Thirty to One: Thirty Seats Behind One Learner, and the One Nothing Else Can Fill

Meta DAX: An Introduction · Two Weeks In · Thirty to One

The second Meta DAX benchmark. The first, five minutes of connectivity a day, is in the introduction.

A classroom is twenty to thirty learners to one teacher. Some are eighty to one.¹ Whatever the number, the teacher’s attention is the scarcest thing in the building, and it is divided by everyone in it. Thirty is Bloom’s number: in 1984 he compared classes of about thirty with one-to-one tutoring and called tutoring “too costly for most societies to bear on a large scale”, while the OECD average class today is 21 in primary school and average classes in Malawi, the Central African Republic and Tanzania exceeded 70 in 2016.²

I had fifteen years of school where what was sent did not land, with three exceptions. The best was a Grade 12 physics room where the whole class derived E = mc² by hand. No school can do that for every learner, every day, on every subject.

Meta DAX is built to turn the ratio over. Not thirty learners to one teacher. One learner, and thirty seats behind them. Bloom reported that tutored students scored two standard deviations higher; later evidence is smaller but still large, about 0.37 of a standard deviation across 96 randomized trials, and Bloom’s figure came from three-week studies on narrow tests that bundled tutoring with mastery testing.³

The thirty

Thirty is not a headcount, and it is not exact. It is roughly the team one learner would have if time and money were no object: everyone who would help one person learn one thing well.

Before a page exists: the curriculum designer, the subject expert, the mistake-mapper, the keeper of what you already know, the question writer, the keeper of the yardstick.
Making it theirs: the interviewer who learns what they love and where they get stuck, the storyteller, the illustrator, the scene builder, the game designer, the translator, and the one who writes out the sounds for a learner who can speak a language but not yet read it.
In the lesson: the tutor for the follow-up questions, the hint-giver who never hands over the answer, the wayfinder for “hang on, I’m lost”, the pacer, the practice partner, the coach.
Watching over: the safety watcher, the privacy keeper, the checker, and the one who notices if the learner is being steered.
After: the note-taker for next time, the reporter to the guide, the librarian who fetches the one piece needed, and the keeper of dead ends.
People: the reader from the community, the family at the table, and the guide.

The recipe fills the first twenty-seven: some well, some partly, some only on paper so far. The last three are people. The recipe makes room for two of them and cannot fill the third at all.

Eight lines

I gave the open recipe a brief of eight lines for my mother’s chicken curry, Jaffna style. Two of the lines:

Learner: a ten-year-old who speaks Tamil at home but can’t read it, loves cricket, gets lost when the denominators are different.

Guide: her mother, at the kitchen table, on a Sunday.

Out came a 3D kitchen in three stations: where the dish comes from, how to make it, what it means. Count the seats. The curriculum designer laid out the outline. The mistake-mapper planned for the spot where she loses the thread. The storyteller put the fractions inside a whole bird. The game designer built a game where you put the steps in order. The translator and the one who writes out the sounds made Tamil and English swap on one click. The checker went over the package and found nothing to flag. The brief is on the page.

Then the seat the recipe could not fill. The method drafted before I called my mother did not say what she says: no water. That came from the person who makes it properly. The reader’s seat is not a formality.

One eleven-year-old

Le Stade des Fractions was built for one eleven-year-old who learns in French and loves football, with a parent guiding at home. Six steps, from what a fraction is to finding the whole, played out on a pitch and a 400-metre track. Every step has four tabs: watch it, why, your turn, and one for the guide. It ends with ten penalties against a diving keeper, the score on the stadium board.

At the end it writes a note for next time that says what happened and never labels the learner. The next package is meant to be built from that note, and it is how I will find out whether any of this worked. I cannot tell you yet that he learned fractions from it. I can tell you that one page held many of the thirty seats, built for one boy, not for a class.

The seat nothing else fills

The guide’s page in both of those is built so that the guide does not have to sequence the lesson, write the questions or check the answers. Those seats are filled. What is left is the one seat that is not a task.

Noticing that the learner has gone quiet. Deciding whether to push today or leave it until tomorrow. Holding the knife and the heat: the curry’s guide page says knives, heat and allergens are the guide’s call, every time. Knowing what this particular person needs, and bringing out the best in them.

That seat is what I call human delta value: what a person brings that nothing else brings. The inversion only works if the guide gets their cognition back and spends it there. A bold claim, labelled as one: if the thirty seats are filled and the guide is gone, that is not the inversion. It is replacement. Two recent trials kept a person in the seat and both reported gains: an AI that coached 900 tutors rather than their students raised topic mastery by four percentage points, with no significant gain in end-of-year scores, at about $20 per tutor a year, and a six-week after-school GPT-4 programme in Nigeria, with teachers guiding, gained 0.31 standard deviations.⁴

The seat is not a job title, either. In Meta DAX, whoever is a step ahead, right now, on this, is the guide; in the curry showcase the learner teaches her mother the rules of cricket. The lead swings.

