0% Vibe-coded software

Why?

There are multiple problems about so called "vibe-coding" and generative AI (e.g. LLMs like ChatGPT or image generators like DALL-E) in general. I'll give a brief overview of the points I see as the most important ones.

Environmental impact

AI is using massive amounts of Energy and water.^[1] This is, of course, also true for normal data centres. Besides that, every piece of computer hardware does consume power. But in the past few years, the amount of resources used for AI has dramatically grown in comparison to "traditional" data centres. And the development is expected to grow further in the next years.^[2]

So why is this a problem? - The problem here is the low yield of useful information when using AI. In comparison, a single AI prompt is more likely to consume a large amount of power for the same prompt.^[1,3] Additionally, the information or product obtained from an AI prompt is usually, as long as you know what you are doing, of a far worse quality than manually researched/created products.

The rate of wrong information in AI answers is hard to determine because every source provides a different number. It is usually around 50% which means that the obtained product is pretty much useless in about 50% of the cases. Thus, just using the search engine of your choice is probably more energy efficient because you get a large amount of raw sources which can potentially be useful - just from one query.

Data sources

A widely spread concern about generative AI is the source of the data it uses for training. The problem is the following: AI uses training data to be able to do anything. This means AI companies need to get massive amounts of data which is licensed under various different licences. The main concern about this is that most, if not all, AI companies (although they may disagree about this) don't care about the licences of their training data which means that they use the data and re-publish abbreviations of it without adhering rules set by the licences.^[4]

Many Artists and Developers in the FOSS community, including me, dislike the way AI companies use our work without adhering to the licence's terms. A first logical consequence is to not use AI for the creation of new work which is exactly what I'm doing.

Code quality

An important reason which also affects users of vibe-coded software is the lack of quality and quality control in those pieces of software. As mentioned above, the answers of LLMs usually contain a large amount of wrong information. This leads to two problems with AI supported coding: On one hand, there is the AI returning code with many security vulnerabilities and bugs. And on the other hand the user (vibe-coder) who is usually focussed on quickly creating a piece of software which at least seems to be working. An additional problem that AI often fails to maintain a readable and consistent structure, especially in larger projects. But this reason is subjective and while many developers report similar issues, there are no studies about AI creating poorly structured code.

What can we do about this?

It's actually not that hard to do something about these issues: A first solution would simply be to carefully check the code returned by the AI. This way it would be possible to quickly generate code and spend the time you saved on even more security checks and debugging than normally necessary. The other option would be to not use AI. The learning process may be a bit longer and harder, but in the end, you may be able to use a programming language on your own. And the ability to read and understand documentation, which is very useful, may grow as an side effect.

Humanism

A last important aspect; and also an interesting one: The art of coding, otherwise known as "the art of googling" is slowly becoming the art of saying "still broken make working" to an AI. While this still is something, it's questionable if this really is a useful skill and whether it wouldn't be more useful to actually learn how to code.

Essentially, the question is the following: Do we really want to give up the production of software because it is easier to just tell an AI to create it for us? I personally think that the value of art and software comes from the person who creates it and not from the actual usefulness of the product. Ideally a piece of software is both useful and a personal creation, but a piece of software which was created by something impersonal (like an AI or a large, capitalist and monopolistic company) doesn't have a personality or a face either. It just has a specific use, which is fulfilled more or less well but certainly without any joy or personal dedication.

Conclusion

There are many reasons why we should or should not use AI for the creation of software and art. But looking at the reasons above, I personally think it isn't justified to drain our environment and to sacrifice the quality and personal touch of our software for a quicker development process.