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

At a glance:

• General Intuition raised $320 million, bringing its valuation to $2.3 billion.
• The startup trains a single agentic model on Fortnite gameplay data, including exact button‑press labels.
• A quadrupedal robot learned to navigate after only eight minutes of real‑world data collected on a city street.

Funding round and valuation

General Intuition announced on Thursday that it closed a $320 million Series B round, pushing the company’s post‑money valuation to $2.3 billion. The financing lifts the startup’s disclosed capital to $454 million, adding to the $134 million it raised at launch last October. The round was led by Khosla Ventures and featured participation from General Catalyst, Jeff Bezos, Eric Schmidt, former F1 champion Nico Rosberg, as well as researchers from Google DeepMind and MIT.

The bulk of the new capital is earmarked for scaling compute capacity through a partnership with CoreWeave, as well as for making the company’s API broadly available by the end of summer. Vinod Khosla, whose firm led the round, likened the upcoming leap in world‑model performance to the “quantum leap” that occurred when reasoning emerged in large language models.

How gameplay fuels the model

General Intuition’s core technology hinges on data harvested from Medal, the co‑founder Pim de Witte’s earlier platform that lets gamers upload and share video‑game clips. Hundreds of millions of hours of gameplay, enriched with precise action labels that record exactly which buttons were pressed and when, form the training corpus for the startup’s spatial‑temporal reasoning model.

De Witte argues that most competitors attempt to infer player actions from raw video alone, a method he deems insufficient. By contrast, the embedded action data lets the model distinguish the “self” from the “environment,” fostering a richer sense of causality that can transfer from a virtual screen to real‑world dynamics.

From virtual to physical: robot demo

During a visit to the New York R&D floor, the author observed an AI‑controlled quadrupedal robot navigating the office after only eight minutes of real‑world data collection on a city street. The robot’s default mode was “exploration,” using a single camera as its eye. Though it occasionally clipped chair legs or bumped into a trash bin—behaviour likened to a toddler learning body awareness—the demo showed that the same brain powering a Fortnite‑playing agent could be transferred to a physical embodiment.

The robot’s rapid adaptation underscores the startup’s claim that gameplay can serve as a scalable shortcut to the massive, expensive data‑gathering pipelines traditionally required for sim‑to‑real transfer. While the demo is impressive, General Intuition remains one of several companies attempting to bridge this gap, and the scalability of the approach remains an open question.

Roadmap and commercial strategy

General Intuition positions its world model—not the model’s output—as a “gym” for developers. The company intends to sell the agentic model itself, enabling customers in gaming, simulation, and robotics to build downstream products. Early adopters could use the API to test robots in digital twins of factory floors, power human‑like bots inside game studios, or send quadrupeds into hazardous environments.

De Witte emphasizes that the startup will not become a self‑driving‑car company; instead, it aims to make it ten times easier for others to create such companies. To that end, General Intuition has launched Nerve, a marketplace where gamers can earn money by labeling data, tele‑operating robots, and eventually contributing to real‑world data collection, thereby creating a data flywheel that fuels further model improvement.

Ethical stance and future outlook

The founder’s humanitarian background informs a clear policy: General Intuition’s agents will not be used for lethal autonomy or other applications that could harm humans. De Witte, a Dutch entrepreneur who previously worked with Doctors Without Borders, stresses that the company will avoid “escalatory” uses of its technology.

Beyond ethics, the startup’s European‑heavy team and its refusal to entertain acquisition offers signal a long‑term vision of becoming a foundational AI model provider—akin to Anthropic or OpenAI—rather than a target for a larger lab. Whether the simulation‑to‑real transfer can hold at scale will be the ultimate test, but the combination of proprietary gameplay data, substantial funding, and a growing ecosystem positions General Intuition as a noteworthy contender in the race to build generalist AI agents.