> The problem, as far as I can tell, is that if all this works perfectly, this will be simply the human-programmed Minecraft we already have, except far more expensive to run. The original Minecraft is already infinitely playable, thanks to the way it randomly generates a 7x-larger-than-planet-Earth landscape with each new game. You can't use generative AI like this to get a new kind of game, only a better simulation of the game you trained it on.<p>Oh come now, that's not true and is showing a lack of imagination, especially from someone who's been covering generative models this long. There's lots of things you can do with an accurate differentiable model of Minecraft - just like there's lots of things you can do with a LLM or a diffusion model beyond just 'generate a random sample'. (Just think about all the things you can do with CLIP steering a diffusion model...) Imagine describing GPT-3 as saying 'as far as I can tell, if this all works perfectly, this will simply be Common Crawl, except far more expensive to read'!