If the exact words we choose when writing matter so much, then why use a non-deterministic LLM that produces slightly different output on every run?