I’ve found LLMs to be very good at teaching. Better than I expected.
You have to tell it what you want out of the session. A kid could just type in the question and get the answer, but that’s not tutoring.
When I use them for learning I tell it my goal, summarize my current status, and then tell it to teach me incrementally with a quiz after each step. You can get fancy and have it reach back into history to throw in extra refresher questions.
It uses very few tokens compared to a coding session so you can use the highest model for this and the usage is negligible. Using the top model available to you is important to keep things accurate. I wouldn’t have trusted this back in the GPT 4 or Sonnet days, but with Opus, Astra, or Fable it’s great.
For longer sessions set up a directory and have it keep notes about your progress, problem points, and what was covered in the past.
This is what I meant about using them in the right way. You can totally setup guardrails but my experience has been that unless a person has good self-control, it's way too easy to just get an answer from it. You have to really be interested in learning to do this, though. Thanks for sharing!
You can make shells around the AI experience that make sure this happens, and you'll have to provide some form of oversight.