> ## Content Index
> Fetch the complete content index at: https://www.hardfun.org/llms.txt
> Use this file to discover other available public pages before exploring further.

# When Small is Beautiful
- URL: https://www.hardfun.org/when-small-is-beautiful/
- Published: 2026-11-17T15:52:00.000Z
- Updated: 2026-11-17T15:51:59.000Z
- Description: Schumacher’s 1968 question about teaching machines is the right one to ask about AI tutors and class size. Four questions before anyone argues about the ratio or the tutor.
- Author: Greg Hayes
- Tags: education, News, video games

I just got home from a meeting of independent school heads and trustees in Tacoma, Washington. The consensus in the room, from the people closest to classrooms, was that smaller classes are better for kids. I mostly agree. The qualifications are what follows and they matter more as AI learning tools get more capable and more common.

The argument about AI tutors is being held in the wrong order. Whether a machine teaches something better than a person depends on what the something is and almost nobody says what it is before they start arguing.

I’ve watched school leaders and parents argue about class size as if the number came first. Boards love a ratio too. It fits neatly on a slide, which rarely says what the teacher needs to do.

Game studios settle a version of this argument before they ship. When Blizzard cut Overwatch from six players a side to five, its game director said players could not keep track of eleven others on the field. Among Us, a slower game of talk and suspicion, went the other way in 2021 and raised its lobby from ten players to fifteen. Nobody who has shipped a game believes the right party size is the same for every game. The number comes from what the players have to do, what the hardware can carry and how much one person can keep track of.

The AI-teaching debate is not new. In 1968 an economist made this point about teaching in a London lecture he later folded into a book most people remember only for its pithy title. E.F. Schumacher’s *Small Is Beautiful* is about economics. The machines of his day were the behaviorist B.F. Skinner’s teaching machines and a university Britain was setting up to teach by television and by mail.

Schumacher found the debate over whether they taught better than people beside the point until someone answered a prior question. “What are we trying to teach?” Some things, he wrote, can only be taught in a very intimate circle. Others can be taught en masse, by broadcast or by machine. The more active and intimate the activity, the fewer people can take part in it. A twitchy game like Overwatch wants a smaller group for the same reason a slower, talkier game like Among Us can hold fifteen.

I’ve spent my career examining and engineering attention in games and media. Now I also help govern a school. From either perspective, goals and constraints come before the design.

**What are we trying to teach?**

Schumacher answers his own question just one chapter later. He sorted problems into two kinds. Convergent problems have answers that converge the harder you work on them. *Solve for x. Conjugate the verb.* Divergent problems do the opposite. The harder you think, the further apart the thoughtful answers get, because the problem is built out of opposites that have to be held at once. His example of a divergent problem was how to educate children. Discipline or freedom. Structure or room. Nobody solves that one. Teachers live it, in a room, with particular kids.

Judgment is the faculty for divergent problems. It is what knowledge becomes after enough practice and the practice isn’t on the worksheet. *Is this worth finishing? Do I ask for help or sit with it for ten more minutes?* There is no answer key for those, for the machine or for the teacher. What the teacher has is the kid in front of them.

The convergent part can go to the lecture hall or the machine and a well-built machine might carry a good deal of it, as the LeapPad did for phonics and Khan Academy does for algebra.

**The order of operations**

The best evidence that the order matters comes from people worried about AI in schools. Last year a team at the MIT Media Lab had fifty-four people write short essays over four months, some with ChatGPT and some with nothing but their own heads, while an EEG cap recorded how hard they were working. Then it switched eighteen of them. The ones who had written unaided and then got the machine remembered what they had written and kept thinking. The ones who had started with the machine and then lost it went quiet on the EEG and struggled to quote a sentence from their own essays. The study is a preprint with a small sample and the authors have asked journalists to stop saying it shows AI makes people dumber. EEG measures engagement, not learning. What the study suggests is narrower, but telling. The machine helped people who already knew how to do the work and set back people who were still learning to.

Eric Horvitz, Microsoft’s chief scientific officer, made the same point from inside the industry on the Alliance for Decision Education’s podcast in September. The most powerful users of today’s AI tools he has seen are people who already know their field well. Put that next to Schumacher and the sequencing is clear. The convergent learning can be amplified by a well-built machine. Learning it still happens to a particular kid in a particular room, with a teacher noticing whether it took.

**The ratio is a proxy**

[Last episode](https://www.hardfun.org/off-task/) I argued that a kid drifting off the assignment sometimes carries valuable information and that a teacher can only follow their detour when they have the space for it, which mostly means time. Here is the rest of that thought. Following a detour means noticing it first. Noticing is the most active and intimate thing a teacher does and it has a scale, the way a squad does in a game.

**Nobody can notice thirty-two detours**

That is what the student-to-teacher ratio really is for. Not better scores. Noticing. The student-to-teacher number schools print is a proxy for whether an adult who knows a particular kid can see what that kid is doing with their time. Attention is the scarce resource in a game and in a classroom. 

Daniel Kahneman (*Thinking, Fast and Slow*) and Gary Klein (*The Power of Intuition*) came to this question from opposite camps and wrote a paper together anyway. They agreed on when a person’s judgment can be trusted. The situation has to be regular enough to learn and the person has to have found out, quickly and truthfully, when they were wrong. A good game supplies both by design. The divergent problems of a kid’s life supply neither. Nothing tells a fourteen-year-old promptly or honestly whether the thing they quit was worth finishing. Somebody has to. The noticing adult is the feedback for problems with no answer key. The student-to-teacher ratio is how much of it there is to go around.

