Re:Language models cost 10x more in some languages than others
by seanmcdirmid on Thursday May 18, 2023 @03:20AM
(#35984965)
Given how Chinese talk about topics using code words to discuss controversial topics online (and avoid low effort keyword censorship), I wonder if an LLM could learn to do that given a censor in the training process?
[ Parent
| on HN ]
Re:Language models cost 10x more in some languages than others
by hunglee2 on Thursday May 18, 2023 @03:48AM
(#35985112)
it's all code words, this is not as conspirational as imagined by non-language speakers. A great deal of normal conversation / interaction is idiomatic given that it is difficult to create new words in logographic languages. Your point is valid though - how does AI learn this? I don't think it can easily, it is akin for an AI to keep up with street slang, it couldn't guess it an advance it would have to be trained in it, which seems it will need greater human attendance than alphabetical languages!