A recent paper introduces a novel framework that integrates physical knowledge directly into loss functions used in machine learning tasks, particularly those involving the prediction and modeling of physical systems such as molecules and spins. This approach is designed to enhance the effectiveness of machine learning methods in scientific applications by embedding domain-specific information into the training process. The framework has been applied to various machine learning tasks, targeting complex physical systems to improve predictive accuracy and model reliability. While the claim that this framework improves machine learning for scientific applications remains unverified, the integration of physical principles into loss functions represents a promising direction for advancing computational modeling in physics and related fields. This development aligns with ongoing efforts to tailor machine learning techniques more closely to the unique characteristics of scientific data.