This is why small is the wrong word. Schumacher said as much. He wrote that the modern world suffered from an idolatry of giantism and added that if it suffered from an idolatry of smallness he would argue the other way.

Phillips Exeter has taught around a table that seats twelve since the early 1930s and the table is oval rather than round so the teacher can see every face. Noticing is not the only thing a circle that size is for. What it does to character development will be explored in a later episode. For now, the right size belongs to the activity.

**Small is beautiful when it is for something. The argument is about what**

**Test scores and what they miss**

The evidence that small classes alone raise test scores is modest and contested and the argument doesn’t need it to be more. Tennessee ran the largest randomized experiment on class size American schools have seen. Eleven thousand six hundred children in the early grades were assigned by lot to classes of thirteen to seventeen or twenty-two to twenty-five and the small classes pulled ahead, more so for poorer kids. Economists have argued about the size and durability of the effect ever since. The Gates Foundation spent close to two billion dollars over nine years making high schools smaller and Bill Gates wrote in 2009 that the effort had fallen short. A lottery-based study of New York’s small high schools later found higher graduation rates. Both things are true.

The most incisive objection comes from Dylan Wiliam, who has spent a career on classroom assessment. Cutting class size, he argues, is the most expensive thing a school can do and it dilutes teacher quality because somebody has to staff the new rooms. A teacher who asks the right question and looks at thirty answers on thirty whiteboards learns what the room knows in a minute, no monitoring of individuals required. He is right about the convergent part and his method is itself a noticing technology. It makes thirty answers visible at once. What it cannot do is follow one kid’s detour.

**Scores measure the convergent part. Small is for the divergent part, which the tests don’t see**

A classroom of thirty can’t be a circle of twelve. And importantly, learning to carry your part in a common task, which kids experience on a soccer field, in a band, in a theater troupe and on an Overwatch team, happens every time a kid chooses a challenge that nobody assigned.

If a machine carries some of the convergent part, the time it hands back is available for the divergent part. In two of the largest trials to date, a two-year study of Khan Academy’s tutor in eighteen Tennessee middle schools and a Maryland experiment with more than two thousand undergraduates, most students barely touched the machine and the gains were small. In Maryland grades fell. The Tennessee researchers concluded that engagement was the binding constraint. Somebody still has to notice when the kid never opened the tool. 

**Four questions before anyone argues about class size**

[Two episodes](https://www.hardfun.org/the-essay-was-never-the-point/) ago I listed six questions to ask before a learning tool enters a classroom. These belong in front of them.

1\. What are we trying to teach and does it have an answer that converges?

2\. Does learning it depend on an adult noticing what this particular kid is doing?

3\. How many kids can that adult notice at once?

4\. If a machine is about to take some of the convergent part, where is the time it gives back going?

Overwatch is five a side because the designers found that eleven other players was more than one person could keep track of. The number was never the point. The activity was.

If you teach, run a school, sit on a board or make games, I’d like to hear from you. What in your building can only be taught in a small circle and how small does the circle have to be?

**Further reading:** E.F. Schumacher, *Small Is Beautiful: Economics as if People Mattered* (1973), especially “A Question of Size” and “The Greatest Resource: Education”; Schumacher Center for a New Economics, [“Writings on Issues of Scale by E.F. Schumacher”](https://centerforneweconomics.org/publications/writings-on-issues-of-scale-by-e-f-schumacher/?ref=hardfun.org); Nataliya Kosmyna et al., [“Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”](https://arxiv.org/abs/2506.08872?ref=hardfun.org), MIT Media Lab preprint (2025); Eric Horvitz with Annie Duke, [“Human Cognition in the AI Era,”](https://alliancefordecisioneducation.org/podcasts/episode-045-human-cognition-in-the-ai-era-with-dr-eric-horvitz/?ref=hardfun.org) The Decision Education Podcast, Episode 045 (2026); Daniel Kahneman and Gary Klein, “Conditions for Intuitive Expertise: A Failure to Disagree,” *American Psychologist* (2009); Dylan Wiliam, *Creating the Schools Our Children Need* (2018); Alan Krueger, [“Experimental Estimates of Education Production Functions”](https://www.nber.org/papers/w6051?ref=hardfun.org), NBER (1997); Philip Oreopoulos and Hannah Low, [“One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment”](https://www.nber.org/papers/w35620?ref=hardfun.org), NBER (2026); Jing Liu et al., [“The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement”](https://docs.iza.org/dp18989.pdf?ref=hardfun.org), IZA (2026); Phillips Exeter Academy, [“Harkness”](https://exeter.edu/node/16?ref=hardfun.org); Video Games Chronicle, [“Overwatch 2 will reduce PvP team sizes to 5v5”](https://www.videogameschronicle.com/news/overwatch-2-will-reduce-pvp-team-sizes-to-5v5-instead-of-6v6/?ref=hardfun.org) (2021); Innersloth, [“Among Us’ 3rd Birthday & 15 Player Lobbies!”](https://innersloth.com/among-us-3rd-birthday-15-player-lobbies?ref=hardfun.org) (2021); Education Week, [“Bill Gates Praises Charter Models”](https://edweek.org/policy-politics/bill-gates-praises-charter-models/2009/02?ref=hardfun.org) (2009); Sarah Garland, [“Small schools, big results?”](https://hechingerreport.org/?p=7492&ref=hardfun.org), The Hechinger Report (2012).