Improving Recommendation Systems and Search in the Age of LLMs

Explore cutting-edge applications of large language models in recommendation systems and search through this keynote presentation delivered at the AI Engineer World's Fair. Discover how LLMs are revolutionizing traditional approaches by addressing three critical challenges: transitioning from hash-based item IDs to semantic IDs for better content understanding, leveraging LLM-augmented synthetic data for improved data quality and augmentation, and implementing unified models to reduce operational costs while maintaining separate systems. Examine real-world case studies from industry leaders including Indeed's implementation of synthetic data generation, Spotify's innovative approaches to content discovery, Netflix's Unicorn unified model architecture, and Etsy's unified embeddings system. Learn about the evolution from early Word2vec embeddings to modern transformer-based architectures and their measurable impact on user experience and system performance. Gain insights into scalable system designs, model architecture innovations, and the future vision of LLM-driven content discovery and intelligent search systems that serve customers at massive scale.