Google will have to match that somehow. I guess the search engine wars are getting hot again.
I think the real challenge is internationalization - it would be a challenge to build GPT-N like models for all the other languages, that work as well as the original one.
Interesting if Google will roll it's own language model for that purpose. Is it possible, that we might get several language models, each one for a specific category of users, or would that approach lead to a loss in generality/quality?
ChatGPT already works suficiently well in my own language (Danish). It can probably be improved, but so far it has worked well enough for several news stations to report on ChatGPT while prompting it in Danish. And schools are also worried about its ability to write essays in Danish, among other things.
I had some fun with it trying out a few Dutch dialects. Not perfect but it seems to be able to translate what it knows rather than rely on knowledge in a particular language only. Likeswise, it can translate programs to different languages. Paste some code and let it translate to Rust, kotlin or whatever.
For Polish, which is one of more difficult western languages, chatgpt works flawlessly. As in - it can write in a nicer style than an average person does.
Interestingly, it has no problem with throwing in words from my language when I forget english ones.
The only two issues I found:
- doesn't do good rhymes when I ask it to write poetry
- when I asked it to generate content that has mixed polish and english words (as if written by a pole who spent the last 20 years in US, and replaces some words with their english counterparts), it was unable to do so. It could only write either clear english or clear polish.
A major advantage Google has is youtube. Most of the content in web for Indian/Asian language is present inside youtube videos. If that content can be transcribed with decent precision youtube can be huge content booster for such language models.
Funny thing here is that in a blind test where the subject doesn‘t know where the result comes from, more than half like the Bing results better. So this seems to be mostly expectation …
I think the real challenge is internationalization - it would be a challenge to build GPT-N like models for all the other languages, that work as well as the original one.
Interesting if Google will roll it's own language model for that purpose. Is it possible, that we might get several language models, each one for a specific category of users, or would that approach lead to a loss in generality/quality?