Sara Beery lectures on out-of-distribution generalization for MIT's 6.7960 Deep Learning course. The session covers why models trained on one data distribution often fail when deployed on another, examining adversarial robustness, distribution shift, and the gap between training performance and real-world reliability. Beery works through the sources of OOD failure and surveys strategies researchers use to measure and improve robustness, connecting the material to practical deployment concerns in computer vision and other applied settings. As lecture 17 in a graduate-level sequence, it assumes familiarity with core deep learning concepts and builds on earlier material on generalization theory. The talk is delivered in a standard classroom lecture format with slides, aimed at students who already have grounding in neural network training and want a deeper look at where and why models break outside their training conditions.