My relationship with AI at the moment
AI and programmers
- My relationship with AI at the moment (you are here)
- We are not Special
Over the past, I don’t know, one and a half years (I guess?) I have been going through all the 5 stages of grief when it comes to AI, it’s actually funny to notice that, in retrospect, I’ve come from the profound depths of hating it to actually thinking: “Oh well, I think this is actually pretty great!”.
This is the place where I will tell you the story of how I discovered what I really liked in programming, how I feel now, and, hopefully, make people understand that maybe, there could be a way out of this.
AI companies vs AI as a tool#
It’s quite common nowadays to be completely crucified when saying that AI is actually a good tool. This is because, at least in my view, people confuse the opinions about “AI, the tool” with “the AI companies”, and I think there’s a big difference.
When I say that I am ‘embracing AI’, it’s actually that I’m in embracing the tool called AI. Not the companies, or the marketing scheme behind those. The main difference here is that I think AI as a subject and as a tool is something good and can really change humanity for the better. If used correctly.
I couldn’t be more against AI companies than I am now. Really. I think these companies are the pinnacle of the worst thing capitalism can do to people.
But, currently, model pre-training is extremely expensive. So, in order to have AI as a commodity in our current economic model, we will unfortunately have to go through the AI technocracy phase. My hope is only that, once it happens, people are aware enough to realize what’s happening and fight back so it’s not terrible for everyone.
I will expand that in another post, but for now, let’s continue.
The AI and me#
I was never against AI (the tool). I always loved new tech, and embraced every new thing that appeared. This very blog is filled with my experiments and tryouts. When AI became a thing with ChatGPT a few years back, I jumped onto the boat almost immediately.
But when I got there it was less than amusing. AI was bad for a long time. Things were completely useless, the code was not only very poorly written but also didn’t work at all, or was filled with a multitude of small bugs.
For a long time, AI for me was basically a glorified Google Search. One that I could stop trying to engineer the search terms and just ask the damn thing that was on my mind and get a fairly ok answer, or at least a list of links.
This was how I used AI as a tool for a long time. And, over this time, I started seeing people boasting themselves about their “amazing new app”, which was ultimately an ad-filled slop, that’s the thing I was always against, AI or not.
Same with people generating slop texts and images and saying that was the future. If that was the future, I didn’t want to be part of it. The problem was not using the AI itself, it was the sloppiness. I often thought: “Do you care this little about the things you show people you don’t even bother to check or polish?”.
I was a heavy defender of “anti-ai” stuff, tried my best to avoid using anything generated by it or even touched by the unholy hands of AI. After all, code that’s meant for humans should be done by humans! I was an extremist.
It was also easy to justify hating it, and hating everyone who liked it, because the output of AI back then was terrible. So, by extension, if you liked AI and its less-than-average output, you were someone who didn’t care about quality, therefore you should also be hated.
But somewhere between that last post and now, things changed… We went from “this is completely unusable” to “this is actually pretty ok…”. When I tried Claude Code for the first time I was quite surprised, not because of what it did, but because it didn’t completely suck! There were usable things there, and it had a plan mode that was not terrible like ChatGPT was. I then I promoted my Glorified Google Search™ to a Glorified Rubber Duck™.
I think this video summarizes a lot of what I think:
At a given moment in the video above, he shows part of the RailsWorld conference where DHH presented a keynote, and I saw myself doing something I never thought I would do in my life: agreeing with DHH… So, writing code by hand is not economically feasible anymore. And then it hit me…
It never was, it was just… The only way to do it.
Consider the chair maker#
This essay is probably the best one I’ve ever read about this topic until now, I have never thought about it through this angle, and I want to share that with you in case you don’t want to read it (even though you should, seriously).
It compares our currently state of things with the industrial revolution. First, it puts you in the shoes of a chair maker. Before the machines, you made the best chairs with your bare hands, with super fancy details, pure woodwork, amazing stuff, a real work of art. Above all, you’re an artisan. But it takes you six months to make one, and it’s expensive.
Once steam machines became a thing, chairs started to be made in industrial scale, but they had no soul, they lacked the mastery. So, two movements came about: the Arts & Crafts Movement and Art Nouveau. While the first focused purely on the human craft and abhorred any machine-made product, the other embraced the new manufacturing techniques and said “maybe we can make beautiful things with this new tool too!”. Guess which one won.
