EdTech's Holy Grail: A Personalised AI Tutor for Every Child

Leveraging AI and Bloom in the continuing quest for quality education at scale.

Growing up, the idea of 1:1 tutoring made my stomach do cartwheels. Being privileged enough to grow up in Shanghai, my parents quickly cottoned on to the fact that kids could be pushed to strive a fair bit harder than was the norm in our native Australia.

A large drive in this realisation was Mary, the mother of my friend Max. May was what some may describe as a ‘tiger mother’.

I still remember the reasons Max couldn’t hang out on any given day after school. Monday was Chess tutoring. Tuesday was Maths. Wednesday, Chinese (Max’s family spoke Chinese at home - could never really wrap my head around this one). Thursday, violin. All this around a schedule already packed with school, sports, homework and social time.

Unsurprisingly for a 7-10 year old, I really didn’t get it at the time. Fast forward 13 years and I realise Mary (and many other Chinese parents of her ilk) were really on to something. While I can’t imagine many tiger mothers themselves are in need of justification for their practices, I found my mind changed by the ideas of Benjamin Bloom.

A pioneer in old-school classroom education, Bloom found that one-on-one learning using the mastery learning method delivered a two sigma improvement in student outcomes. A study from Harvard in 2004 utilising these methods found that individually tutored students performed 200% better on a standardised maths test.

While standardised testing and graded scoring systems have been used for proving the efficacy of Bloom’s methods, the methods themselves (fortunately) rely little on standardised testing or benchmarking.

I will discuss in more depth below how I envision education might change as students evolve from having search engines at their disposal to all-knowing answer engines. What is important here is that much of the utility of Bloom’s model illustrated above maintains its utility, with ‘remember’ being perhaps the only exception in the short-to-medium term.

Now, back to Max. Without a doubt, Max benefited from some derivation of Bloom’s theory. One-on-one tutoring and constant formative assessment would have rapidly sped up his feedback loops and put him miles ahead of regular classroom learners like me. However, Max’s (and many other under-pressure early childhood learners) educational adventures were all in pursuit of arbitrary benchmarks that led to fixed endpoints.

His maths tutor wanted to see him max out his IB scores so he could get into a top-tier University. Even non-classroom exploits were subject to rigorous numerical standardisation. Aiming for a high ELO at the chessboard. Maxing out to Grade 9 with the Violin. So on and so forth.

However, as briefly alluded to above, with the growing prevalence of LLMs (i.e answer machines) and generative content for all types of multimedia, the Max’s of this world could be forgiven for thinking that all those early childhood tutoring hours may have been spent in vain (many may also beg to differ). I am not going to propose how we redesign educational curricula to suit this changing paradigm. Instead, I am going to focus on the immense value that this 1:1 tutoring methodology can have under any kind of new system that may emerge. That is a topic for groups of people much smarter and more powerful than I.

In the rest of this piece, I am going to try and outline what the early outlines of  permissionless, frictionless  protocols for global Bloom-style tutoring may look like and how you, the ambitious would-be founder, might bring them to life.

The hope is that by improving access and reducing the cost of delivery for personalised education we will have a world full of people educated to the full in whatever fields they so choose.

Early-Childhood Education Today: Old Models & New

Education is a massive spectrum, so this article fill focus on the niche of early childhood education. I do this because I feel like it's the one segment that has still yet to really experience much deviation from historical norms.

MOOCs are very popular for mature education and COVID popularised online delivery for college-level courses. Khan Academy and similar models have had a lot of success among middle school and high school students. Yet, with the exception of a few pioneering startups targeting very privileged niches, there are very few startups working on truly universal solutions for every child to level up as early as possible.

Firstly, why is early childhood education so important? Some of the benefits of receiving quality education early are obvious. It gives students more confidence in their abilities. It provides a baseline for thinking skills that is compounded upon before foundational education essentially gives way to specialised execution around the age of 22 (at least for the majority of students in our current world).

