Vincent :coffeecup:
@unlambda Google translate uses LLMs and has for quite a while. At the inception, "AI" as a term wasn't sexy or vogue.
They've been improving for the same reason that other LLMs are improving. The algorithms are getting better, the hardware is faster, and there are more data to train on. I don't know if they've switched to Generative AI, but there's no need to as that would likely make the translations worse.
Translation LLMs are the only "AI" I find useful, too.
Rich Felker
@vincent @unlambda This is another place I don't want LLMs. They ruined Google Translate like all their other products.
It used to do real NLP parsing-based translation, where the output sounded robotic but precise to the parsing of your input, and fairly obvious if it got something wrong.
Now it uses LLMs so the output sounds convincingly natural and it's hard to tell that the meaning is wrong or that it introduced bigoted shit into your words like assuming a person's gender based on their profession.
Brian Campbell
@dalias @vincent It does feel like it's gotten better to me; but I share the concern about it getting better because it's hallucinating.
That's one of the big problems with software as a service; you can't just compare against old versions to cross check.
Maybe it's time to check out and see if there are any worthwhile open source local models for this task. Would be good to work on de-googling.
Rich Felker
@unlambda @vincent That's how LLMs fool people. The whole optimization problem they're solving is making something that *sounds right*. Worse results are accepted by people who don't understand and account for this just because they "sound better".
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.