The continual evolution of wireless communications technologies has profoundly changed our daily lives. Mobile Internet represents a recent revolution that has been enabled by the seamless mobile broadband access of 4G LTE networks. It in turn has empowered various groundbreaking mobile applications, such as online-to-offline commerce, mobile payment, mobile social networks, and so on. Meanwhile, with the recent revival of artificial intelligence (AI), a plethora of intelligent mobile applications emerge, e.g., augmented/virtual reality, mobile robots, autonomous driving, etc. Given the limited onboard computing resource, power supply, and sensing capability of mobile devices, these applications will critically rely on effective communications to access external computing resources (e.g., cloud and edge servers) and exchange data (e.g., sensing data, intermediate features, or AI models) for better user experience, perception capability, robustness and safety. This calls for a paradigm shift in wireless networking, from “data-oriented communications”, which maximize data rates, to “task-oriented communications”, where the data transmission is an intermediate step and is to be optimized for the downstream task. This talk will introduce new research problems of task-oriented communications for mobile intelligence, which demand integrated designs of communication strategies and machine learning models. It will present a pragmatic end-to-end design approach based on an information bottleneck principle and neural architecture design. Moreover, it will also introduce promising tools for further theoretical and algorithmic investigations, which are recently developed in structured high-dimensional estimation and deep learning.