Intelligence is learning to act.

Everything a model says can be revised. Nothing it does can be taken back. That asymmetry is the alignment problem in its final form, and it is the only problem we work on.

We started a research lab around one question: what has to be true before a system is permitted to act. Not whether it is capable. Not whether it means well. What can be established, bounded, and proven in advance.

It's called Phase Portrait. In dynamical systems, a phase portrait is the map of every trajectory a system can take, including the boundary between the paths that end well and the paths that do not.

We treat this as an engineering problem. Our team and our roadmap are aligned around solving it and nothing else.

We are a small team in Palo Alto, from Google DeepMind, xAI, AWS AI, and Coinbase Security.

We are backed by Lightspeed and Bain Capital Ventures, among others.