Nobody is talking about current LLM. And, we don't know if gradient decent will be involved, since the systems don't exist yet. Maybe it all be some runtime gradient decent. If there's an optimization problem involved in any part of the process, and numbers are involved, there's a near guarantee it will include gradient decent.