Just read that "Large language models speak 320 languages. They think in one" What? No way.

Thanks, AI people, for taking the time to talk about the multilingual that needs to be multicultural, but please oh please, don't spread misinformation, no matter how catchy it is.

LLMs do not internally operate in a language at all. They're all about vector spaces (embeddings). Tokens from different languages are mapped into numerical representations. This is no language, no interlingua.