Eventually models will likely get their creativity by:
1. Interacting with the randomness of the world
and
2. Thinking a lot, going in loops and thought loops and seeing what they discover.
I don't expect them to need humans forever.
Agreed, by some definitions, specifically associating unrelated things, models are already creative.
Hallucinations are highly creative as well. But unless the technology changes, large language models will need human-made training substrate data for a long time to operate.
That will likely always be the case. Even 100% synthetic data has to come from somewhere. Great synopsis! Working for hire to feed a machine that regurgitates variations of the missing data sounds dystopian. But here we are, almost there.