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

Continue. story About John's talk, we need to talk about existing benchmarks and new proposed ones.

1. Atari 200 M (DeepMind classic) Existing approaches give a superhuman level at a "month of screen time", that is, the bench is already "almost solved" and the next step is needed. 2. Atari 100 K The goal is to learn as much as possible 2 an hour of gaming experience. Algorithms like BBF/EfficientZero make progress but are sensitive to image and action delays. 3. Sequential Atari Catastrophic Forgetting Consistent mastering 50+ games without losing skills. It's not working very well yet. (By the way, it is forbidden to throw "task-id" so as not to cheat.) 4. Real-Time Atari via Robotroller The bottom line is latency stability and pixel work on real hardware. In the end, it turns out ~180 MS lag from camera to action; assessment by front camera and OCR account (play) 5. Unbounded Video Wall (project) Research driven by curiosity in a mix of passive and active content. There is no external "episode" and a clear final score. Carmack calls all of his tests "open" because Keen Technologies plans to put the code and hardware in open source so the community can replicate and improve the results.

If we take stock, AGI is simply a larger LLM. You need an agent who learns on the fly, does not forget the old and acts in real time. Games remain the best “physically safe” bridge between simulation and the real world, but metrics should encourage learning speed and delay resistance, not just the final score.

• The main challenge is methodical. To avoid highly specialized tricks, Carmack offers tough but replicable benchmarks and built-in cheating protection.

Progress in these areas will show how far we really are from AGI and give a more honest picture of how to move in its direction.

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