Education Unbundling
opportunities in a post-AI tutor world
Economic growth is predicated upon two real elements: technology improvements and labor upskilling. Technology improvements pose less of an issue for economic growth going forward. AI even if we are stuck with GPT-4 is good enough to grow productivity comfortably in a large swath of industries. And then of course, there’s always El Segundo.
What I am fairly concerned by for the future of economic growth is the labor upskilling element and everything downstream in education. And even more than that, I fear a world that removes human agency in the process of automation. Traditionally, education has occupied these two features: create human agency and equip students with the skillset needed to flourish in the workforce and daily life.
Unfortunately, K-12 increasingly feels like its in a bad state and post-secondary education and labor upskilling is far from the ideal.
And if you’re expecting the state of public education to rebound post-Covid or reconfigure itself for a new world, it’s unclear that writ large the current structure can adapt. On the one hand, about 82% of public education costs are fixed costs. That already creates quite a bit of problems for creating adaptation to new paradigms (or even admitting that the model is quite broken as it stands). On the other, all the minutia are controlled by local school districts and boards. Any wholesale change across the US is immediately untenable.
These two factors combined with the wider “opt-out of the public schools” phenomenon create a sort of spiral. Fixed costs on a decreasing student base yield a catastrophe in the making and it’s not clear when it becomes fatal.1
One of the ways to get cost control is to consolidate and as a friend pointed out to me this week, that’s indeed what has happened. The number of school districts over the past 50 years has continued to shrink. But this can’t go on indefinitely and whatever your opinion is on the importance of public schools going forward, it’s hard to ignore that much of the action is happening outside the public school sector.
There’s a wide range of people focused on this whole issue, both from a technology standpoint and a philosophical one. It’s almost impossible to write anything about education without revealing something of your own bias, so let’s just get mine out of the way. Here are some key considerations that I weigh:
K-12
A purely digital education is a fundamentally broken endeavor. See Covid. We ran that experiment. And even in the instance where it doesn’t break down academically, it breaks down socially.
Education has to aim higher than information transfer. I’m an Aristotle guy: education is about formation. I’m bearish on AI tutors for formative matters, while extremely bullish on their ability to enable educators to focus far more on the formation of the student. We traditionally indexed on the humanities as a good bar for useful formation. But what may actually be needed is a redefinition of the humanities altogether not as discrete subjects but as something primarily for formation itself.2
Educational “Georgism”:
We should ask the hard question: what is the value to parents of having a place to drop their kids off per day vs. the value of the education? What is the teacher currently spending the majority of their time doing? Teacher frustration nets out to “I love my kids, but I am spending 80% of my time on non-education problems (with enough success stories to make me stay).”
In my view this is a large part of the exhaustion on all sides with the current education system. Childcare + education are bundled, the bundle is a bad deal for all sides as the system has regressed substantially in providing this package and people are absolutely fed up. I have no real opinion on how to alter this inside the public school construct. As such, this will primarily focus on alternative education paradigms.
Testing as an aptitude gauge is most likely here to stay, but if we can gauge an AI’s aptitude by making it pass these tests, could we eventually reverse this and gauge a student’s aptitude solely from its engagement with LLMs? This would get us substantially closer to personalizing a student’s trajectory, and provide far more continuous feedback on skill/educational attainment.
Within the current paradigm, can we reduce the amount of rote work that teachers have to do? Magic School is doing some incredible stuff with large adoption amongst teachers.
Certification/Licensing in Labor Markets
What does it even mean to upskill? Do workers actually care? Do their stated preferences differ from their actions in actually taking pathways to expanded economic mobility? Can technology even solve this?
Imagine that you could create an LMS with LLMs for specific vocations and accreditations that guaranteed: would anyone use it? Is the issue actually just credentialism? In some sense this was the promise of education marketplaces and other platforms: you could increase the reach of teachers. Did this pan out?
Said differently: if Khan Academy and the first wave of the internet didn’t fix issue in certifications or educational outcomes writ large, why do we think LLMs will? The proponents of LLMs will focus heavily on the personalization element, the detractors on technology’s current impact on the classroom.
Nonetheless, as Byrne Hobart has written, the personalization element of LLMs might be the piece that finally unlocks the Information Age for new educational approaches altogether.
