General Intuition bets $2.3B that video games can teach AI to act in the real world

General Intuition has placed a giant wager on a deceptively simple idea: if an AI system can learn how people move, react and make decisions inside video games, it may also learn enough about space, causality and timing to operate in the physical world. The startup, spun out of Medal, is now armed with a fresh $320 million financing round at a $2.3 billion valuation and a growing roster of heavyweight backers who believe gameplay data can become a shortcut to advanced AI agents.

Inside the company’s New York research and development office, that theory is not just a slide-deck promise. The startup is showing a model that can control a game character for hours, a simulated world model that behaves less like a flashy demo and more like a training ground, and a quadruped robot that wanders through the office under the guidance of the same underlying system. The company argues that the bridge from Fortnite-like environments to real-world robotics is not a long shot, but the next logical step in pre-training.

The pitch is ambitious, the capital is enormous, and the unanswered questions are equally large. Can a model trained on game footage, button presses and human reactions truly generalize to robots, drones and industrial settings? Can “intuition” be learned at scale from digital play? And can a startup built on proprietary gamer data create a durable edge in a race that includes some of the biggest players in AI and robotics?

What General Intuition just raised and why it matters

On Thursday, General Intuition confirmed that it had closed a $320 million round at a $2.3 billion valuation. That brings the company’s disclosed funding to $454 million after an earlier $134 million round at launch last October.

The latest investment was led by Khosla Ventures and drew participation from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers affiliated with Google DeepMind and MIT. General Intuition says most of the new capital will be deployed to expand compute capacity, with a plan to continue pre-training the next generation of its model on infrastructure provided through a deal with CoreWeave. A portion of the money is also earmarked for broadening access to the company’s API by the end of the summer.

That funding profile is notable for two reasons. First, it signals investor confidence in a thesis that remains largely unproven at commercial scale: that gaming data can serve as a scalable substitute for the expensive real-world data used to train many robotics systems. Second, it reflects broader enthusiasm around general-purpose agents, an area attracting capital because it promises software that can observe, decide and act rather than merely generate text.

Key facts at a glance

From Medal clips to machine intuition

General Intuition was created out of Medal, the video-game clip-sharing company co-founded by Pim de Witte. Medal’s business gave the team something most AI labs do not have: a massive archive of gameplay paired with the exact actions taken by the player. That matters because, according to the company, a video alone does not capture the full structure of decision-making. Button presses, timing and reaction sequences provide a richer record of intent.

De Witte and his co-founders, Eloi Alonso, Adam Jelley and Vincent Micheli, have built the startup around the idea that this action data can support spatial-temporal reasoning — the ability to understand movement through space over time. In General Intuition’s view, that is the missing ingredient for models that can move from digital environments into robots, vehicles and other embodied systems.

The startup’s argument is that many rivals are trying to infer actions from pixels alone. General Intuition says that is not enough. Human gameplay, it contends, contains a valuable trace of what players saw, how they decided and how they reacted. That action trace may be the key to producing models that understand not only what is happening, but what should happen next.

De Witte argues that the company is building what he sees as a new phase of model pre-training: a system that can interpret game-state information, take action in a virtual environment and respond to physical dynamics in ways a text-only model cannot.

Inside the demo: a robot, a world model and a game agent

The company’s office demos are meant to make an abstract claim feel concrete. One screen showed an agent moving through a Fortnite-like game for what staff said was 100 continuous hours. Nearby, a large quadruped robot with a single camera eye roamed the office floor. The company says the same underlying “brain” drives both systems.

That robot, which appeared to navigate by exploring its environment, occasionally nudged furniture or bumped into a bin in a manner that made its limitations obvious. But General Intuition says the goal is not perfection on day one. Instead, the demo was meant to show that the same model can adapt across embodiments: from a game controller, to a simulated environment, to a physical machine.

According to the company, it took only eight minutes of real-world robotics data to fine-tune the quadruped model. That data was gathered on the street rather than in the office, a detail that underscores General Intuition’s emphasis on using sparse but high-value real-world examples to sharpen a system that has already learned from much larger volumes of game data.

What the company says the model learned

In a separate demo on a laptop, the startup showed a world model that generated a simulated environment frame by frame, rather than relying on a traditional game engine. When asked to move through it, the system behaved in a way that suggested it had internalized some basic rules of the physical world.

