20
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 10 Mar 2025
20 points (100.0% liked)
TechTakes
1765 readers
94 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 2 years ago
MODERATORS
Google Translate having a normal one after I accidentally typed some German into the English side:
What's the over/under on an LLM being involved here?
(Aside: translation is IMO one of the use cases where LLMs actually have some use, but like any algorithmic translation there are a ton of limitations)
On the left side within the text box there's a sparkle emoji... so I guess that means AI slop machine confirmed
More seriously though, Google Translate had odd and weird translation hiccups for a long time, even before the LLM hype. Very possible though that these days they have verschlimmbessert^1^ it with LLMs.
^1^ Just tried it, google translate doesn't have a useful translation for the word, neither does DeepL. Disappointing. Luckily, there are always good old human-created dictionaries.
Wait didn't Google Translate used to have a feature where you could type in improvements? I don't see it now so they might have gotten rid of it...
Aside: my favorite human-created dictionary is Kenkyusha's New Japanese-English Dictionary. I have a physical copy and it's around 480,000 entries across nearly 3000 pages and paging through it I just feel "yes, now this is a dictionary". It's so big that I might have to give it away or leave it with a friend if my plans of immigrating work out.
Anecdotally, greek <-> english stuff seems to be deteriorating also.
How the fuck can a hallucinating bullshit machine have use in translation
To the extent that machine translation was already a bullshit machine I guess. Language learning I sometimes get totally desperate if there's some grammar construct I can't figure out. A machine translation sometimes helps me know what to look up, or at least move on to the next sentence.
Anyway this isn't a position I believe strongly in. It's iffy for sure and none of these companies ever share their quality evaluations or put "probably nonsense" warnings on the output or give you an option.
Translation is definitely mostly pretty good, but I think I still prefer the older style with broken grammar to LLMs making up well formed plausible sentences that are completely wrong.
Also the results of translating back and forth and on and on are a lot less interesting, though in exchange it is fun to type stupid nonsense into it.
Really what I want is both:
A list of words and their individual translations. Parts of speech, pronunciation, and any relevant conjugations, tenses, etc. How the sentence is put together grammatically / vocab wise basically. Google Translate stinks for this you have to type in fragments of a sentence and hope for the best. This is what I'm usually after since my goal is to learn a language, not have it read to me.
A computer's best guess about what a sentence means as a whole. In case I'm terribly confused and it happens to be accurate enough for me to figure it out from there.
Google Translate focuses on #2 over #1. e.g. it doesn't make a very good dictionary / grammar reference.
Machine translation was the original purpose of the transformer architecture, and I guess it was unreasonably good at it compared to the existing state-of-the-art RNNs or whatever they were doing before.
LLMs are a development from machine learning of the sort that works, and translation was always squishy. Just a straight LLM can do surprisingly well at translation, though if I run "surprisingly well" through a translator it comes out as "impressive demo, not so great product." Tools like Google Translate are a complex array of stuff, not just one approach.
Translation is a good fit because generally the input is "bounded" and stays on the path of the original input. I'd much rather trust an ML system that translates a sentence or a paragraph than something that tries to summarize a longer text.
lol i can replicate this glitch for glitch