Anthropic SRE Alex Palcuie, speaking at QCon London, shared real-world experiences using Claude for incident response. While AI excels at log analysis and pattern detection at scale — including catching a fraud case that would have been missed — it consistently confuses correlation with causation, misdiagnosing a broken KV cache as a capacity problem. Claude produces readable but shallow postmortems that miss true root causes and lacks institutional knowledge of long-running systems. Palcuie warns that over-reliance on AI may atrophy human SRE skills, and invokes the Jevons Paradox: AI-driven efficiency may increase software complexity and thus incidents. His conclusion is that human SREs remain essential and should keep being trained.
Sort: