Why AI Can Write Code But Can’t Keep It Running (Yet)

影片總結

Grepr CEO Jad Naous says AI cut a six-month production project to 10 days, but autonomous cluster operations remain unsolved.

Grepr CEO Jad Naous tells a16z General Partner Martin Casado that observability can consume 5–20% of infrastructure spending, while delivering diminishing reliability gains. Grepr first built a plug-and-play telemetry filter that can reduce data costs by 90%, and is now using detected changes in logs to guide AI troubleshooting. Naous describes coding agents as a major productivity boost: using Claude Code, he shipped a complex data-lake feature in 10 days, versus an estimated six months previously, with human review still essential. For operations, he says AI works best when systems surface unusual changes and the agent investigates possible causes; ignored early warnings at customer sites later became incidents. The guests agree that running distributed systems is harder than writing code because production behavior depends on workloads, rare failures, and knowledge scattered among people. Naous sees promise in AI-driven investigation, but both say fully autonomous cluster operations remain an open challenge.

章節

1. 0:00From Stanford to founding Grepr: Jad’s SHA-3 work, AppDynamics growth, and observability research
2. 7:23Why observability got so expensive: microservices expanded telemetry while reliability gains flattened
3. 10:15How AI is changing engineering: adaptable engineers gained a 10x edge through better AI workflows
4. 14:02How AI is changing engineering: distributed-system complexity still depends on human judgment
5. 18:25Jad’s AI stack and workflow: WSL, Opus 5.5, and three to four agent sessions
6. 22:55A six-month project in ten days: Jad built data-lake autopromotion for production
7. 24:55Can AI reason about operations? Query benchmarks help, but production experiments remain necessary
8. 30:10How Grepr catches incidents early: new log patterns and warnings that preceded incidents
9. 35:01How Grepr catches incidents early: scanning a petabyte of logs to surface new patterns
10. 38:35How Grepr catches incidents early: AI tests causal chains from telemetry
11. 42:16Why security is the harder problem: adversaries exploit rare states and operations failures
12. 45:47Why security is the harder problem: code descriptions cannot replace running systems
13. 49:05Why security is the harder problem: AI’s gap between digital training and real-world work
14. 53:33What comes after code: real-world AI takes longer, while Grepr customers seek 90% lower Datadog bills