Hunting for educational innovation will require looking outside of the current educational paradigm altogether.
That leaves us with microschools, homeschooling, private schools, and online programs as the places where we would expect experimentation to appear. And it turns out these approaches are all growing dramatically.
While the information taught in K-12 differs substantially in scope from a work accreditation, I think there’s a common design pattern that is worth thinking through and that seems sort of inevitable:
Create stellar AI tutors that are supported by teachers, instructors, parents, guardians, and more.
Create more liquid “colearning” spaces.
And this yields the right starting point:
If AI tutors are successful, and actually capable of creating the subject-specific outcomes we desire, then the cost of useful information transfer is going exponentially down. This isn’t a panacea for all the challenges of education, so we should also consider what arrangements will yield increased skill, social, and professional acquisition. Because we can now unbundle education outcomes from specific real estate in a specific system, we can theoretically create new real estate environments more suitable for driving specific skill and social goals.
Primer and Moonrise are equally fascinating for figuring out what this looks like.
Combine these two approaches and you may have a path to education in the 21st century that is cheaper, better, and more rewarding for all involved.
But first some quick math on why this works:
A Better Kumon
Kumon and its kin are tutoring centers mostly designed to help students with test prep. Kumon charges about $200-$300 a month, netting your student about 8 hours of tutoring per month.
And in fact, the actual work doesn’t look all too different from this slick demo:
Kumon nets out to $37.50 an hour. Assume for a second, that Kumon and AI tutor have roughly the same outcomes and are used for the same number of hours a month. Well, the AI tutors on the high end are netting out to about $29 a month at the high end.3 That means we are quickly looking at minimum a 10x cost reduction on the pure tutoring side.
It’s impossible to overstate how dramatic of an improvement this is. And we aren’t even at GPT-5 yet.
Even independent of the the quality of education stemming from this deep personalization, AI tutors will create dramatic surpluses for parents.4
So what if instead we used our $170-$200 a month surplus in tutoring spend to be newly allocated to educational goals outside of pure information transfer? Or stated differently: where will this surplus flow?
One answer that businesses like Moonrise are proposing are new co-learning centers.
Moonrise’s value proposition is incredibly slick. For $250 a month, drop your kid off for up to 12 hours a week in a center chockfull of craft supplies, 3-D printers, microscopes, Legos, and more.
If we think through the point from earlier, that formation is not simply about information transfer, there will always be an inherent need for spaces built for this sort of skill and relational development. We don’t have to over-dramatize this: kids at minimum need spaces to develop agency, collaboration, explore new skills, and make friends in spaces filled with adults/mentors that they trust and love.5
But there’s even more room to envision that this educational surplus could be combined with other parent spend for all sorts of new educational businesses. What does it look like to provide new skill-based learning facilities? Here I’m envisioning sports academies, dance facilities, and more that all come with the same co-learning + AI tutor + skill acquisition development.
You might simply call these “schools,” but the difference here is that the information transfer that AI tutors are capable of consists of 1-2 hours a day done at home or in a small group, and the kids are instead given free reign to explore different skills at their own pace with instructors standing by for private lessons, paid classes, and more.6
If the economics work here, parents would now roughly pay $6500 a year across educational and social axes - well within reach for the majority of the middle class.7
The unbundling and rebundling potential combined with latent demand for these offerings leads me to believe this is one of the larger opportunities in real estate development. It wouldn’t surprise me if the story of the 2030s is a redevelopment of the physical world to enable more liquid environments for kids, parents and far more.8
While this sounds great in theory for kids, does it scale to professionals focused on certifications, licensing programs, or even ?
In many ways, yes.
What’s needed for the vast majority of certifications and labor programs is information transfer + skill acquisition.
You can see this distinction even in how programs currently price themselves - a classroom hour is going to be priced differently than an hour spent in the shop.
Here too, AI tutors can unlock a 10x reduction in the cost of classroom hours with corresponding membership programs with real estate focused on skill acquisition.
Hackerspaces are an early form of these membership-based co-learning spaces where adults can get access to tools and materials to experiment/build.
Any surplus spend created from more cost-effective and successful AI tutors can in theory get reallocated to membership based hubs.9 In terms of structuring certifications, the AI tutoring companies should be the ones thinking about this and either partner or build these centers as ways to get “hours toward training” requirements done.10
While regulation may prevent some of the full impact initially, there is plenty of low-hanging fruit across all sorts of fields. Any sort of vocation with a specialized associates degree or license could in theory be adapted to these sort of AI tutors + skill-acquisition based real estate.
