Hyperpersonalised N=1 learning

For decades, formal education has resembled a Procrustean bed—a system that stretches or cuts students to fit a rigid mold, regardless of their needs, talents, or pace.

Every child is expected to learn the same material in the same way, at the same time, and at the same speed, in lockstep with their peers. Struggle too long with a concept, and you’re left behind. Master it too quickly, and you’re bored and disengaged (likened again to the Procrustean bed).

Either way, the system moves on without you. This design flaw isn’t just a pedagogical inconvenience—it’s a systemic failure. One of the most interesting ideas that tries to update our model of learning which has remained fairly consistent for over a 100 years has been the Bloom 2 Sigma problem. I first encountered this as a neatly positioned question in Patrick Collinson’s blog—

Educational psychologist Benjamin Bloom found that one-on-one tutoring using mastery learning led to a two sigma(!) improvement in student performance. The results were replicated. He asks in his paper that identified the “2 Sigma Problem”: how do we achieve these results in conditions more practical (i.e., more scalable) than one-to-one tutoring?

In a related vein, this large-scale meta-analysis shows large (>0.5 Cohen’s d) effects from direct instruction using mastery learning. “Yet, despite the very large body of research supporting its effectiveness, DI has not been widely embraced or implemented.”

In 1984, Bloom published his now-famous study showing that students who received one-on-one tutoring performed two standard deviations better than those in conventional classroom settings. In other words, the average tutored student outperformed 98% of students in traditional classrooms. This wasn’t a marginal improvement—it was a disruption.

But there was a problem—scalability. Personalized tutoring doesn’t scale. For decades, this “2 Sigma Problem” remained unsolved.

Bloom’s findings showed what was possible—but the educational system had no feasible way to implement his insights at scale. And so the problem sat, gathering dust, while the world marched on with industrialized schooling.

Complementing Bloom’s insights is the idea of mastery learning—the notion that students should achieve a deep understanding of a topic before advancing to the next. Mastery learning is common sense: if a student hasn’t truly grasped fractions, throwing algebra at them is educational malpractice. It’s like building the top floors of a skyscraper on a foundation made of Jenga blocks.

Mastery learning respects the fact that different students require different amounts of time and practice to learn a given concept. But traditional education can’t accommodate this; it marches at a fixed pace, leaving some students overwhelmed and others unchallenged.

With mastery learning, it reduces the floor, and increases the ceiling. As Behn Kuhn puts it in his blog—

School is a closed-world domain—you are solving crisply-defined puzzles (multiply these two numbers, implement this algorithm, write a book report by this rubric), your solution is evaluated on one dimension (letter grade), and the performance ceiling (an A+) is low. The only form of progression is to take harder courses. If you try to maximize your rewards under this reward function, you’ll end up looking for trickier and trickier puzzles that you can get an A+ on.

The real world is the polar opposite. You’ll have some ultra-vague end goal, like “help people in sub-Saharan Africa solve their money problems,” based on which you’ll need to prioritize many different sub-problems. A solution’s performance has many different dimensions (speed, reliability, usability, repeatability, cost, …)—you probably don’t even know what all the dimensions are, let alone which are the most important. The range of plausible outcomes covers orders of magnitude and the ceiling is saving billions of lives. The habits you learn by working on problem sets won’t help you here.

Simply put, the diagnosis of why we have challenges to unlock a rapid acceleration in learning capabilities was because (a) it’s impossible to situate one teacher for every student and that’s not scalable, and (b) age-wise demarcation of students and teaching wrongly prioritises the learning for the average leaving out the smartest or the less brighter students who are already farther behind.

But with AI…

Particularly with language models, offers something genuinely new: scalable personalization. Which would mean that we could finally take a dig at solving the question which Patrick Collinson had posted.

And we’re already seeing examples — Take Duolingo. Its gamified language platform uses reinforcement learning to adapt lessons in real time, tracking what you struggle with and dynamically adjusting future content. It’s not just reactive—it’s personalized curriculum design at scale. And it aligns well with the principles of Mastery learning, you’re not in cohorts based on your age, you’re in a cohort based on your current skill level.

Or consider Khan Academy’s Khanmigo, an AI-powered tutor built on top of GPT-4. It provides Socratic questioning, step-by-step feedback, and individualized learning support that mimics human tutoring with surprising fidelity. It doesn’t replace teachers; it augments them, giving students instant, always-available guidance tailored to their needs. We have various other examples — Sana Labs, which uses machine learning to personalize corporate training, and Socratic by Google, which uses AI to help students solve homework problems by guiding them toward understanding rather than just giving answers.

These systems are still primitive compared to a human tutor—but unlike human tutors, they scale.

They’re always available, always patient, and—importantly—capable of operating at the student’s ideal zone of development, where challenge and skill are perfectly matched.

With this combination (hyperpersonalised AI-led tutoring, with mastery learning pedagogy), I think more students get to have more practise hours enduring a state of flow. Psychologist Mihaly Csikszentmihalyi described flow as the mental state where people are fully immersed in a task, balancing skill and challenge so precisely that time seems to disappear. It’s where the best learning happens.

Traditional education rarely produces flow. It’s optimized for efficiency, not engagement. But AI-driven personalized systems, by continuously tuning the difficulty and style of instruction, can engineer flow experiences. And they can do so for everyone, not just the fortunate few who get access to elite tutors or high-end private schools.

Perhaps the most promising shift isn’t just the individualization of pace, but of path. Traditional education forces every student down the same curriculum, in the same order. But AI systems can support exploration. A student fascinated by music theory can go deep on counterpoint and harmonic analysis while brushing up on algebra in parallel. Interest-driven learning no longer means abandoning core subjects; it means integrating them into the learner’s context.

The main drawback of the earlier systems were that they were solely to optimize test performance (SAT, IITJEE etc), and we ended up with a glorified Skinner box rather than a real learning tool. But with these new systems, that learns with you over years—tracking your strengths, surfacing connections between subjects, curating resources, asking probing questions when your curiosity spikes, and laying off when you’re tired or disengaged — That’s not a fantasy anymore.

We’re in early days, but the architecture is taking shape.