Can AI safely automate medical decision-making tasks to improve patient outcomes? In this talk, the presenters will share the challenges in the development and translation of medical AI, and how they are being addressed through a blend of innovation in algorithm development, dataset curation, and implementation design. They will first talk about self-supervised learning methods for medical image classification that leverage large unlabeled datasets to reduce the number of manual annotations required for expert-level performance. Then, they will discuss open benchmarks that can help the community transparently measure advancements in generalizability of algorithms to new geographies, patient populations, and clinical settings. Third, they will share insights from studies that investigate how to optimize human-AI collaboration in the context of clinical workflows and deployment settings. Altogether, this talk will cover key ways in which we can realize the potential of medical AI to make healthcare more accurate, efficient and accessible for patients worldwide.