Action-Labeled Gameplay Enables Control
General Intuition
This is a data quality argument, not just a data scale argument. Medal does not just supply frames of gameplay, it supplies clips tied to player actions and game events, which means a model can learn that a button press changed position, aim, or score instead of merely seeing that something happened on screen. That makes the training signal closer to imitation and control data than ordinary internet video, which is exactly the gap physical AI companies are trying to close.
-
Medal is large enough for this to matter. Medal says it has more than 15M monthly active gamers and expects 2.5B+ clips this year, while another company page describes billions of decision labeled clips and more than 1B hours of video tied to the pipeline. The point is not just volume, it is repeated action outcome pairs at internet scale.
-
That is a different asset from DeepMind's games stack. Genie 3 and SIMA 2 are built to generate interactive worlds and test agents inside them, but DeepMind also notes limits in action space and multi agent interaction. General Intuition starts from logged human play with concrete controls and outcomes already attached, which is a cleaner supervisory signal for learning behavior from demonstrations.
-
The closest robotics analog is the broader race to own proprietary interaction data. Physical Intelligence emphasizes fine tuning with small amounts of robot data, while other physical AI companies are building flywheels around field deployments, glove demonstrations, or teleoperation. Medal gives General Intuition a native source of human action traces before it ever owns a robot fleet, which is unusual among independent model labs.
The next step is turning this gameplay advantage into a bridge from virtual actions to physical control. If General Intuition can keep extracting structured action labels from Medal and then connect them to robotics fine tuning or fleet data, it can move from being a world model company with good synthetic environments to a control model company with a proprietary behavior pretraining stack.