Building a real-time AI tutor for 5-year-olds

Elizabeth here, co-founder (and clinical child psychologist). Fair question. Catalin's post was about the engineering (he is my co-founder and our CTO), so the learning side got short shrift :).

Here's what it actually looks like for a child. Say a 6-year-old is reading a story out loud (I will use a reading example here). The tutor is listening to every word. When she stumbles on "chick," it doesn't just tell her the word; it decides, based on her history, whether to break it into sounds, point back to a pattern she's seen before, or let her wrestle with it a moment longer because she's close. If she misses the same pattern twice, that digraph shows up woven into her next story. If she reads fluently but can't tell the character what happened in the comprhension conversation after the story, she gets another text to work on comprehension instead of just pushing harder words. The instructional approach isn't novel or new, it's what a good teacher does, grounded in the science of learning. We run evals on the interactions and real subject matter experts are grading and annotating the behavior. What's new is doing it responsively, for one specific child, on every turn.

On engagement: I'd push back a little on the framing that engagement and learning are separate things (anyone on our team will tell you this is a drum I have beaten for years). A disengaged child learns nothing, no matter how good the pedagogy is. But we're not optimizing for time-on-screen. The lessons and sessions are bounded. The engagement work exists so the child stays in the productive struggle zone long enough for the teaching to happen.

Why AI: it's not that AI "needs" to be the solution. In fact, a great human tutor is better, full stop, but it has never scaled. A classroom teacher with 25+ kids teaches to the middle. This is the first technology that can make real-time, child-specific teaching decisions, which is what tutoring actually is. More on the pedagogy here if you're curious: https://www.ello.com/our-teaching-approach)

My school district is beginning to roll back on computer use for kids, after having gone all in on putting almost all student work on laptops (and heavily relying on learning apps) during the pandemic. They are now offering guidance to teachers and parents about when and how to limit computer usage. They also banned student cell phones a couple years ago and students are never allowed to access social media on school property. All that to say, I think my district (and other similar ones) are struggling so hard to balance good active learning against unhealthy amounts of screen time that they will reach this same place with AI fairly quickly, within a year or two I think, and start wanting to set restrictions and place guardrails.

I personally don’t know anyone who’s worried about their kids falling behind because of lack of AI knowledge. I know lots of parents worried about how screen-centered life is for kids.

No, in my opinion as a parent this is lunacy. By the time my little kids "need" AI or the Internet in a meaningful way the AI labs will have either delivered the dream of "just talk to the computer" in which case there's nothing to learn to use, or they will be smoldering Pets.com-esque craters. Either way, for us right now it's books and a little bit of "old fashioned" movies and video games; no algorithmic feeds, all human-created content, and a focus on enjoying narratives and experiences together. We can look things up together on the Internet if we need to, and if that routes to an LLM Dad may groan a little but it's OK. Mostly we learn by trying things out, making guesses and talking about them, or looking in books.

I understand the calculus may change with middle school and up, but I still think that despite the "rat race" dynamic of grades and homework, kids who learn to think the pre-2023 way will come out ahead in the long run, even if it's only in life satisfaction.

[dead]

Thank you. The segment showing a child reading text on the screen which highlighted a word they had difficulty with seems like it could be a useful learning interaction. How does your system follow up in that case? Have you studied this type of interaction?

That's the only moment in the video that gave me a sense of what it might be like for a child using this system.

In the blog post you say:

> Imagine a custom story about dragons this week, ice princesses the next — woven with the letter blends your child needs to practice right now.

Have you considered using an automated orchestration system to deliver literature that already exists? This example seems like an opportunity to introduce children to really thoughtful literature like Astrid Lindgren's Pippi Longstocking stories but I'm deeply skeptical that generating the stories with an LLM would inspire a similar experience.

Are there other examples of your platform from the perspective of a child using it? I think those are both interesting cases: 1) interactive feedback on a subject they are making an effort toward mastering, and 2) trying to deliver information when it seems relevant. I'd like to know more about how you are approaching these things and other aspects of the learning process.