AI Infrastructure
with Operating Systems

The proliferation of large language models, AI agents and embodied intelligence is reshaping the interaction paradigm between humans, devices and computing systems. This shift calls for rethinking OS primitives, kernel abstraction, and system software stacks for AI-native workloads. Today's mainstream operating systems are still designed around application-centric abstractions and human-driven interaction, which creates fundamental mismatches for AI-native workloads: the lack of unified management abstraction for heterogeneous accelerators, inefficient GUI-based agent execution, coarse-grained permission mechanisms for cross-application data flows, lack of system-level resource orchestration for heterogeneous edge-cloud inference, and insufficient deterministic guarantees for mission-critical intelligent scenarios. Addressing these mismatches requires re-examining core infrastructure concerns — CPU/GPU/NPU scheduling, memory management, I/O and data movement, virtualization and isolation, runtime and serving systems, and compiler-system co-design — through an OS lens.

This tutorial aims to bring together researchers, engineers and practitioners from operating system, cloud computing, edge AI, agent systems and embedded computing communities, to discuss the architectural evolution of operating systems for the intelligent world. We will examine core challenges including heterogeneous system resource management, LLM serving, agent-native system abstraction, end-cloud collaborative inference runtime, deterministic convergence of intelligent workloads, and security & privacy for pervasive AI agents, and share practical experiences from real-world deployments on cloud, servers, wearables, and smart terminals.

The tutorial will be followed by a banquet, and we warmly invite you to join us for further discussion with speakers and attendees.