Large Language-Style Universal Models for Short Video Recommendations

Explore a cutting-edge ray-based solution for large-scale recommender systems in this 30-minute talk from Anyscale. Dive into the design of a new system for Josh, India's leading short video platform with 85M daily active users, 170M monthly active users, and over 100M videos. Learn about the innovative approach inspired by large language modeling, featuring a single big model with sparse inputs and outputs, early and late fusion for different outputs, and scaling through a universal graph search algorithm generalizing HNSW. Discover the advantages of this method, including simplified vertical integration and end-to-end content training from raw user actions. Examine empirical results showing double-digit improvements in engagement metrics. Access the accompanying slide deck for visual aids and additional information. Gain insights into Anyscale's AI Application Platform and the Ray open-source framework for scaling AI workloads, including Generative AI, LLMs, and computer vision.