A $2.3 Billion Bet: Video Games as the Key to Training Tomorrow's Robots

When a Virtual Player Becomes a Real Robot

At General Intuition's New York office, two parallel scenes play out: on a computer screen, an AI agent has been playing Fortnite nonstop for a hundred consecutive hours. In an adjacent corridor, a four-legged robot guided by a single camera roams its surroundings, learning to avoid chairs and trash cans — much like a child discovering for the first time where their body ends and the world begins.

What's striking isn't the robot or the virtual agent in isolation, but that a single mind controls both. This idea is the core of a massive bet that is attracting billions of dollars in funding today.

A Funding Round That Cements the Company's Status

General Intuition announced the closing of a $320 million funding round, pushing its valuation to $2.3 billion. Total disclosed funding now stands at $454 million, including the $134 million seed round announced last October.

The round was led by Khosla Ventures, with participation from General Catalyst, alongside prominent investors including Jeff Bezos and Eric Schmidt, as well as researchers from Google DeepMind and MIT.

From Gaming Clips to Understanding the World

The company's story begins with Medal, a platform owned by founder and CEO Pim de Witte (age 31) that allows players to record and share gaming clips. The platform has accumulated hundreds of millions of hours of gameplay, which formed the foundational data for training General Intuition's model on spatiotemporal reasoning — that is, how to move and navigate through space and time.

But the secret lies not in the video clips themselves, but in a hidden layer within them: precise recordings of every button a player pressed and the exact moment they pressed it. De Witte believes most competitors try to infer actions from video alone — an approach he considers insufficient. This granular behavioral data gives the model a unique ability to distinguish between "self" and "environment," deepening its understanding of causality and interaction with the world.

A Training Environment, Not an Entertainment Product

The company has developed what it calls a "world model": a simulation environment generated frame by frame, rather than relying on traditional game engines. Experiments have shown that this model develops a deep understanding of the surrounding physical rules: walls are solid, stairs are for climbing, and shadows shift with the movement of the sun.

Yet this simulated environment is not the final product — it's the "training gym," as the team calls it internally. The real product is an action model capable of operating in the actual world, whether inside a game, inside a factory, or on a busy street.

As evidence of its generalizability, fine-tuning the model to operate the four-legged robot required just eight minutes of real-world data — collected in the street, not inside the building where the robot later navigates.

Why Investors Trust This Bet

Vinod Khosla, founder of Khosla Ventures, believes the behavioral data derived from gaming is the key differentiator that sets General Intuition apart:

• The emergence of "intuition" in AI: Khosla compares this moment to the emergence of reasoning in large language models, calling it an equally significant qualitative leap.
• Unrivaled proprietary data: Hundreds of millions of hours of gameplay paired with corresponding action data give the company a competitive advantage that is difficult to replicate.
• Scalability: While most robotics approaches require costly and slow field data collection, gaming provides a cheap and practically limitless training environment.

Next Steps and Challenges Ahead

The majority of the funding will be directed toward expanding compute capacity through a partnership with CoreWeave, with a focus on pre-training the next version of the model. A portion has also been allocated to making the API more broadly available before the end of the summer.

Challenges remain, however: General Intuition is not alone in this race, and proving that world models perform reliably in real-world applications at scale is a test that has yet to be passed. Neither the founders nor their investors appear to be in a rush to demonstrate this through an acquisition — they are betting that what they are building could become the backbone of a new generation of embodied AI and simulated worlds.

✦ بقلم فريق دروب أيديا