I see a couple of major practical reasons why game (engine) devs are under no threat from this even if it gets better in the future:

1. Scale. Like all things AI, this is not going to scale well. This doesn’t generate code, 3D models and textures, both making games and playing them requires running the model. So if you want a game to have a persistent environment where the world behind you doesn’t get regenerated into something different after taking 8 steps, the context window is going to get real large real fast. And unlike programmed games, you can’t make choices about what’s worth remembering and what isn’t, what can be kept on persistent storage and is only loaded when it becomes relevant etc., because it’s all one big, opaque blob of context, generated by a black box; you either have it remember everything or it becomes amnesiac in a way that makes it useless. Memory availability also isn’t increasing at a rate where this becomes a non-issue any time soon.

2. Control. Manipulating the world though a text prompt gives a lot of control, but it’s also very course. It’s easy enough to tell it that you want a character that can run and jump, but how fast does it run? Does it accelerate and decelerate or start and stop instantly? Does it jump in a fixed arc or based on the running speed and duration of the jump button being pressed? How far and how high? You’re going to run in the limits both of what you can convey and what the language model will understand pretty quickly. And even when you can get it to do exactly what you want, it would have been faster and more practical to manipulate values directly or use a gizmo place things. But there’s no way to extract and manipulate those values, because again: big, opaque blob of context.