MIT's Deep Learning course (6.7960, Fall 2024) turns this session over to a guest lecturer covering large language models in practice. The lecture works through prompting strategies, chain-of-thought reasoning, and instruction-tuning, framing these techniques through the lens of in-context learning: how a model trained only to predict text can be steered, at inference time, to follow instructions or reason step by step without additional gradient updates. The talk builds on the course's earlier neural network foundations to explain why these behaviors emerge and what they reveal about how transformer-based LMs represent tasks internally. Delivered as a 77-minute session, it sits within MIT OpenCourseWare's public release of the full course, aimed at students who already have some background in deep learning architectures and want a focused treatment of how modern LMs are prompted and adapted.