However, there are an entire suite of auxiliary benefits lying below the surface. Among them:

• Improved knowledge of non-vocational life skills (i.e. knowledge around self-care, health, behavioural matters etc.)
• Social development. Digital solutions may not be a silver bullet for this, but they can definitely serve as facilitators.
• Reduced need for remedial support in later years.
• Parental empowerment. Universal solutions that impose less of a burden on primary carers provides them more time to do other things whilst also ensuring children are on the right path.

The above graphic is striking. Even in developed parts of the world, the academic returns from early childhood education are underwhelming to say the least. There could be any number of reasons for this. Poor teaching quality,  lack of student clarity around the future utility of education, lack of global accessibility for all types of students all play a role.

How can we obfuscate these away to ensure that a) every student is able to have a baseline level of knowledge to apply to further learning and b) they are prepared for new forms of education that are optimised for different endgames than getting the highest SAT/IB/GAMSAT/{insert standardised test of your choosing}?

The Market Today

According to Business Research Insights, the global market for early childhood education currently sits at US$258.94bn per annum.

It is important to note is that ‘early childhood education’ in this context also captures businesses that comprise entirely different startup markets, including childcare, nurseries and recreational programs.

That being said, it does show the magnitude of the problem. After all, it is a market that will count every child born from now onwards as an end customer.

For the context of this piece, the focus of market analysis will be on those incumbents already trying to provide some form of ‘tutor for every child’ solution or augmenting the student experience learning optimised for developing problem-solving skills.

See below (and zoom in):

While the stated aims, ambitions, target audiences and delivery methods of all of these businesses differ (in some cases, substantially), they all leverage the magic distribution powers of the internet to reach anyone anywhere at any time.

Khan Academy and YouTube have been the big drivers of self-paced, independent and often extracurricular learning in the internet age. Businesses like Method Schools (and other schools integrating Khan Academy into their workflows) have matched independent learning philosophies with the socialising benefits of the in-person classroom.

Synthesis has built a solution for cultivating genius by putting incredibly gifted kids in the same virtual spaces as one another and having them hammer away at problems. Even if it sounds, looks and smells like an elitist, selective private school, the mission of ‘genius creation’ is an incredibly important one. Geniuses are, after all, the people who push the dial forward and change the way the world thinks.

Imagine Worldwide has mode universal access possible - without the need for internet access!!!

As much as we can continue complaining about state school systems and declining standardised test scores (largely unimportant), we must acknowledge that we have lived in a blessed era for learning.

But there is still so much to improve. That fact should be brain shatteringly exciting for parents, schools and entrepreneurs the world over.

The Problems to be Solved in Primary Education

Misdirection. Regardless of form of delivery, downstream ‘customers’ of the K-12 education system (e.g. universities, employers) are still looking to judge and evaluate candidates based on rankable scores. This needs to change. There is an element of roundedness required to be able to communicate and cross-pollinate ideas. Beyond this, however, passion and commitment to a given path are far better predictors of success and more conducive to the kind of creative genius that pushes the dial forward. How can education systems be designed that start to develop this edge at the same time as building roundedness?

One idea may be to borrow Google’s “20% Project” system. Early childhood students are given 80% of platform time to pursue whatever kind of novel curriculum displays the best efficacy at achieving roundedness. The other 20% is dedicated to mentoring and teaching students to take a passion project from idea to delivery.  Many educational systems try and bake this kind of project work and design thinking methodology in at middle years with the likes of science fairs and other kinds of personal projects. Why not earlier?

Access & Equity. Geography (even at the city level), regulation, language and a million other factors all play a role in downstream outcomes. Again, the benefit of the internet’s great equalising effect is going to be accelerated in coming years and, optimistically,  obviate some of these problems.

Some issues remain, however. Many still lack access to the internet. Many live in environments where top-down regimes censor content that would teach students about the world. The internet problem, like a few that are to be outlined, doesn’t lie strictly within the domain of educational platforms. However, the content platform may. Even if there may still be some overhang in inequality if information is censored, there is still substantial room for improvement when students gain access to personalised learning assistants that can re-package all available resources to help them learn what is required for success on the global stage.

