General Intuition raises $320M to train AI agents in video games

A New York startup is betting that the path to capable AI agents in the physical world runs through millions of hours of video game footage — specifically, gameplay clips that capture not just what players see, but exactly which buttons they press and when.

General Intuition announced Thursday it raised $320 million at a $2.3 billion valuation, bringing total disclosed funding to $454 million since launching last October. The round was led by Khosla Ventures, with backing from General Catalyst, Jeff Bezos, Eric Schmidt, and researchers at Google DeepMind and MIT, according to details first reported by TechCrunch.

From Medal clips to spatial reasoning

The company spun out of Medal, a platform where gamers upload and share gameplay clips. Those hundreds of millions of hours of footage became the training dataset for General Intuition's model, but CEO Pim de Witte says the real advantage isn't the video itself — it's the action labels embedded in each clip that record precisely what inputs a player made at every moment.

De Witte argues that competitors trying to infer actions from video alone miss critical information. His team's approach teaches the model spatial-temporal reasoning by connecting visual information to specific actions, helping it distinguish between "self" and "environment" in ways that improve its understanding of causality.

During a demonstration at the company's New York office, a quadrupedal robot navigated the space using a single camera, occasionally bumping into furniture like a toddler still learning spatial boundaries. The same model powering that robot was simultaneously playing a game resembling Fortnite for 100 hours straight. The robot required just eight minutes of real-world data for fine-tuning — data collected on the street, not in the office where it operated.

The world model as training gym

General Intuition has built what it calls a world model: a simulated environment generated frame-by-frame rather than rendered by a traditional game engine. The company views this not as the product itself, but as a training environment — internally called "the gym" — where the agentic model learns before deployment.

The startup plans to sell the agentic model as an API, positioning itself as an ecosystem enabler like Anthropic or OpenAI. De Witte says the model works on anything controllable with a game controller or keyboard and mouse, and the company has tested it on drones, driving simulations, and other devices beyond the quadruped.

Why it matters

Most approaches to training embodied AI agents require enormous amounts of expensive, slowly-gathered real-world data. General Intuition's thesis — that gameplay with embedded action data can serve as a scalable shortcut — could dramatically accelerate development timelines if it holds up at scale. The company's proprietary access to Medal's data represents a moat that attracted investors willing to bet on a generational company rather than a quick acquisition target. However, whether simulation-to-real-world transfer can work reliably across diverse physical environments remains an open question the industry hasn't fully answered.

The vast majority of the new funding will support compute scaling through a deal with CoreWeave, with a portion allocated to making the API more broadly available by summer's end. General Intuition currently serves a handful of customers in gaming, simulation, and robotics.

The company has also launched Nerve, a jobs marketplace where gamers can earn money through data labeling and robot teleoperation using their existing setups. De Witte, who is Dutch and employs a largely European team, has drawn clear ethical boundaries: the technology will not be used for lethal military applications, though search and rescue missions are acceptable.

Vinod Khosla, whose firm led the round, compared the potential breakthrough to the emergence of reasoning in large language models. "In world models, I think the quantum leap is the emergence of intuition in the AI," he said. "The human action data and reaction data you have in games is the key part to the emergence of intuition."

These details were first reported by TechCrunch.