Given Khan’s bold predictions, it’s no surprise that two years later, when his organization’s AI tutor, Khanmigo, wasn’t getting used by students nearly as much as he and his team had expected, it became a major education headline. In the flurry of commentary that followed on my LinkedIn feed, I came across a telling critique from Audrey Watters, one of edtech’s most outspoken critics. Watters saw Khanmigo’s lackluster performance as the latest example of a troubling pattern: that when edtech products fail to deliver, their proponents blame teachers, blame students, and then double down on pushing their solutions onto schools.

Watters makes some fair points about the often taxing burden of edtech on schools, and her comments led me to wonder why that pattern keeps repeating. I’ve come to think the answer is less a story of bad edtech actors and more a flaw in the kind of money behind new products and school models in education. That money is almost always impatient money, and its impatience collides with how slowly real change in education unfolds.1 That collision pressures even good ideas to oversell themselves before they’re ready, and the overselling can do more to set back educational change than the solutions themselves do to advance it.

A mismatch of timelines

Two kinds of funders supply most of the money behind new ventures in education: venture capital, which backs for-profit startups, and philanthropic foundations, which fund nonprofits, new school models, and the research around them. Although their target metrics are different, both are impatient, and that shared impatience becomes a major challenge for the most significant education innovations.

Venture capitalists make their money when the companies they back grow fast enough to produce a big exit, usually an acquisition or a public offering, within the five- to seven-year window a typical fund allows. What they care about is growth. Whether a company changes its industry along the way matters only if that’s what it takes to grow. In education, the fastest way to grow is to sell something that fits the way schools already work: a tool a teacher can add to a lesson, or a product a district can add to its software stack, without changing anything else. That’s why so much venture-backed edtech plugs into conventional classrooms—the pattern I described in my last post—and why any transformation edtech startups pitch tends to be narrow.

Philanthropic foundations look like they ought to be more patient given they aren’t chasing a financial return. But they face their own pressure to show results quickly. A foundation’s program officers and directors often need to show, year by year, what their funding achieved, and many of them will move on from their roles within a few years. They need results they can point to while they’re still in the job and while their grants are still active. So even though a foundation isn’t after a 100x return on investment, it is after its own form of results on a timeline that looks a lot like venture capital’s—often in the form of higher test scores, widespread adoption of a proven product, or a school network that’s scaling. The metrics are different, but the clock is similar.

But here’s the frustrating reality: not every industry transforms at the speed of Silicon Valley. Innovation history is full of tectonic shifts that happened on a much more drawn-out timeline. Transcontinental rail travel was overtaken by commercial air travel, but that change took roughly four decades. The telegraph gave way to the telephone, but that leapfrog took over half a century.

Likewise in education: the limiting factor on transformation isn’t just imperfect technologies; it’s the time required to develop new educational models enabled by those technologies. That kind of redesign puts educational change on a timeline measured in decades, not school years, which is a timeline that neither venture capital nor philanthropic funding is structurally willing to wait for.

Predicting the speed of transformation

My former Christensen Institute colleague, Horace Dediu, dug into cases of disruptive innovations that played out across different time horizons and landed on a valuable insight: the presence of pre-existing modular interfaces determines the speed of disruption.

Hang with me while I explain how this works. The adoption of innovations in any sector tends to roughly follow an S-curve: adoption creeps along slowly at first, then climbs to an inflection point where growth is at its fastest, before leveling off as it approaches the whole addressable market (see Diagram 1).

Dediu found that when a new technology can just plug into an existing ecosystem, it spreads quickly because adoption doesn’t require redesigning the ecosystem to work with it. That means a narrow S-curve. But when there are no such interfaces, when the technology only works if you also design and build the system it lives in, change moves at a crawl, resulting in a much wider S-curve (see Diagram 2).

You can see this play out with familiar innovations. Mobile apps spread quickly because they plug into infrastructure that already exists: app stores, cloud computing services, and smartphones already in billions of pockets. In contrast, the airplane and the telephone were slow precisely because the new technologies needed new surrounding systems before their advantages could be fully realized. Telephones needed thousands of miles of wire and telephone exchanges. Airlines needed airports, radar, radio communication, and air traffic control. Inventing, refining, and building out those surrounding parts of the system took time.

Education is similar to early air travel and telephones. Most new instructional models aren’t something you swap into a conventional classroom with a little training, the way you would a new textbook. They require rethinking the teacher’s role, the use of time, the assessment and feedback systems, approaches to student motivation, and the expectations families bring to school, all at once. That kind of redesign of schooling may also require new teacher development pipelines, new college admissions systems, new curriculum, new post-secondary pathways, and new state policies that allow alternative approaches to schooling, alongside new software designed specifically for helping facilitate the new models. The whole surrounding value network needs to be rebuilt, which is exactly what makes for a long adoption timeline.

Now consider how the funding timeline and the adoption timeline collide. In K–12 schooling, the most reliable way to get the steep adoption S-curve that funders want is to conform to school systems as they exist today. A new model that requires rebuilding the system can’t produce that curve, however promising the model may be. If it needs a decade or more to take off, the window in which its funders need to see scaling revenue or scaling impact closes long before the scaling arrives.

That leaves the people trying to innovate in education with two paths. They can build products that plug into conventional schools and grow quickly, even though, as I argued in my last post, those products are likely to deliver only marginal gains. Or they can build new models that might deliver much more impact in the long run, but that will struggle in their early years to find funding.

The cost of overpromising quick results

There’s an additional problem created by these timeline mismatches. They foster a temptation for innovators to lean too heavily into messaging, marketing, and salesmanship. Companies selling plug-in products are tempted to pitch their point solutions as revolutions. Meanwhile, emerging models with the potential to transform schooling feel compelled to overpromise near-term results. Both forms of overselling corrupt the signals that tell us whether an innovation actually works.

To see how this happens, consider the usual path that moves innovations with long time horizons from the margins to the mainstream. When an innovation requires building a whole new system, it doesn’t arrive looking better than the status quo. Instead, it gets started on the margins, serving either people the mainstream isn’t serving at all or people who are glad to accept something rough and unproven because they value its auxiliary benefits. Early telephones didn’t work over the long distances served by telegraphs, but they let users reach someone across town without sending a messenger and waiting for a reply. Early air travel was more dangerous and expensive than travel by train, but when it came to carrying lightweight mail, speed mattered more than reliability and cost per pound.

So how do these new systems go mainstream? They improve slowly over time, and as they improve, people in tiers of the mainstream start migrating from the old to the new. Because no one is forced to switch, organic migration is the clearest signal that the new system is proving its value.

Overselling breaks that signal, often through hype and sometimes through mandates. Hype manufactures the appearance of demand by making a product look popular before it has proven useful. District mandates push a product onto teachers and students who didn’t choose it, so the licenses districts sign say nothing about whether the product actually works on the front lines of education. In both cases, adoption numbers fail to reflect whether the innovation delivers value.

What’s worse, this overselling actively lengthens the runway it was trying to shorten. When marketing creates the expectation of a narrow S-curve and the narrow S-curve then fails to materialize, what follows is the familiar backlash: disillusioned educators, discouraged funders ditching their strategies, and a tarnished reputation for a whole category of innovations.2 When the field resets, it starts two steps back from where it could have been with patient progress. Slow-but-steady work becomes even harder to fund after a cycle of dashed promises.

