Coding expertise is going to collapse from AI reliance
Claims that AI coding assistants will erode real programming expertise are drawing comparisons to past shifts like calculators in math and compilers replacing assembly. Commenters describe how large language models let even non-experts generate large volumes of “vibe coded” software, raising concerns about code quality, loss of hard‑won intuition, and a future shortage of people who can debug or reason about complex systems without AI. Others argue that expertise will simply move up a level of abstraction, with the most valuable engineers focusing on architecture, problem selection, and using AI as a tightly controlled tool rather than a replacement for understanding.
Skill Atrophy vs. Normal Tool Progress
- Many argue LLMs will erode “coding intuition” by removing the cognitive friction that used to build deep understanding, similar to concerns about calculators and GPS.
- Others counter that this is just another abstraction step, like moving from assembly to high‑level languages or from hand drafting to CAD; some skills decline, new ones arise.
- Disagreement hinges on whether LLMs are “just another abstraction layer” or qualitatively different because they are non‑deterministic and operate at the level of intent, not just translation.
Education, Juniors, and Learning “Friction”
- Several commenters stress that beginners need years of coding without AI to develop logic, taste, and mental “muscle memory.”
- Studies cited in the thread suggest novices say they use AI like a tutor but actually use it as an answer machine, undermining learning.
- Some educators are experimenting with AI that quizzes the student on their own code or generates textbook‑style explanations and exercises.
Code Quality, Safety, and Complexity
- Many report “vibe‑coded” AI output as “functional enough” but structurally poor, especially when models iterate over their own earlier code.
- There is concern about huge AI‑generated PRs that humans can’t realistically review, leading to unmanageable, opaque systems.
- Others say good engineers can harness LLMs to refactor, simplify, and build better architectures faster; well‑structured codebases become powerful amplifiers.
Role of Human Expertise
- Strong consensus that LLMs still require experts to:
- Frame the right problems.
- Impose constraints and architecture.
- Decide when code is “good enough,” safe, and maintainable.
- Skeptics worry that if fewer humans gain this expertise, future systems will depend on tools no one really understands.
Economic and Labor Impacts
- Some foresee many “average” coders being replaced or devalued, with demand concentrating on high‑level generalists and system thinkers.
- Others expect a shift similar to spreadsheets: many more non‑experts building “good‑enough” software, with specialists maintaining underlying platforms.
Usage Patterns: Assisted vs. Autonomous
- A divide emerges between:
- “Guided coding” (LLM as autocomplete, explainer, reviewer) seen as highly productive and educational.
- “Agentic/headless coding” (tickets → agents → code) seen as risky, hard to review, and a driver of slop and skill decay.