The Next Horizon of System Intelligence (www.sigops.org)

A group of systems and AI researchers has launched a blog series calling for community discussion and contributions on "system intelligence" — the idea that generative AI (LLMs and related models) can move from exogenous assistants (autotuners, ML-based optimizers) to becoming endogenous, self-evolving components of computing systems. They argue this shift could reshape how we design, evaluate and trust systems: instead of engineers hand-crafting interfaces and knobs, systems would accept high-level, declarative goals, monitor their own behavior, reason about trade-offs, and generate policies, code, and configurations autonomously. The authors ground the vision in recent AI capabilities: pretrained models capture system fundamentals (cache coherence, consensus, scheduling), in-context learning adapts to scenarios, advanced reasoning helps interpret complex logs and interactions, tool use enables live system observation (profilers, CLIs, APIs), and code generation plus natural-language summaries let models implement and iterate on designs. They also flag critical research questions — new abstractions and correctness principles for self-evolving systems, safety/interpretability limits, how to train the next generation of system engineers, and whether AI can autonomously design and formally verify massive systems. The series invites principled debate, empirical stories, and tooling best practices to chart this frontier for the SIGOPS and broader AI/ML communities.