Action-Model AI General Intuition Raises $220M at $6.2B Valuation

Its valuation jumped 2.3x in three months. General Intuition says it has raised $220 million in new funding at a $6.2 billion valuation — just three months after a $320 million Series A valued it at $2.3 billion, which itself followed a $133.7 million seed round last October.

General Intuition spun out of game-clip platform Medal, and it builds something different from the “world models” the rest of the AI industry is chasing: “action models.” The distinction is sharp. World models generate virtual worlds outright. Google DeepMind’s Genie 3 turns a single text prompt into a real-time, walkable 3D world at 24 frames per second and 720p resolution, and can alter weather, time of day, or objects mid-session through what it calls “promptable world events” — with objects that leave the frame retaining their state when they reappear, for a few minutes of consistency. World Labs’ Marble turns a single image or a 360-degree photo into an explorable 3D space that exports to game engines or robotics simulators. Both are focused on generating the world itself.

General Intuition’s action models do the opposite — they don’t build a world at all. Given an environment, an embodiment, and a goal, they decide what to do inside that world and act on it. With this announcement, the company said it’s opening its models up externally for the first time, trained on billions of action-labeled videos and capable of acting in real time in environments they’ve never seen. How much strategic value world models now command was underscored just two days earlier, when AMD announced it would acquire World Labs for $8.2 billion, with founder Fei-Fei Li moving to AMD as EVP and chief scientist. General Intuition is betting on the opposite side of that trade — generating action, not generating worlds.

The training data still comes from Medal, which the company says is the fastest-growing company in games, on track for 3 billion uploaded videos a year. General Intuition illustrates the scale of that data with a comparison: more “accidents” happen in simulation in a single day than occur across the entire United States, and at any given moment more people are holding a game controller tied to a steering wheel than Waymo has cars on the road. Its logic is that the inputs and outputs people use to play games or teleoperate robots — vision in, control actions out — are fundamentally the same, an approach the company compares to how Tesla trained FSD, just applied across a far broader range of actions and environments.

Founder Pim de Witte built and ran Medal for nine years. At 14 he ran one of the largest private RuneScape servers, an experience he later connected to humanitarian work with Médecins Sans Frontières, co-building the crisis-mapping tool MapSwipe. He spun General Intuition out of Medal last October with researchers Eloi Alonso, Adam Jelley, and Vincent Micheli, handing Medal’s CEO role to Ken Colton to focus fully on General Intuition.

The new round was backed by Valor Equity Partners, Atreides, 776, Point72, Khosla Ventures, and General Catalyst. The company said the funding will go toward growing its team in New York and Europe.

The Competition

Competition in “AI that understands the physical world” remains fierce among world-model labs. World Labs, founded in 2024 by Stanford’s Fei-Fei Li, raised $1 billion before AMD agreed to acquire it outright for $8.2 billion, ending its run as an independent startup. Odyssey, founded by self-driving veterans Oliver Cameron and Jeff Hawke, raised $310 million at a $1.45 billion valuation to become a unicorn, with Amazon among its investors. Decart raised $300 million at a $4 billion valuation and struck a partnership to run real-time world models on AWS’s Trainium chips. NVIDIA has its own open-weight world model, Cosmos, meaning major chipmakers themselves are now racing to lock down world-model capability. Where those companies build and sell the generated world itself as a product, General Intuition treats world models as an internal training tool only, focusing instead on generating the action that happens inside them.