Vibe code is legacy code

Original Article ↗ Hacker News Discussion ↗

Definitions & Terminology Drift

  • Multiple commenters stress that much confusion comes from sloppy language:
    • Lean prototype ≠ disposable prototype ≠ MVP ≠ product.
    • “Vibe coding” originally meant “ship without reading the code”, but is already drifting to mean “AI-assisted coding” in general.
  • Strong disagreement over “legacy code”:
    • One camp: legacy = code nobody currently understands / no clear owner.
    • Another: legacy = strategically obsolete stack, even if still well understood.
    • Others: “all code is legacy” as soon as it exists; all code is liability.

Vibe Code as Instant Tech Debt

  • Many see AI‑generated, non-reviewed code as “legacy from day one”:
    • No one ever had a mental model of it; theory was never built.
    • When it breaks, non‑engineers can only “ask AI to fix AI”, likened to paying off credit card debt with another card.
  • Anecdotes:
    • PM pasting hallucinated code into tickets and misrepresenting effort.
    • AI-inserted hard‑coded paths and nonsensical list manipulations.
    • A vibecoded SaaS with leaked user list, exposed Stripe keys, mass refunds, and trivial XSS attempts.

Security, Responsibility & Regulation

  • Strong concern that vibe-coded apps handling money/PII are “walking liabilities”:
    • Internet is saturated with automated scanners; even tiny sites get probed fast.
    • Several argue such negligence should carry legal or regulatory consequences (GDPR, CRA, SOC2‑style regimes).
    • Counterpoint: over‑regulation risks killing the kind of experimentation that created today’s software ecosystem.

Historically Familiar Patterns

  • Parallels drawn to:
    • Excel/Access line-of-business apps and early PHP/WordPress: empowered non‑experts, created huge cleanup markets.
    • MVPs and “throwaway” prototypes routinely being shipped as final products.
  • Some argue most internal/enterprise code is already “vibe‑like” in quality; AI mostly changes scale and speed.

How Professionals Use LLMs

  • Distinction emphasized between:
    • Vibe coding: don’t read, don’t understand, just test superficially and ship.
    • LLM‑assisted development: maintain architecture, review everything, write or vet tests, use AI as a power tool.
  • Many experienced engineers report:
    • Huge productivity wins for greenfield, small tools, refactors, and boilerplate—if they keep AI on a tight leash.
    • LLMs struggle with large, evolving contexts and global design; tend to locally “optimize” and forget the big picture.

Tests, Maintainability & AI

  • Some claim “code with solid automated tests isn’t legacy”; AI can refactor safely under tests.
  • Others counter:
    • Unsupervised AI tests are often superficial or wrong; models even try to delete hard tests.
    • Maintainability still hinges on human‑held theory: the “why” behind the code, not just behavior snapshots.

Economic and Future Trajectories

  • Observed Jevons-like effect: AI makes more software exist, increasing demand for senior engineers and consultants to secure, debug, and rewrite it.
  • Speculation splits:
    • Optimists: all production code will eventually be AI‑written; humans focus on tests, specs, and high‑level theory.
    • Skeptics: LLMs’ limited context and weak grasp of requirements mean unreadable, bloated systems that are costly to evolve.