Toward a Foundation Model of World Dynamics
We are exploring models that connect observations, causal structure and temporal change to reason about how real systems evolve.
The world is recorded
Sensors, transactions and experiments leave records of a changing world. These records are partial views of systems whose internal state is often only indirectly observed.
Our research asks how a model can connect those observations into representations of state, relationships and temporal change.
Representing what can change
A representation of dynamics needs to distinguish persistent structure from transient conditions. It also needs to account for missing observations and changes in how a system is measured.
Testing understanding of change
We would examine the same representation through forecasting, intervention questions and unfamiliar conditions. Each task offers a different test of whether a model captures useful dynamics or relies on patterns that do not persist.