Dawn of
recursive intelligence.
We study how intelligent systems can improve themselves — and how to make the result small, fast and open enough to reach everyone.
What we work on.
Four questions guide everything we build.
Recursive self-improvement
How can each generation of models help create a better next one — learning from its own results, its own mistakes and its own discoveries?
Efficient intelligence
How small can a capable model be? We aim for models that run offline on ordinary hardware, from phones to browsers.
Reliable answers
A model should know the edges of what it knows — and say so — rather than sound certain when it is not.
Breadth
One approach across every modality: language, vision, audio, multimodal, structured data, robotics and the sciences.
The future is self-improving. Our job is to build it carefully, and to share it openly.
Research with us.
We welcome collaborations with researchers, institutions and builders.