so there's no flash<br>of the wrong palette before React hydrates. Mirrors the useTheme hook<br>(storage key "rw-theme"; classes rw-theme-dark|light|ascii). -->

General Intuition raises $320M at $2.3B valuation for gameplay-trained AI agents - RuntimeWire

RuntimeWire

You're browsing RuntimeWire with JavaScript disabled. Articles and<br>navigation work fully. Interactive features — search, comments,<br>and newsletter signup — require JavaScript.

Why it matters

General Intuition's round shows where frontier AI financing is moving: away from scraped text alone and toward proprietary action data that could train agents for physical tasks.

Pim de Witte has turned Medal's archive of gamer behavior into a $2.3 billion venture bet: General Intuition said Thursday it raised $320 million to train AI agents on action-labeled gameplay, with Khosla Ventures leading the round and General Catalyst participating.

That distinction matters because the headline number being passed around is not the cash raised. The financing is $320 million; $2.3 billion is the valuation reported by TechCrunch. Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers at Google DeepMind and MIT also participated, according to the report. General Intuition says the round brings disclosed funding to $454 million, following the roughly $134 million seed round General Intuition announced in October 2025.

De Witte, 31, is not pitching General Intuition as another video model company. He is pitching General Intuition as a model provider for agents that can perceive, predict, and act across virtual and physical environments. The origin story is Medal, the game-clipping platform de Witte built before General Intuition. Medal gives General Intuition the asset investors are paying for: not only gameplay video, but the player inputs attached to it, recording what a human did and when.

That makes General Intuition's round a data round as much as a model round. The capital is going mainly toward compute and pre-training the next model, TechCrunch reported. General Intuition also says a commercial API is already in use with first partners in games, simulation, and robotics, with broader availability planned after a selective rollout.

The bet is that games solve a robotics data problem

General Intuition's central claim is simple and hard to prove: games may contain enough structured human action to teach models some version of spatial-temporal intuition before they touch the physical world.

On General Intuition's site, General Intuition says its models learn from "action-labeled video datasets" and that Medal users upload billions of gameplay clips each year. TechCrunch reported in October that Medal had about 10 million monthly active users and 2 billion videos uploaded per year. Those clips are valuable because they pair visual context with intent, action, and consequence. A model can see the scene, the player's input, and the result.

That is the gap General Intuition is trying to exploit. Text and image models learn from observations. Gameplay gives General Intuition observations plus behavior. De Witte told TechCrunch that most competitors infer actions from video alone, while General Intuition uses the input labels embedded in game clips. In his framing, the model is not just watching the world. It is learning what an actor can do inside one.

The hard part is transfer. A model that can navigate a game-like scene has not automatically solved robotics, drones, autonomous vehicles, or factory simulation. General Intuition showed TechCrunch an AI game agent that had been running for 100 hours and a quadruped robot fine-tuned with eight minutes of real-world robotics data. Those demos support the direction, not the conclusion. General Intuition has not disclosed revenue, pricing, exact customer count, or production-scale evidence that the sim-to-real transfer holds outside controlled demonstrations.

General Intuition is selling agents, not the gym

The more interesting strategic choice is what General Intuition does not want to sell. General Intuition is building world models, but TechCrunch reported that General Intuition treats the world model as an internal training environment, or "the gym," rather than the end product. The commercial product is the agent or API access to agents trained inside that environment.

That separates General Intuition from a set of world-model efforts that are being framed around generated environments. Google DeepMind has Genie, World Labs is building interactive 3D worlds, Decart has shown Oasis, and Runway has published work on using general world models for robot policy evaluation. General Intuition's angle is narrower and potentially more defensible: use Medal's action data to train agents that can operate inside games, simulations, robots, drones, and other controllable systems.

That is also why Khosla's involvement is telling. The market has rewarded frontier AI teams for compute access and research talent, but in...