AI for coding is still playing Go, not StarCraft

AI coding tools excel at isolated, well-defined problems but struggle with large, messy codebases and distributed systems. Drawing parallels between AlphaGo's success in Go versus AlphaStar's delayed mastery of StarCraft 2, the complexity lies not in intelligence but in managing vast configuration spaces, imperfect information, and real-world chaos. Real software engineering involves debugging infrastructure, handling observability data, and navigating non-deterministic failures. To advance AI capabilities, we need proper benchmarks and sandboxes that test performance on complex, distributed systems with realistic workloads and production scenarios.