Robots Will Be Trained on Video Games - Pim de Witte, General Intuition

Pim de Witte of General Intuition, the team behind the Medal game recorder, argues that models built to act in space and time need a different design than LLMs, which work turn by turn. His models take in frames and send out actions continuously, the way a gamer plays. Because game controllers already run most robots, Medal's game clips (about a billion videos a year and over 1.5 trillion action tokens a month) become training data, and one model can play games at a high level and control more than 10 types of robots. He also previews early access, accessibility uses and hackathons in New York and London. Chapters: 0:00 Intro 0:41 Why acting is different from predicting 1:54 Starting with Medal, the game recorder 2:27 Game controllers already run robots 3:28 How LLMs and action models differ 5:21 Use the controller, skip the motor math 6:21 1.5 trillion action tokens a month 7:14 World models and environments 8:55 Why games are so action-dense 9:52 One model, 10+ robots 12:03 What you can build on top 13:36 Accessibility uses 14:54 The team and hackathons 17:10 Q&A: how it differs from JEPA Pim de Witte https://generalintuition.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!