Скачать или смотреть Barbarians at the Gate: How AI is Upending Systems Research (October 2025)

Title: Barbarians at the Gate: How AI is Upending Systems Research (Oct 2025)
Link: http://arxiv.org/abs/2510.06189v2
Date: October 2025

Summary:
This paper explores the transformative potential of AI-driven research in systems, focusing on automating algorithm discovery. It argues that system performance problems, characterized by reliable verifiers, are well-suited for AI-Driven Research for Systems (ADRS). Through case studies using OpenEvolve, the paper showcases ADRS's ability to discover superior algorithms in cloud scheduling, load balancing, SQL query optimization, and transaction scheduling. The paper discusses best practices for guiding algorithm evolution and implications for the systems community, advocating for a shift in human researchers' focus to problem formulation and strategic guidance.

Key Topics:
AI-Driven Research for Systems (ADRS)
Algorithm Discovery
Systems Research
OpenEvolve
Machine Learning
LLMs
Cloud Scheduling
Load Balancing
SQL Query Optimization
Transaction Scheduling

Chapters:
00:00 - Introduction
00:18 - Barbarians at the Gate
00:31 - AI-Driven Research for Systems
01:15 - Key Takeaways
02:11 - Why Systems Research?
03:38 - PhD Student Time Allocation
04:02 - Open Evolved Framework
04:35 - Prompt Generator
05:05 - Solution Generator
05:38 - Evaluator
06:18 - Feedback Loop
06:39 - Solutions Selector
07:27 - Results
07:35 - Mixture of Experts
08:59 - Cloud Spot Instances
09:47 - Multi-Region Savings
10:44 - Transaction Scheduling
11:49 - The Power for Researchers
12:11 - What Goes Wrong
12:21 - Runtime Errors
12:39 - Search Failures
13:16 - Algorithm Failures
14:17 - Best Practices
14:35 - Starting Point
15:47 - Evaluated Design
16:30 - Big Picture Implications
17:31 - Limitations
18:14 - Weak Verifiers
19:11 - Open Challenges
19:50 - More Agentic LLMs
20:37 - Guiding Objective Function
21:09 - Statement of Arrival
21:35 - Compounding Acceleration
21:55 - Challenge to the Community