Reading it has been eye-opening to me. AI is a tool, it exists now. We cannot “undo” it. So, let’s make the most of it by implementing it into our work, and making beautiful things with it.
More than that, it’s no use fighting a tool. For example, today we are using IDEs, auto completion, compilers. I bet the computer engineers from the 60s would look at us with disgust because we were not using perforated cards or writing machine code by hand.
I would say it’s way easier (and more democratic) to write high level code than writing assembly, I’d definitely choose IDEs over it at any time if given the choice.
Soon after I started using Claude, I couldn’t stop! I was a code machine! But then I had a sudden realization: “I think I haven’t written code at all the past weeks” this is unacceptable! Is it, really?
Much like the chair maker, I was in denial that I wasn’t exercising my craft, and I was terrified, one of the few things I do really well is code, if I can’t code, what’s left for me? Who am I if I can’t code? Am I still a chair maker if I don’t bash the wood and sculpt perfect lion arms anymore?
After a while, it all started to wind down, and I realized that my joy in programming didn’t come solely from writing the code. I still loved the process, of course, but I did an exercise, and I want you to do the same. Think about the coolest project you’ve done so far. Ask yourself: why was it so cool?
I think, at least most of the time, the answer will not be “The part where I opened VSCode and typed 5000 characters nonstop for 3 hours straight”. For example, mine is my Enigma machine, the coolest part wasn’t writing the code, it was that now I had a clone of a machine that was used in the second world war working in a browser and I knew how it worked, I knew how to send encrypted messages using it! How cool is that?
So, maybe the joy of programming people keep talking about is not about the programming at all, but about the problem solving part of it, the understanding. Maybe my teachers were right when they said:
We are not paid to write code, we are paid to solve problems others can’t in ways they don’t understand.
I stopped to rethink all the things I have done in the past 6 months:
- Completely recreated this blog
- Refactored my personal website completely
- Solved a big pain I had, with OpenSynapse
- Solved another one, with this other tool
- Fixed my touch sensor
- A lot of small things that were bugging me a lot on my Linux rice
Some of these projects are so old, that the post-its I had for them in my kanban board were already losing the color. And I managed to do them all in 3 months, tops. I use them all every single day, I would never have had time (or patience) to do that if it wasn’t for AI. I would just have lived with those problems long enough until someone else did it.
In the end, it made me realize that I like to build things (more on that), I like to solve problems. How I solve them is also part of the joy of solving it, but not as much as actually solving it. I went from fully despising AI and everyone who uses it, to understanding that it can be a useful tool if you know what you want to do. It can be even better if you understand what you want to build.
And this ‘understanding’ part, I think, might become problem.
The problem we already had#
I recently came across a term: Cognitive Debt. Basically, it defines the idea that, now we are creating so many stuff, so fast, and switching context so much that our brains can’t follow along and we end up using (or depending) something we don’t fully understand – which is kind of ironic because we really do not understand LLMs and we are currently very dependent on them.
I can share a story that’s happening right now, as I write this essay. Like I said, I recreated this blog and all those things from the list before. Do I know exactly how they work? Absolutely not! But that’s ok, I wrote a GraphQL parser in 2017 and I also didn’t know exactly how it worked then (and I typed every single character in it). So it’s normal for people to forget about things they did, even when doing it, we can’t hold this much context.
The problem was that I had to preview this post on a browser to make sure abbreviations were being rendered correctly, and I realized that I have no idea how any of the pipeline worked, because I didn’t write it… I am also on a plane, with no WiFi, so I had to do like the old romans and go do that disgusting thing called reading the code 🤮.
I eventually figured it out, but it was only possible because, when I prompted this blog into existence, I knew exactly what I wanted to build. More than that, I knew how I wanted to build it.
So I prompted not only for the final result but I also gave it the guidelines for coding. The architecture, the module separation, what tools to avoid, and what to use.
The same is not true, for example, for my patch on the touch sensor, I know a total of zero C, let alone firmware or driver programming, I have no idea what that thing is doing, and it’s not really something that sparks joy in me.
Vibe coding can be a joyful thing, but vibe debugging surely isn’t. And this is the first thing about my relationship with AI at the moment:
I try my best to understand stuff or, at least, document every small detail so I can understand it later when needed
Do I think vibe coding is a problem? Depends on the thing you’re vibing.
