Beautiful problems deserve beautiful typesetting

I teach mathematics to the kids in my neighbourhood: informally, in evenings, up to A-level. Much of my preparation is spent looking for problems. Beautiful problems specifically: ones that illustrate a technique with no waste, that surprise the reader, that yield to a clean argument once the right idea arrives. The hunt is its own pleasure.

The supply is uneven. The best books are extraordinary: Engel’s Problem Solving Strategies, Andreescu’s number-theory volumes, Coxeter and Greitzer’s Geometry Revisited. But they go out of print, they live behind paywalls, and they are hard to share with a fifteen-year-old who has a phone, intermittent internet, and nothing else. Online the situation is worse. Mathematics on the web is generally badly typeset; the formatting fights you, and the problems sit alongside cookie banners and adverts.

While I have been looking for problems, AI researchers have been building datasets to train machines on the same material. MathNet from MIT CSAIL. MATH from Hendrycks et al. HARP. Twenty thousand olympiad-level problems, openly licensed under CC BY 4.0 or MIT, with full attribution to the original publishers, sitting in JSON files designed to be read by a model.

But these are problems written by human setters for human contestants. By the time they reach a training pipeline, the human reader has been forgotten.

OlympiadHQ is an attempt to surface that material back to humans. Twenty thousand attributed problems, no account required to read, no subscription, no advertising. Each problem carries its source, its country, its competition, and its licence. The typesetting is TeX Gyre Pagella for prose and KaTeX for mathematics. Search is full-text. Mobile is first-class.

It is a side project. I would not have built it without the AI tools I now use daily for engineering work; they made an evenings-and-weekends effort tractable that otherwise wouldn’t have been.

The broader thought is simple. AI datasets are infrastructure built largely on the work of human authors. When the infrastructure is open, surfacing it back to human readers, well-presented and properly attributed, feels like the right thing to do.

Beautiful problems deserve beautiful typesetting. That is the project.