Ask HN: Can a LLM trained on two languages in isolation translate them?

When training data does not contain any mixup / translation between the two at all. Will it still be able to do translation between those two languages? From my very layman understanding it shouldn't at all, specially if the character set is totally different because the relation between the tokens between those two languages will be zero. For example, if training data was English and Chinese books and none of the texts used other language in its own context, will the translation ability of such LLM be zero? Or will it learn to convey the meaning and gain translation ability?