Fortunately, we are on path towards a vision of education that actually leaves no child behind.

Distribution of Funding. Parallel to the issue of access is the issue of funding. Under the current classroom model, children in less privileged areas with worse classrooms get less funding, which gets them even worse classrooms. This is unfair and unsustainable for a world that thrives when rising tides lift all boats.

Parental Engagement. Parental involvement in education has historically been a harbinger for improved outcomes. Better attendance rates, increased accountability and ‘gap filling’ for areas that may not have been explained or understood in the classroom are all some drivers for this. Less appreciated in the formal data is the serendipitous, non-classroom learning that happens around the dinner table.

The flows of information enabled by the internet are slowly eroding this advantage, but in terms of the day-in, day-out parental engagement remains an important factor. How can systems be designed that don’t penalise students for having parents who can’t be as engaged as others? Better yet, how can systems be designed that allow parents of less pedigreed educational backgrounds to support their children to the fullest (this probably warrants an entire market breakdown of its own and borders the childcare market very closely).

Diverse Needs Accessibility. As hard as the traditional classroom may try, it is nigh on impossible to properly accommodate special needs students without creating some kind of impediment to the rest of the classroom. More individualised learning platforms almost remedy this by default. What kind of features will these platforms leverage to best capture these long-tail cases?

New Enablers

A spoiler for those reading that have yet to use ChatGPT or any kind of consumer-ready NLP interface. Tomorrow’s search engines aren’t search engines. They will instead be answer engines.

For those who have been following along, this presents an obvious quandary for entire education systems that judged children on their ability to source and communicate answers.

This will have a few catalysing effects for the future of education, including but not limited to:

Extreme Personalisation.

AI, and personal intelligent assistants, will vastly accelerated the degree of personalisation in any given service we have access to. For those wanting to learn more about how, I will once again (and likely not for the last time) recommend this brilliant article on autonomous agents from Matt Schlicht.  Education will be no different.

Personalisation is a key ingredient to the Bloom two-sigma effect. However, personalised teaching agents are going to take these effects yet another step further. Rather than gradually learning about a pupil’s learning preferences and most effective modes of presentation over weekly, hour-long sessions, the teachers of the future will have an innate knowledge of what kind of information the student is best at processing.

Want to learn Python? Today, you would likely leverage some one-size-fits-all open-source resource like Harvard’s CS50 (which is now taught by an AI) or any one of the amazing YouTube channels teaching the programming language. In the future, the novice is more likely to learn from an intelligent assistant that not only has a complete understanding of the language, but also a better knowledge of your preferred learning patterns than you do.

There will be no point where you feel disengaged or unmotivated by the learning process. You will be assigned formative projects that are uniquely interesting to you and your reasons for learning Python rather than ticking boxes to check your progress against some arbitrary skill tree.

Entertainment through Customisation.

In 12 or more years of schooling, I can imagine every reader’s educational sample size is large enough to include at least one case of a subject being ruined for them by a charisma-less, over-demanding or plain mean sop of a teacher.

In a world of autonomous, personalised and digital delivery, it is almost impossible to imagine this scenario playing out again.

Even with tools available to us today, I could have a digital replica of Morgan Freeman’s voice teaching me a physics course prepared by Carlo Rovelli. Or any combination of voice, knowledge, mode of delivery and subject. Don’t like it? We’re not at all far away from assistants that will adapt almost instantaneously to your preferences. (An interesting philosophical twist: by being able to make any content seem delightful, if every student had an agent we could theoretically determine exactly how skilled resource allocation occurs - scary thought).

The early success of artificial human replica protocols like Character AI and Replika (subject of one of the great Not Boring posts) indicates the market readiness of these kinds of products. What remains to be done in the space in order for a fully autonomous education assistant is a way to monitor student progress and self-assign tasks based on these insights.

Intelligent Mentoring & Assessment.

