This work offers a discussion on how computational mechanics and physics-informed machine learning can be integrated into the process of sensing, understanding, and reasoning of physical phenomena. A foundation in physics can leverage interpretability, data efficiency, and generalization of the models sought for the dynamics of complex physical systems. Consequently, this synergy results in promising approaches to develop world models that are capable of performing accurate and reliable simulations (reasoning) in low-data regimes. Among the possible alternative formulations, we highlight how thermodynamics offers a general framework to construct inductive biases, demonstrating its potential in applications where physics-consistent predictions are essential.