The measurable path to slow-and-steady progress

Education needs new funders who are willing to make big bets on a longer time horizon, a case Michael Horn has made in arguing for alternatives to venture capital. But patience by itself isn’t enough, because a longer timeline creates a problem of its own. How do you tell the difference between an innovation that is slowly climbing the early part of an adoption S-curve and one that is just cruising along a flat line (see Diagram 3)? Patient capital that can’t tell the difference might let a bad idea devour funding for years with nothing to show for it in the end.3

New innovations need intermediate wins along the way to long-run transformation: markets where they can produce real evidence that they’re climbing. For airplanes, the journey to long-distance passenger air travel climbed a ladder of progress whose rungs included airmail, express freight, crop dusting, aerial photography, land surveying, bush flights to remote Alaska, air shows, chartered flights for executives, and military combat. These applications gave planes opportunities to prove their value and continue improving long before they were ready for long-distance passenger travel.

Likewise, technologies and models that could eventually transform education may need to run through applications such as summer camps, afterschool programs, exam prep, private tutoring, on-the-job training, learning for hobbyists, professional skills certification, novel elective classes, homeschooling, and microschools before they transform whole-day, mainstream schooling.

Conclusion

Smart, patient capital doesn’t write checks and then wait for miracles. It looks for ladders where each rung gives an early innovation the opportunity to win on its own terms. Genuine adoption in early, small markets is the signal that an innovation could have a route to bigger impact. And even if the climb stops short of transforming the whole sector, every rung it reached still delivered real value to real learners. Moving up an improvement ladder is also the way out of the pattern Watters named. Early results get delivered in markets that want them, rather than being oversold and underdelivered in the heart of public education. That kind of progress may seem frustratingly slow to people who are impatient for radical transformation. But in the end, it may be the only real route to substantive change.

Preview image: clock movement from Diderot and d’Alembert’s Encyclopédie (Paris, 1760s–70s), drawn by L.-J. Goussier and engraved by R. Bénard. Public domain, via Wellcome Collection.

Then in August, a newly released randomized controlled trial delivered an even stronger verdict: for the students in the study, the AI tutor layered on top of Khan Academy’s existing practice platform added little beyond what the platform already did on its own. As the researchers put it, “access was nearly universal but engagement was thin.”

Khan Academy responded by working to improve the product. After piloting changes with select districts during the school year, it began rolling out a rebuilt Khanmigo to all district partners this summer. The new version is proactive rather than passive, so that the tutor offers help in the moment students need it instead of waiting to be asked. Khan Academy also worked with a team of six engineers from Google to add interactive diagrams that Khanmigo generates with Google’s Gemini models when it detects that a visual could help a student with a math or science problem.

In response to the news about Khanmigo’s low usage, edtech skeptics jumped to point out that this was entirely predictable. One prime example was Dan Meyer’s pointed “obituary” of Khanmigo, which noted that the AI tutor had enjoyed extraordinary advantages such as early access to OpenAI’s models, endorsements from national, state, and local officials, connections to some of the wealthiest people in the world, and Microsoft backing, yet it still fell far short of the expectations its creator, Sal Khan, had set for it. If Khanmigo couldn’t make it with all of that going for it, Meyer asked, “what hope then should the rest of the edtech industry place in chatbot tutors?”

It’s a fair question, but I think skeptics are drawing the wrong conclusion from that failure. Yet so are the optimists who look at AI capabilities and expect an impending revolution in education.

How well AI tutors can improve student learning depends on more than just the abilities or limitations of AI. Looking at this problem through the lens of disruptive innovation theory leads me to predict that AI tutors will only deliver marginal gains at best when plugged into conventional classrooms. Dramatic learning gains will require redesigning classrooms and schools with AI as one key piece in new learning models that rethink both instruction and student motivation.

The oversimplification of teaching

Around the same time the Khanmigo commentary was circulating, someone in my family forwarded me a Washington Post article that ranks professions by their vulnerability to AI displacement. A graphic at the top of the article lets you search for different professions, and according to the graphic, most teaching roles were rated as highly vulnerable. The person sharing it thought I’d be excited to see this new evidence that AI was about to disrupt education. But I pushed back, frustrated by the article’s seemingly narrow and naive view of teaching.

Washington Post graphic of professions ranked by vulnerability to AI displacement (washingtonpost.com)

At face value, the job of a teacher can seem pretty simple to someone outside of education: explain facts and concepts at the front of a classroom, answer questions, hand out assignments and tests, then grade students’ work. If that’s your notion of teaching, then AI seems to be on the cusp of transforming education. It can explain concepts more clearly than most teachers, adapt to individual students’ misconceptions in real time, provide instant feedback, and never get tired or frustrated. If teaching is just information transfer done well, of course AI is going to outdo humans any day now.

But that framing is blind to some of the major factors that actually determine student learning. One such factor is motivation. As Khan himself observed regarding the study released in August, “What this [new research] has put in huge spotlight—not that it’s a shock to educators—is how much motivation matters.”

Real learning takes effort, and students don’t put in that effort unless they are motivated to learn. But motivation doesn’t come just from making lessons fun or interesting. It’s also deeply tied to relationships, belonging, and status within a community. The developmental psychologist David Yeager has shown that the goal that drives much of adolescent behavior is the quest to earn status and respect from others. A young person’s fulfillment of that goal is deeply intertwined with the culture of their school, the culture of their classroom, and their relationships with the adults and peers around them.

Think about what motivates a student who earns straight A’s. It’s not some intrinsic beauty of a letter written on a piece of paper or the satisfaction of earning points in an arbitrary point system. It’s what that grade signals—to parents, to peers, to coaches, to the student’s own sense of who they are and how others see them. The grade means something because it operates inside a community and a set of relationships where it carries social weight. Remove that community and the relational context, and the grade means almost nothing.

A better AI tutor can explain fractions more clearly. It cannot make a student care about fractions. And caring—being embedded in a community and a set of relationships that make academic effort feel meaningful—turns out to be most of the problem.

Plug-ins versus new models

Adding an AI tutor like Khanmigo to a conventional classroom may improve learning at the margins, especially as Khanmigo’s developers continue to refine their product. But I think the bigger gains from AI tutors will likely come from new models of schooling and instruction designed from the ground up to both leverage AI and enhance student motivation.

Our dominant approaches to education were developed long before AI, and their structures (the schedule, the grouping of students, the teacher’s role, what counts as success) were never designed with today’s technologies in mind. Those same structures shape how instruction is delivered and what motivates students to engage with it. An AI tool added to a conventional classroom can improve the instruction, but the motivational structures around it stay as they were. In contrast, a new model of schooling can treat instruction and motivation as one design problem, with AI as one piece of the design.

The history of factories offers a useful illustration. In the late 19th century, factories were optimized for distribution of mechanical power. At the center of the factory was a large steam engine that drove all the individual pieces of the machinery through a branching system of drive shafts, gears, and belts. The friction inherent in this system meant that the power declined rapidly as distance from the steam engine increased. Therefore, the machines that required the most power were placed closest to the steam engine in the middle of the factory.

A factory in Schaffhausen, Germany, with machinery powered by drive shafts and belts (theoildrum.com)

As electrification swept the industrial world in the early 20th century, some factory managers started replacing their large steam engines with large electric motors. But the new technology didn’t meaningfully improve factory output. Electric motors were less noisy and didn’t produce smoke. But they also broke down more often than their steam-powered predecessors.

