The article examines the generalization of Physics-informed neural networks (PINNs) beyond training domains, emphasizing predictive accuracy and adherence to physical principles. It challenges the role of overparametrization, suggesting it may promote overfitting rather than enhancing generalization in scientific machine learning.