Ben Evans on Nostr: A: Adding more parameters to a LL model makes it "better" B: Continued ingest of the ...

A: Adding more parameters to a LL model makes it "better"

B: Continued ingest of the general Internet corpus "improves" general models.

Opposition: What if neither of those hypotheses is true? What if neither A or B comes to pass?

What if:

~A: Adding more parameters adds nothing in terms of utility;

~B: Increased output of LLMs pollutes general input sets.

C: Subsequent LLMs are even worse than current generations.