Distinguished Seminar Series on Data Science & Artificial Intelligence - "The Power and Vulnerability of Multimodal LLMs" by Dr Liangliang Cao
Research Seminar
-
Date
14 Nov 2025
-
Organiser
Department of Computing
-
Time
11:00 - 12:00
-
Venue
Online via Zoom / PQ306
Speaker
Dr Liangliang Cao
Summary
Modern AI has grown so powerful that it has transformed how humans interact with machines. In the first part of this talk, I will share my experience of this evolution over the past fifteen years—from winning the first ImageNet Challenge in 2010 to recently co-leading Gemini Live at Google DeepMind in 2025—and trace how large-scale data and multimodal learning have driven progress across text, image, and speech domains. In the second part, I will discuss the intrinsic vulnerabilities of these systems, including data and model limitations, as well as the growing challenge of power consumption. I will argue that despite the costly industrial race for ever-larger models, there remain profound research opportunities for academia to make AI more efficient, reliable, and beneficial to human life.
Keynote Speaker
Dr Liangliang Cao
Principal Engineer and Director
Gemini Team
United States
Dr Liangliang Cao received his PhD from the University of Illinois at Urbana–Champaign, his MPhil from the Chinese University of Hong Kong, and his B.S. from the University of Science and Technology of China. After completing his PhD, he joined the IBM Watson Research Center and later co-founded Switi Inc., a startup acquired by Google in 2018. He subsequently founded and led the Google Cloud Speech Modeling team (2018–2021), served as the model DRI for Apple Intelligence (2023–2024), and most recently worked as a Principal Engineer and Director on the Gemini team at Google DeepMind (2024–2025). Before the pandemic, Dr Cao served as an adjunct professor at Columbia University and also at the University of Massachusetts Amherst. He is an IEEE Fellow and also a recipient of the ACM SIGMM Rising Star Award.
