Named after the space of every possible computation.
A ruliad is Stephen Wolfram's term for every rule, run every possible way. Today's models sit in one small corner of it. We started the lab to look around the rest.
The next leap won't be a bigger model. It will be a better learner.
Scale has taken today's models a long way, but the way they learn has barely changed. We think the next generation will be defined by four properties, and our research is aimed at each.
- Continual
Models should keep learning after they ship.
Today's models are trained once and then frozen. A better learner updates from experience as it goes, without forgetting what it already knows.
- Efficient
Learning should take far less data than it does today.
People pick up a new skill from a handful of examples; models need millions. Closing that gap takes new learning algorithms, not just more tokens.
- Grounded
Models should learn how the world works, not just how text reads.
That means perceiving, remembering over long horizons and acting, so that what a model knows is tested against something real.
- Interpretable
We should be able to see why a model does what it does.
A model you can't inspect is hard to trust and harder to improve. We want progress that comes from understanding, not only from tuning.
Automating AI research, starting with our own.
We're building a research engine that runs experiments, reads the results and decides what to try next. Our researchers build it and work alongside it, and every result it produces makes the next search faster.
Read our research →We're building a focused team of researchers and engineers.
We look for people who can take an idea from a paper to a working experiment, and who care more about what's true than what's impressive.