Abstract: While the rapid advancement of Large Models has driven unprecedented successes in capabilities, efficiency, and real-world applications, a fundamental question remains: do we truly understand how they work? Establishing a solid theoretical foundation for these highly black-boxed systems is no longer just an academic curiosity, but a crucial prerequisite for developing safe, interpretable, and controllable AI. This article argues that mathematically, large models are multifaceted entities, acting simultaneously as function approximators, probability samplers, information compressors, and universal Turing machines. To fully capture this complexity, the study of AI mechanisms must transcend isolated viewpoints. The author proposes a continuous methodological spectrum that draws analogies from four fundamental scientific disciplines:•　 Mathematics delineates the theoretical boundaries and absolute capacities of models under idealized assumptions through rigorous proofs.•　 Physics extracts empirical laws (e.g., scaling laws) and statistical patterns via controlled experiments and synthetic data.•　 Biology conducts mechanistic “anatomy” to decode internal representations, neural circuits, and functional modules within real-world models.•　 Psychology explores cognitive boundaries and emergent behaviors through black-box testing, prompting, and input-output interactions. These four paradigms are not mutually exclusive but deeply interconnected and mutually reinforcing. By bridging abstract theoretical proofs with empirical behavioral observations, this continuous spectrum provides a holistic framework for AI mechanism research. Ultimately, this paradigm shift will transition the field of large models from empirical engineering into a rigorous, cumulative scientific discipline.