The transformation came roughly three decades later from factories that were designed from the ground up around what electricity made possible. Instead of clustering equipment around a large central motor, they began putting smaller motors in individual pieces of equipment and then organizing the whole factory based on the flow of production. At first, electric-powered equipment was not as reliable as the old system. But as the new technology dissolved the constraints imposed by the old system, factory productivity grew to two to three times what it had been previously.

Khan Academy’s response resembles what those early factory managers did. Making the tutor proactive is a reasonable fix for the engagement problem the research identified, and it may well increase how much students learn from the platform. But it is a better motor bolted to the same factory floor: Khanmigo remains a tutoring layer added on top of the conventional structure of a class, with the same schedule, the same grouping of students, and the same relationship between teacher, student, and content. A tutor that prompts students can make help easier to reach, but whether students want to engage still depends on the relationships and incentives in the structure around it.

Where to look instead

The debate between edtech optimists and edtech skeptics tends to focus on whether technologies like AI tutors can actually improve learning. Implicit in most versions of that debate is an assumption that AI tutors will succeed or fail based on how they improve classrooms as they exist today. The optimists are watching the technology and assuming the transformation will come from the technology itself. The skeptics are watching the technology fail to transform student outcomes and concluding that the transformation isn’t coming. Neither is paying much attention to the places where people are actually redesigning the model.

Those places exist, though they don’t look impressive by conventional school metrics, which is what you’d expect of early-stage models. They include some versions of microschools, hybrid learning models, competency-based programs operating outside the traditional school calendar, and learning communities built around entirely different assumptions about what school is for. These are the early attempts at designing factories around what electricity actually makes possible, rather than the old factories that simply swapped out the steam engine. Whether any of them find good ways to use AI as one piece of a design that addresses instruction and motivation together is the open question to watch.

I don’t claim to know which particular models on the margins will prove successful. My hypothesis is that any dramatic learning gains from AI will emerge from new models of schooling at the margins of K–12 education, not in conventional classrooms. So if you’re looking to conventional schools to see whether the transformation is happening, you’ll likely miss the story as it unfolds.

As a note, we drafted these over the summer before the renewed wave of reporting on Alpha School began. The purpose of these posts isn’t to justify or criticize Alpha or dive into specific details. The purpose is instead to use sound, predictive theory to understand the macro decisions those connected to Alpha are making and what they may portend.

Is Alpha “disruptive”?

There are a lot of hot takes to that question right now.

Its boosters say yes—and can’t wait for Alpha to transform education.

Its skeptics say sure—if disruptive means that it’s going to cause a whole lot of trouble.

Others call it disruptive meaning something closer to what the word has come to signal in startup marketing decks: that Alpha is new, brash, and shaking things up.

What nearly all of these answers share is a loose grip on the theory of disruptive innovation itself. When you use disruption in the context of the actual theory, rather than as applause or as an insult or as a synonym for “novel,” the interesting thing about Alpha is not whether the school is disruptive. The school is likely not a disruptive innovation. But there is an element inside the school that could enable disruptive innovation.

Are Alpha’s schools disruptive?

As a school, Alpha does not fit the disruptive innovation pattern.

Disruptions, in Clayton Christensen’s use of the word, start at the bottom of a market or outside it entirely, with a cheaper, simpler, more accessible, and generally less impressive offering that established players are happy to ignore. Then, through the use of a technology enabler, it climbs while retaining its simpler, more affordable, more accessible value proposition.

Alpha is mostly doing the opposite. It is expensive and positioned at the high end. A $65,000 school that markets prestige and selective-college outcomes is, by design, not sneaking in at the bottom of the market. Even though it eschews the elaborate facilities and amenities of most top private schools, by dint of its price point, it’s largely competing head-to-head with them. It isn’t starting at a disruptive foothold those schools choose to ignore.

One could argue that Alpha will be focused on opening more campuses to serve more students rather than claiming its value comes from exclusivity like elite private schools. Yet it’s unclear that its business and operating model can easily become more accessible. Its highly paid learning guides—the adults who motivate students and curate and facilitate afternoon projects—and low student-to-adult ratios seem core to its value proposition.

Yes, there is a possibility that Alpha will successfully follow Uber’s playbook, which effectively used Uber Black cars as a marketing device to capture attention and trust, generate high-margin revenue to fund early development, and then quickly launch lower-priced offerings that disrupted the supply-constrained taxi market. Because the value of a school is complex and opaque—and hence that value often gets judged by the characteristics of the students the school serves, not the value it adds—there’s rationality in launching at the high end to establish a brand as a “good” school. It’s then possible that as Alpha’s technological core improves and matures, it could allow Alpha to do something similar to Uber. Alpha’s plan to launch a low-cost, fully virtual version of its software, for example, suggests it may consider that strategy.

It’s more likely, however, that the release of its software signals something else. Alpha’s technology enabler—its Timeback platform that oversees its morning 2-hour learning cycle—may be core to a more plausible story of potential disruptive innovation that doesn’t involve its own school, but instead unleashes a wave of low-cost private schools that together could prove disruptive. Taking a detour through a question that has haunted edtech for two decades—why has technology promised so much and delivered so little—will help us understand the potential and some of the possible limitations for this strategy.

Why the last wave of edtech disappointed

A decade ago, a wave of adaptive learning software arrived with claims that look a lot like Alpha’s. Programs that met students at their level, advanced them through content at their own pace, and personalized practice were going to transform achievement. Districts bought in, the transformation mostly did not arrive, and the backlash that followed hardened into a skepticism that still colors how many educators hear any new technology claim.

We think there is a more precise and useful way to make sense of the failure. Illustrative of the lack of results is what Laurence Holt, a Senior Advisor at XQ Institute and author of The Science of Tutoring, has documented: a pattern he calls the five percent problem. Studies of certain edtech programs find that they can produce real learning gains, but only when students use them as intended. The kicker? In these studies of edtech usage in schools, only about five percent of the students ever reach that threshold of faithful use. As a result, when a study finds weak average effects across schools that bought a program with evidence of producing learning gains, it is not really measuring whether the technology can work if used. It is measuring how few students in fact used it to its recommended dosage.

There are of course many other reasons why edtech has failed—faulty incentives and procurement challenges that result in edtech with lousy instructional design getting adopted; a focus on pacing guides over mastery; the distractions that result from introducing edtech; and more. But they all center around the challenge outlined in Holt’s research. The failure of the last edtech wave was not a failure of software per se. It was a failure to redesign the surrounding system of the schools themselves to effectively take advantage of the technology.

An analogy from aviation captures why redesign matters. When jet engines arrived, simply replacing a propeller engine with a jet engine did not automatically transform the capabilities of an airplane. The new technology created possibilities that conventional airframes were not designed to exploit. Realizing those possibilities required changes to the aircraft itself: pressurized cabins, stronger fuselages, swept wings, and other features suited to higher speeds and altitudes. The engine was the prominent innovation, but much of its value came from redesigning the aircraft in conjunction with what the new technology made possible.

Likewise, the last wave of schools bolted new instructional technologies onto an educational model built for a different era and then wondered why student learning didn’t take off. They adopted the engine without substantially redesigning the airframe along with it.