In the next 20 years, better education should not be measured by how much we can improve on the median NAPLAN/SAT/whatever standardised score applies where you’re from. Instead, it should be a more accurate representation of how well a student’s learning & development matches their reasons for beginning that learning path in the first place.

A large but silent problem with early childhood education today is that kids have no idea why they’re learning things. They can establish pretty quickly why learning their native language is so important because before long they are using it extensively everyday. Maths, less so.

Being able to better explain and constantly mentor kids on why they are learning certain things and how it applies to them is a massive boon for engagement. Then comes the problem of “assessing” without demotivating.

Assessment will continue to trend towards true games of understanding. In a world of personalised learning, it is highly possible that modes of assessment and benchmarking are purely tailored to the individual.

‘Amy’ might be highly motivated and driven by the need to display some kind of arbitrary 90% competency on her biology exam. ’Adam’, on the other hand, just likes biology so he can dissect frogs (humanely, of course). Both want to go into surgery.  With personalised assistants, Amy can have a black & white test written up for her whilst Adam can receive live feedback on his technique. Both will have separate paths to becoming top percentile surgeons crafted based on what keeps them engaged to improve everyday.

Predictive Analytics.

As the volume of training data grows for our teacher AIs, it will become better and better at understanding which combinations of skills, interests and personalities are best fits for certain paths in life. Based on this, the AI can continuously offer up recommendations and probabilities of success to help the student evaluate any given study or ‘career’ choice as they progress through any autonomous educational pathway.

For example, I may be a 6 year old (I’m not) with a particular affinity for languages who has developed a precocious interest in quality movies. My personal assistant will be able to present me with a constant stream of options and enrichment activities based on the paths that it may think I enjoy. In a world where humans still want free will, this assistant would leverage these predictions to maintain my illusion of free will and never force me down any given path. It would be able to tell that mathematical research might not be for more but producing and prompting generative cinema just might be.

It can communicate with other agents to direct me to other students around the world who may just be fated collaborators based on our respective interests and skills.

Importantly, it will also provide a rich wealth of information as to how children foresee the world, what they want to do in it and allocate resources as such.

Imagine how much we will learn about education when we can see our own skill trees mapped out from where we are now to the very genesis of our education.

Given that all of the above are very genuine possible inputs to the way children are educated in the future, how might they combine with current best and accepted practices to form a generational commercial education opportunity?

A Vision for Edtech’s Holy Grail

Based on the problems outlined above and the realistic restrictions that will be imposed by regulations in the foreseeable future, I view the above illustrated feature set as being table stakes for a globally scaled, high efficacy early childhood education unicorn.

Personalised learning assistants will be the hub of any model that wins. The key points of the section above - personalisation, intelligent mentoring, quality analytics and their combined efficacy in delivering children towards successful post-education (this term feels dystopian) outcomes - will be the core drivers of who wins education in the age of machine intelligence.

Beyond these factors, the winner will likely be the first developer to make these things as fun to interact with as it is to actually go in and see your favourite teacher. On this note, it will likely look, sound or interact with you in an entirely dissimilar fashion to ChatGPT.

In-Person local hubs will augment this heightened learning experience with the socialisation that early childhood education systems have nailed since, well, forever (see Exhibits A and B below).

Play will remain a necessity. ‘Schools’ of some form as we know them today will step in (almost ironically) to fill this need and desire. Maybe the teachers of yesteryear fill this void, sitting in the same supervising chairs as they do today. Maybe we move closer to the ancient vision of a village raising a child, whereby one community facilitator manages and is rewarded for taking pastoral care of the kids as they go about their work and play.

Regardless of how it plays out, the demand for physical spaces for both childcare and co-educational purposes will remain. The economics of these spaces will remain exclusive and rivalrous. This means opportunity. There will be money and reputation on offer for providing rewards for those operating these spaces.

Acting as key components of these local hubs will be human mentors. These human mentors will not be as smart as the AI teachers that the children possess in the palm of their hand. But, as they have forever, they will be important role models and touchpoints for how the ideas and goals being set by the digital teaching assistant are manifested in the real world.

