#story Let the Barbarians In: How AI Can Accelerate Systems Performance Research Audrey Cheng, Shu Liu, Melissa Pan, Zhifei Li, Shubham Agarwal, Mert Cemri, Bowen Wang, Alexander Krentsel, Tian Xia, Jongseok Park, Shuo Yang, Jeff Chen, Lakshya Agrawal, Ashwin Naren, Shulu Li, Ruiying Ma, Aditya Desai, Jiarong Xing, Koushik Sen, Matei Zaharia, Ion Stoica UC Berkeley 2025 https://lnkd.in/gr2WKM-M For decades, improving the performance of computer systems—making databases faster, clouds more efficient, or networks more reliable—has depended on human intuition, experience, and painstaking trial and error. This paper argues that a new kind of “barbarian” is at the gate: AI systems that can explore vast design spaces far faster than any individual researcher ever could. The key insight is surprisingly simple. Many systems research problems come with built-in referees. If you propose a new scheduler, load balancer, or query engine, you can test it in a real system or a simulator and immediately measure whether it performs better or worse. That makes systems performance an ideal playground for AI-driven research, where models can repeatedly generate ideas, test them, learn from the results, and try again. The authors call this loop AI-Driven Research for Systems (ADRS). Using several open-source ADRS tools, they apply this approach to ten real-world problems, from cloud scheduling across regions to load balancing in large AI models and accelerating SQL queries with language models. Across these case studies, the AI-generated designs don’t just imitate human solutions—they often match or even outperform the best techniques developed by experts. Beyond the results themselves, the paper draws practical lessons. It shows how much guidance AI needs to be productive, how detailed feedback should be, and why careful evaluation is critical to avoid being fooled by apparent gains. The message is not that AI replaces systems researchers, but that their role is changing: less time spent hand-crafting every optimization, and more time spent framing the right problems, defining good benchmarks, and steering the search. There is no one-size-fits-all recipe yet. But the authors suggest that as these tools mature, systems research may increasingly resemble directing an intelligent swarm of explorers—letting the “barbarians” roam the design space—while humans focus on judgment, insight, and long-term direction.