Why AI hasn't replaced software engineers, and won't

Original Article ↗ Hacker News Discussion ↗

Capabilities and Limits of LLM Coding

  • LLMs can now scaffold full apps (web, mobile, infra-as-code) and handle non-trivial tasks (reverse engineering firmware, security scripts, Terraform/Ansible, Frida, etc.).
  • They excel at boilerplate, wiring, and “text-shaped” backend problems; much weaker on nuanced UX, visual taste, and edge cases.
  • Models still make basic logical mistakes (e.g., string-sorting dates) and will confidently pursue wrong paths without realizing it.
  • Effective use requires strong human guidance, testing, and review; unguided “vibecoding” produces large volumes of brittle, low-quality code.

Greenfield vs Maintenance and Complexity

  • For solo devs and tiny teams, LLMs make previously impossible 50–100k LOC projects feasible; time shifts from typing to planning, reading, and QA.
  • In larger systems (many engineers, hundreds of features), complexity, coupling, and invisible constraints dominate. LLMs don’t manage this well yet.
  • Greenfield prototypes are largely automatable; long-term maintenance, integration with legacy systems, and security remain hard and human-intensive.

Replacement vs Augmentation

  • Some report concrete replacement of “several developers” for small startups and pet projects; others argue these are net-new projects that would never have hired engineers.
  • Many see AI as a force multiplier: one good engineer + agents can replace a much larger mediocre team, especially on new projects.
  • Consensus: weak/“ticket shuffler” devs are most at risk; strong engineers who can specify, review, debug, and architect around agents gain leverage.

Decision, Domain Knowledge, and Accountability

  • Core value shifts toward:
    • Deciding what to build (product thinking, domain knowledge).
    • Verifying and being accountable for what ships (tests, security, compliance).
    • Understanding the codebase and business context deeply enough to navigate trade-offs.
  • Non-technical staff are already building internal tools with AI, but quickly hit complexity walls and become de facto undertrained developers.
  • Organizations still need a human “fall guy” and cannot yet offload legal/operational responsibility to AI; vendors offer no liability guarantees.

Economic and Labor-Market Effects

  • Some expect “3D printer/CNC moment”: far more small bespoke tools, fewer traditional dev roles, lower average pay, higher pay for top performers.
  • Others note historical patterns: automation tends to expand total software demand, not shrink it, though specific roles and titles shift.
  • Concern that AI will commoditize coding, hollow out mid-skill roles, and concentrate rewards among a smaller set of high-skill “AI shepherds.”