''' I still remember the exact moment the idea hit me. I was watching my nephew, a bright kid, get frustrated with a math app. The app was "adaptive," but it was just adjusting the difficulty. It had no idea he was bored, not struggling. It was a cold, unfeeling system. And I thought, "We can do better."
People are always asking me how our AI tutor feels so... human. It’s the question I get most often, and I love it. Because the answer isn't a simple one. It's not just a clever algorithm or a massive dataset. It's a new way of thinking about how we build learning systems. This is the story of how we did it.
The Ghost in the Machine
Building a truly adaptive AI tutor is more art than science. At my first company, RemoteTeam, we learned that building for people meant understanding their workflows, their frustrations, and their moments of joy. At MovieLaLa, it was about understanding taste and emotion. We’re applying the same principles here, but at a much deeper level.
Our goal was to create an AI that could understand a student's emotional and cognitive state in real-time. Was the student excited, confused, or losing focus? We needed to know.
This meant going beyond right or wrong answers. We had to build a system that could read the subtle cues. We use a combination of facial sentiment analysis via the webcam (with permission, of course), the speed and pattern of a student's clicks, and even the hesitation in their mouse movements. It sounds like a lot, but it’s all part of building a complete picture.
The Code Behind the Curtain
So, how does it all work? The core of our system is a custom-built transformer model we call "Cognitive Flow." Think of it as a super-smart prediction engine. But instead of predicting the next word in a sentence, it's predicting the student's next mental state.
Here’s a simplified look at the logic:
def get_next_action(student_state):
if student_state.emotion == "confused" and student_state.performance < 0.6:
return "offer_hint"
elif student_state.emotion == "bored" and student_state.performance > 0.9:
return "increase_difficulty"
elif student_state.time_on_task > 300 and student_state.emotion != "engaged":
return "suggest_break"
else:
return "continue_lesson"
This is a massive oversimplification, but it gets the basic idea across. We’re constantly taking in data, updating the student’s state, and choosing the best next action. It’s a continuous, real-time loop.
We also use a technique called Bayesian Knowledge Tracing. It’s a fancy way of saying that we model what a student knows and how likely they are to forget it. This allows us to create a personalized learning path that’s always challenging but never overwhelming.
The Data-Driven Breakthroughs
Of course, none of this would be possible without data. Lots and lots of data. We’ve processed over 2 million learning sessions. And the insights have been incredible.
For example, we discovered that students who physically nod their heads while watching a video explanation are 78% more likely to answer the follow-up questions correctly. We never would have guessed that. Now, our AI can recognize that nod and use it as a positive signal.
We also found that a short, 15-second break with a fun animation can dramatically reduce frustration. It seems obvious in retrospect, but seeing it in the data was a huge breakthrough. It’s these little human touches that make all the difference.
The Hardest Part
The biggest challenge wasn't the code; it was the content. How do you create a lesson that can be remixed and reconfigured in a million different ways? We had to break down every subject into its smallest possible components. We call them "knowledge atoms."
Each knowledge atom is a tiny piece of information – a definition, a formula, a historical date. And each one is tagged with a ton of metadata: difficulty, prerequisites, related concepts, and more. This allows our AI to assemble lessons on the fly, perfectly tailored to the student’s needs.
It was a massive undertaking. We had a team of 20 subject matter experts working for over a year to create our initial content library. But it was worth it. It’s the foundation that our entire system is built on.
This is Just the Beginning
I’m incredibly proud of what we’ve built. But I’m even more excited about what’s next. We’re just scratching the surface of what’s possible with truly adaptive AI.
Imagine an AI that can not only teach you calculus but also help you write a song. Or an AI that can help a CEO practice for a big presentation, giving real-time feedback on their delivery and body language. That’s the future we’re building.
So, the next time you hear someone say that AI is cold and impersonal, tell them they just haven’t met the right AI yet. '''
Frequently Asked Questions
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.
How can I apply this thinking to my own situation?
Start by identifying the core principle behind the opinion, not the specific example. Then ask yourself: does this principle apply to my context? If yes, test it in a small, low-risk way before going all in.