Hi there, my name is Justin Skycak, I'm the Director of Analytics & Algorithms at Math Academy, I developed all of our quantitative software, and I'd be happy to answer your questions.

1. The AI is more like an expert system that emulates the decisions of an expert tutor with regard to what tasks a student should work on at any given point in time (what should the student learn next, what do they need to review). There's a knowledge graph that encodes structural relationships between thousands of math topics (such as prerequisite relationships, but also other types). And then there's an algorithmic reasoning system that looks at a student's answers, overlays them on the knowledge graph, figures out what the student knows (and how well they know it), and decides what learning tasks are going to move the needle most given their personal knowledge profile. The decision-making is inspired by cognitive learning strategies such as mastery learning, spaced repetition, interleaving, minimizing associative interference.

2. Yes, spaced repetition is a core part of the system. Each student has a personalized spaced repetition schedule that adapts to their performance on each topic, and when choosing what topics a student should review or learn next, we're always trying to implicitly "knock out" as many due reviews as possible to maximize learning efficiency. (For instance, if a student is due for a review on one-step ax=b equations, we can implicitly "knock out" that review by having them learn two-step ax+b=c equations instead.)

From a quantitative standpoint, the spaced repetition model was one of the more challenging (but equally fun) parts to build. You normally think of spaced repetition in the context of independent flashcards, but in a hierarchical body of knowledge like mathematics, it gets really complicated because repetitions on advanced topics should "trickle down" to update the repetition schedules of simpler topics that are implicitly practiced (while being discounted appropriately since these repetitions are often too early to count for full credit towards the next repetition).

Our spaced repetition model not only accounts for implicit "trickle-down" repetitions but also minimizes the number of reviews by choosing reviews whose implicit repetitions "knock out" other due reviews (like dominos), and calibrates the speed of the spaced repetition process to each individual student on each individual topic (student ability and topic difficulty are competing factors).

Very cool! Thanks for the reply.

It would be awesome if you could also add GPT-4 as a kind of helpful tutor. Not sure if you're already experimenting with that.

> Our spaced repetition model not only accounts for implicit "trickle-down" repetitions but also minimizes the number of reviews by choosing reviews whose implicit repetitions "knock out" other due reviews (like dominos), and calibrates the speed of the spaced repetition process to each individual student on each individual topic (student ability and topic difficulty are competing factors).

That's super interesting and definitely one of the issues I faced while building anki cards for math classes I took in undergrad. Thanks again!

FYI this is the guy that wrote the comment you replied to: https://www.justinmath.com

He's true math nerd, and definitely not a marketer. In fact, I don't think a single marketer works at the company. It's a bootstrapped labor of love and I've been following their journey for almost a decade.

I get the skepticism, but some things are legit.

I'm sure there are math nerds that work in marketing.

Your whole blog post comes across as an advert, that may not have been your intention but that's what it looks like to me, legit or not.

Don't be nasty.

Especially ridiculing a tech person giving detailed and interesting answers as a "marketeer".

"Director of Analytics"