AI-assisted coding will change software engineering: hard truths

Original Article ↗ Hacker News Discussion ↗

Future trajectory of AI coding tools

  • Debate over whether current LLM limits are temporary or near a plateau.
  • Optimists expect rapid, possibly exponential improvement and serious job displacement within 3–10 years.
  • Skeptics argue we may be near a local maximum (data limits, diminishing returns from scale) and that big further gains are not guaranteed.
  • Some compare this to other tech that improved fast then plateaued (planes, cars); unclear if LLMs are early “biplanes” or already near maturity.

Current capabilities and the “70% problem”

  • Many report LLMs are great for boilerplate, API usage, and simple scripts, but consistently wrong or incomplete on the final 20–30%.
  • Tools often hallucinate APIs, miss small syntax details, or get stuck cycling between different broken versions.
  • Effective use requires micromanagement, verification, and domain knowledge.

Impact on developers, skills, and careers

  • Concern that juniors will skip “hard parts,” leading to fewer truly competent intermediates/seniors and atrophy of existing expert skills.
  • Some see a future where experienced engineers become more valuable precisely because they can supervise and repair AI output.
  • Fears that employers will demand AI use to justify lower pay and fewer engineers.

Code quality, maintenance, and security

  • Widespread worry about “AI-generated slop”: superficially impressive demos that fall apart in edge cases.
  • Anticipation of more security bugs and brittle “house of cards” systems needing expensive cleanup.
  • LLMs struggle with large, idiosyncratic brownfield codebases and internal patterns.

Frameworks, languages, and abstractions

  • One view: AI will encourage custom frameworks per company, hurting skill transfer; another: LLMs will push consolidation around popular ecosystems they know best.
  • Some argue languages are “verbose and stupid” and real progress would be better DSLs and formal methods, not just AI autocomplete.
  • Others stress that programming is about precise communication; natural language remains too vague.

Agents and automation

  • Heavy skepticism about “agent” hype: many see them as glorified scripted workflows with an LLM front-end, far from autonomous Jarvis-like systems.
  • Comparisons made to NFT/metaverse hype cycles.

Practical uses today

  • Common replacements: Stack Overflow and basic search, especially for library boilerplate and quick Python/pandas/matplotlib tasks.
  • Single-line or small-snippet completion inside IDEs seen as a good balance; full in-editor generation often degrades architecture over time.
  • Tests generation and deep refactors are cited as areas where current tools underdeliver.

Ethics, incentives, and product quality

  • Strong concern over unethical training, labor exploitation, environmental costs, and “paying to train your replacement.”
  • Expectation that productivity gains will fund more features, not better quality, reinforcing existing “enshittification” trends.
  • Reference to automation paradox (Bainbridge): more automation may reduce human practice while increasing the need for expertise during failures.