Vibe Coding: first reaction
The emergence of the term Vibe Coding (coined by Andrej Karpathy) at the beginning of this — 2025 — year and the endless stream of demonstrations showing the rapid creation of web applications using AI services like Lovable, Bolt, Replit, etc., were likely met with mixed feelings by many.
On one hand, there was undeniable progress: nothing like this had been seen before. On the other hand, the uniformity of these examples was very unsettling.
Each time, a very quick creation — in a matter of ten or twenty minutes — of simple, yet fully functional, applications was demonstrated based on straightforward, one or two-paragraph prompt requirements, often even recorded by voice. Clearly, these tools were perfect for small websites, web application prototypes, and simple MVPs.
What’s next?
However, no one demonstrated what might happen to these applications afterward: how suitable they were for further growth and to what extent these, or any other AI tools, could support such growth.
Will we be able to maintain the current pace of development, or at least not significantly slowed down, during the product’s further evolution, or will we have to return to the “do it yourself” method?
Not to mention the question of whether this miraculous AI toolkit could somehow be applied to existing, so-called brownfield projects, which virtually all existing codebases consist of.
Ultimately, people wanted to see something more serious than mere “wow effects,” which, obviously, won’t sustain anyone for long. The question arose about the possibility of using LLMs to manage long-term software development processes.