I tried coding with AI, I became lazy and stupid

Original Article ↗ Hacker News Discussion ↗

How People Are Actually Using AI to Code

  • Several modes described: “vibe coding” whole features from short prompts, using AI as super‑autocomplete, rubber‑ducking / explanation, or doing long, structured back‑and‑forth design sessions.
  • Many say AI works best on small, well‑scoped tasks (100–200 LOC, boilerplate, glue code, simple Rust/React bits) or for reading unfamiliar code and summarizing docs.
  • Some senior devs claim they now “only program using LLMs” but always review, refactor, and keep architecture decisions human‑driven.

Perceived Benefits

  • Faster first drafts, fewer keystrokes, and removal of tedious work (CRUD, React boilerplate, parsing helpers, UI they’d never have built otherwise).
  • Helps juniors or non‑experts tackle languages and projects they previously found too intimidating.
  • Frees some to think more about architecture and higher‑level design instead of low‑level details.

Risks: Laziness, Lost Understanding, and “Slop”

  • Many echo the article: letting AI design and write large chunks leads to poor “bird’s‑eye” understanding and painful maintenance.
  • Reports of convoluted codebases, stylistic mess, security issues, and “mountains of barely working slop,” especially in open source contributions.
  • Some reviewers notice teammates firing off AI‑generated PRs, relying on others to catch obvious bugs, optimizing for perceived velocity.

Prompt Engineering and “You’re Holding It Wrong”

  • One camp insists skillful prompting/context is essential; AI should be treated like an overeager junior whose work must be guided and checked.
  • Critics push back that “you prompted it wrong” is unfalsifiable and shifts blame from real model limitations; deterministic, predictable behavior is still lacking.

Productivity Studies and Experience Gaps

  • A frequently cited study on experienced OSS devs reported a 19% productivity decline with LLMs.
  • Others cite different studies showing 25–55% gains, especially for junior developers.
  • Several suggest: AI seems more beneficial for less experienced devs; for seniors the picture is mixed and highly workflow‑dependent.

Proposed Middle Ground / Best Practices

  • Use AI as a force multiplier, not an architect; never ship code you don’t understand.
  • Keep your own checklist of pitfalls (security, correctness) and fold recurring issues back into prompts or project docs.
  • Restrict AI to well‑defined tasks, write design docs first, review every diff, and sometimes deliberately code without AI to avoid skill atrophy.