AI tutoring is a real and useful product category. Sal Khan’s Khanmigo and Synthesis’s tutor are both genuine attempts at Benjamin Bloom’s 1984 challenge of giving every kid one-on-one mastery instruction. The studio doesn’t compete with them. Tutoring optimizes for accelerating a kid through an existing curriculum. We optimize for something Seymour Papert called constructionism: the kid as the maker of new things, with AI as a collaborator, not an assessor. Most kids need both at different moments. We picked the one we’d be best at.
What AI tutoring is designed to do.
The reference point for almost every serious AI-tutor product is Benjamin Bloom’s 1984 paper, “The 2 Sigma Problem.” Bloom and his graduate students had run a series of studies comparing one-on-one mastery tutoring against conventional classroom instruction. The students who got one-on-one tutoring with mastery learning scored, on average, two standard deviations higher than the conventional group. Two sigma. The top 2 percent of the classroom curve became the median of the tutored group.1 Bloom’s problem was that you couldn’t give every kid a tutor. Tutors are expensive. School budgets aren’t built for it.
AI tutoring is the thirty-year-late attempt to solve that. If you can spin up a competent one-on-one tutor for each kid at near-zero marginal cost, the 2-sigma effect becomes a policy lever. Sal Khan has been making that argument since his 2011 TED talk on Khan Academy, and his 2024 book Brave New Words spells out the AI version of it: Khanmigo as a patient, Socratic tutor that walks each kid through an existing curriculum at their own pace.2 Synthesis is doing roughly the same shape with a different pedagogy, oriented around live problem-solving sessions.
What the tutor optimizes for is mastery of an existing body of material. The job is to take a kid who doesn’t know algebra and turn them into a kid who does. The curriculum is fixed and known in advance. The kid’s job is to absorb it. The tutor’s job is to find the gaps in the kid’s understanding, scaffold around them, and verify mastery before moving on. This is a real and valuable job. It works.
What AI tutoring doesn’t address.
There’s a different category of thing kids do, and the AI-tutor frame doesn’t describe it. When a kid sits down to make a game, write a story, or build a website, there is no fixed curriculum to master. The artifact doesn’t exist yet. What needs to be learned is whatever turns out to be in the way of finishing the thing. The kid isn’t a student receiving knowledge. They’re the author of a project that doesn’t yet exist, with a half-formed taste guiding them and an unknown number of unknown problems ahead.
In that frame, AI as tutor is the wrong shape. A tutor evaluates. The kid is being measured against a known answer. But there’s no known answer to “does this game feel scary enough.” That call belongs to the kid. The tutor’s default move (here’s a hint, can you try again, let me explain the underlying concept) doesn’t help when the question isn’t about a concept. It’s about the kid’s own intent.
The other thing tutoring is built around is private practice. The kid and the tutor work through problems. There’s no audience. There’s no published artifact. The output of a tutoring session is the kid’s improved understanding, which is invisible until it shows up on a test. For some learning, that’s right. For making, it’s wrong. The thing that makes a kid’s game real is that someone else plays it. Mastery of a curriculum doesn’t ship to a URL.
A tutor measures the kid against a known answer. There’s no known answer to “does this game feel scary enough.” On where the tutoring frame stops fitting
None of this is a criticism of Khan or Synthesis. They’re solving a different problem and they’re solving it well. The criticism is of the category-mistake where every AI-for-kids product gets rolled up under “tutoring” and judged by the same yardstick. The studio is not trying to be a better tutor. It’s trying to be a different kind of thing.
The constructionist alternative.
The frame for what we are building has a name. Seymour Papert introduced it in his 1980 book Mindstorms, then elaborated it in a 1991 collection with Idit Harel: Constructionism.34 Papert had been Jean Piaget’s student in the 1950s. He took Piaget’s idea that kids build their knowledge through their own mental constructions and added one move: kids learn most durably when they’re building something shareable. A castle in the sand, a robot, a program, a game. The thing the kid makes becomes the scaffold for the thing the kid learns.
