In another thread I likened the current era of AI to playing with PovRay in the 1990s to render "Amazing raytraced 3D graphics!" at a snails pace, with pixels crawling across the screen left-to-right, top-to-bottom over hours or days. Some of use got very excited even back then, because it gave us a peek at a likely future or real-time photo-realistic 3D graphics. We're there now, a mere three decades later.

I expect the pace of AI research to be vaguely similar. In three decades, we'll have interactive generative VR for sure. In the meantime, we'll hit a lot of fun milestones too. I suspect the first games that are real-time AI generated are not that far off. It's surprisingly easy to train AIs when there's simulation available as the ground truth, especially when that simulation can be modified to output per-pixel embeddings, not just noisy RGB sensor measurements like with video input training data.