In 2004, Sal Khan was tutoring his cousin Nadia in mathematics over the phone. He started posting short explanatory videos to YouTube so she could review concepts on her own schedule. Within a few years, millions of strangers were watching those same videos. By 2012, Khan Academy had become a nonprofit serving tens of millions of learners, and Sal Khan was giving a TED Talk about “the world’s free virtual school.” None of this was planned. It emerged from a simple observation: a good explanation, once recorded, can be replayed by anyone, anywhere, forever.
That observation sounds obvious in retrospect. It was not obvious at the time. For most of human history, access to quality instruction was a function of geography, wealth, and social class. The best teachers lived in the best cities, taught in the best schools, and served the students whose families could afford them. Everyone else made do. The idea that a kid in rural Bangladesh and a kid in suburban Connecticut might watch the same calculus lesson taught by the same person was, until very recently, not a practical possibility. It was a utopian premise.
What Khan Academy Actually Proved
The lasting significance of Khan Academy is not that it made education free. Libraries did that. It is that it proved the demand. When quality instruction was made freely available, without application process or geographic constraint, people showed up in numbers that surprised everyone — including Khan himself. The implicit assumption embedded in elite educational institutions has always been that rigorous learning is something most people neither want nor can handle. Khan Academy — and later, the MOOCs that followed — shattered that assumption empirically. Hundreds of millions of people, across every income level and every continent, demonstrated that the appetite for learning was never the bottleneck. Access was.
Khan Academy also proved something subtler: that the traditional classroom model conflates two things that do not have to be conflated. A teacher’s time in front of a room has historically served double duty as both instruction delivery and individual support. Video-based learning split those functions apart. A recorded explanation handles delivery. A human teacher, freed from lecturing, can spend their time on the part that actually requires a human — noticing which student is confused, asking the right question, offering the kind of encouragement that a video cannot. This inversion, which educators now call the “flipped classroom,” was not Khan Academy’s explicit goal. It was an emergent consequence of making good instruction infinitely replicable.
Where the Revolution Stalled
For all its reach, Khan Academy left significant problems unsolved. The most fundamental is the problem of personalization. A video explaining polynomial long division is the same video for every viewer. It cannot slow down for a student who is still shaky on basic multiplication. It cannot skip ahead for a student who already understands the concept. It cannot notice that a particular student tends to make sign errors under pressure and quietly adjust the practice problems accordingly. The record-and-broadcast model democratized delivery but could not democratize the tutoring relationship — the one-on-one exchange that research consistently identifies as the most effective form of instruction.
Benjamin Bloom documented this problem in 1984 in what has become one of the most cited findings in educational research. His “two-sigma problem” showed that students receiving one-on-one tutoring from a skilled instructor performed two standard deviations above the mean compared to students in conventional classroom instruction. Two sigma is an enormous effect. It is the difference between an average student performing like a student in the 98th percentile. The implication was stark: we know what works. We simply cannot afford to give it to most people. One-on-one tutoring at scale is economically impossible if it requires a human tutor for every learner.
The second unsolved problem is completion. MOOC platforms launched in the early 2010s with enormous enthusiasm, and then researchers published the completion rates: typically two to ten percent. This was not a reflection of the courses’ quality. It was a reflection of a structural reality: without accountability, feedback, and the sense that someone is watching and cares whether you finish, most adults will not see a course through. Free and available is necessary but not sufficient. Learning also requires a relationship, even if that relationship is with a system rather than a person.
The Next Frontier
AI-powered adaptive curriculum is the first technology that plausibly closes these gaps. A well-designed AI tutoring system can, for the first time, approximate what Bloom documented about one-on-one instruction — adjusting difficulty in real time, identifying gaps before a learner can articulate them, offering hints calibrated to what the learner has already demonstrated they know. The feedback loop that previously required an attentive human can now, in a meaningful number of cases, be approximated by software. This is not a claim that AI tutors are equivalent to great human teachers. It is a claim that AI tutors are far better than no tutor at all, which is the realistic alternative for most learners globally.
The implications are significant. If even a fraction of the two-sigma effect can be reliably delivered by software, the educational impact at scale dwarfs what Khan Academy achieved with static video. We are not talking about making a good explanation available to everyone. We are talking about making an adaptive, responsive, personalized learning experience available to everyone — something that has never existed outside of elite private tutoring and the most well-resourced school systems in the world.
AI also addresses the completion problem, though in a different way. Adaptive systems can detect when a learner is struggling and offer a different path instead of letting frustration harden into abandonment. They can shorten the distance between effort and feedback, which research shows is one of the strongest predictors of persistence. They can surface the most relevant content for a learner’s specific context rather than requiring every learner to sit through the same sequence regardless of what they already know. None of this guarantees completion. But it removes several of the structural reasons why people quit.
The Mission We Are Trying to Continue
AESOP AI Academy exists at this intersection. The subject matter — AI literacy — is among the most consequential skills a person can develop right now, and among the most unevenly distributed. The people who understand how AI systems work, what they are good at, where they fail, and how to use them effectively are acquiring a durable professional advantage. The people who do not are not simply missing a tool. They are increasingly on the wrong side of a decision-making asymmetry that compounds over time.
That asymmetry follows familiar lines. It tracks wealth, geography, and educational background in ways that should make anyone who cares about equity uncomfortable. Access to the best AI tools, the most sophisticated prompting techniques, and the clearest frameworks for working alongside AI systems is concentrated in the same population that has always had privileged access to the best education. This does not have to be permanent. But it will not self-correct. Someone has to build the curriculum, make it freely available, and design it to be genuinely useful to people who are not already experts.
We are trying to do that. The courses here are not designed for students who are already embedded in technology careers. They are designed for students who are curious, who are capable, and who have not yet had anyone explain how these systems actually work in a way that respects both their intelligence and their time. We use AI to build adaptive, rigorous, genuinely educational content. We make that content available without cost or credential requirement. And we do it because we believe the evidence Khan Academy assembled two decades ago still holds: when good instruction is freely available, people show up.
The Work That Remains
There is no shortage of humility required here. Khan Academy set out to build the world’s free school and produced something genuinely extraordinary — and it still could not fully solve personalization or completion. The early MOOC platforms believed scale alone would transform education and learned that accessibility is necessary but not sufficient. Each generation of education technology has overpromised and underdelivered on the timeline, even when it genuinely moved the long-term arc forward.
AI-powered adaptive education is not exempt from this history. The tools are more powerful than anything that came before. They are also new, imperfectly understood, and capable of failing in ways we have not yet fully mapped. The honest position is that we are early in learning how to use these tools well for genuine education — not content delivery, not assessment automation, but actual learning that changes how a person thinks and acts. That is harder than it looks from the outside, and it is the project we are committed to.
Sal Khan started by tutoring his cousin. The goal was just to help one person understand math. The fact that it scaled the way it did was not a product of ambition. It was a product of having the right idea at the right moment and making it freely available. We are trying to hold onto that instinct — start with whether the learner actually understands, build from there, and resist the temptation to confuse volume for impact.