I agree that the results will be different, and certainly a very narrowly trained LLM for conversation could have some potentials if it has proper guardrails. So either way there's a lot of prep beforehand to make sure the boundaries are very clear. Which would work better is debatable and depends on the application. I've played around with plenty of fine tuned models, and they will get off track contextually with enough data. LLMs and procedural generation have a lot in common, but the latter is far easier to manage predictable outputs because of how the probability is used to create them.