The implication for curriculum is unusual. In constructionism, the curriculum is whatever the kid happens to need to know in order to finish their project. There is no pre-fixed sequence. A kid making a side-scroller needs to figure out collision detection. The collision-detection lesson lands because the kid’s sprite is currently walking through the wall, and they want it to stop. The motivation is the unfinished artifact. The teaching is whatever clears the next blocker.
Mitchel Resnick has spent his career at MIT’s Lifelong Kindergarten group continuing this work, building Scratch and articulating the “four Ps” (Projects, Passion, Peers, Play) as the design grammar of constructionist environments.5 Yasmin Kafai and Quinn Burke’s Connected Code extends the same argument to networked, shareable making.6 The body of work is forty-five years deep. The studio is built on top of it.
In the constructionist frame, AI is not the assessor. It’s a collaborator the kid steers. It proposes changes. The kid decides what stays. The kid’s job isn’t to demonstrate mastery of a known body of knowledge. It’s to build something they care about, and to develop the working theory of model behavior, the deciding-not-accepting habit, and the iteration loop that we’ve written about at against prompt engineering. None of those are tutoring goals. All of them are durable.
What the difference looks like in product form.
The clearest way to see the difference is to compare the surface the kid is looking at when the AI says something. An AI tutor’s surface is a chat where the AI explains, asks Socratic questions, evaluates the kid’s response, and decides whether mastery has been reached. The studio’s surface is the ChangeDisclosure card: here is the change Inkie proposes, here is what file it touches, decide.
The surface tells you what the product believes the kid is. A tutor surface treats the kid as a learner being evaluated. A partner surface treats them as a maker being supported. Those are different products. Both are reasonable. The choice between them is a choice about what kind of thing you want your kid to be doing in the third hour of using the software.
When each is right.
The honest answer to the “which AI product” question is that most kids need both. Tutoring is right when there’s a known body of material that has to be learned. The kid is behind in algebra. The kid is preparing for a placement test. The kid wants to read Spanish by summer. Those are real problems with known answers. A patient AI tutor that walks the kid through mastery is a better answer to them than anything we ship.
Constructionism is right when the question is “I want to make a thing.” A game. A story. A site for the kid’s soccer team. A short film. There is no curriculum to master. The thing has to be invented and shipped. The kid is the author. The AI is the partner who proposes moves the kid can accept, modify, or reject. Different shape, different product.
What we’ve noticed in cohort is that the families who succeed with the studio have usually figured this out. Their kid does Khan in the morning when they need to learn fractions. They open the studio in the afternoon when they want to build a side-scroller. Nobody confuses the two. The tutor isn’t failing when it doesn’t help with the side-scroller. The studio isn’t failing when it doesn’t teach fractions. They’re different jobs. Run by different software. With different design assumptions about what the kid is for.
We picked constructionism because the pedagogy is forty-five years old, well-studied, and underbuilt in the AI era. Most of the AI-for-kids capital is going to tutoring. Almost none is going to constructionist tools that take AI seriously as a partner instead of an assessor. The studio is our bet on that gap. If you want to feel the partner-not-tutor distinction in your hands, the mini-studios at tellandshow.ai/try take about three minutes. The thing that’s different is the question the AI is trying to answer.
References
- Benjamin S. Bloom, “The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring,” Educational Researcher, vol. 13, no. 6, 1984. The foundational paper on the mastery-learning effect.
- Salman Khan, Brave New Words: How AI Will Revolutionize Education (and Why That’s a Good Thing), Viking, 2024. The book-length statement of the Khanmigo argument; the 2023 TED talk preceded it.
- Seymour Papert, Mindstorms: Children, Computers, and Powerful Ideas, Basic Books, 1980. The founding statement of constructionist thinking, written at the MIT Media Lab.
- Idit Harel & Seymour Papert (eds.), Constructionism, Ablex, 1991. The collection where the term gets its full theoretical treatment.
- Mitchel Resnick, Lifelong Kindergarten: Cultivating Creativity through Projects, Passion, Peers, and Play, MIT Press, 2017. The continuation of the constructionist line of work at media.mit.edu/groups/lifelong-kindergarten.
- Yasmin Kafai & Quinn Burke, Connected Code: Why Children Need to Learn Programming, MIT Press, 2014. On constructionism applied to networked, shareable making.