Century
An independent AI research lab
Modern AI systems are the most capable artifacts ever produced by trial and error. They write, translate, prove, and predict — and yet our understanding of why they work remains shallow. Century is an independent research lab organized around that gap. We study the structure of current models to understand what training actually builds, we finetune and train open-weight systems to test that understanding directly, and we work toward models that reason from first principles rather than from recall. Our long ambition is to turn machine reasoning into a reliable instrument for the hardest engineering problems people attempt: launch systems, structures at unprecedented scale, and the physics that constrains both.
1.Purpose
Century exists to understand, and then to build with what it understands. We did not start with a product thesis or a five-year roadmap, and we are deliberate about that: in a field where the central objects are barely understood, a fixed plan is mostly a way of committing to today’s misconceptions. Our method is to follow the problems — run the experiment, look closely at what the model actually does, and let the results decide what comes next.
What is fixed is the standard. An explanation counts only if it predicts something; a capability counts only if it survives contact with a problem that matters. Everything else is provisional.
2.Research directions
Our work proceeds along three lines, each feeding the others. They are described in full in our research agenda.
2.1The structure of trained models
A trained model is a found object: gradient descent built it, and nobody was consulted about the internals. We study current models — primarily open-weight systems we can run and dissect locally — to understand what their computations are made of and how capability is organized inside them.
2.2Training and finetuning open weights
Understanding that cannot be acted on is trivia. We finetune open-weight models, and are building toward training runs of our own, so that every hypothesis about how models work can be tested the only honest way: by changing the model and seeing what happens.
2.3Reasoning from first principles
Today’s models largely predict what an answer looks like. The deep problem we care about is training models that derive — that work from conservation laws, constraints, and definitions the way a physicist does, so that their conclusions can be trusted in domains where no precedent exists to imitate.
3.Applications
The problems we ultimately want to serve are the ones with no room for plausible-sounding error: rockets, and structures at scales no one has yet built — megastructures in the literal sense. These domains are governed by physics that is unforgiving and design spaces too large for human iteration alone. A model that genuinely reasons from first principles would change what is buildable.
We hold this ambition honestly: it is a long-horizon goal, not a near-term claim. The research in Section 2 is the path to it.
4.People
Century is looking for people who studied physics or mathematics because they couldn’t help it, and who want their work aimed at the hardest problems available. If that is you, read how to join.
hello@century.sh — we read everything sent to this address.