Interpretability and generative capability in generative models are fundamentally two complementary aspects. A highly interpretable model typically learns the true underlying generative mechanisms behind data, such as physical laws, causal relationships, or explicit structures. As these mechanisms are inherently stable and universally applicable, such models can reliably generalize beyond training data, producing more reasonable and robust samples with fewer generation failures. In addition, a highly controllable and powerful generative model implicitly or explicitly captures genuine and effective underlying rules. The ultimate goal of training generative models should extend beyond obtaining high-quality samples to exploring and understanding the underlying generative mechanisms of phenomena. When a generative model demonstrates controllability and scalability with respect to a dataset, it indicates the model has genuinely learned the mechanisms that generate the data. This opens up a paradigm in scientific research, enabling the discovery of underlying principles through observational data reconstructed by generative models, particularly when these models exhibit controllability and scalability. Leveraging powerful nonlinear mapping, efficient iterative training, and structured interpretability, artificial intelligence holds the potential to uncover and understand rules and principles currently beyond human knowledge.