Schools treated the software as something you could plug into the existing architecture of a classroom the way you would add a new app to a phone. Drop the program in, assign students to it a few times a week, and expect results. But the technology and the classroom around it were interdependent. The software’s effectiveness depended on how class time was structured; on what the teacher did while students were on the computer; on how student effort was motivated; on whether the culture of the room treated the program as central or as a babysitter for the back half of the period; on curriculum pacing, assessments, and grading; and so forth. Plugging a genuinely interdependent technology into an architecture that was never redesigned to receive it is a recipe for exactly the mediocre results the last decade produced. That was the logic behind our work over a decade ago detailing blended-learning models: that the model itself had to change for technology to be of any use in bolstering student outcomes.

What Alpha built instead

A second aircraft example parallels Alpha’s strategy. When Lockheed’s Skunk Works built the SR-71 Blackbird, the fastest air-breathing aircraft ever flown, they faced a budget problem on top of an engineering one. The plane was already extraordinarily expensive. Designing a bespoke engine from scratch would have made it more expensive still. So they did not. They took existing Pratt and Whitney engines, technology that was already good enough at the raw job of producing thrust, and they poured their genius into everything around the engine, most famously the movable inlet cones that tamed airflow at three times the speed of sound. The integration in other words was the innovation, not the engines. The choice to buy rather than build it was as much about up-front cost as about engineering.

Alpha seems to have followed a similar playbook. Its two-hour learning block originally contained the equivalent of off-the-shelf engines. The underlying instructional software products were not, for the most part, proprietary AI-powered breakthroughs in how content gets taught. They were existing adaptive-learning tools, the same general class of software that the last edtech wave deployed to such underwhelming effects. Custom-building the underlying learning software from the ground up would have been expensive and slow. Fortunately, the existing off-the-shelf software allowed Alpha to make the SR-71’s choice: take software that is already good enough at the raw job of delivering instruction and practice and spend money instead on the custom layer wrapped around it, the part that makes off-the-shelf engines perform like something far more advanced.

Alpha, however, went farther than most schools by redesigning the “airframe”—the rest of the systems around the software. The school rebuilt the role of the adult by stripping out teacher-led instruction and pouring its guide’s entire energy into motivation and coaching. It rebuilt the structure of motivation, with the bargain of students getting their time back for fun activities in the afternoon if they focus and do their morning work. It also created a currency system that we described in our earlier piece for rewarding students for buckling down and making progress on software-based learning. It rebuilt the culture of effort, so that focused work on the software became the organizing expectation of the morning rather than one activity among many. It built a proprietary AI-powered platform, Timeback, that sits on top of adaptive learning applications to monitor students during their two-hour learning blocks and provide feedback to both students and guides on students’ engagement and efficiency. These choices are a coherent answer to the exact problem that sank the last edtech wave: Alpha redesigned the “airframe” to fit the “engine.”

Alpha’s new challenge

For the families Alpha serves currently, based on parent testimonials and the school’s self-reported outcomes, it seems that arrangement has, on balance, worked.

But according to Alpha’s leaders and a variety of other reports, Alpha is increasingly building its own digital curricular products. Why might it be spending lots of time and money to do so?

Our theory-informed view is that as Alpha has sought to expand—both in the footprint of its own schools and through offering Timeback to other schools—it’s likely that its original software and curricular architecture is hitting a wall in just how much better it can get. And it must get better, as it seeks to serve students who likely have more-demanding needs than its original set and schools outside of the direct control of Alpha itself. In the language of disruptive innovation, Timeback must go up-market to take on more complicated problems.

Why might Alpha’s software and curricular architecture have hit a wall in terms of just how much it can improve? Our suspicion is that it was the lingering unreliable interdependencies between the off-the-shelf software and Timeback.

The outputs of third-party learning software are noisy. Those products were built to deliver their own instruction and report on their own terms, not to hand Timeback a granular picture of what students know and where their misconceptions lie. Alpha can see that a student finished a module in an outside program. It has a much harder time seeing why the student struggled, or what to serve up next.

The assessment layer compounds the problem. Alpha did not build its own assessments. It uses off-the-shelf instruments and pulls their results into Timeback to help build each student’s plan, which is more than most edtech platforms have ever managed. But an assessment instrument administered a few times a year is a coarse tool for a system that wants to make daily decisions. It tells you where a student ended up. It says very little about what happened in between.

And as Alpha seeks to serve students in more-demanding circumstances, that visibility and feedback into student performance is something to improve upon. Another theory from Christensen’s work—the Theory of Interdependence and Modularity—helps us understand why Alpha is gradually moving away from third-party edtech curricular products.

When an innovation is new, its developers often rely on existing products for other parts of the system that surround the innovation, because in the early stage you can’t afford to custom-build everything. Combining a new innovation with off-the-shelf components is fine when you’re serving a context where the alternative for your customers is either something at odds with what they really want or nothing at all. But as a company pushes into more demanding applications, it discovers that its product isn’t yet good enough for what those customers expect.

According to the Theory of Interdependence and Modularity, getting good enough means the parts of the system have to be designed together, tuned to each other, and optimized as a whole. The pieces become interdependent, and you can’t swap one out without rethinking the others. Only later, once performance is more than good enough and the interfaces between components are well understood, can a system become modular again, with clean standards that let anyone mix and match parts from different suppliers without the whole thing falling apart. The dominant product architecture in an industry tends to shift back to a modular state once consumers’ needs are overserved and they are looking for more affordable customization.

The Ford Motor Company evolved in that sequence. Early in the Model T era, Ford bought steel and many components from established suppliers. But as it pushed to scale and streamline production, dependence on outside companies made it harder to control quality, costs, and the flow of materials. The inputs and final output were just too interdependent without clear and definable modular interfaces. Ford therefore brought more and more production in-house, which culminated in building the enormous River Rouge complex. Here Ford made its own steel, glass, castings, and many other components. Later—beginning around the transition from the Model T to newer models—the company moved to rely increasingly on independent suppliers. The reason? It needed to focus increasingly on scale and cost control, which outweighed the need to control the interaction of every single component in a car, which had become stable and good enough. With the basic car architecture stable and good enough, Ford could clearly define what it needed from outside providers and ensure that they met its standards. From the perspective of consumers, they were increasingly looking for customization. Any car so long as it was black—the early motto of the Model T—was falling out of favor.

Alpha is in the beginning stages of a similar progression—from off-the-shelf components to creating nearly all of its own digital curriculum in a highly interdependent architecture with Timeback to push the bounds of its software and serve students in more-demanding circumstances. To keep pushing the frontier of performance and reliability, Alpha has to own the interfaces it was previously borrowing. This is expensive, but if Alpha hopes to scale Timeback, likely necessary.

The bet beneath the school

This context brings us back to the argument around whether Alpha is disruptive.

Let’s start with two of Alpha’s choices that we and many other educators find hard to defend on pedagogical grounds. Guides who know their students well are forbidden from stepping in to teach a concept when a child is stuck, and the morning’s academics are walled off so completely from the afternoon’s projects that neither informs the other. In our earlier companion piece, we treated these decisions as real tradeoffs, which they are. But understanding the above analysis helps us understand why Alpha made these tradeoffs.

These weren’t pedagogical decisions viewed through the eyes of serving one student or optimizing the transfer of skills, but those of an organization seeking to dramatically improve the pedagogy of its core technology. Both the choice to not allow a guide to instruct and to not connect core academics to the afternoon’s projects allow Alpha to create a cleaner feedback loop in the interface between the student and the digital curriculum.

