How this got here
It starts with a particular kind of A: the essay that is fluent, well-cited and plausible, handed in by a student who cannot summarise its argument two weeks later. The grade is real and the learning is a fiction, and nothing catches the difference until an exam does.
The design question that follows is not “how do we make an AI tutor more accurate” — it is “how do we make one that refuses to be an answer machine, in a way a motivated student cannot talk it out of.” Everything in the product falls out of that question: the attempt gate in the API rather than in a prompt, the hint ladder that costs something, the second model reading every reply. The longer version is at /company/story.
A tool that refuses the answer is not for everyone, and the people it is not for leave in the first week. That is deliberate. How large that split actually is, nobody here knows yet — this is new, and a number that has not been earned will not be printed.
What we’re not
We’re not trying to be the biggest edtech company. The best edtech companies at scale have consistently turned into surveillance products or ad platforms — because those are the business models that scale to hundreds of millions of low-ARPU students. We’re building for the students and families who will pay $10/mo for a real tutor. That’s a smaller market. It’s the right one for us.
We’re also not trying to replace teachers. Learn mode is at its best alongside a good teacher — because the reasoning practice it forces is exactly what turns a lecture into a durable skill. Where there’s no teacher available (homeschool, self-learning, the middle of the night on a problem set), Learn mode does its best work solo. But it’s a supplement, not a substitute.