a16z Podcast: Why AI Writes Code Faster Than It Keeps Systems Running

SPEAKER_1 (0:01)

The world of AI coding looks digital and tidy at first glance, but this episode asks a harder question: what happens when that same intelligence has to keep real systems alive? That's the exact challenge Jad Naous tackles. His journey is fascinating—Stanford, MIT, startups, infrastructure, venture capital—and now he's building Greer, focused on observability and reliability. What's driving that focus? He found that engineering teams were spending 5% to 20% of their infrastructure budgets on observability, but what they were really trying to buy was reliability. Greer's answer? Extract signal from noise earlier. As Jad puts it, We plug into the data stream, reduce costs by 90%, and feed those signals into an AI layer that troubleshoots and fixes issues. And the coding landscape itself has transformed remarkably fast. Jad says AI has moved from simple autocomplete to something strong engineers now fold into daily workflows. He's using persistent command-line sub-agents and Claude 3.5 Opus to ship massive projects far faster than before. How much faster are we talking? He built an auto-promotion feature for a data lake in about 10 days—work that once would have taken six months.

But here's the catch: "The disparity between average and great engineers has multiplied 10x, " because breaking down complexity is still a human advantage.

That brings us to the core distinction he makes between code and operations.

Jad defines "syntax" as the code you write, and "semantics" as the actual runtime behavior of the system. And that gap matters enormously. Distributed systems depend on shifting traffic, workloads, caching, locking—factors that models don't automatically understand. AI can optimize well when it has clear benchmarks, like SQL tuning, but it has far less native knowledge when the problem involves memory, resilience, or production tradeoffs. So what's the most promising path forward? Jad argues it's not traditional root-cause analysis, but anomaly detection. AI is best used to spot the needle in the haystack, inspect the signal, and help verify whether an incident is real.

Full autonomy over complex clusters? Still unsolved. The broader takeaway is actually optimistic, though. Jad pushes back on fatalism, saying, "It is not fatalistic to believe that; it is the opposite."

AI will keep expanding into domains where the solution space is constrained and the data matches the task. But when the work depends on the physical world, or on systems with too many unpredictable states, the challenge gets much harder. Exactly. For now, coding is ahead of operations, and the race to close that gap is just beginning.