Vibe Coding Is Not the End of Programming — It's the Beginning of a Harder Conversation
Site Owner
发布于 2026-05-19
AI collapsed the gap between intent and implementation. Now the real question: what does that make the programmer? A clear-eyed look at vibe coding, verification skills, and who actually survives.
Vibe Coding Is Not the End of Programming — It's the Beginning of a Harder Conversation
Last month, a senior engineer told me something that stuck: "I shipped more in the past three months than I did in the past year. And I understand less of what I shipped."
He wasn't proud of this. He was worried.
Welcome to vibe coding — the practice of describing what you want in plain English, watching an AI generate the implementation, iterating until the output feels right, and shipping it. No PR review of the internals. No one quite knows how the error handler works. But it runs, and it works, and the sprint is done.
It's real, it's spreading fast, and it's forcing a redefinition of what it means to be a software engineer.
The Middleman Is Gone
For forty years, programming has been a translation problem. You have an idea in your head. You write it down in a language the computer understands. The compiler or interpreter converts it to machine code. The gap between "what I mean" and "what the machine does" has always been the core friction.
AI collapsed that gap — not perfectly, but dramatically. Today, a product manager with no formal CS training can build a functioning web app. A researcher can automate their entire data pipeline. A startup founder can go from idea to deployed product in an afternoon.
This is not hyperbole. It's the new normal.
The question everyone's avoiding is: so what does that make the programmer?
The Skill That Matters Is Now Verification
#Vibe Coding#Agent#AI编程
Here's what I've observed in high-performing engineering teams in 2025: the engineers who are thriving are not the ones who write the most code. They're the ones who know whether the code is right.
Not can write — can verify.
When an AI generates a authentication module, the valuable engineer is the one who catches that it doesn't handle token refresh correctly. When an AI scaffolds a data pipeline, the valuable engineer is the one who notices the edge case where null values silently propagate. The AI can write. The human has to know if the writing is correct.
This is a fundamentally different skill than writing code. It's closer to debugging, code review, and domain expertise combined. It's about having a mental model of what the system should do, and being able to interrogate whether it actually does.
Vibe coding shifts the bottleneck from generation to evaluation. And most engineers have not been trained for that.
The Confidence Problem
There's a subtler issue nobody talks about enough: vibe coding creates engineers who can't debug.
When you write every line yourself, you develop an intuition for where things break. You know the module you rushed at 11pm has rough edges. You remember the weird gotcha you worked around in the database migration. That institutional memory is gone when an AI rewrites the module from scratch.
Now your code is a black box you've approved but don't deeply understand. When it breaks in production at 2am — and it will — you're starting from zero.
This is not an argument against vibe coding. It's an argument for deliberate partiality: know which parts of your system you want to understand in detail, and write those parts yourself. Use AI for the rest, but don't confuse "shipped" with "solved."
What the Market Is Actually Paying For
Let me be direct about the economics: if your value as a programmer is translating requirements to syntax, your market is shrinking fast. That's not a knock on you — it's just a description of the task that AI just got very good at.
The value that's appreciating: systems judgment. Knowing which tradeoffs to make. Understanding the business domain deeply enough to know what "correct" even means. Designing the architecture that lets a team of ten engineers move fast without stepping on each other. Owning the relationship between the system and the humans it affects.
These are not things AI does well, at least not yet. They're things that require context, experience, and accountability that no model has earned.
The Engineers Who Will Struggle
Here's my honest take on who's going to have a rough few years:
The mid-level engineer who learned to code by memorizing patterns. If your mental model of programming is "I have seen this kind of problem before, I apply the known solution," AI is going to eat your lunch. That's exactly what transformers do well. Your competitive moat was pattern matching. It's gone.
The "I write the code, someone else deals with the consequences" crowd. If you've built a career on implementation without ownership, vibe coding will expose that. When anyone can produce working code, the people who can be trusted with the decisions around it become more valuable, not less.
The "I don't need to understand it deeply, it works" crowd. This used to be a workable strategy. It's not anymore. Working and understanding are diverging — and the gap is where incidents live.
The Engineers Who Will Thrive
On the flip side:
The T-shaped engineer who goes deep in one domain and broad everywhere else. Deep enough to catch subtle bugs and design solid interfaces. Broad enough to communicate across teams, understand the business context, and ask the right questions before writing anything.
The systems thinker who cares about the whole. Not just "does this module work" but "does this module work in the context of everything else that's changing." AI generates local optima. Humans navigate global tradeoffs.
The person who treats AI output like a junior developer: useful, needs supervision. The best engineers I've seen using AI tools treat it like pair programming with someone brilliant and careless. They maintain good boundaries. They verify. They understand what they're delegating and what they're not.
This Is the Third Transition
Programming has been disrupted twice before.
The first transition: Assembly to high-level languages. Fortran, C, Python. Did programmers become obsolete? No — they became more productive. The total number of programmers grew.
The second transition: Manual coding to framework-heavy development. Rails, Django, React. Did programmers become obsolete? No — they shipped faster. The surface area of what a solo engineer could build expanded dramatically.
The third transition — AI-assisted code generation — is different in character, not in direction. It's not replacing programmers. It's replacing a specific cognitive task within programming: translating intent to syntax.
That task was always the easy part. The hard part — understanding what to build, why, and how it fits — is not going away.
The Real Risk Is Not AI Replacing You
The real risk is engineers replacing themselves — not through AI, but through adopting AI uncritically. Shipping fast, understanding slowly, and waking up one day to find that the system they maintain is opaque to them.
The engineers who'll be fine are the ones who treat vibe coding as a power tool, not a replacement for judgment. Who stay curious about the internals even when they don't write them. Who understand that shipping working code and building reliable systems are related but distinct disciplines.
Vibe coding is not the end of programming. It's the end of pretending that writing code was ever the hard part.