On the flipside of our geographically local hubs are geodesic, online interest hubs. For those who are hearing the word geodesic for the first time, I redirect you to where I first learnt the word - Balaji Srinivasan’s Network State. A lot of the ideas herein are designed for working professionals but ideologically make sense through the lens of a stateless education as well.

So what does the geodesic interest hub look like in the case of our children’s personalised learning assistants. With the aid of data privacy protocols, our autonomous buddies will be able to generate a zero-knowledge interest graph for each child and geodesically match them to one another. From here, virtual classrooms, project groups and workshops can be created (and possibly be facilitated by human experts) to allow these pupils to benefit from collaborating with like-minded peers from around the world.

For the sake of ensuring that we don’t set out an unstoppable and ineffective education machine on the world, some reinforcement mechanisms are required.

The first will be human feedback mechanisms to ensure that these digital assistants and their methods. Anyone who has seen any movie even slightly pertaining to the future of AI knows that what can go wrong will go wrong. Just as in today’s large language models, there will be an extreme demand for human feedback providers in models that could be dealing with and educating literally millions of children.

There will likely be demand for program auditors, ensuring that the pathways that digital assistants choose for a child do not interfere with other human goals. To put this in a stark example, we do not want our aspiring surgeon Adam to be testing out on real cadavers at age 7, even if it is most relevant to what he wants to do.

There will also need to be some kind of mechanism to ensure that assistants are not being overly deterministic. Say, for example, that assistants receive some kind of incentive or reputation boost based on their ability to place students into certain professions. Obviously, some professions are easier than others. This warps the incentive of the assistants to deterministically direct children down these roads and essentially throw entire economies out of whack as resources are wildly misallocated.

Firstly, this presents an opportunity for humans to be involved in reputation scoring and incentive design for these assistants. Secondly, it shows that economists are still important. Ensuring that children’s true interests and desires are respected whilst still maintaining some baseline level of skill distribution across different important areas is imperative to the future of education.

Sidenote: There is an entirely separate, multi-billion dollar opportunity here for creating systems that make it easy for non-technical people to observe, test and provide feedback to any kind of model.

Additionally, some kind of global progress tracker is needed to continuously monitor the efficacy of our hypothetical Bloom Two-Sigma AI. This could be an autonomous system that plays the role of the proposed economist above. With deep analytics available on the ways in which autonomous teachers are directing students, humanity will need to be able to visualise the longer term effects of this on resource allocation and adjust models accordingly.

What is the endgame?

This is a big philosophical question. In today’s world, the endgame of education is to get a good job and be able to interact and collaborate effectively with others.

What ‘work’ means is going to be very different in 10 years time. In this sense, what should change about the ‘endgame’ of education is that students are able to create their own self-fulfilling ‘work’ that leverages the tools available to create meaning in their own life. The parts about collaborating and interacting with others shouldn’t change a bit.

What might an MVP look like?

The most obvious first step is to build a Character AI style ChatGPT-wrapper (I feel yucky just writing this) specialised for speaking in the style of a primary school teacher.

In essence, what I am trying to say is that the most likely MVP could just be a Mr.Rogers chatbot.

From here, the primary issue to solve is student evaluation. How can you attribute gains in their learning to their digital versus human teachers?

An interesting growth tactic to spread this to the public conscious might be to have live streamed kids hackathons (with parents consent, of course). Each child is equipped with their own personal assistant to assist with demonstrable project work. This would serve as a live action demonstration of the efficacy of problem-solving centric learning methods and also a fun way to see the creative ways in which children leverage AI to bring their ideas to life.

If this sounds far fetched or might not have a market, I am willing to bet that you are wrong. Case in point - the most recent Scripps National Spelling Bee reeled in 9.2 million viewers.

Where do you go from here?

Before winning the distribution game and breaking into state-level educational systems, the commercial elements of a protocol like this would likely have to begin at the extracurricular level.

Synthesis’ ‘Teams’ solution presents a handy playbook for how to test these ideas among trial groups. From the initial hype-generating tactics, create a waitlist of parents keen to explore using these tools with their kids.

