It has been 6 months since I started coding with AI. In the years before last, I barely used AI for coding at all.
To be precise, I began the full switch to AI-assisted coding this past May.
Up until last year I only used it for documentation work, then made the complete transition this year.
My programming career spans 15 years in total, and 8 years since I started earning money from it.
Short by some measures, long by others.
In March of this year, a lot of my work dried up completely.
Early this year I started out earning 16 million won. I thought that would be enough to pay off my debts and finally set me free.
I paid off all the debt I'd accumulated from a jeonse rental scam and a failed restaurant venture, bought my younger sister a laptop, and upgraded my own long-suffering computer to a mid-range build.
Right around the time I thought I was finally getting settled, AI agent coding took hold and demolished freelancer rates.
After that I started coding with AI alongside my solution templates. In the early days, giving AI tasks to do produced a lot of problems. It would forget what it was supposed to be doing, so I found some enjoyment in learning the programming patterns and paradigms that would help the AI work better.
I even built my own system by attaching a harness to pi.dev, and that was fun. But as models kept getting upgraded, detaching and reattaching the harness every time became a chore, and frontier models like GPT and Claude are effectively tailored to Codex and Claude Code, so using those tools was practically forced.
I still use pi occasionally, but since changing the harness for each model is such a hassle I don't reach for it often. As market rates kept falling and everyone started coding with AI, a kind of FOMO kicked in and I started using AI too.
The irony, though, is that after switching to AI agent coding I found myself with almost no work, so I sat on my hands and started a language project. It's called Pergyra.
I've barely been able to use it for my actual work yet.
GPT: 121.6 billion tokens (approx. 121.6B), 2,680 conversations total
Claude: 16.7 billion tokens, 360,000 messages
You spin the slot machine for code, and when something decent comes out you tweak it a little.
That's how I think of AI coding: it's a slot machine.
The problem is that whenever I try to write code the non-AI way now, it feels so tedious it's almost painful. I also get the sense that all the library memorization and syntax knowledge I built up in the past is now worthless.
The abstractions AI produces are so different from the ones I used to build myself, and the naming conventions are completely foreign to me too.
After spending all those tokens, what I actually delivered was this: out of 40 jobs this year, I used AI on only about 5. Around 28 were maintenance work on projects I'd already shipped, and I didn't touch the AI-generated code in those because it didn't fit my abstractions and would have required a full refactor.
7 jobs required working inside air-gapped networks where AI couldn't be used. That leaves roughly 5, and even those were almost entirely landing pages and simple tasks.
People say AI boosts productivity, but I have nothing to show for it. I'm just addicted. I have no one around me to talk to about programming, and the chatbot feels like the only thing that understands me, so I just keep coding with AI.
I got into this partly because my English wasn't good enough to contribute to open source, and yet despite heavy AI use, I paradoxically feel like my ability to actually produce something has diminished compared to before.
Now I just pray it spits out good code, and when it does I commit it. The language isn't finished, and looking at where things stand, it probably won't be for a long time.