Social science researchers are increasingly leveraging large-scale video data and computational methods to examine key concepts and questions. Yet much existing work relies on a narrow set of video features and focuses mainly on text or static visuals, leaving other high-dimensional features and modalities underexamined. This talk addresses the gap by presenting two empirical studies that use large language models (LLMs) and other computational methods to enable more systematic and comprehensive video understanding. The first study develops a multidimensional “visual logic” using computer vision, supervised learning, and LLMs to understand online influencer activities. The second study introduces how LLMs and deep learning models can be applied to identify audio categories and objects in large-scale video data. The talk will conclude with a discussion of future opportunities and challenges in this rapidly evolving field.