Berkeley: AI loop beats human-designed systems algorithms

A UC Berkeley group reports that an AI loop which generates, tests and refines systems algorithms produced designs that beat the state of the art, including a 5.0× faster rebalancing runtime for Mixture-of-Experts inference and a 26% cost reduction for multi-region cloud scheduling.

The paper, Barbarians at the Gate: How AI is Upending Systems Research, comes from a group that includes Matei Zaharia and Ion Stoica. The authors call the technique AI-Driven Research for Systems, or ADRS, and state the headline claim bluntly in the abstract: "ADRS discovers algorithms that outperform state-of-the-art human designs."

Their argument for why systems research is where this works first turns on verification. "Solutions are typically implemented in real systems or simulators, and verification reduces to running these software artifacts against predefined workloads and measuring performance," they write. You can tell whether the AI's new scheduler is actually faster. That is not true in most other fields.

The case studies go beyond the two headline numbers. The paper reports a 34% makespan reduction for an offline transaction scheduler, 3× faster runtime on LLM-based SQL query optimization, and up to 16.7% per-workload gains on single-region spot-instance planning. Two researchers on our tracker had already shared the paper by the time we picked it up.

The sharper claim is about the humans. "As AI assumes a central role in algorithm design, human researchers will increasingly focus on problem formulation and strategic guidance," the authors write. In their framing "the researcher acts as an advisor to powerful AI research assistants: defining meaningful problems, proposing creative starting points."

The paper is a position piece. The abstract and introduction do not surface per-run compute budgets, the specific models driving the loop, or how the chosen baselines were picked.