Phillip Isola, teaching MIT's 6.7960 Deep Learning course, moves past the standard forward-pass model to ask how systems can be made to think harder at inference time rather than just training time. He covers beam search for navigating large output spaces, chain-of-thought prompting that lets a model reason through intermediate steps, in-context learning where examples in the prompt shape behavior without retraining, and test-time training that adapts a model on the fly. Search-based techniques round out the lecture as a way to squeeze better performance out of a fixed trained network. Running 83 minutes as part of the Fall 2024 MIT OpenCourseWare series, the lecture treats inference as its own design problem, distinct from architecture and optimization, with concrete methods for each approach rather than a survey of buzzwords.