I have a theory that it should have a very different “personality” (probably more like writing style) depending on language because it’s an entirely different set of training data

In English chatGPT is rather academic and has a recognisable style of writing, if you’ve used it a bit you can usually get hints something was written by it just by reading it.

Does it speak in a similar tone, with similar mannerisms in other languages? (where possible, obviously some things don’t translate)

I don’t know a second language well enough to have natural conversation so I’m unable to test this myself, and may have worded things awkwardly from a lack of understanding

  • velox_vulnus@lemmy.ml
    19·
    7 months ago

    Bard can’t speak my language Tulu. I teach it some phrases in romanized letter, because the script is yet to be added in Unicode and might take another ten years. So far, only some improvements.

    Here’s an example:

    Prompt: “Eerna poodar enchina?”

    Exact translation: “Your name what?”

    Normalized translation: “What is your name?”

    Output:

    “Eerna poodar enchina?” is a Tulu phrase that means “Whose house is this?”. It is typically used to ask for directions or to find out who owns a particular house.

    Here are some other ways to ask for directions in Tulu:

    “Namma poodar enchina?” (Our house is where?)

    “Nanna poodar enchina?” (My house is where?)

    “Enda poodar enchina?” (Their house is where?)

    “Enda poodar ninna?” (Whose house is this?)

    “Enda poodar yen?” (Whose house is that?)

    Expected output:

    “Eerna poodar enchina?” is a Tulu phrase that means “What’s your name?”. My name is Bard, an AI assistant.

    As you can see, everything about this is so wrong. It also mixes Kannada with Tulu. And none of this is intelligible - it’s all gibberish. They’re not even the same language.

    Bard also sucks at Hindi/Urdu translation. I’ve asked it to translate “Bheegi Bheegi Raaton Mein”, which is a pretty old Bollywood song, but it just fails. And even if it does after carefully crafting the prompt, it fails to capture the beauty of the lyrics in the actual source. I’m not good with French, although I’ve learnt about it before, so I’ve not bothered trying.

    • PlexSheep@feddit.de
      5·
      7 months ago

      How can a language work, if it’s signs are not in Unicode? That sounds incredibly tedious.

      • velox_vulnus@lemmy.ml
        8·
        7 months ago

        For now, folks are using the Kannada script. But I don’t use it, because, let’s just say that it is kinda similar to the French-Occitan situation. By the way, here is the developing script that will be added to Unicode soon.

      • emergencyfood@sh.itjust.works
        4·
        7 months ago

        Not all languages have scripts, and not all scripts are on Unicode (yet). So in this example Tulu speakers would use either the Kannada or the Latin script, both of which are on Unicode.