 million funding round was led by Khosla Ventures, with notable participation from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers from Google DeepMind and MIT. The majority of these funds are earmarked for scaling compute capacity, specifically through a deal with CoreWeave, and for pre-training the next iteration of their model. A portion will also go towards broadening the availability of their API by the end of summer, positioning General Intuition as an ecosystem enabler, much like Anthropic or OpenAI. Their goal is not to build specific end-products, such as self-driving cars, but to provide the foundational AI that makes it significantly easier for others to develop such technologies.

Pim de Witte, influenced by his three years in humanitarian work with Doctors Without Borders, has instilled a strong ethical framework within General Intuition. The company has drawn a clear line: its agents will not be used to harm humans, explicitly rejecting applications in lethal autonomy. While Silicon Valley shows increasing interest in military applications, de Witte and his predominantly European team, including Chief of Staff Brianna Martin who publicly resigned from Palantir over its work with ICE, are committed to non-escalatory uses, prioritizing applications like search and rescue missions. This ethical stance reflects a broader concern for the societal implications of advanced AI.

Further addressing the impact of AI, General Intuition launched Nerve, a jobs marketplace designed to empower gamers. This platform allows individuals to leverage their existing setups to earn money, starting with data labeling tasks and potentially progressing to robot teleoperation. De Witte views Medal's user base as a demographic highly susceptible to AI-driven job displacement and aims to provide them with a direct stake in the evolving AI landscape.

General Intuition envisions building a "data flywheel" where their API, once more broadly available, will be tested across diverse use cases—from simulating robots in factory digital twins to powering humanlike bots in gaming studios or navigating hazardous environments with quadrupeds. While a quadruped is their first real-world physical embodiment, the model has been successfully tested on drones and in driving games, adaptable to any device controllable via game controller or keyboard/mouse. The strategy is to prioritize customers who can offer valuable real-world data, thereby diversifying the embodiments the generalized foundation model serves and advancing research. Despite impressive demonstrations, the ultimate scalability of this simulation-to-real-world transfer remains an open question that General Intuition, backed by its unique data position and investor confidence, is actively working to solve.