Some languages will be hard to train for properly: eg Serbian has 7 declensions (nominative, dative, genitive...), gender and number embedded in noun forms, ver

Some languages will be hard to train for properly: eg Serbian has 7 declensions (nominative, dative, genitive...), gender and number embedded in noun forms, verbs also embed gender and number on top of tense, and this explosion in complexity makes it impossible to find the data covering all possible combinations. Using ChatGPT in Serbian is like talking to a really bad... bot from the 90s. It mostly feels like two bad machine translation steps to and from English in between. So, there is a whole NLP world out there that's badly covered by OpenAI, which means lots of options for advances in ML. Not that I find there aren't plenty of improvements to be made to ChatGPT English models too.