#llm_scaling — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #llm_scaling, aggregated by home.social.
-
How many times have we seen breathless endorsements of some sort of weird LLM use case, only to see it revealed as dishonest hypejockeying by people with a vested interest mere weeks later?
We're in the double digits.
Just stop.
Don't fall for it.
Anyone can be bought.
It doesn't matter their pedigree.
It doesn't matter their credentials.
Scientists are always in need of funding, and can be made to endorse whatever nonsense if it means their funding is secure.
Just stop.
Staaaaahhhhp.
Do.Not.Repeat.The.Lie.
-
Wow, thank you. Not my field, but this certainly makes for gripping reading.
Paper by #PVCoveney and #SSucci, from University College London.
"In a regime of highly diminishing returns from scraping more data, the spurious correlations overwhelm the true ones. The low scaling exponents are the result of a near-tie between the exponential growth of resources and the exponential decay of the true correlations with data size."