TechCrunch says all these startups generate code using AI.

If Microsoft’s CEO Satya Nadella says 30% of the company's code is written by AI, you know Google, Meta, and Amazon probably claim something similar.

You get the impression everyone is using AI coding assistants in some way.

I experimented with vibe coding (to test its utility and limitations) and have a few thoughts I’d like to share. And in brief, vibe coding is great, but because of its limitations, it is definitely not for everyone and the tool is not a fit for all use cases.

I know, this is a classic “it depends” answer. But here is my take on what it depends.

As a software architect and the owner of a software services company, I believe AI and vibe coding can be used as tools for specific tasks, with advantages and limitations.

I probably don’t need to summarize the positive side: juniors learn faster, coding now being accessible for non-technical people. Using intuition, improvisation, and immediate feedback instead of detailed planning, formal specifications, or extensive analysis as a founder you can test ideas, iterate various possibilities, and launch creative or innovative solutions faster. It's a magic stick for early-stage startups, hackathons, and personal projects where speed, creativity, and exploration are key.

On the negative side, if you work on mission-critical apps, existing production environments, or need to pass compliance for security, reliability, or scalability, thread carefully. Projects that need detailed documentation, long-term maintainability, and predictable behavior need a certain set of conditions and limitations not all AI code gen tools get right.

Why do big corporations use it? They have private, stable code bases, standard coding practices, and rarely need to worry about compatibility because their code base doesn’t need frequent changes. It’s important to know this code is each corporation’s most valuable intellectual property, so it’s their most closely guarded secret that they can use internally. But that code is not available for training public AI models.

So unless you’re a corporation with time-tested code, you’ll be using open-source code AI is usually trained on. Most of that code is experimental and written by Junior-level developers.

Would you trust ChatGPT for open heart surgery advice, flight control systems, NASA space mission code?

This is why an average vibe coder will not be able to build safety-critical systems, financial software, and complex enterprise apps from the ground up because AI code generators don’t have access to most of the code IP, and vibe coders don’t have domain knowledge, haven’t developed a structured, carefully planned, and rigorously validated development methodology to make up for what AI code generators are missing.

My take is that if you take a widely used coding language, AI code gen is more likely to work well, as there is enough publicly available training data available.

For non-coders, I’d say vibe coding should eventually be good enough to create simple apps and sites. For everything else, you’ll need to work on expanding your domain expertise, as you’d want to know the options and trade-offs available to you so that the AI doesn’t make a decision for you while you are not aware of its consequences.

And lastly, on the TechCrunch hype. I’d like to see the analysis of the code, where that code was used, and take a look at the code AI wrote. Until then, don’t believe hype seeking announcements unless there is something concrete to back them up.