Games

It will always have to be fed human art. LIMs, LAMs, and LLMs break down extremely rapidly when they are fed their own output, because the mistakes become exponentially cumulative. And if you think that eventually these models will have absolutely no errors that require human intervention to fix, you clearly have no background knowledge in the actual programming makeup of these instruments.

It's not a problem that more compute can solve, it's a fundamental issue of epistemology that we haven't even been able to truely solve for the human experience, let alone a completely alien and non-biological experience.