← back to problem page
Whitepaper · v0.1·progress & abundance·open to refutation

An optimism.fun request for startups

Pedagogy at scale

One-on-one tutoring works far better than classrooms. AI may make it affordable for every child.

Published

2026-04-24

Authors

optimism.fun

Status

Draft · v0.1

License

CC BY 4.0

§0weekly drop · week 1

Editor’s drop

Published 2026-08-20 · optimism.fun

blackpaper · the problem

Benjamin Bloom measured it in 1984: a student who gets one-on-one tutoring instead of a normal classroom performs two standard deviations better. That is the difference between an average student and a student in the 98th percentile. Bloom called this "the 2 sigma problem" because the tutoring worked and nobody could afford to give it to every child. One tutor per student does not scale. It never has. The result is what the World Bank calls "learning poverty": the share of 10-year-olds in low- and middle-income countries who cannot read and understand a simple sentence. Before the pandemic that number was 57%. After two years of closed schools it rose to roughly 70%. Seven in ten children in the countries measured cannot do the single thing school is supposed to teach first. UNESCO's Global Education Monitoring Report puts the number of undereducated children worldwide at 1.5 billion. That is not a rounding error inside the 2.4 billion school-age children on Earth, it is most of them. Layer on the adults who never got the schooling their labor market now requires, and the population affected runs into the billions. The World Bank's meta-analyses of returns to schooling put the earnings cost at roughly 20% of lifetime income for a child who does not reach basic literacy and numeracy on schedule. Multiply a fifth of a lifetime of earnings across a billion and a half children and the number is not measured in millions or billions, it compounds across generations, because undereducated parents produce the next cohort of undereducated children. The world spends about $6 trillion a year on education (UNESCO). That is not a funding problem in the naive sense, it already ranks among the largest expenditure categories on the planet, ahead of most countries' entire GDP. The problem is what the money buys: it mostly buys teacher salaries and classroom seats delivering the same fixed-ratio, one-to-thirty instruction that Bloom's research already showed is two standard deviations short of what is achievable. The $400 billion or so of that spend classified as EdTech (World Bank) is a rounding error on the total and has not moved the learning-poverty number in the direction it needed to move. This is a solved problem in the narrow technical sense, we know exactly what works, and a completely unsolved problem in the delivery sense, because the thing that works has never been affordable at population scale. Every other quest on this site assumes a population capable of acquiring, applying, and building on explanations. This is the quest that determines whether that population exists.

