The air-interface problems are characterized by high-dimensional decision variables, dynamic wireless environments, and significant task diversity. Directly applying AI techniques developed for other domains incurs substantial storage and computational overhead, making it difficult to deploy on resource-constrained wireless edge. Incorporating domain-specific knowledge of communications into DNN design has been a long-standing pursuit in the field of wireless AI. Deep unfolding offers a systematic way to leverage optimization algorithms, yet it suffers from limited generality and high inference complexity. Recently, another emerging trend is to draw inspiration from physics-informed machine learning (PIML) and uncover the “physical knowledge” in wireless systems. However, wireless communications possess their own unique principles and knowledge, and the methodology for integrating communication information into the design and training of intelligent air-interface differs fundamentally from those of PIML. This talk addresses the key challenges faced by air-interface AI—scalability, generalizability, and multi-task capability—and presents several systematic approaches to discovering and exploiting wireless priors for designing DNNs that learn wireless policies efficiently.