rexi
https://itsfoss.com/falkon-browser/
https://en.wikipedia.org/wiki/Falkon
"Falkon (formerly QupZilla [5]) is a free and open-source web browser developed by KDE. It is built on the QtWebEngine, [6] [7] which is a wrapper for the Chromium browser core. [8] Both KaOS and openMandriva Lx use Falkon as their default browser..."
Nivex ? ?
@craigmaloney @mozilla Cosigned.
Maxi 10x 💉
@craigmaloney You really don’t get it. Mozilla never developed AI software for servers. Their STT voice model efforts and the translator are local software.
Maxi 10x 💉
@craigmaloney Not true. They have incorporated the translation AI, which runs completely locally.
Maxi 10x 💉
@craigmaloney How is this contradictory to "Mozilla builds local AI software"?
Rich Felker
@SnoopJ We've seen that even if you say things are "AI generated", people generally do not understand what that means about accuracy; they attribute intelligence to it - because that's right there in the deceptive marketing name "AI". You really need for it to *not sound like* it was written/translated with intent by an intelligent being, or people will infer that it was.
SnoopJ
@dalias oh, totally agree that it's an ethical priority to clearly present translations as what they are.
For my taste, I'm OK with the current Firefox default behavior of prompting before translation. I would be incandescent if they switched that to automatic by default, even if the models were really great.
Rich Felker
@SnoopJ Not that it's more achievable, but that it's less dishonest, and not actively dishonest. Making something that intentionally looks polished and deliberate in the target language, but that's constructed out of a model of existing writing in the target language rather than understanding of and faithful attempt to convey the meaning of the original, is dishonest and a disservice to the user.
SnoopJ
@dalias ah. is it fair then to say that you're saying transliteration (or maybe "transcription" is the better-fitting language jargon?) is a more achievable goal than translation? I can certainly agree with that
Rich Felker
@SnoopJ By actually parsing the grammar and having robotic, formulatic transformations that intentionally *don't* sound natural in the target language but preserve clues about the structure in the source language.
SnoopJ
@dalias sure, granted. I just don't know how you get that kind of nuance without tossing the idea of automation out the window entirely.
Rich Felker
@SnoopJ Ideally you translate in a way that fully preserves ambiguity. But the question you're asking is different from the topic at hand, which is about the "AI" model confabulating meaning out of *its* statistical expectations based on limited local context rather than the editor or reader getting to fill in any missing meaning from a much more specific context and know what was filled in vs explicit.
SnoopJ
@dalias are there *any* translation models that don't end up with bias? I mean, I hear what you're saying, but if zero-bias is the goal, I don't see how you get there from here with any approach, automated or no.
Even human translators introduce bias. To pull out a favorite example, how one decides to translate «Aujourd’hui, maman est morte» can influence the reading of an entire novel: https://www.newyorker.com/books/page-turner/lost-in-translation-what-the-first-line-of-the-stranger-should-be
Even a rule system can be biased, e.g. how do you represent the はvs. が dichotomy?
Rich Felker
@SnoopJ @rysiek @craigmaloney It's NOT a good approach. It pulls biases from the training data into the output.
SnoopJ
@dalias @rysiek @craigmaloney AIUI the Firefox models AREN'T based on "LLMs" in the colloquial sense.
Many of the models *do* use the Transformer architecture, but the size of the model weights¹ is several orders of magnitude away from the state-of-the-arse junk currently separating gullible VCs from their investorbux.
Language-pair translation is a legitimately good approach and Mozilla et al. should be applauded for making it available locally.
¹see `models/prod` in https://github.com/mozilla/firefox-translations-models
Rich Felker
@rysiek @craigmaloney Machine translation got so much worse when it started to be based on the predecessors of LLMs rather than parsing grammar. Slightly klunky & robotic text with clear meaning got replaced by confabulations of gender and similar based on biases in training sets.
Rich Felker
@rysiek @craigmaloney But it would be so much better if it wasn't based on "AI".
Michał "rysiek" Woźniak · 🇺🇦
@craigmaloney eh, I am a pretty vocal critic of "AI", but the local, in-browser translation tool Mozilla is developing in Firefox is something I can really get behind.
And that's of course based on "AI".
Edit: apparently I was wrong and Mozilla is in fact talking about focusing on generative AI, not just stuff like language translation. That's absolutely a crap move. 🤦♀️
scrottie (he/him/they)
@craigmaloney @mozilla Starting to think the solution to the fuckup Mozilla CEO problem is to stop having CEOs. That's clearly not working.
Rich Felker
@rexi @craigmaloney @mozilla Lost me at "wrapper for the Chromium browser core". That's a non-starter.
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.