Ranting about AI development
I wanna rant a little bit about my work life and how it affected me. I don't claim it's universal, and I don't claim it's unfixable, but it's what's on my mind right now. And I want to bring it out of that cage. It would be very interesting to read it a couple of years from now. I owe this to my future self.
First, I have to explain what my current situation is, how I work, and why I work this way.
I've barely coded by hand over the last few months. First, it was the shutdown of the internet in my country. I tried to code a little bit during that time, but you can't really do much without the internet. Then it was AI. When we started our big project, the new AI models had been released and, finally, they were actually something. I knew I had to start using AI as much as I could. I knew that it wasn't enough to use it as a mere tool; I had to use it as an employee. I knew it would be impossible to finish the project within the given timeline if I cared about maintaining my agency over the code itself, even if I had the experience to do it quickly. AI had suddenly become extremely good.
And I knew it was coming. I didn't expect it to happen this soon, but I knew it was inevitable. At some point, AI would've looked good enough. It would've fooled you into feeling like its output was usable, and that's all the market has ever cared about—under communism, socialism, capitalism, anarchism, or anything else. It's not a political standpoint; it's the collective human nature to pursue what can make the fastest money. Because you need to get fed 3 times a day. Like a hummingbird, you need to always look for food; rest, and you'll die. Unless you've secured a constant stream of food, it's very hard to care about perfection. There's a reason why most philosophers were rich people. Quality is either a luxury or a spiritual devotion.
But that's not all. Just like any other job, the nuances of a job are what define it. From the outside, most jobs look simple, doable. It's easy to feel like you can steer a ship with zero expertise, but it's the age of specialization and the devil's in the details. To know why you're wrong, you need to be knowledgeable enough, to see enough details to realize their scale. A single detail never matters. The accumulation of them is what changes your perspective. And most people don't have the will or patience to sit through your lecture about the nature of your work. And those who need your expertise... well, they need its output, not the work itself. That's how you'd be hired or fired: based on output.
Even though I had so many reasons why it was a bad choice to go all in on AI, I accepted the new nature of my work (I couldn't just go find another job, you know) and started perfecting its usage. At the beginning of the project, I tried to have so much agency. I designed an architecture for the frontend, read the code, and continued gradually. At some point, I realized two things at the same time:
- No one can read all this code on such a tight deadline. It turned out to be at least 160K lines of code. No one can read that. And I'm pretty sure reading it is not enough. As I explained in my book, code review has three levels. And for code that can contain very silly mistakes, you need the third level of code review: the harshest one. You have to be really focused on reading it and put so much time and effort into it that using AI would become nearly meaningless. All the time saved by using AI would turn into time spent reading the code. Later, I realized that DARA's 2024 report also confirms this and shows that code review doesn't affect code quality and that the number of bugs still stays roughly the same. It's because we, as programmers, have taken code review for granted, because we rarely needed to be this paranoid about a piece of code. We knew someone had intentionally made and reviewed it, and we just had to find some niche things, such as coding conventions or things like that, which AI is already pretty good at.
- I was genuinely not finding bugs anymore. One reason is that I didn't already have much knowledge about writing websites. I had to be much, much more knowledgeable to be able to find bugs in working code that follows many of the best practices. I either had to go very slowly and use the same AI to understand each change fully, which—as I explained—reduces your productivity so much that it makes using AI meaningless; or I had to let go of this and find the bugs by just testing things out, like a game tester; or go make websites by hand for months and use that knowledge to vibe-code, which is not a very financially wise choice.
This ended with me letting go of code review as well. And it makes perfect sense to me. Code implementation, code review, and code debugging/refactoring all go hand in hand. Even while you're coding, you're constantly reviewing it and fixing its bugs on the fly (e.g., fixing typos). It's absurd to think coding is complete without also doing code debugging, refactoring, or reviewing. I have talked about this enough in my book and twitter, I'm not gonna go into details here.
It's also worth mentioning that right now, I'm mostly using GPT Luna 5.6 (xhigh). I've chosen 5.6 because, for this model, it's the one that hallucinates the least, as far as I've experienced (As I'm writing this, SWE-sweep benchmark proved that 5.6 Luna xhigh is the 5th best model, after Fable 5.1, so I guess I was right). And I've chosen Luna because it's the only model I can use on a $20 plan without running out of tokens. Sometimes, to tackle big problems or problems the model has gotten stuck on, I switch to Sol 6.1, but I do it very rarely so I don't run out of tokens.
The Situation (It's crazy)
This is where the shit begins. It could be something unique to me and a few people, but I don't care whether it's unique or not. All I care about is how to mitigate it, and I've yet to find a solution.
For me, the problem with using AI is that it completely annihilates my daily routine. Back then, I could just grind for 10 hours, even more, and still have so much time to waste. At the end of the day, if I didn't waste the day on bullshit, I never felt like I hadn't done anything. But that's rarely the case now. I'm working 30 hours a day and I'm still struggling either to finish the job or to feel like I've done something.
