Phillip Isola devotes this MIT 6.7960 Deep Learning session to the practical craft of getting neural networks to actually work, rather than to new theory. He runs through opinionated, hard won tips on debugging training runs, diagnosing why a network refuses to learn, choosing learning rates and initialization schemes, and reading loss curves for early warning signs of trouble. The lecture leans on anecdotes from his own research practice, treating deep learning training as an empirical, almost experimental skill rather than a purely mathematical one. Topics include overfitting checks, sanity tests for new code, and the small habits that separate a model that trains cleanly from one that silently fails. Delivered to MIT students as part of the Fall 2024 course, it is aimed at anyone who already knows the basic architecture of neural networks and wants to know how working researchers actually get them to converge.