Novelty grabs people's attention. A system based on the statistical analysis of past content won't do novelty. This seems like a very basic issue to me.

Novelty itself is easy, the hard part is the kind of novelty that is familiar enough to be engaging while also unusual enough to attract all the people bored by the mainstream.

Worse, as people attempt to automate novelty, they will be (and have been) repeatedly thwarted by the fact that the implicit patterns of the automation system themselves become patterns to be learned and recognised… which is why all modern popular music sounds so similar that this video got made 14 years ago: https://www.youtube.com/watch?v=5pidokakU4I

(This is already a thing with GenAI images made by people who just prompt-and-go, though artists using it as a tool can easily do much better).

But go too soon, be too novel, and you're in something like uncanny valley: When Saint-Saëns' Danse macabre was first performed, it was poorly received by violating then-current expectations, now it's considered a masterpiece.

A system that digs out undiscovered mechanisms to drape novelty on based on where the statistical analysis says it’s already been, that would do it though.

Sounds horrifying. The TikTok algorithm is scary enough

We can sit and imagine horrors the whole week. It has no bearing on the capabilities of machine learning.

If there’s a clear distinction between LLMs looking for patterns in text and LLMs looking for patterns in patterns in text, I’m interested in seeing it better and understanding it.