I'm not too impressed with these examples. I mean it's impressive that it can build code, but it's a very long way off building entire games that you'd want to play.
The human still has to provide the creative input - the concept of the game, the rules, and multiple steps of guidance to get anywhere (when it works).
The code is sometimes generated correctly, but only if a) it's fairly trivial and b) the model can generalise from something else it's seen. There's squillions of maze examples out there so models build slight variations on this easily enough, but as soon as you inject some originality, the model flounders.