whitepaper · the proposal

The unlock is not a new pedagogical theory. Bloom's finding has stood for forty years. The unlock is that a large language model can now hold a patient, adaptive, infinitely repeatable one-on-one conversation with a student at a marginal cost near zero, which is the one input Bloom's method needed and never had. What would actually have to be built: not a chatbot bolted onto a curriculum, but a tutoring system that (1) diagnoses what a specific student does and does not understand from their actual answers rather than a fixed lesson plan, (2) explains a concept multiple different ways until one lands, the way a good human tutor does, (3) holds enough state across sessions to know a student's history rather than starting cold every time, and (4) is honest when it does not know something rather than confidently wrong, which matters more in a tutoring context than almost any other AI application. Khan Academy's Khanmigo is the most visible attempt at this today; it is real but early, and the learning-poverty numbers have not moved yet, which is the honest caveat here, not a settled success story. The buildable-now case rests on three things converging at the same time: LLMs cheap enough to run a full tutoring session for cents rather than dollars, enough evaluation infrastructure (learning-gain measurement, not just engagement metrics) to tell a tutor that works from one that just feels good, and enough device penetration in low- and middle-income countries (smartphones, not computers) to reach the population that needs it most, which was not true a decade ago and is true now. **The market, worked from real numbers.** UNESCO puts the number of undereducated children at 1.5 billion (Global Education Monitoring Report, https://www.unesco.org/gem-report/en). The realistic price point is not the $6 trillion global education budget, that money is locked into salaries and buildings a tutoring product does not replace. The realistic comparable is Khan Academy's own Khanmigo enterprise pricing: districts buy in starting around $10 per student per year at scale (2026 pricing across multiple vendor comparisons), separate from a $4/month ($44/year) individual-family tier that is not the scale price point. 1,500,000,000 children x $10/student/year (district-scale AI tutoring, Khanmigo enterprise floor) = $15,000,000,000/year. That is a $15 billion/year addressable market for AI tutoring specifically, not the $6 trillion education-spend headline this site's problem page cites for context, and not the individual-consumer $44/year tier, which would imply a larger but less realistic $66 billion figure that assumes every affected child's family can pay directly rather than a government or district paying at bulk rates. $15 billion/year is the conservative, defensible number: it is what it costs to put a real AI tutor, at real enterprise pricing, in front of every undereducated child on Earth, once, per year. Two honest uncertainties. First, $10/student/year is a floor price observed in wealthy-country enterprise deals; delivering the same service in the lowest-income markets where the 1.5 billion figure is concentrated will likely require a lower price point or subsidized delivery, which would shrink the market-size figure and increase the capital-required figure simultaneously. Second, nobody has yet shown at national scale that an AI tutor moves a learning-poverty statistic, only that it moves individual learning-gain studies; the entire premise depends on that translating from lab-scale to population-scale, which is a real if, not a settled fact. Conjecture, open to refutation: the mechanism (Bloom's 2-sigma effect, now deliverable at near-zero marginal cost) is well established; the claim that deploying it changes the world number, is not yet proven and is exactly the kind of claim this site exists to track as it resolves.
§1abstract

The four-axis ranking

We rank humanity’s most important problems on four quantifiable dimensions — quantity of humans affected, severity per person, current solution quality, and addressable market size — and package each as a proposal in the spirit of Musk’s Hyperloop Alpha. This document is the proposal for pedagogy at scale. Every number below is sourced and tagged with confidence. Every ranking is a conjecture, open to refutation.

Quantity · humans affected

1.5B

high

Severity · 0–1

unknown

Current solutions

3.0 / 10

med

Market size · TAM

$6.0T

high
§2problem statement

What we are trying to solve

Benjamin Bloom's 1984 research showed one-on-one tutoring raises student performance by 2 standard deviations versus classroom instruction. The problem was cost: one tutor per child is infeasible at population scale. AI tutoring changes the economics. 1.5B children are undereducated globally (UNESCO); billions more adults lack skills they could have acquired with better pedagogy. This is the problem Deutsch writes about most directly in The Beginning of Infinity, better explanations compound across generations.

§3why it persists

The gap between the world and the world that is physically possible

Today: 1:1 human tutoring produces +2σ outcomes vs classroom (Bloom 1984) but is unaffordable for most. Median schooling under-delivers; global learning poverty ~70% in low-income countries.

Current solution quality is rated 3.0 / 10 (med confidence) — meaning there is substantial unclaimed ground between what exists and what is possible. estimated — 1:1 tutoring (+2σ per Bloom 1984) remains gold standard but unaffordable for most; LLM tutors emerging but unproven at scale.

§4existing alternatives

Who is already working on this

5 entities are currently working on this problem across public markets, private companies, and research orgs. Each is evidence the market is real; none has obviously solved it.

Duolingo

public · USA

Language learning reaching 100M+ monthly users. Expanded into math, music, and literacy.

$14.5B

Replit

private · USA

Browser-based coding platform. 100M+ learners worldwide; AI agent for app-building compresses learn-to-ship cycle.

$3.0B

Speak

private · USA

AI conversational tutor for language learning, unlimited speaking practice at marginal cost.

$1.0B

Synthesis

private · USA

AI tutor for kids (math, reading). Attempting to close the 2-sigma gap for ages 7-14.

undisclosed

§5proposed direction

If we solve this, here is the world we get

After · 10 years

Every learner has access to AI tutoring matching the best human tutors, in any language, at near-zero cost. Bloom 2-sigma becomes the default, not the exception.

