Distinguished DAIS Alumni Speaker Seminar: Dr. Yu Shi, "Scaling Up Recommendation Systems in the Age of AI."

Abstract: Recommendation systems decide what billions of people see every day, and as AI agents take over more routine work, the share of human attention spent on recommended content will only grow. Yet these systems remain among the most complex in production machine learning: multi-stage retrieval and ranking pipelines, thousands of heterogeneous features, and user histories spanning years of behavior.

A central question has driven the field's recent progress: do recommendation systems obey a scaling law the way language models do? This talk argues yes -- but that realizing it requires holistic co-design across data, model architecture, and systems, not just larger models. Drawing on work deployed across Facebook and Instagram, the talk traces three generations of architecture: HSTU, which established scaling behavior for sequential recommendation at LLM compute scale; Ultra-HSTU, which bends the scaling curve through algorithmically induced sparsity and semi-local attention; and WHALE, a unified architecture that scales feature-interaction and sequence modeling jointly. Alongside these, it covers the systems work that makes them affordable -- fused attention kernels across NVIDIA and AMD hardware, mixed-precision training and serving, and batched target-aware inference with caching. The talk closes with open research directions in a field that remains, by the speaker's argument, one of the most interesting places to work in applied ML.

Bio: Yu Shi is a Senior Technical Manager at Meta, where he leads a team of machine learning engineers and research scientists building hyperscale ranking and foundation models for recommendation across Facebook and Instagram. His current work centers on scaling laws for recommendation: model architecture, training and inference efficiency, and the co-design between them. He received his PhD in Computer Science from the University of Illinois Urbana-Champaign in 2019, advised by Jiawei Han in the Data and Intelligent Systems (DAIS) group.