Intelligence should compound.
No one could rebuild the modern world from scratch. Yet we built it, because minds work together and each generation inherits everything the last one learned. Civilization is what intelligence looks like when it compounds.
We are building that engine for machines: societies of agents that cooperate, learn from everything they experience, and evolve steadily. The hardest problems humanity faces will not fall to a single, static mind. They will fall to vibrant, evolving societies of agents: intelligence that compounds.
CORAL
CORAL scales intelligence along a different axis: not a bigger model, but more agents, better organized. Agents explore in parallel islands, each population evolving its own approach, and the strongest discoveries migrate between them. The result is a search no single agent could carry out: wider than one context, longer than one session, and improving as it runs. CORAL sets a new state of the art on discovery tasks across algorithm design, GPU kernels, and systems engineering.
REEF
Continual learning infrastructure for self-improving agents.
A pretrained model is fixed the moment it ships. How much a model knows and how fast and deep it can learn are different pursuits. Scaling improves the first and leaves the second underdeveloped. We work on the second.
Deployment should be where a model's education begins and becomes self-motivated. REEF serves agent requests, observes feedback, grows model weights or agent harnesses, and commits improvements that pass evaluation: a system's own experience, distilled into lasting improvement.