Large Language Models for Automatic Cloud Incident Management
summary
A system predicts root cause and mitigation at incident creation by fine-tuning foundation LLMs and evaluating against production data; fine-tuned DaVinci-002 outperforms zero-shot GPT variants, trained on a root-cause corpus spanning up to ~124k incidents and a 23.5K mitigation set, and validated with human reviewers (over 70% rate recommendations ≥3). Earlier empirical work analyzes 152 high-severity incidents to prioritize dependency failures and human-heavy mitigations, and the platform moves toward multi-modal inputs, plugins, and chain-of-thought validation.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:24System Innovation
- 1:20What is AI Ops
- 2:29Service Reliability
- 3:38Service Reliability Opportunities
- 6:36Findings
- 8:35Emergence
- 8:52Challenges
- 9:49Research Background
- 10:57Fine Tuning
- 12:17Evaluation
- 12:46Results
- 13:33Feedback
- 14:44Conclusion
- 18:20Thank You
related talks
13mIntegrating Large Language Models into Security Incident Response
26mBecoming AI engineers
46mHow Large Language Models Are Reshaping the Cybersecurity Landscape
40mIRonMAN: InterpRetable Incident Inspector Based ON Large-Scale Language Model and Association miNing
33mLessons Learned Applying Large Language Models in Healthcare
38mNext-Gen Detection: Harnessing LLMs for Sigma Rule Automation
36mModernizing Incident Response with LLMs, RAG, and the MCP
22mFrom the Trenches: Automated Testing of Large Language Models