The real problem#
Cyrus Lopez wrote this interesting piece about our generation being the last one that actually knows how something works. Even though that article is not pointed at AI specifically, we are probably the last generation that had to go through a dial-up modem, or to blow game cartridges, etc.
And this is being passed on to the age of AI. We might as well be the last generation that builds software and actually understands the code behind it…
Like others have already pointed out, and I agree, the problem with AI is the cognitive surrender. Where you completely surrender your thinking to it, you don’t care or don’t want to even know how it works. This is what I think is the problem we are having right now.
Going back to the two projects from before. Cognitive debt happens on the first case. Where I didn’t know how a thing worked, but I had enough knowledge to go after it so I could solve the issue, it’s not like I don’t care, it’s that my brain was just lagging behind.
Cognitive surrender, on the other hand, is the second case, where I don’t know what’s happening, and I don’t have enough knowledge to even go after and learn what’s happening because I didn’t care to ever wanting know how it worked in the first place.
When vibing is used for small things that won’t mess up people’s lives, or won’t store any sort of personal data, it’s usually fine. We have this mantra that every code we type must be perfection but, going back to the chairmaker essay, usually what people want (or use) is not the “perfect” solution, it’s the one that works. And AI is just highlighting that even more, but the fact is that: not every code needs to be good.
But building things that will have a big impact on the lives of others is a great case to not use AI like that. You wouldn’t like to hop on a plane to discover it’s GPS is entirely vibe-coded and get lost in the Sea.
We must have standards on, at least, the most critical things we build. This is where I draw the line. The use of AI can be nice for small projects that will have a small impact on your life, but as soon as you are dealing with others, it becomes a matter of responsibility.
The code is written by AI, but if something goes wrong it’s your ass on the fire. That’s my second stand with AI:
You need to be responsible for the things you prompt into existence.
So, where do I stand#
After all this, where do I stand? I think I stand in the place where I should’ve stood years ago. For too long I spent debating tools and processes, while the real thing I liked was there all the time, and now I can finally go for it!
But, to think as a hobby is something I won’t allow to happen to me. I want to think. I want to do hard things. I want to improve myself. If I could delegate all the things to AI, what’s the point of even doing anything?
Above all things, I am a curious soul, that will never cease to be amazed by the weird joy of solving problems.
Footnotes#
-
I actually kinda hated myself for a bit for doing that in the beginning ↩
-
In simple terms. Pre-training is when you give a model its core data to understand the world. This usually requires all the human knowledge combined, which comprises petabytes of information and thousands of servers with parallel processing to be able to be done in a reasonable time. ↩
-
Or maybe they’re expensive because everyone is doing it? ↩
-
This is sort of happening, you can see multiple people already doing it. BigTechs pushed AI so hard down the throats of customers that people started to fight back on their own. ↩
-
When research mode and NotebookLM came out, I was in awe. That alone saved SO MUCH TIME for me ↩
-
And, for some stuff, I still am. For example, I hate when people paste AI back at me without reading anything. ↩
-
AKA: switched from my OpenAI feudal lords to my Anthropic feudal lords ↩
-
Creator of Ruby on Rails ↩
-
Technically, not the ONLY because we always had nocode and lowcode, but these were so bad that I am not even counting it ↩
-
If you’re also in this loop, it’s no use to dwell in it. Remember that our profession wasn’t really a thing until 30 or 40 years go. And AI engineer didn’t exist until a few months ago, in 20 years the AI engineers will also be thinking about what’s left for them when the next thing come about. Relax. You were many things before being a programmer, and you are still one, just a different kind of programmer. ↩
-
Being completely honest, writing the code was one of the most boring parts of it. It was nice to see how it worked inside, but that’s basically it… Nothing fancy. ↩
-
Race Inspired Cosmetic Enhancement, it originally came from drag racing. Eventually it became a term for “customizing your ::insert thing here:: until it becomes another thing”. In the Linux community, it’s widely used to describe heavily personalizing your OS. ↩
-
Yes, I have one. And yes, it’s physical ↩
-
This project was so old I had everything planned, really ↩
-
Magnetic tapes are making a comeback now since SSDs are super expensive. The future generations might actually be able to listen to cassete tapes after all ↩
-
This is pretty expected with AI, we already had debt like this using frameworks before AI was even a thing. It’s manageable if you are on top of it ↩
-
This, on the other hand, was not expected. We wouldn’t expect something we never wrote to work, so this is a new thing ↩