[1/2] John Carmack Reveals AGI Future: Robots, Videogames and AI Agents... (AI column)

Very interesting. speech John Carmack, co-founder id Software, ex-CTO Oculusfounder Keen Technologies. Interestingly, John's Keen Technologies team includes 6 Richard Sutton, the godfather of reinforcement learning. By the way, Jonah all my peers know him as the creator of the game. Doomthat came out 1993 year:) He gave this presentation at the Upper Bound conference. 2025 mid-May. Carmack describes himself as an “engineer-turned-researcher” and considers the creation of AGI “the most important thing he can do today.”

Speaking about the report itself, John suggests abandoning the LLM race, since he does not believe that the “blender”-pre-training of transformers will lead to a real understanding of the world; he cares about the interactive flow of experience. He suggests testing hypotheses on games that look like a clean polygon. Moreover, commercial games are “unbiased” environments, since researchers cannot secretly adjust the environment to the algorithm. And the goal is a fast, continuous, multitasking adaptation of the agent, close to how a cat or dog learns in the real world.

The key ideas of the speech are such 1. "First science, then engineering." Carmack criticizes the RL community for "chasing numbers" and calls for looking for fundamental learning patterns, not just customizing hyper-parameters. 2. Real-time RL vs turn-based simulators In the real world, the environment does not wait for the agent to make a decision. 150–200 Delays "break" many algorithms. John demonstrates how it doesn't work on a robotroller booth that physically moves the Atari joystick 3. Beginning mistakes and lessons from Keen Technologies Here, John reflects on his mistakes as a researcher and highlights the following points: It was too deep into low-level C++/CUDA. Attempted to use the Sega Master System instead of Atari, making it difficult to compare with previous researchers. Executing code on laptop GPUs was a bottleneck - the team moved to remote clusters 4. The video wall is an endless learning environment. The idea was to combine a passive video sequence. (TV/YouTube) interactive games in a continuous tape, where the agent chooses what to study. 5. The transition from simulation to reality Even a simple transfer from an Atari emulator to a real monitor + camera requires addressing I/O, lighting and image drift issues. In general, the problem becomes much more interesting.

Continuation with the description of benches and general conclusions in next post.

#AI #AGI #ML #Software #Engineering #Architecture