The model recognized walls as obstacles, ladders as climbable structures and shadows as something that changes with the sun’s position. That kind of behavior is central to General Intuition’s pitch: if a model can internalize the logic of virtual worlds, it may be easier to adapt that logic to robots moving through factories, streets or disaster zones.

Still, the company is careful about how it frames the world model. It says the model itself is not the final product. Internally, the world model is described as “the gym,” a training environment that supports the broader agentic system the company ultimately wants to sell through an API.

Why gameplay data is the startup’s strategic moat

At the heart of General Intuition’s strategy is a proprietary dataset that is difficult for rivals to reproduce. Medal’s user base has generated hundreds of millions of hours of gameplay footage. But the company says the more important asset is the pairing of that footage with action labels — the record of what players did and when.

That distinction matters because it creates a training set that links perception to action. In practical terms, the model is not simply watching games; it is seeing how humans respond to changes in the environment, how they navigate space, and how they sequence tasks. General Intuition believes that this is the kind of information that can teach machines to behave more naturally in the world.

Vinod Khosla, whose firm led the round, said in a phone interview that he sees a parallel with the development of reasoning in large language models. In his view, just as reasoning marked a leap for language systems, intuition may become the comparable leap for world models. He pointed to the value of action and reaction data from games as the ingredient that could help produce that capability.

Khosla said he was drawn to the company’s vision and its rare data position, describing the emergence of intuition in AI as a potential step-change comparable to reasoning in language models.

Why the investors are buying in now

General Intuition’s backers appear to be underwriting a long-term platform play rather than a near-term product bet. The startup’s technology is compelling in demos, but the bigger question is whether simulation-to-real-world transfer can scale reliably. That issue has not been settled by the broader field, and many robotics approaches still depend on slow and costly real-world data collection.

Investors are betting that General Intuition can reduce that burden by using gameplay as a pre-training source, then layering in smaller amounts of physical-world data for specific embodiments. If that works, the company could become a foundational provider for developers building agents, robots and simulated systems.

The startup is already selling to a limited set of customers in gaming, simulation and robotics. The API expansion planned for later this summer could broaden that footprint and provide additional evidence of whether the company’s model is useful beyond internal demonstrations.

How the new funding is likely to be used

• Expand compute infrastructure for larger-scale pre-training
• Deepen the partnership with CoreWeave
• Open the API to more customers
• Test use cases across gaming, robotics and simulation
• Collect new data from a wider range of embodiments

A business built around agents, not end products

General Intuition wants to position itself as a platform company, not a vertical product maker. De Witte says the startup does not intend to build a self-driving car business, for example. Instead, he wants to provide the model layer that makes it easier for other teams to build their own autonomous systems.

That strategy echoes the market positions occupied by major model providers across the AI industry: sell the core intelligence, let others build the applications. In General Intuition’s case, the applications could range from gaming agents to warehouse robots to digital twins used for industrial training.

The company’s flexibility is one of its selling points. De Witte has said the model can work on any system controlled by a game controller or keyboard and mouse, and the team has already tested it in driving games, drones and other devices. The broader goal is to create a foundation model for embodied action that can move fluidly across environments.

If the thesis holds, General Intuition could become a key infrastructure layer for companies wanting to simulate, test or deploy autonomous systems without spending years building their own data pipelines.

Ethics, defense and the company’s red lines

General Intuition is not positioning itself as a neutral AI supplier indifferent to how its technology is used. De Witte says the company has a clear restriction: its agents will not be developed for use in harming humans. That stance is especially notable at a moment when defense spending and military applications are drawing more attention from AI companies and investors.

De Witte, who spent seven years in humanitarian work including with Doctors Without Borders, says that background informs the company’s ethical line. He argues that he does not want General Intuition to become part of an escalatory arms race or help normalize lethal autonomy.

De Witte said he does not want the company to contribute to systems designed to kill people, arguing that doing so would feed an escalation cycle he wants no part of.

He added that search and rescue applications are different. In his view, using AI to help locate people in danger or support humanitarian response is a legitimate and constructive use case. That distinction places General Intuition among a growing set of AI startups trying to define principled boundaries for dual-use technology.

