Unbounded Labs is an artificial intelligence research team. We build state-of-the-art models for narrow domains — the problems that are too specific, too strange, or too small to hold a frontier lab’s attention, and where a focused model can beat a general one outright.

Our first is Bart, a language model trained exclusively on text written before 1930. No modern hindsight, no anachronisms — a mind assembled entirely out of what was knowable at the time. Building it meant constructing a vintage corpus from scratch, designing period-appropriate evaluations, and adapting post-training techniques to a world that had none of them.

Historical language modeling is the first domain, not the only one. The method generalizes: a corpus built on purpose, an architecture tuned to it, and evaluations that measure the thing you actually care about. We publish what we learn as we go — dead ends included.

Read our research