You can’t build a moat with AI (redux)

Original Article ↗ Hacker News Discussion ↗

AI as (Non-)Moat

  • Many agree: “using AI” isn’t a defensible moat; models commoditize quickly and everyone can bolt them on.
  • Real defensibility still comes from classic sources: workflow integration, proprietary data, distribution, compliance, support, and overall product quality.
  • Some argue even those aren’t strong moats individually (“table stakes”), but in combination (plus brand) they can add up.
  • Foundation model builders may have moats via capital scale and infra; most downstream apps do not.

UX, Control, and “Magic” AI

  • Strong backlash against “magical” AI UX that guesses intent, feels opaque, and fails badly for non‑typical users.
  • Google Search is cited as an example of systems optimizing for “what it thinks you want,” degrading precision and trust.
  • Several commenters explicitly prefer deterministic “tools, not agents”: predictable controls, clear errors, and user autonomy.
  • Others note that for fuzzy tasks (e.g., natural-language document processing), giving high-level instructions to an LLM is genuinely powerful.
  • There’s a deeper split between dev‑style, low-level control UIs vs. “just solve my problem” UIs for nontechnical users; AI can worsen or improve both depending on design.

Hype, Ads, and Perception

  • AI marketing (e.g., Super Bowl‑style enterprise ads) is widely seen as vague, cringe, and focused on “numbers go up” messaging for executives.
  • Some think there’s still little compelling consumer-facing AI; value is mostly enterprise and infrastructure.

Models vs. Applications and Broader ML

  • Commenters stress that “AI” is more than LLMs: CNNs, vision, fraud detection, medical imaging, accessibility, games, and upscaling are cited as real, profitable uses.
  • LLMs are compared to databases/compute: foundational primitives that matter, but not differentiators by themselves.
  • Choice of specific model can matter for niche tasks (e.g., storytelling), so “just use the latest model” is questioned.

Moats, Competition, and Regulation

  • Analogies: AI might evolve like CPUs—initially commoditized, later dominated by a few players as capex and complexity explode.
  • Others compare it to self-driving: long, expensive slog; many vendors die, some survive with limited but valuable capabilities.
  • Data, user base, and network effects (e.g., TikTok) remain strong moats; AI mainly reinforces existing ones.
  • Patents (“…with AI”), regulatory capture, and government alignment are seen as likely tools to manufacture artificial moats.

Cursor / AI Code Editors

  • Cursor’s success is debated: some see it as proof you can build a big business with AI integration; others say its AI isn’t the moat, the product and UX are.
  • Concern that Microsoft can replicate these features in VS Code and undercut on price, underscoring the fragility of AI-only “advantages.”