Proteins accomplish their functions as part of dynamic complexes and through transient protein-protein interactions that regulate signaling networks underlying cancer, autoimmunity, and neurological diseases. These interactions remain among the most challenging targets in structural biology, particularly when involving disordered or flexible partners. We have developed AFsample1–3, enhanced sampling extensions of AlphaFold that expand the conformational states accessible to AI-based structure prediction beyond single-state models. We will continue developing AFsample towards AlphaFold 3, focusing on rare conformational states and incorporating experimental structural templates to guide predictions of multimers and large complexes. A critical unresolved question is whether AI-generated conformational ensembles reflect true physical protein dynamics. To address this, we will use Arrhenius to generate a ground truth library of microsecond-scale molecular dynamics trajectories across a diverse set of proteins below 300 amino acids, spanning different fold classes and dynamic behaviors. At this size regime, our monthly allocation enables approximately 5 µs of simulation per protein target, allowing us to characterize ~10 proteins per month and build a benchmark set of over 100 proteins over the project period. Systematic comparison of these physics-based ensembles against AFsample-generated ensembles will allow us to evaluate the degree to which deep learning methods faithfully reproduce protein conformational landscapes and identify where AI-based sampling can be further improved.