Structure
Physical systems contain states, causes, invariants, symmetries, constraints, and uncertainty. A successful model must learn relationships more durable than surface correlation.
A frontier AI research company
Intelligence that can explain the world.
Physics AI is building models that go beyond statistical prediction. By adapting the explanatory knowledge of physics, we aim to create AI that can generate deeper knowledge in every domain.
Explore the thesisExplanatory model 01 testing
01 / The thesis
Prediction is not explanation.
Modern language models are extraordinary pattern learners. They can reproduce much of what humanity has written, but fluent prediction can imitate understanding without possessing it.
An explanation does more. It identifies a mechanism that makes many observations follow, tells us what would happen under intervention, and exposes itself to evidence that could prove it wrong.
We believe the next major advance in intelligence will come from models that can participate in this process—not merely retrieve existing knowledge, but propose explanations, derive their consequences, criticize them, and replace them with better ones.
The goal is not an AI that always has an answer. It is an AI that can help create better explanations.
02 / How knowledge grows
Train the whole loop.
Models should learn from the full cycle by which knowledge improves.
- 01
Observe
Represent the phenomenon, the context, and the uncertainty.
- 02
Conjecture
Propose candidate mechanisms, including ideas not present verbatim in the training data.
- 03
Derive
Turn each explanation into precise, independently checkable consequences.
- 04
Test
Compare those consequences with simulation, experiment, or other verified evidence.
- 05
Criticize
Search for counterexamples, contradictions, hidden assumptions, and better tests.
- 06
Revise
Preserve what survived and replace what failed with a better explanation.
Better answers are useful. Better processes for creating knowledge are transformative.
03 / Why physics
Reality provides the reward signal.
Physics is a demanding and unusually productive training ground for explanatory intelligence. Its ideas must survive mathematics, counterfactuals, simulation, and contact with the world.
Counterfactuals
A physical explanation should say not only what happened, but what would happen if an initial condition, force, boundary, or intervention changed.
Falsifiability
Dimensional consistency, conservation laws, mathematical derivations, simulations, and experiments provide ways to reject persuasive but wrong answers.
Transfer
Our research question is whether the habits learned here—causal abstraction, criticism, and revision—can improve reasoning far beyond physics.
Physics is the proving ground. Explanatory intelligence is the destination.
04 / Research agenda
Building toward explanatory models.
We are investigating a set of connected problems. These are research directions, not claims of solved capabilities.
Physical representations
Develop internal representations of physical state, mechanism, invariance, and uncertainty that support reasoning across problems and scales.
Explanation generation
Train models to construct candidate mechanisms that are compact, causal, and capable of explaining more than the examples seen during training.
Criticism and falsification
Build environments and verifiers that reward a model for deriving discriminating tests, finding failures, and revising its own ideas rather than defending them.
Explanatory transfer
Measure whether explanatory abilities developed in physics transfer to unfamiliar scientific, engineering, and general reasoning tasks.
New evaluations
Design benchmarks that distinguish correct answers, curve fitting, and memorized derivations from genuine counterfactual and explanatory competence.
Help build intelligence that can explain.
We want to work with exceptional researchers and builders across artificial intelligence, physics, mathematics, philosophy of science, and systems engineering.
Start a conversation