Post #2458732
2026-05-12 18:27 UTC
@anonymouspl@expressional.social @u0421793@toot.pikopublish.ing @cabel@social.panic.com @Landa@graz.social @futurebird@sauropods.win @StarkRG@myside-yourside.net
LLMs can be really good at recognizing patterns. They don't really uncover the meanings of an unknown language, they just make the patterns much easier for us to recognize (for animal sounds, things like "this is where a word ends" and "this is a different sentence"). And we are the ones to try to uncover the meanings, since we are the ones who have a knack for ascribing meaning to patterns
LLM based translations are bad because the LLMs are forced to try to ascribe the meanings themselves, from "patterns" that are already plainly written (as words), and then rewrite the meanings into a different set of patterns. It's the opposite work
This is similar to how they can be a great tool for medicine research too
Replies (1)
-
@StarkRG@myside-yourside.net 2026-05-12 18:59
@soc I can see the underlying architecture, deep learning transformers, being useful in quite a lot of fields, including medical research, and that's what it sounds like you're taking about. LLMs, though, large language models, don't seem to have much real use. I even doubt they'd be of much use to linguists studying a language since all of its data gets stored in incomprehensible and mostly inaccessible ways. @anonymouspl@expressional.social @u0421793@toot.pikopublish.ing @cabel@social.panic.com @Landa@graz.social @futurebird@sauropods.win