Finally got around to watching this.
This got SO CLOSE to my own dream software: real-time world creation.
Imagine a game that created itself AS YOU EXPLORED IT. It would take data from real places, other games, etc, and throw it at a super-AI that would then generate more world on the fly. We're talking machine learning based on real places, not just random procedural stuff.
I imagine an MMO game where there are literally no limits to where you can go. The game could take place in a city, initially, but someone could hop in a car and drive past the city limits and the game will start filling in the surrounding country-side, and even generating a neighboring city. Then the players take it from there and breathe life into them by, say, moving into the houses and apartments, taking up jobs in the commercial buildings, etc. You'll feel like you're exploring a real place, which also makes it perfect for VR! All you have to do to determine what content is generated is to pick what initial data is thrown at the AI (i.e. sample images, 3D data, etc).
I actually think this is within our reach with machine learning in its current state. The issue is server space. Whatever content is generated would need to be stored and preserved for consistency, and that data is potentially infinite. We're talking Google-size server space here. But the end result would be a video game completely unbounded in where, and how far, you can go.
