A16Z partner argues AI coding success does not guarantee operational reliability

Martin Cassato, an A16Z general partner, and Jeb Naus debated whether AI models trained on code can effectively manage production operations. Cassato argued that operations are fundamentally harder than coding because the state space of distributed systems is vast and heavy-tailed. He noted that rare failure states are not well-represented in training data, and the history of systems engineering has relied on over-provisioning rather than smart algorithms because predicting complex runtime behavior is difficult.

Naus countered that while the operational state space is large, the process of debugging is manageable and can be learned by AI if given the right signals. He believes AI can effectively run experiments and verify causal chains to troubleshoot issues, similar to how humans use runbooks. Cassato remained skeptical, viewing AI as a universal function approximator that works best when the solution distribution matches the training data, such as code syntax. He argued that real-world operations involve unpredictable external loads and human knowledge gaps that make economic equilibrium harder to reach.