I am not sure what you want from me here.
Yes, I do contest that "that a lot of LLM brilliance comes from datamining creative thoughts/output from humans and that the brilliant insights go down when there isn’t much in the way of such material".
Because I think, the same is the case with humans. Most of the cultural artifacts we produce is crap, a bad copy of natural original. And the "brilliant insights of humans" are achievable by models running on higher temperature.
I think the proponents of the theory need to explain by which mechanism the actual loss of information supposedly occurs (on the probability distribution of possible LLM outputs). Is it averaging? Added randomness? Preferential skew? To me, it is rather vague, to the point I don't see how it's different from what humans have done for centuries.
Or the opposite, show how those "brilliant insights" from humans manage to survive in the sea of cultural crap otherwise produced in human culture. Perhaps a specific example would help.
This is the pure incredulity that I was talking about.
- Humans have done this for centuries
- These are all cultural artifacts
- ?
- It’s all the same to me
People have claimed—and this is a widespread theory—that a lot of LLM brilliance comes from datamining creative thoughts/output from humans and that the brilliant insights go down when there isn’t much in the way of such material. Further they claim that the eventual convergence towards LLM-only content (once humans presumably give up) will not generate the same quality of output. In fact it will deterioate.
Maybe someone would like to contest that. But that should be done directly. Instead of making pedestrian statements like:
> Humans have been trained on "human-generated data" (cultural artifacts) for centuries, and quality is not down.
Which is borderline just a rhetorical gotcha.