Explore the growing divergence between machine learning theory and practice in this distinguished lecture that examines how theoretical approaches often make pessimistic predictions despite empirical success. Discover why much of modern AI lacks substantial theoretical guidance and learn about conservative approaches to bridging the theory-practice gap through reformulated theoretical questions. Understand how machine learning theory's focus on worst-case computations and data distributions may miss opportunities to leverage favorable structures in real-world datasets and algorithms. Examine specific examples from generalization and privacy research that demonstrate instance-specific analysis approaches versus traditional worst-case scenarios. Gain insights into algorithmic fairness, data privacy, and game theory applications that could help close the gaps between theoretical predictions and practical performance in machine learning systems.