Requests for startups · 2 concrete companies to build

The Bloom 2-sigma tutor

1:1 tutoring beats classroom instruction by two standard deviations and has been unaffordable for 4,000 years. Build the AI tutor that delivers it for the marginal cost of inference, in any language.

why now
Models crossed the threshold where they can diagnose a misconception and adapt — not just answer — at conversational latency and cost.
shape
A tutoring product (not a chatbot) with a real pedagogical model: mastery tracking, spaced retrieval, Socratic dialog, and outcome measurement against standardized gains.
success
A median student using it gains the Bloom 2-sigma effect, reproducibly, on independent assessments.

Credentialing that routes around the diploma

Employers want proof of skill; the degree is a $1.5T noisy proxy. Build the verifiable, adversarially-robust skills-assessment standard that employers actually trust.

why now
AI both threatens take-home assessment integrity and enables proctored, adaptive, cheating-resistant evaluation at scale.
shape
An assessment + credential issuer with employer-side validity studies, designed from day one to resist AI-assisted cheating and to be portable across employers.
success
A meaningful share of hiring uses the credential as a primary signal, decoupling opportunity from tuition.

full rubric + framing on the Requests for Startups page.

§6cost & scale

What the market can pay

The world is already paying $6.0T per year against this problem (global education spend (UNESCO Global Education Monitoring Report); EdTech slice ~$400B and growing; high confidence).

A successful solution does not need to capture more — it needs to redirect a meaningful slice of existing spend, plus the latent willingness-to-pay implied by the severity score above. The cost ceiling for a real solution is bounded by this number; everything cheaper is dominated, everything more expensive is a non-starter.

§7safety & considerations

What could go wrong, and how we know we are not wrong

Section in progress

Failure modes, ethical considerations, and the conditions under which this whitepaper would be falsified are being authored as the weekly cadence ships. The Deutschian commitment: every claim above is a conjecture; we publish the conditions under which we would update. Unpublished sections stay empty. A Monday newsletter is not promised. Subscribe to get the upgrade, or contribute on GitHub.

§8suggested investors

Who would back this

Capital allocators with a stated thesis or deployed portfolio in this domain. This is a starting list — Exa Websets enrichment will expand it to direct check-writers per company.

Grant

Emergent Ventures

Fast grants. High-variance, unconventional, talent-first.

Grant

1517 Fund · Medici Project

Minimum $1k, no strings, no equity. A five-minute video is the whole application.

Fellowship

O'Shaughnessy Fellowships

$100k for creators across domains, art, science, history, startups.

Venture Capital

Rise Fund (TPG)

Growth-stage impact investing with rigorous Impact Multiple of Money scoring.

§9voices

What the thinkers say

“Bad explanations waste billions of childhood hours. Good explanations compound across generations. Pedagogy is an epistemology problem.”

— David Deutsch · Physicist & Philosopher

“Asks whether we can dramatically improve the quality and speed of learning. Cites tutoring and personalized instruction as high-leverage unsolved problems.”

— Patrick Collison · Co-founder & CEO, Stripe

“Education quality is the compounding lever. Progress requires transmitting knowledge at higher fidelity and lower cost to each successive generation.”

— Jason Crawford · Founder

“Finding and developing talent is an underrated problem. Most selection is bad, most mentorship is worse, and the tails of ability are where civilizational value compounds.”

— Tyler Cowen · Economist & Writer
§10sources & criticism invite

Where this is wrong, tell us

Every number on this page carries a source and a confidence tag. Every section open to refutation. If a citation is wrong, a number is stale, or a conjecture is unfounded — file a correction.

corrections → use the feedback widget in the nav · open issue at github.com/adamtpang/optimism.fun

Published whitepapers

One problem.Published essays, not a Monday promise.

The essays that exist are listed at /whitepapers. A new essay is not promised every Monday. Weeks with no essay are empty on purpose. Nothing here invents a back catalog.

optimism.fun uses your email address only to send low-volume project updates through Resend. You can unsubscribe with one click at any time; read the privacy policy.