I gave it a shot. It's interesting for sure, and I appreciate just how many more languages there are to choose from than DeepL. However, as they say in the blog post they use LLMs for translation....
I gave it a shot. It's interesting for sure, and I appreciate just how many more languages there are to choose from than DeepL. However, as they say in the blog post they use LLMs for translation. Which would be fine on the face of it, but I think because the AI has restrictions on curses and swears the translation is Ned Flander-ized sometimes.
In some cases the translation is just as vulgar as the input, in some cases it's censored with asterisks and in others it's just mistranslated to be less vulgar, "damn" instead of "f**k" etc. It's not very consistent.
Considering that it happily spits out many standard curses, I'm guessing that it's not necessarily restrictions like swear filters or instructions in the system prompt, but rather that its...
Considering that it happily spits out many standard curses, I'm guessing that it's not necessarily restrictions like swear filters or instructions in the system prompt, but rather that its training data contained censored words as well as uncensored ones.
I personally think that translation is one of the few suitable use-cases for LLMs (they do it pretty well and support an incredible amount of languages), so I'm glad to see services like this being implemented instead of shoving AI in random places.
This feels like another instance of Kagi actually wanting to be an AI company... I recently came across this blog post and this announcement reminded me about it again.
This feels like another instance of Kagi actually wanting to be an AI company...
I recently came across this blog post and this announcement reminded me about it again.
I gave it a shot. It's interesting for sure, and I appreciate just how many more languages there are to choose from than DeepL. However, as they say in the blog post they use LLMs for translation. Which would be fine on the face of it, but I think because the AI has restrictions on curses and swears the translation is Ned Flander-ized sometimes.
In some cases the translation is just as vulgar as the input, in some cases it's censored with asterisks and in others it's just mistranslated to be less vulgar, "damn" instead of "f**k" etc. It's not very consistent.
Considering that it happily spits out many standard curses, I'm guessing that it's not necessarily restrictions like swear filters or instructions in the system prompt, but rather that its training data contained censored words as well as uncensored ones.
I personally think that translation is one of the few suitable use-cases for LLMs (they do it pretty well and support an incredible amount of languages), so I'm glad to see services like this being implemented instead of shoving AI in random places.
This feels like another instance of Kagi actually wanting to be an AI company...
I recently came across this blog post and this announcement reminded me about it again.