Sure, LLMs can't learn. I'm saying that systems built around LLMs can simulate aspects of what we might call "learning".