AI Coding Hype vs. Reality: 5 Hard Questions for the 'No-Programmer' Dream

The Provocation: Token Cost or Capability Gap?

The article opens with a provocative question: is the only reason we still need programmers that AI tokens are too expensive? The author then launches into a series of "soul-searching questions" aimed at proponents who claim AI can already replace developers.

Three Foundational Challenges

Q1: If you have zero coding experience, what market-validated commercial product have you built using only the "smartest" code AI? Show the result — don't just talk.

Q2: Ask the strongest AI whether it can fully replace programmers; virtually none answer "yes." So who is right — you or the AI? If the AI is wrong, its reasoning isn't strong enough to replace programmers; if it's right, the debate is moot.

Q3: Countless tutorials teach "how to use AI well," but when asked what they've actually shipped, most go silent. A few point to GitHub repos that aren't even half-finished. Is the failure the AI's or the human's? Either way, the discussion ends there.

Five Engineering Reality Checks

The author then confronts the "harness engineering / spec-driven / TDD / MCP / Skill" crowd with concrete, everyday engineering problems:

Q1 — Complex Binary Data Structures: For a module whose inputs and outputs are intricate binary structures, can AI reliably construct those structures?

Q2 — Subjective Quality Evaluation (Audio, Image, Animation): Take a spectrogram as output. Without human perceptual validation, how can AI judge correctness? Even if you stare at the output, how much useful information can text alone convey?

Q3 — Verifiable Correctness of Complex Computations (e.g., PDE Solvers): If AI writes a missile trajectory program and its own test case says the missile lands at point A instead of point B, how do you prove whether the test case or the program is wrong? You'd have to recompute the physics yourself.

Q4 — Impossible or Contradictory Requirements: A user with no information-theory background demands lossless compression of a 4K commercial movie to 100 KB. How do you, as the prompter, judge that the requirement is impossible before wasting cycles?

Q5 — Test-Case Reliability and "Lucky Guess" Syndrome: Designing a thorough test suite (covering all branches and edge cases) is often as hard as writing the code. Why trust that AI-generated tests are flawless? Optimization comes from measurement and understanding, not from AI confidently suggesting "use a more efficient data structure."

Critique of Buzzword Inflation

The author observes a recurring pattern: every era has its jargon-chasers. Yesterday it was "prompt engineering," today "context engineering," tomorrow "MCP," "Skill," "Harness," "Agentic Workflow." What could be explained simply — "feed the model some context and wrap a control flow around it" — gets packaged as a new industrial revolution. The vocabulary density rises, but practical delivery doesn't necessarily keep pace. Some invent job titles and methodology papers before shipping a single working project.

Author's Stance: Boundaries, Not Replacement

The author is not anti-AI coding; they use it daily. The core argument: success lies in drawing a clear line between what AI does well and what humans must handle — not in chasing 100% AI-generated code. Pure AI development today is either marketing-driven or limited to trivial CRUD apps. The world hasn't changed: the tools are still tools, and experts remain experts.