I admit that I talk to it in English the most (as it's my native tongue) but I had read here that it has gotten noticeably worse in other languages, even making grammar mistakes in linguistically-close-to-english-and-sufficently-present-in-training-data western european languages like German or French.

I do enjoy translating things into Sindarian or other conglangs from literature though.

I had a theory that giving it the custom instructions in latin might give better results, but i haven't had time to figure out how to benchmark something that's so non-deterministic and black-boxy

> Aren't LLMs potentially great to ward off language death due to their significant abilities to transpose ideas into different languages?

What do you mean?

Not GP, but in my case, I find my conversations with ChatGPT or Copilot often being in my native language. Whereas with Google, official docs, and StackExchage it's only and always English.

Obviously, the code I write in the end is English. But explaining a domain in a native language - especially when that domain embeds cultural things - helps a lot. For example, there's a massive difference between "add_high_VAT()" in a Dutch Context from a US context. Even add_VAT(lookup_VAT("FR", "low")) demands quite some domain-knowledge about EU tax system. Which I can express in English, but is much easier in my native language.