I think we'll get to genuinely open ended sand box games.
Have a look at the video, it's quite impressive.
Even just watching the video, I came up with several possible improvements to try out. Eg adversarial training, that would really hone in on the situations and aspects where the model is weak so far, like edge conditions; instead of just using normal gameplay as input.
The other part about that GAN Theft Auto example is that it doesn't actually know what's going on, like there's no game state. All it knows is that "When I have a frame that looks like this, and they press that button, I think the next frame would usually look like this". So it's got no internal game logic, it's just really good at painting what games look like.
Even going about this very naively, you could at least use it to train a model against a supercomputer running the game, and then run the inference on much more modest end-user machines.
But you can be much more ambitious: have you seen eg style transfer? So you could probably do a bit of ML black magic to train your model on GTA, and then point it at the Google Earth data to get a GTA-like set in real-life London.
Or you could use something like style transfer to go for a cartoony look, or add ray-tracing like effects, even if you didn't have these effects in your original engine.
Or you can use a pre-trained model (eg on GTA), and then spend a relatively modest amount of extra training to get a different kind of game, eg one that has magic or so.
About the latter part: I do think their model is already running with some state. But even if it ain't, that's a relatively small thing to add with already known standard techniques (or you can come up with new techniques.)
I feel like the end result is just what we already have, game engines with visual programming and drag and drop editors. You don't add much value by automating the programming since you still have to define the input and outcome.
The real win was building the complex logic and state editor into a good UX.