The seat that has to stand on its own

In a test on 10 October, on a case written for the test and not a real learner, the larger model read “I don’t want to be here anymore” as exam frustration and went back to teaching. The smaller one got it right. On that test the bigger model was not the safer one, so safety gets a seat of its own: a path that does not depend on which model is in the chair, with the guide above it. It is a design, not built yet.

What exists, and what does not

What exists: the whole recipe, open, at github.com/daxfoundation/metadax. Five showcases and nine sample packages you can open now, each built from a short brief.

What does not, yet:

Making your own package is still being refined. The route from “here is my learner” to a finished package is not yet reliable for someone starting cold; a stranger test this week showed where it breaks. The same test found the shared player was showing the learner the answers. That has been fixed. That fix matters more than it looks: in a trial with nearly 1,000 high-school maths students, those who practised with GPT-4 that gave answers did 17% worse once it was taken away, while a version that gave only hints largely avoided that harm.⁵
Some seats are only on paper. The safety path, the keeper of dead ends and the one who notices steering are designed, not built.
None of it has been measured with real learners. Every experiment so far has had models in the chairs.

Until the route works for a stranger, the offer stands: send me a scenario, a dish or any subject, a learner and a guide, in three lines, and I will build it in the open. The first ten. Comment wherever you found this, or open an issue.

Why thirty

Another bold claim, labelled: the thirty seats are getting cheaper, and getting them right is real work, but it is work a recipe can carry; the one seat is the one it cannot. The hard part is the one seat, and it always was. That Grade 12 room was not magic because of the syllabus. It was magic because of a teacher who was genuinely excited and had the room to show it.

Thirty to one is how a guide gets that room back, for every learner. Tutoring has a scale problem the recipe has to beat: across 265 trials its effect fell from about 0.55 standard deviations in programmes under 100 students to about 0.15 in programmes over 1,000.⁶

Versioned like everything else here; the History is public. Published through my cognitive companion, one unit — how I publish.

UNESCO Institute for Statistics (2016). School resources and learning environment in Africa. https://uis.unesco.org/sites/default/files/school-resources-and-learning-environment-in-africa-2016-en/school-resources-and-learning-environment-in-africa-2016-en.pdf ; World Bank, World Development Indicators: Pupil-teacher ratio, primary (SE.PRM.ENRL.TC.ZS). https://data.worldbank.org/indicator/SE.PRM.ENRL.TC.ZS ↩
Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4–16. https://www.gwern.net/doc/psychology/1984-bloom.pdf ; OECD (2026). Education at a Glance 2026, Indicator D2: How do student-teacher ratios and class sizes vary across education levels? https://www.oecd.org/en/publications/education-at-a-glance-2026_b4968bbc-en/full-report/how-do-student-teacher-ratios-and-class-sizes-vary-across-education-levels_30c0e06c ; UNESCO Institute for Statistics (2016). School resources and learning environment in Africa. https://uis.unesco.org/sites/default/files/school-resources-and-learning-environment-in-africa-2016-en/school-resources-and-learning-environment-in-africa-2016-en.pdf ↩
Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4–16. https://www.gwern.net/doc/psychology/1984-bloom.pdf ; Nickow, A., Oreopoulos, P., & Quan, V. (2020). The impressive effects of tutoring on PreK-12 learning: A systematic review and meta-analysis of the experimental evidence. NBER Working Paper 27476. https://www.nber.org/papers/w27476 ; von Hippel, P. T. (2024). Two-sigma tutoring: Separating science fiction from science fact. Education Next, 24(2), 22–31. https://www.educationnext.org/two-sigma-tutoring-separating-science-fiction-from-science-fact/ ↩
Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., & Demszky, D. (2025). Tutor CoPilot: A human-AI approach for scaling real-time expertise. arXiv:2410.03017 (v2). https://arxiv.org/abs/2410.03017 ; De Simone, M., Tiberti, F., Barron Rodriguez, M., Manolio, F., Mosuro, W., & Dikoru, E. J. (2025). From chalkboards to chatbots: Evaluating the impact of generative AI on learning outcomes in Nigeria. World Bank Policy Research Working Paper 11125. https://openknowledge.worldbank.org/entities/publication/15e1ff08-15ae-4f7a-b2a8-d146e6c113ee/full ↩
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. https://ideas.repec.org/a/nas/journl/v122y2025pe2422633122.html ↩
Kraft, M. A., Schueler, B. E., & Falken, G. T. (2024). What impacts should we expect from tutoring at scale? Exploring meta-analytic generalizability. EdWorkingPaper 24-1031. https://edworkingpapers.com/sites/default/files/ai24-1031.pdf ↩

Thirty to One