Extreme-scale workflows play a crucial role in increasing scientific productivity by helping scientists orchestrate scientific campaigns. With the rapid advancement of artificial intelligence (AI) and its growing application in scientific research, the community is witnessing both the integration of AI tasks into scientific workflows (workflows for AI) and the augmentation of existing workflows with AI capabilities (AI for workflows). Given the early stage of these emerging topics, significant gaps remain in AI reasoning for science, high-performance computing (HPC) system design, workflow orchestration, and cross-community collaboration. This article synthesizes perspectives from leading researchers across HPC, AI, workflow systems, and domain sciences, drawn from the First International Symposium on Artificial Intelligence and Extreme-Scale Workflows (AIExScale’25) held in conjunction with Supercomputing 2025, to identify four key challenges and outline promising directions for advancing AI workflows at extreme scale.
I report on the design and construction of HiroSim, a modular, deterministic, closed-loop simulation stack for control software development, built substantially through a conversational interface to a large language model. Our interest is less in the productivity anecdote than in a structural question: which properties of a simulation framework make it tractable for an artificial intelligence agent to extend it correctly? I argue that the features that make a stack pleasant for a human engineer—locality, declarative composition, strong typing, reproducibility, and documentation kept adjacent to code—are precisely those that make it amenable to machine authoring. I illustrate with a worked case study, a three-body vehicle built end to end from natural-language prompts, and discuss where the abstraction leaked and what is recommended to others building tools meant to be extended by both humans and machines.
Decades of research have gone into classical compiler design, producing sophisticated frameworks that transform high-level programs into efficient machine code. Quantum computing faces the same challenge: translating quantum algorithms into hardware-executable instructions. A classical-first approach extends established compiler frameworks with quantum concepts, treating quantum processors as specialized accelerators within larger computing systems. This article presents multilevel intermediate representation (MLIR) as an effective foundation for this approach and demonstrates its use in the Munich Quantum Software Stack, which connects high-level applications to diverse quantum hardware through a modular, high-performance computing (HPC)-integrated pipeline. MLIR dialects support multiple abstraction levels while providing built-in parsing, transformation, and optimization infrastructure. Quantum programs are represented through complementary dialects: one explicitly encodes the circuit's dataflow graph, while the other mirrors established languages for seamless tool interoperability. The resulting framework, the Munich Quantum Toolkit Compiler Collection, supports compiler passes for both flat circuits and structured control flow. It is available at https://github.com/munich-quantum-toolkit/core.
