An AI research group. We build novel datasets and study the capabilities they make possible.

What a model can do is largely settled before training starts, by the data it sees. We think many of the open questions in AI are, underneath, questions about data: what to collect, how to shape it, and what a model does with it.

So we work from both ends. We construct datasets for capabilities that don't exist yet, and we take apart capabilities that do, to understand what data would have produced them.