Trivially, because any two of rarely produce the same "next token"!

An ensemble of LLMs trained identically would generate the same next token(s) forever. But we don't - we generate different sequences.

We are not LLMs.