What we lost the last time code got cheap
The author ran medical transcription infrastructure during the outsourcing era and sees the same cost-migration pattern in AI codegen: production gets cheap, comprehension becomes the scarce resource.
What Matters
- Heartland Information Services was a top-tier US medical transcription provider, running hybrid US/India teams where downtime had surgical consequences.
- Offshoring moved knowledge to the wrong timezone; AI codegen goes further—no human ever held the intent at all.
- Joel Spolsky’s 25-year-old claim that reading code is harder than writing it is now structurally unavoidable, not just culturally true.
- Post-outsourcing survivors invested in shared context, code review, and documentation as first-class engineering—not proximity-dependent byproducts.
- Prediction Machines framing: when a core input gets cheap, value shifts to complements; in software, the complement of production is understanding.
- [HN: @pmmucsd] Teams running agents are capturing architectural decisions in a decisions.md file, updated by CLAUDE.md/AGENTS.md whenever the agent must ask a human—intent that code never holds.
- [HN: @wiremine] The honest caveat: much human-written code also lacked context; the pre-AI baseline wasn’t clean, complicating the loss narrative.
