I wonder if LLMs trained specifically for this purpose can perform well with "forgotten languages".

I know I'm simplifying a lot, but all this deciphering isn't it just some kind of pattern matching?