Many people believe that AI programming is a liberation for developers, but based on my experience with vibe coding over the past year, this claim is debatable.
I recently built an app for a friend and planned to put together a user manual for him in the next couple of days. While creating the manual, I noticed a tiny bug, so I took the opportunity to do a round of optimization. Before I knew it, a whole day and night had passed.
The most exhausting part during this process wasn't writing prompts or brainstorming ideas—it was testing. It was comprehensive testing from morning to night, top to bottom, left to right. Even for a minor interaction change, I had to retest every single path because I knew there would be unexpected issues. And indeed, that was the case.
I've developed this "awareness" purely from stepping into too many pitfalls. Forget whack-a-mole; I've even encountered situations where modifying some UI ultimately led to an entire business process being deleted. So every time during the final delivery phase of software, I have to test over and over again.
This "verification anxiety" stems from the fact that vibe coding cannot guarantee that every output will effectively converge, because the nature of LLMs means they will still do things outside the constraints.
Traditional applications are deterministic systems, where the only uncontrollable factor is user behavior. As long as the business process is correct, the result will inevitably meet expectations. But LLMs are different—they are non-deterministic systems based on natural language. And natural language inherently has ambiguity; the same word can mean different things to me and the model at different times and places.
For over a year, many self-media outlets have been hyping that AI will replace programming, that coding is cheap. Indeed, the cost of writing code has significantly decreased, but correspondingly, the cost of verification has become a heavy burden. The inversion of development costs and testing costs will redefine business processes.
To ensure that these probabilistically generated codes don't crash in some environment, I have to pay a hefty "trust cost." Even if I've already given high-level commands to my various AI "generals" to implement enough safeguards (I don't know if the AI actually did it, but I've issued the orders), I know there will be plenty of pitfalls waiting for me. To prevent the app from crashing in front of users, I have to let it crash in front of me first.
The software industry is still a relatively easy field to converge and verify, since code is a relatively deterministic form of expression. But this shift in trust costs will inevitably spread to all industries as AI becomes more widespread.
Once this "cost shift" leaves the software world and enters the real world, the problems become extremely serious. For example, in fields like healthcare, law, or embodied intelligence, you can't manage things with Git like you do in programming, there's no undo, and "hot fixes" are impossible. AI can produce a seemingly professional diagnosis or legal document in a second, but the effort people have to spend to uncover hidden, potentially fatal logical deviations might be greater than if they had written it themselves.
The technical debt generated by AI will drown every field that aims to "improve quality and efficiency with AI." When generating functionality becomes nearly free, ensuring that the AI's output is "trustworthy" becomes the most expensive luxury of this era.
The future will be a long-tail era of infinitely extended trust costs. People's past experiences and intuitions from stepping into pitfalls will become the hardest assets.
AI will lower the barriers to entry across all industries to an extreme, but it will also make the ceiling impossibly high. Business processes will shift from "solving problems" to a new paradigm of "proving that problems have been solved." The judgment to identify all kinds of pitfalls in the vast sea of AI-generated content is the truly invaluable wealth a person can possess.