If guides taught, Alpha could no longer tell whether a student’s gain came from Timeback or from the adult who intervened. If the afternoon’s messy, social, hard-to-measure learning bled into the morning’s metrics, the signal would further blur. Holding everything but the software constant helps understand what the software itself is actually doing—which allows Alpha to better improve it over time.

Building and tuning that custom layer is enormously expensive, and the conditions Alpha enjoys are favorable: families who likely reinforce learning (intentionally or not!) at home plus a wealthy CEO and expensive tuition to fund the engineering. The school, with its high price and its advantaged families, is perhaps a mechanism for paying to develop the technology under ideal conditions. The students are, in the language of disruptive innovation, relatively undemanding customers in terms of raw academic need, which is exactly the population against which you would want to refine a still-maturing product before exposing it to harder cases. Every cohort of high-paying families is, among other things, a round of funded research and development for the Timeback platform underneath.

This helps understand why the platform, and not the school, is where the most credible case for a disruptive strategy lies.

As we’ve discussed, Alpha has already been playing with offering its Timeback platform in other settings—as a standalone virtual platform and for other schools. If Timeback can be made to be “good enough,” then it could become a robust product offering for schools that adopt certain principles from Alpha’s model. And these schools could cost significantly less than Alpha’s $65,000 flagship. That is the move that would matter. A technology enabler—Timeback—that lets a low-cost school deliver strong mastery learning without having to independently solve the integration problem from scratch is the kind of thing that genuinely could climb the market the way real disruptions do.

With that said, we see at least two things that could keep Alpha’s potentially disruptive strategy for Timeback from working out.

The first is hype. Alpha’s leadership team has made expansive public claims, in the manner of a Silicon Valley launch, and the gap between a polished demonstration under ideal conditions and a product that works in the hands of strangers is where a great many education technologies have died. Selling the promise before the product is ready is its own kind of risk, especially in a field as allergic to overpromising as education has become.

The second doubt runs deeper. It concerns whether the integration Alpha has achieved can survive being modularized. Today Alpha controls everything: the students, the families, the guides, the use of time, the culture, and the conditions. That total control is what makes the clean feedback loop possible. But the whole disruptive promise depends on eventually pulling the platform out of that controlled environment and plugging it into other schools with other students, other cultures, and far less favorable conditions. The question is whether the integrated software that Alpha builds will be able to be a modular component within someone else’s school and can perform when transplanted—or whether its performance is inseparable from the more integrated schools in which it started.

To be clear, it’s possible and advisable that Alpha put some non-negotiables around the schools that adopt Timeback: mastery-based learning, a clear motivational model, and the like. But Alpha’s results may owe less to the platform than to two other things that travel poorly. The first is the pieces of the instructional model that live in a school culture, not in the technology. The second is admissions and the families themselves, whose social and educational capital quietly fills gaps the software never had to address. If Alpha’s success depends more on these factors than on a finely tuned and scalable software platform, then the technology sold to a school down-market could arrive without Alpha’s most important ingredients.

According to one early case study, the results on paper at least don’t look promising. Unbound Academy opened in Arizona in the fall of 2025 as a public charter school built on the two-hour learning model, with Alpha co-founder MacKenzie Price as its board president. Unbound is free. By law it cannot screen applicants. In its charter application it promised the state 60 percent proficiency in math against a baseline of 34 percent, and 65 percent in reading against a baseline of 40 percent. Its first year came in at 10 percent in math and 28 percent in reading, below the Arizona state averages in both subjects and below the baseline it had projected for its own students.

There are already some questions about these results—students may have entered even lower than what the school projected, for example. And just one year of results from one school is certainly not a final verdict, as Unbound could post better numbers as it finds its footing. First years are hard everywhere. But the distance between what the projection and what Unbound delivered is wide enough to take seriously. They also fit within the precise set of concerns we’ve raised. It is also, notably, the same conclusion that Alpha appeared to reach last spring when it pulled back on rolling out Timeback to many potential partner schools.

What we are actually watching

This story doesn’t have an ending yet. The fair reading of Alpha, through the lens of disruption, belongs to neither the booster nor the skeptic. The school itself is unlikely to be a disruptive innovation.

But the school may be a clever instrument for funding the development of a wave of education entrepreneurs that could be—and the school’s leadership could be bringing the requisite marketing attention to it. The design choices that read as pedagogical flaws through one lens look different when you see them as the disciplined behavior of a company refining a platform.

Alpha’s decision here is worth appreciating. The usual way to build educational technology is to sell into schools you do not control and then try to improve the product from a distance, while every partner school runs a different model, staffs it differently, and holds a different theory of what the software is for. That arrangement produces exactly the muddy signal that doomed the last edtech wave.

Alpha instead built the schools, hired the adults, set the culture, and ran the platform inside conditions it could hold steady. Owning the whole system is expensive, and it limits what a company can learn about contexts unlike its own. But it is a far better way to find out whether a technology works than negotiating with two dozen partners who each want something different.

There is one more thing sitting in the decision to build its own school and software that cuts against the most common criticism of Alpha. The usual objection is that a school costing $65,000 a year serves only families whose children were going to be fine anyway, and that whatever Alpha proves cannot be claimed as a contribution to anyone else’s kids. That objection has real force when it comes to who benefits today. But it misses something about who is absorbing the risk. Education has a long and unhappy record of trying unproven models out on the students with the least ability to walk away, financed by public dollars that could have gone to something that already worked. When those experiments fail, the children who bore the cost had no say in the matter and little recourse afterward.

That is not what is happening here. The money financing this bet is Liemandt’s own, and the tuition sustaining it comes from families who looked at an unproven model, understood that it was unproven, and chose it anyway. If two-hour learning turns out to be less than advertised, the people who paid for that discovery are the ones who volunteered to fund it, and the rest of us will have learned something at no cost to our own students. To be clear, that calculus changes the moment a model like this moves into public education, which is exactly why Unbound deserves the scrutiny it is getting. But for the phase Alpha is in now, having a wealthy CEO and willing families underwrite the experiment is not a mark against the project. It is the most defensible way to run one.

Whether Timeback ever escapes the test range, performs in ordinary conditions, and reaches the students who would benefit most is the thing actually worth watching over the next several years. That is a far narrower and more answerable question than the noisy argument now raging about whether Alpha is the future of school or a mirage for the rich. It is also, we suspect, the question Alpha’s own leaders care about most, whatever their marketing says. The school is the part everyone is arguing about. The platform is the part that, for better or worse, might help shift the broader education landscape.

As a note, we drafted these over the summer before the renewed wave of reporting on Alpha School began. The purpose of these posts isn’t to justify or criticize Alpha or dive into specific details. The purpose is instead to use sound, predictive theory to understand the macro decisions those connected to Alpha are making and what they may portend.

Alpha School, the startup microschool network known for using AI to manage students’ academic content in just two hours a day, has been grabbing the limelight in education over the last year. But with newfound prominence has come a growing range of criticisms.

One line of critique says the model looks impressive only because it serves the children of wealthy, highly educated families; its results would never replicate for low-income students who lack the resources and support that Alpha’s families take for granted. Another, for example, says that whatever Alpha does for test scores, it shortchanges the deeper work of education—the building of genuine understanding of core content and its connection to skills—and produces students who can perform on an assessment without having been taught to think.

These angles of critique deserve serious consideration, particularly as Alpha seeks to expand its model. But the criticisms also share an assumption: that a good school is one that serves everyone well and attends to every part of a child’s education.

