Private beta · invites in small batches

Name a topic. Get a course you can touch.

Ark One writes the course, then hands you the scene: a graph you drag, a circle you turn, a board you move. You answer by doing, and it checks every answer against the scene itself.

The list is the only way in right now. We write the moment your batch opens.

Scroll

Try a level.

Three questions from a maths course, run the way the app runs them: move the scene, commit, get checked. Nothing to type.

  1. 1 It plans the course Chapters and levels, each around one small win.
  2. 2 It builds the scenes A graph, a circle, a board, cast for the exact question.
  3. 3 It checks every answer Against the scene's own engine, before you ever see it.

Whatever it is you're stuck on.

There is no catalogue to run out of. Name the thing, in your words, and the course is built for that.

Five minutes at a time.

Levels stay short. Each one opens by teaching and ends on a single small win, so you finish something every time you pick up your phone.

The best way to learn was never a secret.

One teacher, one student. Everyone who has ever taught knows it. The problem was only ever that there weren't enough teachers to go round.

Before we start. When you hear “AI tutor”, what's the first thing that worries you?

Pick one. We'll come back to it.

A maths lesson · thirty students, one teacher, fifty minutes

100 seconds of the teacher, per student, per lesson

  1. Here is a maths lesson.

    Thirty students, one teacher, fifty minutes. It is the most common room in the world, and it is where nearly everyone learns nearly everything.

  2. Now divide the attention.

    Fifty minutes across thirty people is a hundred seconds each. A good teacher spends it well. There is simply no more of it.

  3. Some students get more, after school.

    Three in ten pupils in England have had a private tutor. It is more common in the best-off homes than in the worst-off.

    Sutton Trust, Tutoring: The New Landscape, 2023.1

  4. Nobody thinks that's fair.

    It isn't anyone's fault. It's what an hour of a person costs, and how many hours there are. Keep those nine in mind.

Where students finish a course · percentile of the classroom

Start End of course Classroomthirty students, one teacher One to oneabove 98% of the class2

Can the second line come from an AI?

Stanford, 2024A lift to the human tutor's line: with an AI beside them, tutors got students to mastery 4 points more often, the weakest tutors 9. Reported in mastery, so no bracket on the axis.5
Harvard, 2025Students with an AI tutor learned more than twice as much as the same lesson in class, 0.73 to 1.3 standard deviations, in less time. Scientific Reports.4
Nigeria, 2024Six weeks of AI tutoring after school, 0.23 to 0.3 standard deviations: the gain of 1.5 to 2 years of ordinary schooling. World Bank.3
Classroom With a tutor

  1. In 1984, Benjamin Bloom measured what a tutor is worth.

    He gave some students the ordinary classroom, and others the same course with a tutor, one to one. Then he watched where they finished.

  2. The tutored students finished two standard deviations ahead.

    The average tutored student ended above 98% of the classroom. Bloom called it the two sigma problem: the best method was known, and there was no way to give it to everyone.

    Bloom, Educational Researcher, 1984.2 Later studies find the gap smaller, and still among the largest in education.7

  3. Forty years later, the problem started to move.

    Three trials, three continents. Two report the gain in standard deviations, so they can be placed on this chart. Neither reaches Bloom's tutor yet. Both clear the classroom.

    World Bank, 2024.3 Kestin et al., Scientific Reports, 2025.4 Wang et al., Stanford, 2024.5

  4. The tutor was the expensive part. It isn't any more.

    That is what Ark One is for. Not a smarter classroom. The second line, for everyone in the first. Including the twenty-one who never had it.

You said: it'll replace teachers.

Fair. A chatbot that writes the essay has done the thinking for you. Ark One has no box to type into. You answer by moving the graph, the circle, the board, and the only way through a level is your own hand.

The Harvard students who learned twice as much did every problem themselves. The AI's job was to stand beside the work, not to do it.

2× what students learned with an AI tutor beside them, against the same lesson in class. Harvard, 2025.4

It can. So in Ark One the AI never marks an answer. Every check runs against the scene's own engine: the graph knows where you dragged the vertex, the circle knows where you left the point, and that is what decides.

The Harvard tutor that doubled learning was handed the correct solutions before it was allowed to teach, for the same reason. The model writes the practice. The maths says whether you're right.

0 answers on Ark One marked by a language model. Every verdict comes from the scene.

The world is short forty-four million teachers by 2030, UNESCO says.6 Nothing here is being replaced. There was never enough of it to begin with.

When Stanford put an AI beside human tutors, the weakest tutors improved the most. Ark One is for the hundred seconds a teacher can't stretch: the evening after the lesson, the section that didn't land, five minutes at a time.

44M teachers the world needs and doesn't have, by 2030. UNESCO, 2024.6

Both sides of the desk.

Ark One is built for students and for teachers. For students it means practice they can trust: interactive, built for the exact thing they're stuck on, and checked against the maths rather than a model's opinion. For teachers it means the part of the job that never fitted into a hundred seconds, done in the evenings and brought back to class.

That is where we start. We plan to ship many more tools for teaching, not only for learning, and to work with schools, colleges and universities directly, so that the standard is set by the people who teach in them.

The way people learn is changing. Ark One is here so that students and teachers cross into that new reality together, and neither is left behind.

Four friends on a hillside at dusk, leaning in over one glowing phone, one of them dragging something on the screen and all of them laughing.

Get in while it's quiet.

Invites go out in batches, earliest first. One email when yours is ready, nothing else.

Earliest on the list, earliest in.