I wonder about the latter. It's known that high quality, one-on-one teaching can greatly improve learning speed. Could near future generation AI learn to be a teacher on par with the best human ones?
Even better abstractions have already radically improved the rate of learning. Try learning math by medieval books written before modern notation: the most basic, middle school things like trigonometry will look like arcane mysteries.
I imagine that the problem with getting LLMs to do it is 1)hallucination and 2)a lack of training examples of teaching through text. We rarely teach purely through text messages, I think this might explain why LLMs almost never ask clarifying questions or use the Socratic method. But it might be possible to RLHF it into doing that.
Can you say something about what your approach is?