Software Development in the Age of LLMs: How Many Teams Still Code Like 2021?

TL;DR

Many engineers continue to code without large language models (LLMs), particularly in regulated industries, government, and niche domains, yet corporate pressure and C‑suite directives are pushing AI adoption across most software teams.

1. Not All Companies Have Adopted LLMs

  • Regulated sectors resist change – Teams building medical devices in the U.S. report no AI usage because legal review and compliance slow any new tooling. (Comment by cowanon77)
  • Government and public‑sector work stays manual – State and federal developers cite a five‑year lag before adopting unproven tools, keeping their workflows similar to 2021. (Comment by bthallplz)
  • Embedded and safety‑critical code often bans AI – Contractors working on government‑bound embedded devices are contractually prohibited from using LLM‑generated code due to unresolved copyright and liability concerns. (Comment by voakbasda)
  • Niche technical domains lack model coverage – Debugging low‑level networking or custom bit‑banging drivers yields little benefit from LLMs because training data rarely covers those specifics. (Comment by mysterydip and dotdi)

2. Corporate Pressure Is Growing

  • C‑suite AI mandates – Some companies have explicit executive orders to reach 70‑80 % AI‑generated code within a year, even though internal policies flip‑flop daily, leaving developers uncertain about what tools are allowed. (Comment by timbaboon)
  • Pilot projects and gradual rollout – Organizations are experimenting with AI in testing, verification, and requirements management before full‑scale adoption. (Comment by cowanon77)
  • Token caps and licensing bottlenecks – Even when LLMs are permitted, token quotas force developers to be strategic, and only a subset of staff receive licenses (e.g., Codex vs. GitHub Copilot). (Comment by francisofascii)

3. Individual Strategies to Stay Human‑Centric

  • Treat coding as mental exercise – Some engineers solve daily LeetCode problems to keep their problem‑solving muscles sharp, regardless of AI assistance. (Comment by andreybaskov)
  • Use LLMs only for peripheral tasks – Developers report using models for documentation, API look‑ups, test scaffolding, or quick code snippets, while reviewing output rigorously (≈80 % correctness). (Comment by ww520)
  • Maintain a “socratic” workflow – The skill shift is from typing code to asking the right questions; this approach still requires deep context knowledge. (Comment by jolt42)
  • Reserve AI for non‑core activities – Independent contractors may employ LLMs for security audits, refactoring, or log analysis, but deliberately avoid delegating architecture and design. (Comment by voakbasda)

4. Mixed Experiences with Agentic Coding

  • Short‑term trials often disappoint – Repeated four‑month experiments with autonomous agents produced buggy, hard‑to‑maintain code, leading many teams to revert to manual development. (Comment by QuadmasterXLII)
  • When it works, it can be impressive – A recent trial with “Fable 5.1” showed better results, suggesting that tool maturity and prompt engineering matter. (Comment by QuadmasterXLII)
  • Code quality concerns – Teams report AI‑generated code inflating test suites, creating noisy diffs, and reducing ownership, which can erode codebase health. (Comment by localhoster)

5. The Job Market Landscape

  • Job ads still vary – Some listings explicitly require LLM experience, while others omit any mention, indicating a transitional hiring environment. (Comment by mikemarsh)
  • Contractors retain more agency – Freelancers can choose whether to incorporate AI, though client contracts may dictate otherwise. (Comment by voakbasda)
  • Potential career pivot – Developers uncomfortable with AI‑centric workflows consider moving to sectors that are slower to adopt or to entirely new careers. (Original post by sph)

6. Outlook

  • Inevitable diffusion – Even in highly regulated or conservative domains, AI tools are expected to become legally sanctioned once copyright and liability frameworks solidify. (Comment by voakbasda)
  • Human expertise remains critical – Core architectural decisions, domain‑specific driver development, and nuanced debugging still rely heavily on human insight. (Multiple comments)
  • Balancing act – Successful teams will likely combine rigorous human review with selective AI assistance, rather than fully automating code generation.

This synthesis reflects the range of experiences shared on Hacker News, highlighting where traditional coding persists, why AI adoption is accelerating, and how developers are navigating the transition.

Sources

Related