Physics of AI

Model the system. Design the learning rule. Test the behavior.

The central problem is not simply making models larger. It is deciding what a system should represent, how it should learn, and how its behavior should be measured when the world is uncertain, dynamic, and only partly observed.

1. Modeling. Build representations of dynamics, uncertainty, memory, and interaction that preserve what matters for prediction and control.
2. Algorithm design. Turn those representations into learning and planning procedures that can improve from data, simulation, feedback, and experience.
3. Physical evaluation. Test whether behavior remains useful under distribution shift, limited information, and real operational constraints.

Research as an engineering loop

We move between theory, implementation, and experiment. A model proposes a structure. An algorithm makes that structure operational. Evaluation reveals which assumptions survive contact with a task.

The site is a working record of that loop: research notes, implementations, failures, and methods as they become clear enough to share.