Why the stance matters in Silicon Valley

Silicon Valley has become increasingly comfortable with defense-adjacent AI work. For de Witte, that trend is unwelcome. He has described the ecosystem’s fascination with defense as something that has spread widely through the industry, and he says his company is intentionally outside that culture.

Part of that difference is geographic and cultural. De Witte is Dutch, many of the team members are European, and the company’s leadership appears to view itself as less embedded in the U.S. startup-security nexus than some of its peers. De Witte has also said he recruited Brianna Martin in part because she had publicly resigned from Palantir over the company’s work connected to U.S. Immigration and Customs Enforcement.

Those details are not incidental. They help define the company’s identity as much as its model architecture does.

What happens to workers in an AI-driven future

General Intuition’s ethics extend beyond what the model should never do. De Witte has also thought about the people who may be displaced by the very systems his company is building. As someone who made money early in life by creating and hosting a private RuneScape server, he says he understands how digital communities can be both economically meaningful and vulnerable to automation.

That concern led General Intuition to launch Nerve, a jobs marketplace aimed at gamers. The idea is to allow people to earn income using the hardware and workflows they already have. Participants start with data-labeling work and can move into robot teleoperation and other tasks over time.

The strategy is both practical and symbolic. Medal’s user base is young and heavily exposed to AI-driven change, and de Witte says he wants those users to have a role in the economy that develops around machine intelligence. In that sense, Nerve functions as both an operational pipeline and a social response to automation anxiety.

How Nerve fits into the broader plan

1. Bring gamers into the company’s data ecosystem
2. Start with accessible tasks such as labeling
3. Build toward higher-value work like teleoperation
4. Give users a way to benefit from the AI shift

The competition: an open field, but a crowded one

For all the excitement around General Intuition, it is operating in a field that has no guaranteed winner. World models, embodied AI and robot learning are all areas where flashy demos have often outrun real-world deployment. The startup’s hardware experiments look promising, but no one has yet demonstrated a universally accepted solution for transferring simulated learning into reliable behavior at scale.

That uncertainty is exactly why the company’s pitch is both attractive and risky. If gameplay data really can serve as a scalable stand-in for expensive physical data, General Intuition could become a powerful supplier for robotics and simulation companies. If not, it risks becoming another well-funded AI effort whose technical promise outpaces operational reality.

The startup appears aware of that tension. Its leaders talk less about finished products and more about the compounding advantages of data, compute and customer feedback. They want to use customer deployments to diversify the types of bodies, environments and tasks their model sees.

De Witte says the company will choose customers strategically, looking for partners that can provide useful real-world data and work closely enough with the team to accelerate learning on both sides.

Why this round may signal a broader shift in AI investing

The scale of General Intuition’s raise also reflects a broader change in how investors are thinking about AI. The market has moved well beyond chatbots and productivity tools toward systems that can perceive, reason and act in physical or simulated environments. That shift is especially visible in robotics, where foundational model providers may become just as important as the companies that assemble the machines themselves.

General Intuition is trying to position itself at that layer. Its pitch is not that it will build every robot, but that it can supply the underlying intelligence layer that many robots, games and simulation systems will share. If successful, the company could occupy a role similar to an operating system for embodied AI, one trained not on static text but on the dynamic logic of human play.

For now, though, the company remains in a proving phase. The demos are persuasive, the dataset is unusual, and the capitalization is impressive. But the market will ultimately judge General Intuition on whether its claims hold outside the office, beyond the controlled environment of staged tests and into the messy unpredictability of the real world.

Timeline: how General Intuition got here

What to watch next

General Intuition now has the money, the data and the attention to push its thesis further. The key questions from here are operational rather than promotional: whether its API finds meaningful demand, whether its model works across enough embodiments to matter, and whether the company can keep gathering proprietary data that compounds its advantage over time.

The most important test may not be another polished demo, but whether a customer can plug the model into a real workflow and see measurable gains. If that happens, General Intuition could move from one of the more intriguing AI experiments in the market to a central player in embodied intelligence. If it does not, the startup will still have illustrated just how far investors are willing to go to fund the next wave of AI beyond language.

Either way, the company has made its position clear. It believes the hidden structure inside games is not entertainment trivia but training data for the next generation of machines. And with hundreds of millions of dollars now behind that idea, General Intuition is asking the rest of the industry to take the bet seriously.

superintelligencenews@gmail.com