large action model

A large action model is an AI model trained to predict the next action instead of the next word. While the chatbot models we're familiar with are trained to continue writing sentences, this kind of model is trained to continue predicting which buttons to press and how to move in a given situation.

Think of it like a coach who has watched hundreds of millions of hours of people playing games. By observing massive amounts of data on when players hit the jump button or change direction, the coach develops an intuition for the right move in the moment — even in situations never explicitly taught. The goal of a large action model is to transplant that same intuition from a game screen into the arms and legs of a real robot.

So the hope is that after learning virtual controls from things like game clips, this kind of model can serve as a bridge, carrying that learning over into the movements of a physical machine with a body.

The article reports that General Intuition "revealed for the first time how its large action model works." The name is easy to confuse with large language model, but what this model handles isn't sentences — it's actions, like button presses in a game or a robot's movements. The key point is that it was trained on gameplay footage along with data recording exactly when which button was pressed.