If AI truly wanted to excel in the world of communication, and many more areas, then engineers should've devoted all of their time into translating langauges. Because Learning a new language is fucking hard, it's difficult because you have to not only understand different forms of grammar but also sentence structure, dialect, how to articulate sentences and phrases automatically from one another.
Some words andhow they are spelled could mean from a different mark and abrasion to a whole different meaning and to a whole different context.
Perfect Example; Japanese Katakana is IMMENSELY different than Kanji and Hiragana, yes they do mean similar but if you fucking translate it to english, there's no ACTUAL ONE TO ONE EQUIVALENT.
Yoroshiku onegaishimasu CAN mean: Nice to meet you, I'm counting on you, pelase take care of me, thanks in advance or I appreciate your cooperation.
It can literally mean several tons of things. But meanwhile English; there's nothing we can say that can equal to that besides those translations. There's no blanket words for context situations.
Another example; German gives us Schadenfreude: pleasure at someone else's misofrtune, but we borrow the word because translating it into English takes a whole ass sentence.
Many different words in general arabic, mandarin, navajo, finnish, turkish each language encodes how we see things completely different. But what did we do with a tool that can make our lives easier?
We obsess over AI making Art, Music, and Videos. Meanwhile, the crux of the barrier of one of the hardest things in our society and existence is literally right there. God what a waste.
15 Comments
foggy@lemmy.world · 12 pts · 4d
Google translate's AI created a intermediary language as a midpoint to translate between all languages like a decade ago.
We're so far past AI and translation. So, so far past it
Onomatopoeia@lemmy.cafe · 0 pts · 4d
The Luddism around AI today is shocking, though I suppose it shouldn't be - it's the same as the old Luddism.
MousePotatoDoesStuff@lemmy.world · 1 pts · 3d
The old Luddism was based, actually.
cy888@lemmy.world · 1 pts · 4d
Luddite? How about using the wrong tool for the job?
jatone@lemmy.dbzer0.com · -3 pts · 4d
that midpoint just the fallout of using transformers. we're really not that far past it.
foggy@lemmy.world · 0 pts · 4d
Were a decade past it.
LLMs are a new technology compared to having AI being able to translate between all languages.
You're empirically incorrect. The worst kind of incorrect.
We are so fucking far past translation.
jatone@lemmy.dbzer0.com · 1 pts · 4d
i build / work on llms. they literally use transformers and attention. we're literally not that far past it. We're basically just optimizing the code paths. All the 'advances' the last 5 years have been routing, context management, and reducing bit size.
You're confusing time with actual progress. the stupidest kind of incorrect: confidently.
foggy@lemmy.world · -1 pts · 4d
Some people just can't stand being wrong:
“I work on LLMs” doesn’t make the chronology stop mattering. Google’s multilingual model was published in 2016 and used an LSTM encoder-decoder with attention. Transformers weren’t introduced until 2017, so its emergent shared representation quite literally could not have been “fallout from using Transformers.”
Empirically incorrect.
And yes, modern LLMs still predominantly use Transformer backbones. That doesn’t mean progress stopped at the architecture any more than CPUs stopped progressing because they still use transistors. Instruction tuning, RLHF, compute-optimal training, synthetic and multilingual data pipelines, multimodality, sparse experts, tool use, and end-to-end speech translation are not all “optimizing code paths.”
Again, you're the worst kind of incorrect here. Empirically.
Aatube@lemmy.dbzer0.com · 0 pts · 4d
Ze agrees that we're a decade past it. Zir argument is that not too much has actually been done in that time. I personally wouldn't say CPUs progressed much over the last decade either even though they're a lot faster now; it's a multitude of optimizations, subjective whether significant, and I'm on that "not that much" side.
foggy@lemmy.world · 1 pts · 4d
Oh you again. Go on and copy and paste my response to some AI as seems to be your norm.
Doing a quick Google, here are some major CPU advancements in tehnlast decade:
Y’all love being empirically incorrect for some reason, but if you have any other questions you want to stage with your ego, feel free.
Aatube@lemmy.dbzer0.com · 1 pts · 4d
that is true. whether we think that means it advanced far is a subjective decision and the other guy's subjectivity (and mine) are on the "incremental" side.
and like, it's Objectively wrong to say I use AI to say things. You never responded to that.
jatone@lemmy.dbzer0.com · -1 pts · 4d
Empirically: training improves the outcomes it doesn't change the capabilities.
ergo: we havent advanced much past translations. since that's empirically what llms do they translate questions into answers. I'm done responding to you its a waste of time talking to a wall.
foggy@lemmy.world · 0 pts · 4d
I'm not going to continue to dunk on someone who clearly can't stand being wrong.
✌️
Edit: homie has gone back and downvoted my old posts. Homie really doesn't like being wrong lol. Join my block list.
theunknownmuncher@lemmy.world · 10 pts · 4d
That's literally exactly what engineers did, yeah.
Translating language is what drove LLM architecture to the design we use now, transformers based on the famous Attention Is All You Need paper. https://arxiv.org/abs/1706.03762
We're a few steps past this, now. Translation is already solved.
AnarchistArtificer@lemmy.world · 4 pts · 4d
They have been focussing on translation — and translation has long been held as one of the jobs that is most likely to be fully automatable; back in 2025, Microsoft published research that argued this, and news media has frequently argued that translators are at particular risk to job loss due to AI.
In practice, as has often been the case with AI, these promises have largely failed to materialise. Although the jobs market has gotten a lot worse for translators, we've seen time and time again that AI just isn't anywhere close to being able to fully replace human translators. Projects that rely entirely on AI end up producing awful translations that just aren't up to scratch, and projects that use human translators but force them to use AI just end up causing those humans to spend longer cleaning up a terrible translation than it would've taken to just do it themselves.
The journalist Brian Merchant (author of "Blood in the Machine") has an excellent article covering a variety of perspectives from translators. If you want an in depth perspective from someone close to the ground, Lucile Danilov is a games translator who has written an excellent piece here that I'd highly recommend.
The fact that they've been trying pretty damn hard to make AI capable of fully replacing translators but are still having such mediocre results leads me to believe that we're not going to see it happen any time soon. All of the nuance and complexity in language that makes it tricky to translate for humans is why it's even harder for a stochastic parrot.