AI makes tech debt more expensive

Original Article ↗ Hacker News Discussion ↗

AI, Tech Debt, and Codebase Age

  • Many agree: AI tools work best on “young, clean, idiomatic” code and struggle with legacy systems full of edge cases and non‑standard patterns.
  • Others argue that what’s labeled “tech debt” is often maturity: accumulated fixes and “scars” from real-world bugs and business quirks, not just mess.
  • Some say AI doesn’t make tech debt more expensive per se; it just delivers less benefit in high‑debt environments.

Speed vs Quality in Early Stages

  • Debate over whether early-stage code “must” be messy to move fast.
  • Several argue high-quality code is actually required for sustained speed; cutting corners only gives a brief boost before slowing you down.
  • Others note that “clean” young code often just lacks edge cases; complexity grows as real users push requirements.

LLM Strengths and Weaknesses

  • Widely seen as strong for: boilerplate, scaffolding, rote transformations, basic tests, small functions, refactors within clear interfaces.
  • Weak for: novel or weird architectures, complex legacy logic, devops/config in fast-moving ecosystems, and deep reasoning about subtle edge cases.
  • People report frequent “demo-level” solutions: missing pagination, incomplete edge handling, overcomplicated dependencies, or superficially correct but subtly wrong code.
  • Many treat LLMs as “smart but unreliable juniors”: helpful, but everything needs review.

Rewrites, “Scars,” and Maintainability

  • Strong caution against full rewrites: old complexity often encodes hard-won bug fixes and business logic.
  • “Scars/warts” are seen as necessary complexity; removing them without understanding why they exist leads to rediscovering old bugs.
  • Good tests and documentation (especially “why” decisions were made) are highlighted as the real enablers of safe refactoring—human or AI-driven.

Productivity, Workflow, and Hype

  • Some report real productivity gains, especially in test generation, yak‑shaving tasks, and cross-language snippets.
  • Others find verification and debugging of AI output negates any time saved.
  • Skepticism about the article’s thesis and about broader AI hype is common: claims of “AI transforming software” are seen as premature, with current tools closer to advanced autocomplete than a solution to tech debt.