This post is talking primarily about agentic AI programming, "vibe coding", but contains my thoughts about generative AI in general.
Generative AI is ubiquitous at this stage. It is something that we cannot ignore as it is touching almost every one of us in some way.
It is an odd thing. Nothing I know of in the recent past affects so many people, and has caused such polarised reactions. There are people who believe that generative AI is the next big thing, leading to AGI, and there are those that believe exactly the same thing but are vehemently opposed. There are also those who object to generative AI on moral, ethical, or environmental grounds.
As with many things it seems, I generally sit squarely in the middle. The things I have passionate feelings about are non-controversial, and the controversial things in life I have no real opinion about. I have stated that I am neither pro-AI, nor anti-AI. I am interested, and almost excited about the technology, but I'm also particularly uneasy the current implementation. This mirrors my thoughts on previous tech-fads, like blockchain. It's a cool idea, but how it's made, packaged, and used, leaves a lot to be desired.
I also won't touch too much on the aforementioned moral, ethical, or environmental complaints against AI. This is not to say that these are not real, they very much are, but these are things that need to be fixed with legislation, licensing, and regulation, and while they are legal and available, people will continue to use them as they are. These systemic issues are far beyond the scope of what I am comfortable talking about. The remainder of this post will be discussing other aspects of generative AI, but do not take this as dismissal of these complaints, just a resigned "if it is going to exist…"
This is where I make a confession. I installed OpenAI Codex. Now, before anyone grabs their pitchforks and starts trying to cancel me, let me explain myself. I am not intending to become a vibe coder. I had no experience with agentic AI, I just knew about the aggressive arguments from each side. I had opinions, but without any experience, I didn't really feel comfortable making any judgement on what they are like to use.
And to be clear, I've not unleashed an AI on any of my existing projects, I'm not going to subject anyone to AI slop. I have specifically picked tasks and ideas that I'd like to do, but will never ever get around to. I will also almost certainly delete any code that was generated, albeit after having read through it to see whether I could learn anything from it. I have given it three projects of different complexities, and different amounts of guidance, to see what it would do. The three projects were a simple Cockpit page for monitoring the health of computers on my network, a data translator/exporter, and a self-hosted compiler.
It can do the job. I was never really doubting that. I still wouldn't trust it to write anything mission critical. I think a lot of the fear is a bit unfounded and based on outdated failure modes which are much less prevalent than they used to be, similar to AI-art generating images with incorrect fingers. There are issues, but most of what it produces is at least somewhat competent, certainly appears at first glance to be on par with a lot of human written code I've come across. I wouldn't trust it to be totally safe for anything mission critical without spending a lot of time auditing what it produces. This is, to a degree, what I feel about any software though.
Having looked through some of the code it generated, I have also found several issues, both in its understanding, and just in the implementation. These could be fixed with more prompting, but that just goes to show that it's far from infallible, despite how confidently it spews data at you.
My concern, which I feel is now more founded, is that generative AI encourages disengagement and reduces scrutiny. I think it is absolutely a tool which could be used by a competent programmer to generate solid work. The way that these tools are set up though, the way that the agents interact with you, they are not designed to encourage you to do that. They want you to move fast and break things. They write tests, and explain things in nauseating detail, to the point where all but the most dedicated programmer will glaze over and just go "yeah sure, continue please". This is not entirely a fault of the programmer, it is exactly the kind of behaviour the interaction model makes easy.
And it is so easy.
I have caught myself looking at other little awkward things and thinking "maybe I could just get Codex to sort that" - these are all things which I both could do, but do not have the time available in the day to do. So at first glance, I technically don't lose anything by getting it to do that. And at the end of the day, I don't audit every piece of software that I install on my machine.
This is a terrifying train of thought. As I stated before, I am certainly not pro-AI. I am not an 'AI Bro'. I am not comfortable with how easily this has changed me from not using it to finding uses for it.
And one more last thought before I finish up. The projects I did use this on are now burnt to me. Similar to when you start cheating in a video game, I cannot now go back and implement them myself. The solution is not really a free thing for me to explore, and I'll always have the AI-generated one in the back of my mind, even if I've long since deleted it, just like a cheated game, you'll always know what comes next, and how easy the cheat was. I was expecting this to happen, which is why I specifically chose projects that I knew I'd never find time to actually work on.
So, have my views on generative AI changed at all? Not really. I still believe that it is here to stay. I still believe we need better legislation to deal with the… everything that comes with it. I now have the experience to say that it is genuinely useful if used carefully. But I am strongly of the view that "carefully" means that it needs someone with patience, meticulous attention to detail, and a high level of skill controlling it. The hype around it is pushing its use in entirely the opposite direction.
I will now remove Codex from my computer, and mark the projects it created for me to review and investigate, before they are also sent to the great bitbucket in the sky.
And, I will re-affirm that while my projects, such as my current Atari Jaguar game, and the contents of this website, might be badly written slop, it is, and will remain, 100% genuine, home-grown, human-made slop.