elilla&, famigerada travesti21h ago

daughter asked me if the following impressions are correct

1) Google Translate is worse now
2) it's worse because it seems to flat out make up stuff
3) this must be because they're using so-called "generative AI", in a way their previous machine learning didn't
4) DeepL is also worse now for the same reason, but less so than Google Translate.

this is also my subjective impression of the state of things, but I have no idea how one would go about proving whether that's correct. it's not like we can get evidence of what's inside the proprietary black boxes

Show threadLilaHexe21h ago
@elilla DeepL has admitted to changing to an LLM. The original non-LLM model is still available... behind a paywall
Show threadhermit
@LilaHexe @elilla do you have a link to this info? it's been a while since i've used deepl very frequently, i want to know if i should stop recommending it...
Mar 23 at 3:41pmWeb
Show threadLilaHexe18h ago
@alive @elilla support.deepl.com/hc/en-us/articles/14241705319580-About-DeepL-language-models
About DeepL language models

DeepL Translator uses language models to generate high-quality and accurate translations. On the web translator, in the desktop apps and via the API, you can choose between the classic and next-gen...

DeepL Help Center | How Can We Help You?
Show threadLilaHexe18h ago
@alive @elilla "While the next-generation language model promises improved translation quality, the classic model still provides users with highly accurate and nuanced translations."

That's one way to say you know your new product is shit
Show threadelilla&, famigerada travesti18h ago
@LilaHexe @alive one more strike for the increasingly widespread model of "LLM slop for the poors, quality stuff for those who can pay"
Show threadGrace Vulpes Alopex13h ago
@elilla @LilaHexe @alive

I'm kinda interested in learning what they were using before the LLM, I thought DeepL was always a deep learning/AI based translator? (hence the name?)
Show threadhermit13h ago
@gracie @elilla @LilaHexe my understanding was that they had a custom neural network based model. unclear to me where the training data for that was, how much more traditional machine translation tech was mixed in, etc, but it worked quite well IMO
Show threadGrace Vulpes Alopex13h ago
@alive @elilla @LilaHexe

yeah, DeepL was always better for me than other MTL software, sad to see it's going to shit now :(