And if it were to work, I’d expect significant reductions in the cost and time to meaningfully enter the workforce leading to far less debt on young workers.
Building Blocks
Given the ability of foundational models, what exactly is there to build?
AI LMS Systems:
Right now all AI tutoring companies are reinventing the Learning Management System even as they articulate what exactly AI tutors can do effectively.
As this paradigm crystallizes, there are a couple possibilities:
A company can build a next-gen LMS meant to gauge whether an AI tutor is effectively guiding a student at the right pace of learning.
Ramifications:
Teachers remain important as managers of this process.
Testing as a means to prove aptitude remains.
LMS systems as traditionally understood, disappear altogether in favor of an LLM simply determining whether a student has proven aptitude.
Ramifications:
AI tutors are pretty much the managers of the educational process and simply given a list of goals that the student must achieve within a certain time period
Testing as traditionally understood disappears with perhaps once a year capstone tests.
Essays and communication skills are elevated as problem solving methods given the constant dialogue with AI tutors.11
In addition to this, I do think we need far more attention put on what real estate buildouts will look like to support these new endeavors and co-learning spaces. Perhaps none of these will ever be venture-scale, but there is a chance given new franchising models. Next week, I’m going to be writing about Opco/PropCo models that might enable the right sort of financing models for these initiatives.
It’s very early, perhaps too early to bet the farm on anything but AI tutors. But I am incredibly optimistic that new real estate buildout will be created and perhaps in short order.
I initially wrote this prior to the OpenAI demo - it’s more fatal than ever as will be addressed later.
Cf. The Diamond Age.
This seems high given Khan Academy’s base tutor without the features above is $4. This + the cost reductions (2x) that OpenAI is passing through at the developer level may set the price point at around $10 a month.
This will get juiced even further by school voucher programs which will begin to feel inevitable in states looking at AI tutors on one side and the inefficiency of current public school systems on the other. These AI tutors are inherently decentralizing: it doesn’t matter if your public school provides AI tutor access if the cost is only $500 or less a year. This puts the public schools in a massive Catch-22. Every inroad you make with AI tutors, will ultimately demonstrate the inefficiency of the current cost structure and ultimately lead to more opt-outs to more personalized education methods. AI personalized tutors inevitably lead to personalization across the rest of the educational axis. The death spiral has arrived and will most likely accelerate.
Everyone is of course quoting Diamond Age in reference to AI tutors, but most people miss the point of that book. Stephenson’s main argument is that technology is inherently destabilizing and leads to a new world where most of the world has retreated into formative practices stemming from their historical culture - namely the Neo-Victorians and the Confucians. But both of these groups ultimately are also culturally stagnant. The customs become rote, the culture becomes navel-gazing, and people are stultified. The Primer, the AI tutor, is ultimately designed by the neo-Victorians to cultivate subversion of and agency in the children whom read it to prevent the heirs of their culture from being formed without agency to ultimately choose their own life path. But Stephenson is also quick to point out that ultimately what unlocks the real ability of education is not simply the Primer, but actually all these other formative structures in her life namely her maternal-like figure.
This will necessarily involve some regulatory arbitrage around homeschooling to avoid accreditation. All accreditation is going to come with mandates around how the education is done. This seems achievable.
This would also disentangle these programs from needing to be in the right suburb for schooling.
The trendline is pretty clear in another important sector: co-working. It seems somewhat natural for this trendline to expand towards co-learning in the near future.
By removing classrooms from the real estate footprint and reallocating that to more tools or stations, you can increase real utilization quite significantly. Alternatively, you can probably buy up industrial real estate and do far cheaper real estate development and simply leave it as an open concept.
The real possibility here is to construct real estate to service myriad categories with minimal instructors needed on site. In turn, that would mean, far healthier EBITDA margins indexed mainly on facilitating self-paced learning. I’d expect AI education companies to think through this and end up with far more automated paths to tracking what’s being worked on and gauging skill acquisition.
This is perhaps one of the most exciting parts to me. My daughter’s AI tutor will practice the family tradition of correcting her grammar at every opportunity.