LLMs are garbage and they add nothing to the browsing experience.
> Are there any reports actually showing degradation because of LLMs rather than post hoc ergo propter hoc?
You can control this option with a setting. It seems like it would be really really easy to just test this. As a result I can't see any reason to doubt this by default or apply a legalistic evidentiary standard when considering it.
> LLMs are garbage and they add nothing to the browsing experience.
The builtin translation feature [0] is LLM-based [1], and that adds a ton to my browsing experience, since it's made web pages in other languages accessible to me.
[1]: According to Wikipedia, "A large language model (LLM) is a language model trained with self-supervised machine learning on a vast amount of text, designed for natural language processing tasks" [2]. The translation code is transformer/RNN-based and trained on raw texts [3, 4], and translation definitely qualifies as a natural language processing task, meaning that the translation feature is LLM-based.
That's a pretty big opinion made as a statement there. I use LLMs for many incredibly useful things daily, an example in the browser is summarisation and outlining.
What evidence do you have that it wasn't tested? Or are you just making an assumption based on one article? Did you do any follow up searching? It might have led you to this:
Hold on... For starters, LLMs made Chinese, Spanish, French websites comprehensible to me. Even the best models like Claude still make some mistakes and miss some nuance, but they largely have genuine understanding of the cultural context and can adapt it. They're qualitatively better than classic machine translation models, and the mistakes will eventually be fixed. I can finally communicate in person's native language and be reasonably sure that most subtlety is preserved. This is something that actually breaks language barriers, defies the Babylon, connects people! If that is nothing to you, I don't know what else is.
Deep search is also extremely useful, it saves me actual days and finds things and cool people I wouldn't be able to find in reasonable time otherwise, a lot of that in isolated communities like discord servers normal search has no insight into. Sure, you have to be aware about things left in the dark, but you also have to be attentive to this with normal search as well.
Video summarization lets me extract the actual message from the pile of clickbait, marketing, and three-hour long interviews - all while pointing at the timecodes for me to verify.
Now, the pressure of bullshit will inevitably shift to LLM training. It's only a matter of time. But that's another question, the tech itself works amazingly well, what the hell are you even talking about?
NMT doesn't "contain" tranformers and deep RNNs, it can use them. LLMs use a transformer architecture, not everything using a transformer architecture is an LLM. NMT can actually use an LLM, but that's not the case according to the documentation you linked, they use a parallel dataset to train their models.
LLMs are garbage and they add nothing to the browsing experience.
> Are there any reports actually showing degradation because of LLMs rather than post hoc ergo propter hoc?
You can control this option with a setting. It seems like it would be really really easy to just test this. As a result I can't see any reason to doubt this by default or apply a legalistic evidentiary standard when considering it.