Judged by that standard, Alpha fails because Alpha has chosen to make tradeoffs—to do a few things well and to neglect other elements.

One question worth asking is whether that standard—that a good school is one that serves everyone well—is the right one to bring to a school like Alpha, or, frankly, any school, in the first place.

The instinct to hope that all schools can serve all students equally well reminds us of the episode of The Simpsons in which Homer’s long-lost half-brother, a car executive, hands Homer the keys to a design studio and tells him to build a car for the average American. Homer, being the everyman, loads it with every feature he can imagine. The result is a bloated, incoherent, wildly expensive machine that bankrupts the company. The joke lands because no one is actually average along every dimension at once. Design for everyone and you build something for no one.

Different underlying logics

Public education has a structural reason to keep building Homer’s car. A district school is accountable to every student and their family that enrolls, to taxpayers, to school boards, to state legislatures, and to a long list of groups each with a legitimate claim on what schools owe children. One way public education progresses is through policy—and policy is the art of compromise among stakeholders who will never fully agree.

As a result, public education has defaulted, almost unconsciously, to a standard in which schools are expected to serve and honor the values of all students whose families have chosen and been able to afford to live within the boundaries of the district. That standard, even though it can produce schools that contain elements of incoherence and fail to be excellent in some respects, also has an understandable rationale for an institution that belongs to the public.

The trouble comes when we apply it, without noticing we are doing it, to a school that was never built on those terms.

Entrepreneurs, like the founders of Alpha, operate on the opposite default.

A startup does not begin by trying to serve everyone. It begins by narrowly defining one problem for one target customer, building the simplest product that solves it, and getting that product into the hands of the customers who feel the problem most acutely. It expects to be relatively bad, at first, at everything else. Only once it has won that first narrow market does it earn the capital, the credibility, and the operational room to expand into adjacent problems.

Amazon sold books before it sold everything, for example. The early narrowness wasn’t a flaw for which to apologize but the strategy itself. Even today, Amazon is clear about the principles around which it organizes and the tradeoffs it makes. That’s why Amazon won’t provide you with an in-person, personal shopper anytime soon.

These two implicit assumptions about public education and entrepreneurs produce two completely different reactions to a school like Alpha.

Through the public-education lens, a school that does some things well and ignores others is a school with holes in it.

Through the entrepreneurial lens, a school that does a few things well and is candid about the rest is a company executing a sound early-stage strategy.

To be fair, there are educators inside districts quietly intrigued by Alpha. They see students moving at their own pace, reaching mastery before advancing, and coming home excited about school, and they wonder what it would take to get some of that in their own buildings.

Nonetheless, we get the wariness voiced by many educators. They have good reason to fear the next overhyped model that blows up in their faces. The field is littered with fads and new companies that proved too good to be true—including AltSchool’s well-funded flameout and Byju’s slide into allegations of fraud.

But Alpha is different in at least one important respect from a microschool like AltSchool. It has a clear and specific school model that has been developed over more than a decade. Its technology is purpose-built to serve that model rather than the other way around. And, despite the hype from its leaders, Alpha’s full school model was never designed to serve every student. So, in our parlance, what Job to Be Done is Alpha built to do, and what does it give up in order to do that Job well?

What is Alpha’s model?

Alpha’s model is, by now, fairly well known. After a school-wide morning rally, students spend the morning on a computer doing core academics and knowledge acquisition—the much-publicized “two-hour learning” block. Students work in focused bursts of around 25 minutes as they move through instruction, practice, and assessment inside Timeback, an AI-powered platform that stitches together a range of digital curricula and tracks each student’s pace, focus, and progress. Because the software meets each student at their level, leans on direct instruction, keeps them active, uses spaced repetition, and refuses to advance them until they have demonstrated mastery above a 90-percent threshold, it compresses into two hours some of what a conventional classroom stretches across a full day.

Those design choices deserve more credit than they usually get. Mastery-based progression, spaced repetition, and careful attention to cognitive load are among the better-supported findings in learning science, and conventional schools mostly do not act on them. A typical classroom moves a whole group through material that follows a fixed pacing guide regardless of who has actually learned it, circles back to content when the unit plan calls for it rather than when forgetting curves suggest, and routinely asks novice learners to hold more in working memory than they can manage.

The educators in Alpha schools are not called teachers, and they do not instruct. Alpha calls them guides. Their job is to meet with students, get to know them, track their progress, and, above all, motivate them. We will tackle the decision to deliberate strip instruction out of their role in our second post, because it turns out to be one of the more revealing tradeoffs Alpha has made.

The compression of learning is a central part of Alpha’s model, because it buys back the resource Alpha treats as most precious: the student’s time. What students do with the reclaimed afternoon depends on which flavor of Alpha they attend, but the throughline in most of the models is the cultivation of life skills that rarely show up on a transcript: teamwork and leadership, public speaking and storytelling, entrepreneurship and financial literacy, grit and the capacity to take honest feedback. In the flagship model, younger students take on projects that range from Tough Mudders to press conferences, while older students pursue a project of genuine personal interest. In the gifted-and-talented model, students simply dig deeper into academics. In the sports academies, they train. And in the entrepreneurial high school, they work to build a business that earns a million dollars before they graduate.

The deal Alpha strikes with students

Driving Alpha’s learning model is a focus on motivation, a flavor of which most educators would call, with some discomfort, extrinsic.

Alpha’s implicit premise is that core academics are not always intrinsically thrilling. Instead they are often like the broccoli you eat because it’s good for you. The school essentially says to stop pretending it’s otherwise and instead strike a deal with students: do the focused academic work in the morning, and the afternoon, along with your time, is yours. Alpha sharpens the deal further with its own currency, sometimes called Alpha bucks, that students earn by meeting goals and can spend on things they want from a school store.

There is a counter argument of course. A good deal of evidence suggests that academics can be intrinsically motivating when knowledge acquisition helps answer questions students find curious, when it’s woven into real-world projects, and when students progress upon mastery and tackle concepts in their zone of proximal development.

But Alpha largely sets that possibility aside, in part because anything resembling inquiry-based learning carries, by Joe Liemandt’s own account, a steep efficiency cost relative to direct instruction, and efficiency is the entire premise of buying back a student’s time. In fairness, there is real evidence behind the efficiency claim, even if we suspect the underlying picture is more situational than Alpha’s framing allows.

(Here’s the aside to our suspicion that you’re welcome to skip: There’s some truth to Alpha’s point, but it also misses a lot of underlying nuance that, in our view, suggests that these are not mutually exclusive categories and that a lot of the answer is circumstance-dependent on the learner’s prior level of understanding and the goals for the learning. Among the nuance: direct instruction is especially important for novice learners, but even they can benefit by making a lesson intrinsically more interesting by starting with a simple but perplexing question or problem and then drawing the student into the learning for direct instruction comprised of a teacher-led lesson, a guided example, and then independent practice; that limited version of “inquiry” can actually make learning more efficient; expert learners tend to benefit from more inquiry as they seek to apply their knowledge and build skills with the acquired knowledge—meaning the goals of the learning are different; a project can be part of that cycle, as it starts with a captivating question or problem or seeming paradox, moves to direct instruction, and, once a learner is no longer a novice, they can tackle the more complex project; and so on. In other words, this is almost certainly a false dichotomy. You can read more about this here and here, for example. Interestingly enough, Alpha’s leaders have stated on numerous occasions the value of being able to offer lessons modified by AI to be in areas of interest for students—meaning you could learn physics through examples based in baseball or dancing, for example. Cognitive scientist Daniel Willingham has written both about the opportunities—and potential perils—of such an approach, but either way, Alpha would presumably be trading some efficiency for some intrinsic motivation if they pursued this).

