Keynotes accepted to WoAIS1 (First International Workshop on AI and Serverless, 2026). See the program for the full schedule.
3:10pm–3:45pm — Keynote 1 (30 min + 5 min Q&A)
Abstract: Serverless is a promising economic model for AI inference: an elastic, shared GPU pool, multiplexed across many models and requests, billed only for useful work. Yet the gap between that ideal and how we serve large models today spans orders of magnitude — and it cannot be closed by adding hardware. This talk argues that the path to far more efficient serverless AI inference is a hunt for the right system primitives: for each key subproblem of the inference lifecycle, identify it, co-design a minimal and composable primitive that exploits a property of the model or hardware, and compose primitives so their gains multiply. I ground this in two systems. ServerlessLLM (OSDI'24) contributes fast multi-tier loading and token migration to eliminate cold starts (10× lower startup latency). BatchGen (OSDI'26) contributes event-driven coroutine scheduling — yield, combine, partition, and KV migration — to keep GPU fleets busy under MoE sparsity and long-tailed generation (up to 9.6× faster inference). I close with the open frontier: the primitives and co-designs we still need, and an invitation to find them together.
Bio: Luo Mai is a Reader (Associate Professor) in the School of Informatics at the University of Edinburgh, where he leads the Large-Scale Machine Learning Systems Group. He also co-leads the UK EPSRC Centre for Doctoral Training in Machine Learning Systems and a UK ARIA project on Scaling AI Compute. University of Edinburgh Research Explorer His research spans the full stack of machine learning systems—from models and data to systems and software—pursuing co-designs that can collectively achieve a 1000× leap in efficiency, scalability, and reliability. More broadly, his work sits at the intersection of distributed systems, machine learning, and data management. His contributions have been recognized with research awards and rising-star honors from both academia and industry.
4:40pm–5:15pm — Keynote 2 (30 min + 5 min Q&A)
Abstract: GenAI introduces a new wave of applications that rely on LLMs to execute business logic. This leads to novel application patterns: code (which may be AI-generated), interaction with LLMs and tools, and preservation of state in semantically rich state management systems. This new class of applications presents novel challenges: stochastic variability of outcomes, autonomy, and new failure modes and security threats Significant advances in models and frameworks have improved capabilities, but do not solve the fundamental system-level reliability problems AI applications face. We argue that meeting this challenge requires a new class of platform-level primitives for agents. In this talk, I identify these primitives across identity and multi-layer permission systems, agent state engineering and isolation, tool management, resiliency, and observability. I examine approaches that our team in IBM Research is currently taking to solve some of these problems and the open challenges that remain.
Bio: Dr. Malgorzata (Gosia) Steinder is an IBM Fellow at IBM Research in Yorktown Heights, NY. With 30 years of experience in computer science research, she leads initiatives at the intersection of Hybrid Cloud and Artificial Intelligence. Her research has shaped IBM's middleware and cloud platforms, including foundational work on cloud management, container services, and cloud security. She has authored over 70 publications, holds 50+ patents, and has served on program committees of multiple top conferences and as an Associate Editor for several journals including IEEE Transactions on Parallel and Distributed Systems. Dr. Steinder earned her Ph.D. in Computer Science from the University of Delaware, where she received the Alan P. Colburn Prize for Best Doctoral Dissertation.