Summary
In this a16z podcast, two longtime colleagues who met in the same Stanford research group discuss why AI has become so good at writing code but still struggles to operate the systems that code runs on. The guest, the CEO of a data infrastructure company, describes how AI coding went from autocomplete to building complex software, including a distributed-systems feature that took about 10 days instead of roughly six months for a single engineer.
The conversation turns to the harder problem of operations. Correctness, availability, and performance in distributed systems depend on workload, access patterns, and traffic, none of which are visible in the code itself. The hosts ask where the training data would come from to teach a model to manage clusters, tune databases, and make tradeoffs such as eventual consistency versus locking or caching versus freshness.
Listeners get a systems-level view of where AI is already delivering 10x to 20x gains on production code, why the real world is messier than the digital environment models learn in, and what it would take to let AI run services reliably. It is a useful listen for engineers, infrastructure leaders, and anyone tracking how far agentic coding can extend beyond writing software.
📺 Source: a16z · Published October 16, 2026
🏷️ Format: Podcast
