> an LLM can't learn without adding new data and a training run.

That's probably the next big breakthrough