Comment by parliament32
6 days ago
> I'm not sure what the solution here
The solution is to use a translator rather than a hallucinatory text generator. Google Translate is exceptionally good at maintaining naturalness when you put a multi-sentence/multi-paragraph block through it -- if you're fluent in another language, try it out!
Google translate used to be the best, but it's essentially outdated technology now, surpassed by even small open-weight multilingual LLMs.
Caveat: The remaining thing to watch out for is that some LLMs are not -by default- prompted to translate accurately due to (indeed) hallucination and summarization tendencies.
* Check a given LLM with language-pairs you are familiar with before you commit to using one in situations you are less familiar with.
* always proof-read if you are at all able to!
Ultimately you should be responsible for your own posts.
I haven't had a reason to use Google Translate in years, so will ask: Have they opted to not use/roll out modern LLM translation capabilities in the Google Translate product?
As of right now, correct.
You are aware that insofar as AI chat apps are "hallucinatory text generator(s)", then so is Google Translate, right?
(while AFAICT Google hasn't explicitly said so, it's almost certainly also powered by an autoregressive transformer model, just like ChatGPT)
> it's almost certainly also powered by an autoregressive transformer model, just like ChatGPT
The objective of that model, however, is quite different to that of an LLM.
I have seen Google Translate hallucinate exactly zero times over thousands of queries over the years. Meanwhile, LLMs emit garbage roughly 1/3 of the time, in my experience. Can you provide an example of Translate hallucinating something?
Agreed, and I use G translate daily to handle living in a country where 95% of the population doesn’t speak any language I do.
It occasionally messes up, but not by hallucinating, usually grammar salad because what I put into it was somewhat ambiguous. It’s also terrible with genders in Romance languages, but then that is a nightmare for humans too.
Palmada palmada bot.
Every single time it mistranslates something it is hallucinations.
Google Translate hasn't moved to LLM-style translation yet, unfortunately
Hard disagree. Google Translate performance is abysmal when dealing with danish. In many cases its output is unusable. On the other hand, ChatGPT is excellent at it.
Google Translate doesn't hold a candle to LLMs at translating between even common languages.
IMO chatgpt is a much better translator. Especially if you’re using one of their normal models like 5.1. I’ve used it many times with an obscure and difficult slavic language that i’m fluent in for example, and chatgpt nailed it whereas google translate sounded less natural.
The big difference? I could easily prompt the LLM with “i’d like to translate the following into language X. For context this is a reply to their email on topic Y, and Z is a female.”
Doing even a tiny bit of prompting will easily get you better results than google translate. Some languages have words with multiple meanings and the context of the sentence/topic is crucial. So is gender in many languages! You can’t provide any hints like that to google translate, especially if you are starting with an un-gendered language like English.
I do still use google translate though. When my phone is offline, or translating very long text. LLM’s perform poorly with larger context windows.