AI, Rockets, and Why Engineering Judgment Still Costs the Same

The popular claim that AI makes software cheap misses what's actually happening: AI cheapens implementation, the act of turning a decision into working code, not the judgment surrounding it. Deciding what a system should do, defining how its parts interact, drawing security boundaries, and evaluating whether generated code actually works remain as expensive as ever. The result is a structural mismatch: AI agents can produce code far faster than humans can meaningfully review it, most of that output is mediocre or subtly flawed, and occasionally something genuinely better than human-written code slips through the flood.

To frame this problem, the piece draws an analogy from rocket engineering: Apollo bought preflight confidence through expensive modeling, static-fire testing, and full-scale rehearsal, backed by immense resources. The Soviet N1 program, lacking those resources, relied on direct iteration — build, fire, inspect, modify, fire again — which produced genuinely superior individual engines. But the N1's thirty-engine first stage was never tested as a complete system, and all four launches failed, because component-level success didn't compose into system-level reliability.

The lesson for AI-assisted software development is that treating every AI-generated artifact as a singular, polished deliverable deserving exhaustive human review is 'Apollo thinking' misapplied to a firehose of generated code. Instead, organizations should adopt the engine shop's model: generate liberally and iterate destructively where failure is cheap, while keeping rigorous, contract-driven scrutiny exactly where integration and system-level meaning are at stake.