Jason Roberts, Part 2: Teach, Don't Entertain · Math Academy

What we talked about.

In part two of his conversation with Sunny Ray, Jason Roberts of Math Academy returns to explain why he believes in teaching rather than entertaining. He describes the reality of running a scaling startup with no external funding, where the focus is on delivering real value to customers rather than impressing investors. Roberts says roughly half of Math Academy's customers are adults, though growth in schools may be shifting that ratio, and that most people find the company through social media and word of mouth. He explains that the product is not an LLM chatbot but an expert-system style knowledge graph of prerequisite topics, with active recall review built on top. He shares a balanced view of AI coding tools, warning about cognitive debt when developers stop understanding their own code. Using gym and video game analogies, he argues that learning requires effort, that engagement should not come before effort, and that the best experience keeps challenge calibrated to the learner. He also claims the system is about four times more efficient than a good math class and points listeners to mathacademy.com.

Math Academy's Jason Roberts argues real learning takes effort, so engagement should never come before it, and fun comes from earned progress.

The questions, and the answers.

In February you said over half your customers were adults learning math for machine learning. What is the number now, and who are they?

It's not specifically machine learning, it's adults. I haven't checked recently, but it's roughly half, give or take, maybe less now because we've been growing more in schools. It's still substantial, and there's a vocal community that wants the math to become expert at machine learning rather than just users of the technology.

How did customers actually find you this year?

We still haven't put real effort into marketing beyond being visible on X, and our director of analytics has started on Instagram and TikTok. Thinking about marketing before your product delivers consistent value is putting the cart before the horse, but technologists also over-correct and avoid it. Most customers still find us through social media and word of mouth.

You said the AI inside Math Academy would be subtle, not a chatbot. What did it turn into?

It's closer to an expert system. Fundamentally it's a knowledge graph of mathematical topics, around 300 to 350 nodes per course, linked by prerequisites. Technology on top decides what to present and when, only after prerequisites are mastered, and it brings back older material through active recall before it fades. It's not an LLM chatbot.

What's one thing you believed in January or February that you no longer believe today?

I've moved toward seeing AI as a real accelerant for coding. Refusing it is like insisting on walking after cars were invented. But handing everything to agents leads to cognitive debt, where you no longer understand your code base. The sweet spot is exercising judgment, taste, and discipline, and engaging with it seriously rather than accepting everything.

You called ed tech a deal with the devil. What is the exact trade a founder makes?

The trade is putting engagement before effort. If you make learning fun and easy so students stay, you lower the bar for what they're expected to do. But learning requires struggle, like going to the gym. Duolingo has nailed engagement, but I'm not sure that makes customers progress a lot. I'd rather be honest that it takes work.

How do you help a student tell the difference between useful struggle and being stuck?

It depends on context, but for high school or most college math, if you're spending more than a few minutes on one problem, you may be wasting time. I'd rather take weight off the bar and get more reps. You want problems at the frontier of knowledge, not too easy and not too hard.

Do fun and hard work have to be opposites? Can progress feel rewarding without making the work easier?

Of course. Think of video games: too easy is boring, too hard is frustrating. The flow state sits where the challenge matches your capability. Overcoming a real obstacle, like solving an algebra problem, is a huge dopamine hit. But you can't fake it. Students know when you praise them for something easy.

What does a 16 year old's education look like if you're right?

Looking at our data, the system is on the order of four times more efficient than a normal good math class. It adapts and optimizes for every individual student. It's like a private tutor versus a class of 30. Parents instinctively know the tutor is more effective, but it's also far more costly.