We hit 50,000 active student users on a Tuesday. I remember because I was in the middle of a board meeting, and I had set up a notification to go off when we crossed the milestone. My phone buzzed, I glanced down, and I couldn’t stop a huge grin from spreading across my face. We had done it. We had built something that was actually helping tens of thousands of students learn better.
But here’s the thing they don’t tell you in the press releases or the celebratory tweets: the journey to that number was brutal. It was a series of painful lessons, costly mistakes, and moments of near-crippling doubt. My MBA was great for learning how to read a balance sheet, but it taught me next to nothing about building a product that people, especially kids, actually want to use. Scaling an AI learning platform taught me that.
I’m not here to give you platitudes. I’m here to give you the real, hard-won lessons from the trenches. If you're building anything in the AI or education space, this is for you.
1. Your Algorithm is Blind Without Context
When we started, we were obsessed with the algorithm. We had a team of brilliant data scientists building a recommendation engine that could predict what a student should learn next with scary accuracy. We tracked everything: time on task, quiz scores, video watch time, you name it. The model was beautiful. The problem? The recommendations felt sterile, almost robotic.
We were tracking what students were doing, but we had no idea why. A student might spend 30 minutes on a single math problem. Our system flagged this as a struggle and recommended easier content. But when we actually talked to the student, we found out he was just incredibly determined to solve it himself. He didn't want an easier problem; he wanted a hint. We were optimizing for completion, not for the actual, messy process of learning.
The lesson: Quantitative data tells you what's happening. Qualitative data—talking to your users, understanding their motivations, their frustrations—tells you why it matters. We ended up building in simple feedback tools: “Was this helpful?” with an open text box. The insights from that little box were more valuable than a million data points.
2. The 'Cold Start' Problem Will Make You Question Everything
How do you personalize an experience for someone you know nothing about? This is the classic “cold start” problem in machine learning, and in education, it’s a killer. You can’t just throw a new student into the deep end. But you also can’t bore them with a 20-minute onboarding survey.
We tried everything. We tried generic “popular” content. We tried short diagnostic quizzes. We even tried letting users pick their own starting point. Nothing felt right. The quizzes felt like a test before school even started, and the generic content defeated the whole purpose of “personalized” learning.
Our breakthrough came when we stopped thinking about it as a data problem and started thinking about it as a human one. What do you do when you meet someone new? You find some common ground. We built a simple, almost game-like experience that asked students to pick 3 topics they were curious about, not just what they were supposed to be learning in school. That gave us just enough of a signal to make a first, decent recommendation. It wasn't perfect, but it was a start. And it felt like a conversation, not an interrogation.
3. If You Can Game It, a 14-Year-Old Will
We introduced a points system to encourage engagement. Students earned points for completing lessons, watching videos, and taking quizzes. We thought we were so clever. We were not.
Within a week, we saw bizarre patterns emerging. Students would open a video lesson, mute it, and let it play in a background tab to rack up points. They would retake the same easy quiz a dozen times to boost their score. They weren't learning; they were farming. They had turned our carefully designed educational tool into a clicker game.
It was a humbling reminder that you can't just slap a leaderboard on something and call it engagement. The motivation has to be intrinsic. We ended up completely overhauling the system. Instead of points, we started rewarding students with access to more interesting content or the ability to customize their learning environment. The focus shifted from an external reward to the internal satisfaction of mastering something new.
4. The Uncanny Valley of AI Tutors is Real and It's Creepy
In one of our early iterations, we created an AI tutor with a name and a face. We thought it would make the experience more friendly and personal. We were wrong. Students found it weird. The canned, overly positive responses like “Great job! You’re a superstar!” felt disingenuous.
There's a fine line between a helpful tool and a creepy digital puppet. An AI that tries too hard to be human just highlights how it's not. It’s the uncanny valley, and it’s a real turn-off. As I see the progress being made at companies I’ve invested in, like Anthropic and OpenAI, I know this will get better. But for now, honesty is the best policy.
We found that students responded much better when the AI was positioned as a smart assistant, not a fake friend. It was a tool, like a calculator or a search engine, but one that was incredibly good at finding and explaining information. We made its language more direct and to the point. No more fake enthusiasm. Engagement actually went up.
5. Teachers Are Your Most Important Users
This was our biggest mistake early on. We were so focused on the student experience that we almost completely ignored the teachers. We thought, “If the students love it, the teachers will have to adopt it.” That’s not how the education system works.
Teachers are the gatekeepers, the champions, and the power users. If they don't understand your tool, if they can't easily integrate it into their lesson plans, or if they feel it threatens their role, it's dead on arrival. We had built a great tool for students, but a terrible one for teachers. The dashboard was confusing, it was hard to track student progress, and it didn't align with their curriculum.
We had to go back to the drawing board. We spent months co-designing a new teacher dashboard with a panel of educators. We made it simple to assign content, monitor progress at a glance, and identify students who were falling behind. We made the teacher the hero, with our platform as their superpower. That’s when adoption really took off.
6. Ethical AI Isn't a Feature, It's the Foundation
When you have data on 50,000 students, you have an immense responsibility. The data you collect and the models you build can have real-world consequences. An algorithm that is supposed to predict a student's potential can easily become a self-fulfilling prophecy.
We had to confront the issue of bias head-on. Our initial dataset was skewed towards students from higher-performing schools. As a result, our model was inadvertently penalizing students from less-privileged backgrounds. It was a gut-wrenching realization.
Fixing this wasn't a simple code change. It required a fundamental shift in our company culture. We invested heavily in data auditing, fairness toolkits, and building a more diverse dataset. We created an ethics board to review new features. It slowed us down, and it was expensive. But there was no other choice. You can't claim to be personalizing learning if you're perpetuating systemic inequality.
7. 'Personalized' Does Not Mean 'Isolated'
My final lesson is perhaps the most important. In our quest to create a perfectly individualized path for every student, we almost forgot that learning is deeply social. Our first version of the platform was a solo experience. A student and the algorithm.
But that’s not how we learn in the real world. We learn by talking to people, by arguing, by collaborating, by teaching someone else. I saw this firsthand building RemoteTeam—the best remote companies are the ones that master asynchronous collaboration.
We started building in social features. Not just forums, but structured, collaborative projects. Students could team up to solve a difficult problem, with the AI acting as a facilitator. They could create study guides together and share them with the community. We saw that when students worked together, they not only learned the material better, but they were also more motivated to keep going.
Building a product that touches so many lives is a privilege. It’s also a heavy weight. The lessons are often painful, but they are the only way to build something that truly matters. The future of education isn't just about smarter algorithms; it's about using technology to create more human, more connected, and more equitable learning experiences for everyone.
Frequently Asked Questions
Are these recommendations still relevant in 2026?
Absolutely. While specific tools and tactics change, the underlying principles remain consistent. I update my thinking regularly based on what I'm seeing in the market and across my portfolio companies.
Which item on this list has the highest impact?
It depends on your stage and context, but in my experience, the items near the top of the list tend to have the broadest applicability. That said, sometimes the less obvious items create the biggest breakthroughs for specific situations.
Can I implement all of these at once?
I'd strongly recommend against it. Pick the 2-3 items that resonate most with your current situation and focus there. Trying to do everything simultaneously is a recipe for doing nothing well.