I might be crazy, but I'm wondering if we'll bypass this in the long run and generate 2D frames of 3D scenes. Either having a game be low-poly grayboxed and then each frame is generated by an AI doing image-to-image to render it out in different styles, or maybe outright "hallucinating" a game and it's mechanics directly to rendered 2D frames.

For example, your game doesn't have a physics engine, but it does have parameters to guide the game engine's "dream" of what happens when the player presses the jump button to produce reproducible actions.