Robots won't learn to act from robot demos. They'll learn from video games. Takeaways from @PimDeWitte's @aiDotEngineer talk on @gen_intuition (spun out of @Medal_TV): - LLMs are turn-based. The real world isn't. GI's models take in frames and output actions continuously, in real time. - Many robots already run on gamepads. So games and robots share the same action space. No need to predict motor torques. - Medal sees 1.5T+ action tokens a month. Pim says that's more in half a day than frontier labs use in total for robotics. - Gamers are action-dense. High APM data beats sparse teleop for fine control. - Result, per Pim: one model plays games at a top level and controls 10+ robot types. Pixels in, actions out. - The moat is real action labels. You can't infer a pilot's rudder input from video. Anthropic bet on code. GI is betting on pixels.