Program

Agenda

Time	Topic	Speaker & Institution
14:00 – 14:45	The Aversion to Systematization in Systems Research	Prof. Timothy Roscoe ETH Zürich
14:45 – 15:05	From Paper to Community and Deployment: Lessons from Building RL Infrastructure	Guangming Sheng University of Hongkong
15:05 – 15:45	AI Systems Research in a Vibe Coding Era	Prof. Kenneth P. Birman Cornell University
15:45 – 16:05	Coffee Break
16:05 – 16:45	Towards A Learning-Directed Operating System	Prof. Aditya Akella University of Texas at Austin
16:45 – 17:30	How Will AI Systems Do Systems Research?	Ant Rowstron CTO at ARIA

---

Speakers

---

---

---

---

---

---

Talk Abstracts

The Aversion to Systematization in Systems Research

Prof. Timothy Roscoe

Frequent concerns in systems research include rising submission volumes, heavy reviewer workloads, overly broad conference scopes, and a growing gap from real-world practice. While discussions often focus on improving peer review, a deeper issue is the field’s resistance to systematizing knowledge. Few papers provide reusable principles, and those that do are rarely cited. This leads to repeated ideas and limited impact. Addressing this requires understanding why systematization is resisted, what benefits it could bring, and how we might begin to build more structured, generalizable knowledge—particularly in areas like operating system design.

---

From Paper to Community and Deployment: Lessons from Building RL Infrastructure

Guangming Sheng

Large-scale RL post-training is emerging as a first-class systems workload, combining distributed LLM training, high-throughput generation, reward and verifier execution, data movement, and evolving control flow. This talk examines RL infrastructure evolution through HybridFlow, its open-source implementation verl, DAPO, and Laminar. HybridFlow/verl enables flexible RL dataflow representation and efficient distributed execution, while verl extends to reproducible recipes, configurable pipelines, and system–algorithm co-design (DAPO). Laminar studies asynchronous RL execution, where system decisions affect policy staleness, convergence, and reward. These cases highlight open challenges in rollout optimization, reproducibility, and validating system optimizations that impact learning at scale.

---

AI Systems Research in a Vibe Coding Era

Prof. Kenneth P. Birman

Pervasive adoption of coding agents like Claude Opus and OpenAI Codex is rapidly transforming software development, especially for teams building applications on complex APIs—work that previously required substantial expertise. This talk examines how this shift changes the questions systems researchers should ask, and how evolving coding practices impact the scope of experimental research. While these trends boost productivity, they risk locking us into ideas current AIs can learn from, leaving them unprepared for fundamentally new paradigms. Using RDMA as an example, I highlight what such shifts demand from researchers and what they warn about future emerging technologies.

---

Towards A Learning-Directed Operating System

Prof. Aditya Akella

Modern applications run on increasingly heterogeneous and dynamic platforms, yet current operating systems still rely on rigid, locally optimized, and weakly coordinated policies that adapt slowly, often making performance and tail latency limited by poor policy choices rather than hardware constraints. To address this, LDOS (Learning-Directed Operating System) treats policy design as a data-driven, system-wide optimization problem, providing rich observability, fast feedback, and coordinated, trustworthy ML-based control, unlike Linux where policies and mechanisms are tightly coupled. The talk introduces three core components: UNUM for system-wide state embeddings and coordinated decisions, Darwin for making ML-driven policies practical by balancing optimality, generalization, and overhead, and C3 for enforcing system-wide and tail-latency guarantees under adaptive control, and concludes with LDOS’s clean-slate design principles and current progress.

---

How Will AI Systems Do Systems Research?

Ant Rowstron

AI is increasingly capable of writing code, running experiments, and reading papers. What happens when AI systems start doing systems research? This talk explores what that means for the field, for researchers, and for the kinds of problems we choose to work on.