We started Kapllan because we wanted to do careful work in a hurried field.

Artificial intelligence, in 2026, has the character of a fire drill. Every quarter brings a new capability, a new benchmark, a new concern. It is easy to confuse motion with progress, and harder still to resist the feeling that one must ship something — anything — before the week is out.

We think the field is important enough that this is the wrong response. The systems being built now will live in classrooms and clinics and courts for decades. They deserve to be built the slow way: with legible methods, honest evaluations, and a willingness to be wrong in public.

Kapllan is a laboratory, not a platform. Our flagship model, Llana, is the result of four years of patient disagreement among a small number of researchers. We publish our assumptions. We publish the cases where the assumptions don't hold. We publish the bits that didn't work.

We are not trying to build general intelligence. We are trying to build a particular kind of intelligence — one that reasons legibly, refuses calibratedly, and remains useful at the edge of its competence. This is less ambitious than some of our peers. We think it is more honest.