Domain expertise has always been the real moat

Original Article ↗ Hacker News Discussion ↗

Domain Expertise as Moat (Disputed)

  • Many agree that deep, tacit domain knowledge (e.g., freight optimization, title insurance, medicine, finance) is hard to encode and still critical.
  • Others argue LLMs already embed “80%” of most domains and let generalist devs get proficient quickly.
  • Distinction emerges between:
    • Public, well-documented domains (chess, generic security cameras, some finance) where models do well.
    • Highly local, regulatory, political, or idiosyncratic domains where they currently fail.

What LLMs Are Good At vs. Bad At

  • Strong at: coding boilerplate, search/summarization, exploring unfamiliar domains, generating prototypes, troubleshooting common errors.
  • Weak at: consistent long-term reasoning, non-trivial specs, deep cross-domain dependencies, avoiding subtle bugs/security flaws, and handling incomplete or shifting requirements.
  • Several report that “agentic workflows” can massively increase volume of output, but not reliably its correctness.

Verification, Specs, and Tacit Knowledge

  • Key shift noted: from “can you build it?” to “can you tell if it’s right?”
  • Domain experts often can judge correctness of examples but struggle to fully specify rules (Polanyi’s paradox).
  • Many doubt non-technical experts can prompt agents precisely enough for complex systems without an engineer-like mindset.

Software Engineering as Its Own Domain

  • Multiple comments stress that systems design, scalability, reliability, data modeling, security, and maintainability are deep domains themselves.
  • “Vibe-coded” apps from domain experts often technically work but have poor schemas, fragile abstractions, and massive tech debt.
  • Consensus that you can’t “QA your way” into quality; architecture matters.

Who Thrives in the AI Era?

  • One camp: senior generalist engineers plus AI become far more effective across domains.
  • Another: domain experts with moderate tooling skill + AI now rival or surpass traditional devs for many business apps.
  • Many foresee new hybrid roles: platform engineers building guardrails for domain users; engineers shifting toward product, strategy, and sales.

Uncertainty and Disagreement on Trajectory

  • Some see an inevitable path to superhuman AI dominating most knowledge work.
  • Others emphasize past over-optimism (self-driving, AGI timelines) and point out current tools still need heavy supervision.
  • Broad agreement on one thing: the landscape is changing fast, and no one has a stable, proven playbook yet.