But set the pedagogy debate aside for a moment, because the more interesting fact is that Alpha has made a definite bet about what moves children to learn and committed to it without apology.

That bet is not as foreign to our own thinking as it might first appear. One of the central arguments of Disrupting Class, written by one of us nearly 20 years ago, was that the erosion of extrinsic motivation in a society tends to drag educational performance down with it, and that the remedy was to make learning more intrinsically compelling. Alpha is proposing something different but based on a similar diagnosis of the underlying problem: that a school can manufacture extrinsic motivation, build its entire model around it, and get results.

What Job do parents hire Alpha to do?

Here is the uncomfortable fact for the skeptics.

A set of parents are not merely curious about Alpha; some are relocating their families across the country and paying as much as $65,000 a year to enroll their children in it.

To understand why Alpha’s parents find it so desirable, it helps to ask what Job parents are actually hiring Alpha to do.

Our research on why families switch schools points to a possible answer. One group that enrolls in Alpha is likely doing so because their child’s academics are already strong enough; they want to create space for their children to explore and not miss out on areas of passion and potential and develop other facets of their character. A second group is likely seeking to escape from a school setting that was failing their kid, with some of those students bored or falling far behind academically and no longer loving school.

As families move to Alpha, test scores carry significant weight. It is a clear currency—a measure that parents believe they understand and can compare. The same goes for showcasing Alpha high schoolers gaining acceptance into highly selective colleges. It resolves the anxiety of choosing something this unconventional. It lets a parent tell skeptical relatives, in effect, that it is fine for their child to spend afternoons on Tough Mudders and startups because the academic results are right there on paper, and the kid is happy to be in school.

That combination—demonstrable achievement alongside a child who genuinely likes going to school—is Alpha’s real north star. Parents, sensibly, hold any school they choose to a do-no-harm standard, and on the test score measures that parents use to gauge that standard, Alpha reassures.

What Alpha gives up

None of this means the skeptics are wrong about what Alpha leaves on the table.

The academics, for all their efficiency, seem thin from the outside. There appears to be little-to-no Socratic discussion during core academics. While students do read books, there appears to be little room for debates over books or current events—the part of the traditional English and social studies curriculum in which messy collaborative argument can turn information into understanding and more nuanced views.

To the extent those experiences happen at all, they happen in the afternoon, walled off from the academic block. Core knowledge instruction in the morning—which, if followed, will help students ace English Language Arts tests, but also has the potential to actually help students build the so-called 21st-century skills and become informed citizens—may seep into the afternoon’s projects, but they aren’t explicitly connected. The projects aren’t designed to reinforce the tech-powered learning of core knowledge. The two halves of the day run on separate tracks by design, despite there being a credible body of research suggesting that the life skills Alpha wants to build transfer more reliably to other contexts when built in the context of doing demanding academic work, not in a separate container.

That apparent thinness likely makes some educators nervous in its own right—and goes to the point that Alpha’s model is not built to serve all students from all backgrounds. But the thinness also raises questions about purpose. What is Alpha actually optimizing for? What does it believe school is for?

One reading is that Alpha’s focus is forming young people who can act with agency and seize opportunity, economic or otherwise. In a world increasingly shaped by AI, it’s equipping them not to serve the system but to have real control and standing within it.

Another reading is less generous. It’s also likely what unsettles many educators. This reading of Alpha’s purpose sees the school’s treatment of the core academics as something just to get a test score—a teaching to the test that educators want to run away from and, it should be said, typically don’t ace. It expresses concerns about the role of AI in Alpha’s learning model—as part of an instructional tool in the morning elevated above teachers and a tool for student creations in the afternoon. And given public education’s role in serving society not just the individual—to preserve America’s democracy and participate in a pluralistic society—it worries about the optimization for the individual over the civic and communal purposes of school.

Alpha as one option in a wider mosaic

So where does this leave us? If you believe a school has to be all things to all people, then every one of these tradeoffs is an indictment. Alpha stands convicted. But that standard, we have argued, is borrowed from a public system built theoretically to try and offer something for everyone. The track record is clear that that system doesn’t do this well—and it is also probably the wrong lens to understand what Alpha is.

The more useful way to see Alpha is as a single, distinct, internally coherent option in a widening mosaic of schooling choices. In a landscape where families can genuinely choose, Alpha does not need to be right for every child, because it does not need to be the only school. It needs to be the right school for the families whose plan for their child it is built to serve, and honest with everyone else about what it is not. Different schools can serve different Jobs, and that pluralism is a feature of a healthy ecosystem rather than a problem to be solved.

Which points, finally, to what we think the fair criticism of Alpha actually is. The problem is not that Alpha is narrow. Narrowness is what lets it be excellent at the Job it has chosen.

The problem is that Alpha does not always present itself as narrow. Its public message tends toward Alpha being categorically better than conventional school, full stop, rather than the more precise and more honest claim: if these are the priorities and values you have for your child, here is the specific Job we are built to do, here is the evidence we do it well, and here is what we deliberately leave to others.

Yes, there are advantages to the simpler message that Alpha trumpets. It has elevated the school’s profile. It has also lifted up the value of mastery-based learning and the fact that core academics don’t need to take all that long—concepts that we, along with many others, have tried to promote for a very long time.

But marketed as a better school for everyone, Alpha also invites the all-things-to-all-people scrutiny it can never satisfy. If it instead marketed itself as a sharp answer to a specific Job, it would be harder to find fault, and much easier to place in the mosaic where it belongs.

We still have other questions. In a companion piece, we’ll tackle one of them—whether the technology underneath Alpha, and the unusual business model funding it, add up to something genuinely disruptive to the wider education landscape. But that is a separate argument. For now it is enough to say that a school built to do a few things well—and being honest about which few—doesn’t have to be something that is feared or loathed. This is what real choice is supposed to look like.

Stay tuned for the next post in this series, titled “Could Alpha School disrupt K–12 education?”

Then reality hit. Boy had I been naive.

Less than a month into the job, I had dropped all my plans for home visits and after-school tutoring. I struggled just to keep my head above water with lesson planning and grading.

Classroom management was a major challenge. Kids in my classes would act out for a whole host of reasons: they were bored, my lessons confused them, they wanted to show off to their peers, or they brought social drama through the door with them. First I’d redouble my efforts at outwitting their games. Then in frustration I’d rally all the stern force within me to prove I wasn’t messing around because I needed to cover my material. But as the tension escalated, I felt guilty knowing that my discipline tactics were less about meeting my students’ needs and more about trying to control their behavior so I could do my job.

To add to the challenge, math is an especially unforgiving subject because concepts build on each other like Jenga blocks—too many gaps in your foundation and the structure can’t hold. For most of my students, there were huge gaps—and the gaps weren’t the same.

For example, one of my students was homeless and often missed school. By the time he reached me, he had racked up so many absences that almost everything in my lessons went over his head. Meanwhile, in the same class I had another student who was exceptionally bright and whose parents made sure he never missed an assignment. He had some weak spots in understanding fractions, but they were hard to notice because he was so far ahead of everyone else.

