The ability to learn from the data, either acquired data on patients or simulated data following a physics model, leads to a new way of thinking about how to acquire k-space data and reconstruct images that go beyond previous technology limited by fixed and usually incomplete models. This chapter will review the application of deep learning to MRI acquisition and reconstruction for efficient, robust, and information-rich imaging. The first topic will be deep learning image reconstruction to accelerate acquisition and improve the efficiency of MRI, including the evolution from compressed sensing to neural networks in the so-called unrolled reconstruction networks, self-consistent networks without explicit data consistency, self-supervised networks without fully sampled references, and generative AI reconstruction methods such as diffusion models. The second topic will be the application of AI to offer a new alternative for fast free-breathing MRI, including auto-navigation and motion-resolved reconstruction in adults and children without the need for anesthesia. The third topic will be how to improve MR fingerprinting using AI, and the last topic will discuss the use of AI to design pulse sequences.