The problem is the non deterministic part An LLM can generate many valid reasons for doing X, including the right one, but you can't know which one of those it

The problem is the non deterministic part An LLM can generate many valid reasons for doing X, including the right one, but you can't know which one of those it will give you