The Craft vs. the Commodity: What We Lose (and Gain) When AI Writes Our Code (syntheticauth.ai)

Two voices clash over the same reality: Højberg mourns programming as a craft threatened by "vibe-coding"—writing specs while LLMs produce the code—arguing this injects non‑determinism, hallucinations, and automation bias into a discipline that has prized predictability and compositional reasoning. He warns that developers skim AI outputs, rely on CI instead of understanding behavior, shift blame to models, and lose the deep "theory‑building" mental models formed by wrestling with code. O'Brien counters that AI generation pragmatically destroys boilerplate and bloat: custom React components, small logging libraries or pagination logic can be produced in minutes, shrinking the moat around packaged libraries and reframing architectural work as rapid evaluation of model outputs rather than prolonged debates—an economic phase change, not just a productivity tool. For the AI/ML community this debate highlights technical and social implications: models must be judged not only by output quality but by reproducibility, provenance, and the way they change developer workflows and skill formation. Practical middle ground: use generation for boilerplate, tests, API exploration and one‑off utilities; preserve human craft for core business logic, novel algorithms, long‑lived systems and architectural decisions. Crucial tooling and norms follow—read and review generated code, add guardrails (provenance, deterministic builds, stronger CI), and train developers to distinguish which code deserves deep immersion versus fast generation.