That can be an issue, although - I think - the European approach will be to block private and semi-private data (like fb posts), but allow public data (articles

That can be an issue, although - I think - the European approach will be to block private and semi-private data (like fb posts), but allow public data (articles etc). In Poland we have a team of people who work on SpeakLeash which is 1TB of polish texts freely available for anyone to train their models on - http://speakleash.org/en/speakleash-a-k-a-spichlerz-english/ http://speakleash.org/en/speakleash-a-k-a-spichlerz-english/ I can imagine models getting their knowledge and intelligence by getting trained on global - english and chinese - data, and then learning country specific languages and viewpoints from way smaller datasets. You can see that already - gpt was trained on way less Polish language material than english, but it’s almkst just as smart and fluent in both. As for cultural quirks - here it’s more about who does hrlf and alignment than the source material.