Code is not the hard part of game development.

Contemporary AI is absolutely trash at things like scene graph management and lighting concerns. It also doesn't do physics, materials, SFX or "fun" very well.

Do not let the apparent visual progress of diffusion models conflate things. Unless we intend to run the entire game via diffusion models (which is certainly possible), I don't think this represents a reasonable baseline against which any amount of AAA art capacity can be compared. The number of units is here is effectively zero. A single AAA artist is infinitely more productive than GPT6 because they can produce and iterate content that the model simply cannot.