This is John Carmack’s session “Keen Technologies Research Directions,” delivered at the Upper Bound 2025 conference in Edmonton and posted on the official channel of Amii, the Alberta Machine Intelligence Institute that runs the event. Carmack is best known for engineering DOOM and the Oculus VR stack; he now runs Keen Technologies, founded to work directly on artificial general intelligence.
In the talk he argues that much modern AI research is too comfortable working in clean, turn-based benchmark settings, and that real intelligence has to cope with a world that does not pause while an agent thinks. He describes a concrete experiment that embodies this view: a physical robot that learns to play games on a real Atari console using a camera to see the screen and a robotic actuator to move the joystick, forcing the learning system to deal with latency, noise, and the messiness of the physical world.
This is a practitioner-level research talk that is light on equations and heavy on engineering judgment. As a primary source it captures, firsthand, one technologist’s distinctive bet on how to make progress toward AGI.
There is no john-carmack person entry in the library yet, so this talk links to the concepts closest to its content.