Vibe coding doesn't replace understanding

Recently someone attempted to recreate the Adobe Suite plus some other programs vibe coding them and I wanted to share my opinion here.

I saw on YouTube that several artists tried it and for the tools that they used daily they could spot right away what was missing and what was worse in terms of quality. For the tools that they didn't use they couldn't distinguish them from the original.

ArtCraft is impressive not because it replaces Adobe or any other tool but because someone could build an approximation of these tools in such a short time. This shows how much the cost of creating software has been reduced with LLMs. But there is a difference between reproducing an application visible features and reproducing the depth and thinking that makes it useful for professionals.

This is what represents best what I see with vibe coded tools: they are as good as your understanding of the problem. Not everyone produces good software with LLMs, those that have seen a lot of software and understand the problems well enough can produce great solutions and they do it faster and better.

However, someone approaching a problem without deep knowledge can generate slop, software that correctly does some percentage of what he wants to achieve and has issues in situations that were not properly considered. Here when I mention slop I don't mean software missing features or having bugs (we humans write a lot of them). What I consider slop is software that appears to solve a problem but breaks down as soon as someone tries to use it seriously.

This doesn't mean that you need to be an expert before starting to build. LLMs make really easy to experiment allowing developers to learn with rapid iterations. Still someone needs to recognize what is wrong, understand the feedback and decide what to improve.

As for the specific case of the ArtCraft suite it got so much attention that probably it set the foundation for others to come and actually develop something interesting and that many will find useful for sure.

Not everything is slop in the land of LLMs and we are seeing more people trying to develop native UIs instead of using electron for everything, improving performance of their apps way faster than before (we did it in rotki finding issues, refactoring legacy code and adding better concurrency!) and approaching solutions that before were not tried just due to the cost being prohibitive.

LLMs dramatically reduce the cost of implementation but they don't remove knowing what to build, whether it works, or how to maintain it over time