Cool thoughts on the AI coding boom: from replacement hype to tech-debt blowups

Skill pyramids and hype, tech debt vs efficiency illusion, layered collaboration, and practical advice for builders and managers.

2026-08-17 · Yuan Liangding

One shift stands out lately: when AI coding tools first exploded, the web shouted that programmers would be replaced. By late 2025, more practical complaints appeared — AI code piles up tech debt; it looks fast, yet the total ledger often shows little net gain.

Entry-level replacement anxiety on one side, real team failures on the other — that contrast already sits inside the industry’s skill structure. This note is a personal reflection, and also a set of usable ideas for builders and managers.

1. Why did “AI will replace programmers” sound so loud?

Skill distribution is a pyramid in every craft: the largest group sits at the base; fewer people remain toward the top. Software is no exception.

Skill pyramid and public voice

Figure · Skill pyramid and public voice: the base feels replacement anxiety first; the top stays cooler because complex work is harder for AI alone.

When AI coding arrived, the strongest shock hit the base. A junior may spend half a day wiring a CRUD endpoint; AI can emit runnable code in a minute. That shock feels personal — “my half-day became one second” — and fuels replacement anxiety.

When the largest cohort speaks at once, it feels like “everyone says programmers will vanish.” That is a statistics effect: loud is not the same as universal replacement. Higher on the pyramid, people stay calmer — the more core and complex the work, the less AI can finish it alone.

2. When the tide goes out: debt and the efficiency illusion

By late 2025, many teams that put AI on full projects hit two clusters: runaway tech debt and efficiency below expectations.

2.1 AI code arrives with built-in debt

Current generators optimize for “probable code that runs now,” not long-term maintainability. Mature developers design for reuse, extension, naming, exceptions, and edges; AI often chooses whatever is fastest — duplication, magic numbers, missing comments, skipped branches, inconsistent style. It ships today; months later, maintenance cost multiplies.

A common story: a team used AI to stack three admin modules and “saved” ten days before launch. During later features they found chaotic names, duplicated logic, and empty exception paths — fifteen days just to clean up, slower than writing it carefully in the first place. Early thrill, later funeral pyre.

2.2 The efficiency illusion: we undercount correction cost

People imagine “drop a requirement, get a finished product.” Reality is prompts, alignment, line reviews, and chasing subtle bugs. Fixing bad AI code is often slower than writing it yourself.

AI development efficiency illusion

Net efficiency Time / project depth → AI-assisted (illusory early win) Traditional baseline
Figure · Efficiency illusion: generation looks fast early; debug, review, and debt repayment pull net efficiency down.

Full autonomy loses control; full line-by-line babysitting removes the point of AI. Teams swing between extremes and discover: typing got faster, while debug, review, and fill-in time grew — net efficiency barely moved.

3. Core issue: collaboration mode, not “AI is weak”

Today’s pain is largely an engineering collaboration problem — we have not settled a sane human–AI boundary.

Expecting AI to own the whole project is unrealistic; refusing AI wastes repeatable labor. Follow the pyramid and layer the work: the more macro and core, the more humans must hold; the more micro and standardized, the more AI can take.

Layered human–AI collaboration model

Figure · Layered collaboration: humans own the macro; AI takes standardized micro work.
  • Top (macro decisions): human-led, AI-assisted — stack, architecture, core rules, quality bar. AI may propose options; decision and accountability stay human.
  • Middle (module design): humans set rules, AI implements — splits, APIs, core flows first; then AI fills code under those constraints. Humans draw the blueprint; AI lays bricks.
  • Bottom (implementation): AI primary, human spot-check — boilerplate, utils, tests, formatting, comments. Spot-check high-risk paths instead of reading every line.

4. Practical advice

For builders: four habits

  1. Skeleton first, then flesh — directories, interfaces, core class skeletons and naming rules before asking AI to fill functions.
  2. Give AI clear rules — naming, validation, exceptions, no magic numbers, comments on critical paths.
  3. Layered review — line-audit core business, data, and security; spot-check utilities if tests pass.
  4. Small weekly debt repayments — light refactors before debt becomes a mountain.

For managers: three reminders

  1. Do not blindly cut schedule — AI saves typing, not design, review, debug, or cleanup.
  2. Use AI by skill tier — juniors: learning aid, must understand every line; seniors: efficiency aid for design and hard problems.
  3. Upgrade what you measure — stop counting lines and feature chips; value architecture control, hard-problem solving, and maintainable systems.

Closing

AI is not a replacement; it is an amplifier. It amplifies junior throughput and bad design debt alike. It can lift disciplined teams and help sloppy teams pile wreckage faster.

Moving from “wow, AI is amazing” to “how do we use it well” is maturity. Worry less about being replaced; decide earlier what belongs to AI and what must stay in human hands.