To AI and back - part 1

I wrote my first AI program in high school (around 2009-2010). I found a tutorial for writing a genetic algorithm to find a list of number that sum to a value (I think). It was written in C++, I knew only PHP, I didn't know any OOP, so the resulting code was atrocious, but after a full Saturday coding, I got it working.

Then in 2012, Coursera launched and the second course there was Andrew Ng's famous Machine Learning class. I got hooked and I rode the MOOC wave, doing all the courses I could from MIT, Berkeley, Stanford and so on.

I spent a whole summer building a receipt parser, from annotating 10000 characters, to training my own OCR engine. I built my first recommendation engine at my job in 2014. And since then, I've been doing AI one way or another, going through all kinds of applications, from computer vision, tabular data, NLP, and so on. I think audio was the only thing I never touched.

I've been doing AI & ML one way or the other for almost 16 years. And a couple of months ago, I started wondering if maybe I should quit AI. I started pondering if this GenAI thing is a bad thing for the world and I'm doing my small part in building it.

One aspect of it was simply mental health. Coding assistants were like gambling for me. One year ago, when Claude Code came out, it was not that good, so it was a "coin flip" if it would get it right. I remember the first time I burned 15$ in a couple of hours, trying to do things with it (this was before subscriptions were a thing). It worked exactly as a variable reward system (like slot machines).

The assistants got better, writing more code with fewer mistakes. And eventually you could control them from your phone. I found it hard to unplug. In the evening, trying to go to sleep? "Just one more prompt bro". Going to the playground with my kids? Check Claude on my phone to see how it's fixing the bug.

That was plain addictive behavior. And I'm a "moderate" case, there are many sad reports where people get AI psychosis, some even committing suicide unfortunately.

Not only that, but using code assistants enabled me to take on more projects at work. I was involved in two startups, plus contract work and I could even work on a fun side project. It's easy to open multiple terminals in parallel and tell Claude to work on different things in each. I could suddenly "write" a ton more code in a day. But, while AI helps with the code (and not just the code), I still have to load in my brain all the projects, I have to be aware the scope, the direction, the larger context and to talk to the other stakeholders. And it's just exhausting. I'm not the only one who says this. Combined with the addictive side of things, that was interfering with my personal life and relaxation, resulting in health issues last year.

But that's just the health side of things. There's another side which I'll touch on in the next post (and I think this time I will get around to writing it much sooner, because I think it's part of my healing journey).