Building the science and systems for self-improving AI.

We structure research so insights can be tested, reused, and compounded.

What we study

Study learning dynamics, representations, and optimization.

Use AI to accelerate the understanding and improvement of AI.Build "meta models" that design AI models like top AI researchers.

Use AI4AI to design AI models for all scientific disciplines.

01 / RESEARCH WORKBENCH

A graph-native workbench for building AI experiments and exploring learning mechanics.

02 / RESEARCH METHOD

A mechanistic auto-research method from observation to hypothesis and intervention.

03 / OBSERVABLE LIBRARY

Inspect, compare, and discuss the internal dynamics of model training.

We are building at the intersection of AI research, systems, and science.