$220 million to train agents on billions of gameplay videos

General Intuition has raised $220 million at a $6.2 billion valuation as it begins opening access to its foundation models. Trained on Medal gameplay videos paired with player actions, the models target gaming, simulation, and robotics.

Another round only months after the last one

General Intuition's valuation has reached $6.2 billion with a new $220 million financing round. Investors include Valor Equity Partners, Atreides, 776, Point72, Khosla Ventures, and General Catalyst.

The financing comes only months after a previous $320 million round at a $2.3 billion valuation. Discussions around the new financing were later reported by TechCrunch in August, with a pre-money valuation of roughly $6 billion under discussion at the time.

General Intuition now plans to expand its compute capacity and grow its teams in New York and Europe. Videos paired with player actions

The distinctive part of the training corpus comes from Medal, the gaming clip platform General Intuition grew out of. The data does not only capture what appears on screen. Videos are paired with the actions players take through their controllers.

General Intuition says it can train its models on billions of these action-labeled videos. According to the company, Medal is on track to reach three billion uploaded videos per year.

Pairing vision with controller inputs is central to the approach. A model can observe a situation, the action selected by a person, and what happens next.

On its website, General Intuition describes two complementary components: action models that determine what to do and World Models that predict the consequences of those actions. From video games to robotics

The next step is transferring those learned representations into environments different from those seen during training.

General Intuition is developing foundation models designed to perceive, predict, and act across virtual and physical environments. The company draws a connection between playing a video game and teleoperating a robot: in both cases, a person observes an environment and uses controls to act within it.

In an earlier demonstration examined by TechCrunch, the same underlying system was used to play autonomously and control a quadruped robot. General Intuition said it added eight minutes of robotics data for the latter task.

Whether that transfer can generalize across more machines and physical situations remains to be demonstrated. World Models trained at scale

Medal's data also feeds General Intuition's work on World Models.

Its research includes MIRA, developed with Kyutai in collaboration with Epic Games. The system learns the dynamics of four-player Rocket League matches from 10,000 hours of data and operates in real time at 20 frames per second.

General Intuition uses scale as a central argument for this strategy. Simulated and gaming environments can expose models to unusual situations at a frequency that would be difficult to reproduce using physical-world data alone.

The company claims that leading foundation models in robotics and World Models are trained on less than 1% of the action data available to General Intuition. The announcement does not provide a detailed methodology for independently verifying that comparison. The models begin moving beyond the lab

The financing coincides with the first commercial opening of the technology. General Intuition says it is beginning to make its models available and has opened an early-access waitlist for partners.

The lab describes them as a new class of foundation models capable of acting in real time across previously unseen environments. Detailed results for evaluating that generalization have not yet been published alongside the announcement.

Initial use cases span gaming, simulation, and robotics, while the additional $220 million is intended to support the expansion of the research and the teams behind it.