I believe what I do is very trivial. With just one month of practice, you'd be able to do 95% of my job. When a bug happens, what do I do? I say, "Hey, this bug happened [possibly a screenshot too]," and wait for it to magically fix it. Then I check again; if it's not fixed, I ask again. If it fails to do so after multiple iterations, I switch to Sol and it'll fix it right away (hopefully). All I know are a few technical words (for example, I know that thing is called HTML. Wow, surprising.), and the rest is just putting my time into it until it works. Rarely do I get to use my actual expertise, and it's mostly for very basic things, such as choosing which programming language to use (which itself is mostly just a matter of personal preference, especially in the age of AI). This makes you feel like you haven't done anything at all by the end of the day, even if you worked on the project for the entire time you were awake.
This feeling was the least problematic one, though, as I could just use the very specialized weapon we have as a species: fooling myself. To try to give value to my work just based on output. The funny thing is that right now I'm producing many times more code and features than before. So yeah, I had to use this concrete measure and evaluate my work at the end of the day based on that.
But the real issue is what using AI for software development actually looks like. You see, despite what AI delusionists want you to believe, there still needs to be someone assigned to create software. Even if the job is trivial, it's still time-consuming; we're not just "one prompt" away from building an actual working website with many users. The models are fine-tuned to one-shot everything, but real work is never like that. Iteration is what a real job is. You create something and you iterate over it many times. 90% of your work is iterating over what you made, changing decisions and other things until everyone in the party is happy with it.
So this is how my day goes:
- I spend a couple of minutes writing a prompt.
- I give it to AI and wait 15 minutes for it to finish making it.
- I go and check the result and give feedback if it apparently failed.
- I wait for another couple of minutes.
- I repeat until it's done.
- Next task.
Sometimes, depending on how big the prompt is, the agent might work for hours. But I still have to be ready in case it finds blockers and stops working unless I provide an answer.
So basically, it's like babysitting 24 hours a day. I always have to be ready to answer it. Many things don't work on the first prompt, either because I didn't think through my choice of words, because I wasn't 100% sure what I wanted, because there were some hidden technical caveats, or because of some bizarre misalignment with the previous technical choices, and so on. I can't focus on anything. When the agent is working on the prompt, I'm 100% free, but I can't put that time toward something else because it's super distracting. It's not like I can just go paint while it's working. My other jobs are also related to programming, and they all need brain power—they need me to be focused. And you can't be focused when your baby cries every 15 minutes. I'm not exaggerating when I say it's like babysitting. Mothers have to be paid for this exact reason, because it keeps them away from being productive. It's not an exaggeration; it's even an underestimation, because it's babysitting with all the positive aspects of it ripped off (it's not particularly easy to feel cute taking care of ChatGPT).
At the end of the day, I have typed for a total of 1 hour. I have produced 20 days of work in one day, but I actually worked for 1 hour and wasted 14 hours on social media because that's one of the only forms of short-form content you can use in this routine. Not only does this make it extremely hard to feel good about yourself and control your dopamine level, but it also means you can't do many of the things you would normally do in your life. I can't go waste 3 hours of my day working on games or doing something personal, because I'm already working for only 1 hour a day, and it's possible that half of that is aligned with those 3 hours, so by wasting that time, I reduce my output by an order of magnitude. So now, not only did I waste 14 hours of my day on social media and work for 15 minutes, but I also produced very little. Every hour feels important while you're wasting 90 percent of it on Twitter (BTW, that's why you're seeing so much tech- and AI-related content on Twitter now, because people get bored waiting for the agent to finish work).
I'm just tired. It's been months since I was in a flow state. I used to work much more than now. It was objectively harder to be a programmer back then. Right now, it's easier per se, but it's much more frustrating to do so. I can't work out because my baby cries in the middle of work, I can't read books because my baby demands attention every now and then, and I can't do a side project because my baby ruins my mood every couple of minutes by making me do some mundane work anyone can do. I can't go out with my friends because my job has been inflated. I can't just "do more work now so my weekend is empty." There's no such thing as "working more." You don't work; you just wait for the next time the almighty agent stops working and you inject the next prompt. The details are just so much.
It's also worth noting that I don't have infinite resources on my computer. When the agent is working and doing heavy tests, I can't just open any application and start doing my own shit. I have only 8 GB of RAM right now, and due to AI and tech companies, I can't afford to buy another one. Isn't that just ironic?
And yet I'd vote any day for AI
I'm already exhausted after typing all this, so I'm not gonna go into the full details on why I believe we shouldn't ban AI. I just think of it as being extremely boomer and anti-technology (and suicidal if you're from USA).
I can't even imagine my own personal life without AI. This reminds me of that one Twitter interaction. An anti-AI person was making fun of herself by showing how unaware she was of what AI and machine learning are, and at some point, someone replied, "You see, this is the problem with being anti-AI. You can't even go ask ChatGPT what AI is; you're stuck in a loop." This is technology, and this is how stupid you'd sound to future people looking back at you.
It's funny to me that people claim to have calculated the "net value" of a technology so widespread that you can't even understand its capabilities. They focus on that one shitty artist who was replaced by AI, but they never mention that mute person who could finally talk because of Neuralink AI being trained on his brain.
But yeah, I do agree this technology comes with so many caveats, depending on your field of expertise and the stage of AI. For me, it's the most toxic relationship I have. But I'm pretty sure its net value is positive, because I'm making money that I couldn't make otherwise, in a country with hyperinflation. I don't know. I'll try my best to fix this life, even though I've been trying to do so for months now and failed. I wanna start a hobby game of my own, just to feel something, but I can't even fit it into my weekly routine. I'm just in a constant stream of struggle.