Reports of code's death are greatly exaggerated

Original Article ↗ Hacker News Discussion ↗

Perception of “code is dead”

  • Many argue code isn’t disappearing; the job of programmers is shifting up the abstraction stack.
  • Repeated “code is dead” narratives are likened to past waves (no‑code, visual tools). The work changes, not the need for code.
  • Some foresee future greenfield work as mostly specs/tests plus a small group of expert “code janitors.”

AI-generated code, quality, and comprehension

  • Heavy use of agents is said to create “comprehension debt”: large codebases no one truly understands.
  • Examples cited: AI-induced outages at large cloud providers and subsequent requirements for human review.
  • Developers report AI producing “mostly OK” code with subtle bugs, increasing the burden on senior reviewers.
  • Others counter that human-written legacy code is often just as bad; AI is “a nail gun,” not the root problem.

Management, hype, and process

  • Some struggle to convince leadership that AI won’t eliminate the need for engineers; optimism about future models often trumps present failures.
  • Suggested strategy: embrace experiments, lead pilots, then surface concrete costs (maintenance, bug tickets, senior time) in business terms.
  • Comparisons drawn to “shift-left” security: noisy hype, mixed outcomes, lasting process changes.

Innovation, creativity, and limits of LLMs

  • Strong view: current models interpolate consensus; they rarely advance the state of the art (e.g., AI-written compiler deemed conventional).
  • Counterview: that’s enough for 99% of work; creativity can emerge via large-scale automated experimentation or reinforcement learning.
  • Debate over whether neural nets can meaningfully extrapolate or just approximate within known regions.

Language, abstraction, and natural language

  • Some argue natural language specs plus AI will replace most direct coding, similar to moving from assembly to high-level languages.
  • Others stress that code remains the most precise, unambiguous way to express complex behavior, especially for critical systems.
  • Classic critiques of “natural language programming” are revisited; supporters respond that today’s systems are qualitatively different.

Economics, careers, and vendor lock-in

  • Concern that AI may reduce demand for “1x programmers,” concentrating work in fewer, more expert roles.
  • Others note business problems are effectively endless and see AI as leverage, not replacement.
  • Significant worry about deep lock-in to specific AI vendors: prompts, workflows, and model-specific behaviors may be non-portable.

Current practical sweet spots

  • Many use AI effectively for:
    • Glue code (OAuth, API integration, boilerplate).
    • Reading docs and wiring unfamiliar systems.
    • Test generation, refactors, simple scripts.
  • For novel architectures, tricky algorithms, or new CRDTs/frameworks, humans still report doing most of the real design work.