To Solve the
Physics of Intelligence
We are an independent research laboratory building AGI that is safe by design, not by chance. We believe alignment is a quantifiable property of a system, akin to entropy or mass.
The Alignment Gap
The current paradigm of AI development is defined by a dangerous asymmetry: Capabilities are scaling exponentially, while control methods scale linearly.
We are building "Black Boxes" larger than our ability to illuminate them. At Metanthropic, we refuse to accept that AGI must be inscrutable. We operate on the conviction that to trust a mind, you must be able to read its thoughts.
Our Thesis
We are closing the gap with three non-negotiable pillars.
Mechanistic Interpretability
We map the neural circuitry of our models. Just as biology maps the genome, we map the "connectome" of our networks to identify where deception, bias, and reasoning failures originate.
Constitutional Design
Safety is not a filter; it is a constraint. We use "Constitutional AI" to train models that self-police their outputs against a set of human-defined principles during the pre-training phase.
Verifiable Reasoning
A model that guesses is useful; a model that proves is revolutionary. We are building "System 2" reasoning engines that expose their chain of thought for human verification.
Operating Principles
Our Charter defines the parameters of our mission. These are the rules we cannot break, regardless of competitive pressure.
Read full CharterBroadly Distributed Benefits
We commit to using any influence we obtain over AGI to ensure it benefits all of humanity, avoiding uses that unduly concentrate power.
Long-Term Safety
If a value-aligned project comes close to building AGI before we do, we commit to stop competing and start assisting them.
Technical Leadership
Policy advocacy is insufficient without capability. To steer the future, we must be on the cutting edge of what is possible.
Cooperative Orientation
We actively cooperate with other research labs. Safety is a shared responsibility, not a trade secret.