You will not. Something like an endless runner would be plausible for an AI, but GTA requires object permanence. An AI that could create GTA-like moment-to-moment mechanics would be a curse to play if it's impossible to actually navigate from point A to point B.

Were any of these realtime captured though? Or just prompts that resulted in videos depicting what these games would be like?

Because if it's just video output based on gaming-related prompts, this isn't anything special.

And if it is playable, how does it know what the controls are or how to organise inputs for cause and effect? Wouldn't you have to give it the prompts to do so? For instance, how could it know shooting things kill them instead of incapacitating them? Or how would it know what the outcome should be for interacting with NPCs in that open-world prompt (viz., do you hit them, talk to them, hug them, kick them, shoot them)?

Lots of questions. I'm sure there are answers in that Twitter thread but I don't have Twitter and am not interested enough to go through the trouble of attempting to read the thread.