As a much lower bar, I'd love it if more of the LLMs used for coding were actually multimodal, accepting images of games as part of the context. In general, LLMs are far better at generating web apps than they are at pure 2D or 3D games that use graphical APIs. There's far less training data, and there's no way to test anywhere close to what you can do with testing on a DOM. In an interactive game, with physics, animations and game logic, the AI just falls on its face because of the complexity.

I realize this isn't the same thing as your point about images as part of training data, but just flagging it in case anyone isn't aware: Claude Code lets you copy and paste images into terminal. I've been designing a "universal history simulator" game for use in my history classes lately, and it is really helpful to be able to make a mockup of a ui change I want and then paste it in, rather than trying to explain it verbally. Also good for debugging graphics issues.