Simulating how atoms and molecules move over time is a central challenge in computational chemistry and materials science. Classical machine learning approaches to molecular dynamics (MD) encode fundamental physical principles directly into their model architectures, most notably energy conservation and equivariance, the requirement that predicted forces remain consistent regardless of how a molecule is oriented in space. These so-called inductive biases have long been considered essential for reliable, physically meaningful MD models. But are they truly indispensable?