I'll probably open source and Show HN the AI tutor I've been working on for my kids at some point, but working on it has given me a little insight into the problem.

The biggest thing is motivation. First off, if Khanmigo requires them to type and read everything, that's going to get tiring fast for most kids. But I don't know how you could do voice in a school setting - mine uses STT/TTS, but with 20 kids in a room, it'd be chaos - STT accuracy and diarization with 2 is already really challenging.

Motivation is helped a bit by following their interest, but it seems like KA is having trouble guiding the kids when they prompt it that way. That was a pretty big issue with mine early on - the kids would talk to it for an hour about whatever topic they were interested in at the time, but it would never branch into something new.

The tutor I'm working on solves it by having a concept graph that covers a lot of learning, from the basics like math, dinosaurs, etc to other developmental topics like 6 year old boundary-pushing humor, and two LLM threads - one that handles the conversational turns, and another one in the background that strategizes and steers the conversational thread by looking at the concept graph connections and considering how ready they are for each, and then injecting steering notes into the conversational thread. Basically system 1 and system 2 thinking. And after sessions, it'll make a basic plan of where to start next time, and what might be interesting to offer up.

I mentioned this in another comment, but I've been really pleasantly surprised at the quality of the tutoring, especially when it bridges into new topics - one of my sons is really into slay the spire, and at different times it’s used that as a launching-off point into probabilities, decision trees, python code of the algorithms he thinks about as he's facing different enemies, and general strategies on different facets, and my other son was really into sharks, which it has bridged into extinct sharks like megalodon, how scientists derive how it looks given cartilage's lower propensity to fossilize, bridging to dinosaurs and their fossils, the K-PG extinction event, how food scarcity filtered for smaller animals like the ancestors of birds, and our small mammalian ancestors. And a whole bunch of other topics.

It's been pretty great in that way, but my biggest open question at the moment is how to get them to engage with it on their own on a more regular basis - they go to it occasionally for random questions, but to get good coverage of that huge knowledge graph would take much more. And fundamentally, I think that human engagement still just has a number of important aspects to it that it's lacking, and I'm not sure if it's possible to replace those well enough.