Many computers and interfaces are deterministic. LLMs are by nature not deterministic and not even non-deterministic the same way on any two invocations given the same prompt and context. Natural language is ambiguous and for many languages very context dependent. It's not the greatest interface for a calculator from which we're expecting deterministic accurate answers.

WolframAlpha is a more impressive front end to a calculator than I've seen out of LLMs. Not only does it show me how it translated my natural-ish language query but it shows me potential alternative interpretations to my question. LLMs by the nature of how training works can't necessarily tell me why and how they interpreted my prompt. The thinking models are better but still not great.