These two students were opposite ends of the spectrum in a class where math skills probably ranged from 2nd to 8th grade. How do you teach a one-hour lesson that addresses that wide variety of needs?

As I struggled to figure out the job, I started to have a sense that it wasn’t just hard because I was a rookie. A lot of what made the job challenging was the design of the system I was operating in. I came to a conviction that there had to be better ways to teach math to middle schoolers. And the more I sat with that conviction, the more I sensed it was really about something bigger: that conventional schooling was falling short in a multitude of ways that went far beyond what any individual teacher could fix from inside the classroom.

So I worked hard, I cared a lot, and I still went home most days burned out and feeling like a failure.

Then one day, I heard something that sparked new hope.

While driving home from school, I was listening to an episode of the Freakonomics podcast that featured a math program in New York City called School of One—later spun out as the nonprofit, New Classrooms. The premise was almost exactly the problem I’d been wrestling with. Take a wide range of students. Pull together a wide range of resources—activities, small group work, teacher-directed lessons, online apps, remote tutors. Use data on each student to build a custom schedule every day. Then, assess at the end of the day and use that data to build tomorrow’s schedule.

As I crawled along the interstate I was yelling in my car. Yes! Yes! This is what I need.

But soon I had a second thought. I don’t see how a program like this ever becomes the standard approach in my school district. I couldn’t articulate it well at the time, but I had a sense that there were systemic barriers to changing education that went beyond just inventing a better instructional model and then spreading it.

I left teaching to attend grad school because I knew I wanted to work on the problem of rethinking schooling, but I had no idea what a job in that space actually looked like. While I was there, I came across a book called Disrupting Class by Michael B. Horn and Clayton Christensen. What I found wasn’t just another vision of what schools could look like. After all, there are many good alternatives to schooling, some of which have been around for a long time. What Christensen’s theory of disruptive innovation did was offer a causal model for explaining how better models would become mainstream.

Getting the theory right

I’ve now been at the Clayton Christensen Institute for thirteen years, and throughout that time I’ve been focused on this question: how does this theory actually operate in K–12 education?

The theory of disruptive innovation was born from research on sectors such as steel mills, hard drives, and discount retail. But a lot of what makes disruption work in other industries either doesn’t exist in education or works differently here.

Christensen had a phrase he kept on a board in his office: Anomalies wanted. The idea was that when reality doesn’t match the theory, you don’t throw the theory out and you don’t pretend the anomaly isn’t there. You wrestle with it. You figure out whether you need to refine your application of the theory, refine the theory, or both.

That’s what I’ve been doing for the last decade—getting to the heart of some of educations’ most knotty anomalies. Education doesn’t line up cleanly with the histories of steel mills or hard-drives where the theory was born. But by studying where and why it diverges from a basic interpretation of the theory, I’ve come to a sharper picture of the dynamics reshaping education.

Through my work wrestling with that complexity, two important insights have become clear:

• First, education can’t be transformed through the reform and change management strategies that so many would-be change makers turn to.

• Second, we are living through a wave of disruptive innovation that will eventually replace conventional schooling with new and better learning experiences for our students.

The first insight runs counter to what most education reformers assume. The reason conventional schools are so hard to transform isn’t lack of effort or ineffective leadership. It’s that existing schools and districts are deeply embedded in a system of longstanding practices, incentives, and expectations that define what counts as quality, what gets funded, and what gets rewarded. Within that web, genuinely different approaches to schooling get pulled by a force like gravity back toward the stereotypical approach to schooling.

Christensen documented this same trap in every industry he studied: the organizations best positioned to improve an existing system are almost never the ones that replace it. Transformation in education won’t come from reforming existing schools from within. It will come from new models that emerge outside that gravitational pull entirely.

Meanwhile, disruptive innovation happens when new technologies enable fundamentally new ways of doing things. At first, those new approaches take hold at the margins and look uninteresting to most people. But then they improve over time to eventually become mainstream.

Those conditions exist in education right now. New learning technologies that didn’t exist two decades ago—like on-demand instructional videos, software that adapts practice problems to students needs, and AI tools for exploring, creating, and getting feedback—are being used by innovators on the margins to invent new models of schooling. They’re still far from becoming approaches to mainstream education, but they’re on that trajectory.

The map I’m trying to draw

My overall goal with this Substack is to give people a better map of the paths that lead to transforming schooling. If you’re an entrepreneur building a new model, a funder deciding where to place bets, a policymaker shaping the rules of the road, a foundation working out a theory of change, a superintendent or principal trying to figure out how your district fits into all of this—my hope is that what I write here helps you locate yourself on that map and act with more clarity about where to put your energy.

The publication is organized into two ongoing series.

• The Map is the primary one: a cumulative argument, post by post, for how K–12 education actually transforms—the mechanics of disruption, the role of value networks, the conditions that let new models take hold, and the long arc by which incumbent systems eventually yield to alternatives. The posts build on each other and are meant to be read as a body of work.

• The Workings is a companion series of shorter pieces that apply the theory to specific cases, current developments, and questions I’m still working through. They can be read independently. You can read more about both on the About page.

A last note, in the spirit of Anomalies Wanted.

I don’t think of this as broadcasting. I think of it as a public conversation. The ideas I’m working with get sharper when they get stress-tested, and the people doing the actual work—in schools, in startups, in foundations, in statehouses—see things I don’t. So if you read something here that resonates, tell me. If you see an anomaly that the theory doesn’t quite account for, especially tell me. If you have a question, a pushback, a story from your own corner of the field, send it. I’d rather have a smaller readership engaged in real dialogue than a larger one consuming casually.

— Thomas Arnett, Clayton Christensen Institute

Two questions, twenty seconds, and I'll know a lot more about who I'm writing for.

Reader Survey

A few notes for skeptics

Disruptive innovation theory is not (yet) a widely used idea among people working in nonprofits, government, and education. Most of its traction, so far, has been in the worlds of business and tech. Given this history, the term sometimes triggers concern among people who see their work as about improving society or helping the marginalized. So, if you’re someone who is already suspicious just seeing the term ‘disruptive innovation’ used in the context of education, I want to note a few things up front.

Disruption is a phenomenon, not an ideology. Clayton Christensen studied a wide range of sectors as a social scientist to understand the naturally-occurring patterns in how new solutions gain traction and how sectors evolve. My aim is to use the theories he developed to offer new insights on how change happens in education. I’m here to make sense of how things actually work, not to cheerlead a “disruptive revolution.” As such, I’ll be candid about both the nuances and the uncertainties in my analysis.

Disruption is not inherently good or bad. On the positive side, disruption can dramatically expand access to things that make people’s lives better. Consider how the assembly line democratized car ownership or how the telephone made it possible to talk with family members across oceans and continents. At the same time, disruption can also have negative side effects. Discount retail has vastly expanded our access to affordable products that make our lives better. But it has also contributed to waste by making cheap products disposable, put some local retail stores out of business, and pushed some local economies toward jobs that don’t provide livable wages. I think a lot about how we can gain the benefits of disruption while also mitigating its potential downsides, and you’ll see that perspective in the things I write here.

I’m not rooting against public education. Realistically, there are some changes we need in education that just won’t come about through trying to reform or change existing schools—be those public, private, or charter. But there are real paths forward for established school systems that want to be part of building what comes next. I’m hoping that what I share here will make their path forward more clear and more doable.