Won’t those models gradually become outdated (for anything related to events that happen after the model was trained, new code languages or framework versions, etc) if no one is around to continually re-train them?
If you are feeding the LLM a report, and asking it for a summary, it doesn't need the latest updates from Wikipedia or Reddit.