If you set the temperature 0, an LLM is also deterministic (same prompt -> same output every single time). We just don't do this because the LLM is not so smart in that mode. But "LLM is not so smart" is changing at Moore's law speeds.
Non-determinism is not an essential property of LLMs. It's an optimization that we've added intentionally.
Non-determinism is not an essential property of LLMs. It's an optimization that we've added intentionally.