2026-04-18 16:27 UTC
@bkuhn@fedi.copyleft.org @arianvp@functional.cafe @tito@social.sciences.re
Beyond English, my Spanish & German are enough to get through some fraction of videos & some amount of the time, I'm occasionally even able to wing basic Portuguese, Italian & if I'm really lucky even French depending on accent, vocabulary, clarity of speech & speed. YouTube, Twitch et al do very poor jobs of recommending things from a mixture of languages, likely in part because all of the extant LLM affairs are dealing with signifiers rather than what's signified. BlueSky & threads weren't particularly good at it all either.
It's not even really just the social media failing. Polyglot text input is even rather clunky. Limiting simultaneously active languages for the sake of completion isn't really worth the cost of access to characters like dotless i's or double-accented o's or g's with breves or circumflexes. Cyrillic is enough less of a focus that there aren't even options to do the sorts of simultaneous completion for a few languages sharing a script that are available for Latin for variations like Belarusian, Kazakh, Serbian et al. Unihan has severe issues, like stroke orderings differing by languages despite not making visible differences in written characters.
The one bright spot is that LLMs have got less confused about how to respond to mixed-language input (this happens more often than even intended, as Latin bits are widely used among European language speakers & English grabs whole phrases from French at a time & sometimes random German words too & ppl often don't realise it's another language they're saying it in).
Replies (0)
No replies.