The air is hissing out of the overinflated AI balloon

Original Article ↗ Hacker News Discussion ↗

Fast Food, Kiosks, and “Easy” Jobs

  • Debate centers on whether AI’s failure at McDonald’s-style drive‑thrus means it can’t do most jobs.
  • Multiple commenters note the cited system is pre‑LLM (2019 IBM, decision trees), so not representative of current tech.
  • Others stress drive‑thru work is much harder than it looks: noise, multiple speakers, regional names, coupons, makeup orders, and real‑time coordination with kitchen shortages.
  • Kiosk/tablet ordering already replaces many order‑takers; some argue apps + QR codes will further reduce need for AI.
  • User experience is polarizing: some hate kiosks (lag, breakage, dark‑pattern upsells); others strongly prefer them for clarity, speed, language issues, and reliable customization.

Ambiguity, Customization, and Human Judgment

  • Long subthread on what “plain” means in fast‑food orders shows how much context and clarification humans handle implicitly.
  • Views differ on whether staff should always clarify ambiguous terms vs. prioritize speed and accept a small error rate.
  • Humans also handle messy edge cases (wrong items on the grill, “ice cream machine down,” partial shortages) that current systems struggle to encode.

LLMs vs Traditional Automation

  • Some argue massive LLMs are overused “gold-rush shovels” where simpler, older automation (rule systems, vision, CNC, self‑checkout) already works fine.
  • Others think LLMs would be strong at constrained menu ordering, especially as chains ruthlessly optimize efficiency.

AI as Tool, Not Magic

  • Several developers say AI is now a standard tool: great for debugging and boilerplate, often wrong but far faster than web search.
  • Concerns raised about unknown true costs: energy, water, pollution, and especially “information pollution” as AI‑generated slop degrades search results.

Bubble, Plateau, and Long‑Term Impact

  • Many see a classic bubble: tech is real and transformative, but companies and hardware build‑out are overhyped, echoing dot‑com.
  • Frequent reference to Amara’s law: short‑term overestimation, long‑term underestimation.
  • Strong disagreement with the article’s claim that AI is “as good as it’s going to get”: most expect continued, but slower, improvement; some think current text‑model gains already look incremental.
  • Consensus: even if the financial bubble pops, LLMs and other AI (vision, specialized models like AlphaFold, self‑driving) will persist and keep reshaping work.