From this waitlist, you can already begin some ragtag form of geodesic online classroom by matching children with similar stated interests and career paths.

Geographically, you could even do this by running after-school groups of children leveraging the early generations of these assistants to collaboratively build projects in local teams. This ticks off proof-of-concept boxes for both play and

Go-To-Market

How do you then turn a group of extremely happy parents who swear by virtual teachers into a dominant force that can weave its way into becoming a dominant player in any education system globally?

Some GTM ideas:

• Cost-plus delivery of agents (generic or personalised) to existing alternative education providers. This includes charter schools, home schooling networks (can be found via Facebook groups and Subreddits)

• Free delivery of generic agents to public schools for testing. The earlier this is done the more time governments will have to process the impact and necessary regulations to protect student interests.

• Replicating popular content creators as teaching agents, Character AI style. Why would parents want their children to learn science from Mr. Boring when you could have a living, adaptive version of Bill Nye with them for an hour per day?

• Establishment of agent APIs as plugins to existing virtual content delivery platforms. How different would the Khan Academy experience be if inbuilt recommendation engines could direct kids along the right skill trees to achieve their ‘end-goal’ learning objectives. What about custom YouTube playlists accompanied by agent-generated projects/missions/tasks to apply learnings from the video transcripts?

Business Models & The Size of the Pie

For manifold reasons, I still believe it is best that the models that lie at the heart of this business proposition remain open-source. Firstly, from a social benefit point-of-view, open-sourcing these models leaves them open to constant accountability and transparency.

It improves market competition, and consequently the quality of output, by establishing incentives for people to fork and build atop these models to make them better for students.

Lastly, it is crucial because it is the best path forward to the lowest cost delivery possible for the highest quality education - the holy grail for a more intelligent generation of humans.

The above is a bit of napkin math on the size of the markets on the table that might be opened up by this next generation of smart teaching assistants. A breakdown of the assumptions:

• This model maintain faith that people will keep core models themselves open-source and free to access. An interesting point that I am excited to see play out with regards to accessibility is how much of the capabilities of these assistants can be preserved for areas without internet access in the short-run, a la Imagine Worldwide.

• Applies assumption of ~750mm children being of primary education age at time of deployment.

• Costs of delivery for in-person groups are based on aiming to provide >50% reduction in delivery costs compared to Australian public primary schools (comes out to ~US$5,225 p.a all things considered).

• Costs of delivery for geodesic groups designed to be more accessible versions of Synthesis. Synthesis most basic delivery method reportedly charges US$180/mo (or $2160 per annum) to enrol students (per Deseret).

It must be noted that the above is very much for of a market map than a revenue forecast for a single company. It would be virtually impossible for a single business to capture all of the in-person market, just as no private school group operates all the private schools in the world today.

The geodesic market is more likely to bend to the power law, but again will still probably see significant competition because of the size of the pie and amount of different ways it can be cut.

These metrics are a bet on education changing significantly to become more stateless and tailored to designing intentional outcomes from virtual delivery.

Final Words

This final piece of wisdom ironically comes from an article title “Why education startups do not succeed”:

“Entrepreneurs think of education as a quality problem. The average person thinks of it as a cost problem”.

Immense personalisation and inbuilt know-it-all factors will make virtual assistants a super-teacher of the highest order. The quality and entertainment-factor of models will be important determinants of who wins in this market, but at a global scale entrepreneurs can not forget the importance of reducing costs of delivery and the many rewards that can be reaped by reaching a wide audience at low cost.

Addendum: Misc. Philosophical Points

• What should be the objective function of early-childhood education? In simpler terms, what do we want today’s children to be when they are adults?
• Society is comprised of generations that, for all their differences, were largely educated under similar regimes. How will inter-generational relationships be affected when a generation is educated by an entirely different species?
• Does having a ruthlessly efficient, autonomous tutor disrupt any serendipitous effects in a child’s learning journey?

Also published here.