Our main goal is to understand what makes data learnable.

Learnability today, in the popular interpretation of intelligence, lives somewhere between collecting more, filtering duplicates, and curating datasets by hand around the models we already have. The field treats it as a solved problem at an unsolved scale.

We think the problem is deeper.

Today’s interpretation of data value is mostly heuristics and scale, where the worth of an example is judged by how much of it we have, how clean it looks, or how much the loss moves. But learning, and evidently, intelligence, seems to be more than this. There is an implicit structure to what makes something teachable, a dependence on the model doing the learning, and a deeper geometry that counting examples alone does not capture.

The issue is that we don’t yet know what it truly means for data to be learnable. Which structure turns into generalization, why it does for one model and not another, whether it carries from one domain to the next. These are not solved problems with clean definitions. They are open questions with enormous consequences for what we build next.

The path to intelligence requires all of these, and more. And as models scale, as compute grows, as collection becomes the last lever anyone trusts, learnability remains a dark map with terrains that have yet to be marked down.

We believe the path to intelligence goes through learnability. It deserves its own research, its own theorems, algorithms, and measures, with applications that make current systems more data-efficient while supporting the future systems that come next.

This is the goal, and it is what excites us: contributing to true, safe intelligence and to the systems that may one day carry it into the world. We are a team of next-generation researchers and builders, fascinated by the unexplored and eager to help guide the world into a new era of understanding.