I use AI now for coding at work. I hate AI with a burning passion, but I use it for one thing: to replace the long hours of internet searching to find out how this-or-that API work in this or that language to do this-or-that task.
In other words, let's say I want to use OpenCV in Python to recognize a square shape in an image. Before AI, I would look up the OpenCV generic documentation, the Python-flavored version of it, figure out what I needed to instantiate / initialize / convert... to do what I needed to do. I knew what I wanted to do, I just didn't know how practically.
Now I just ask the AI server sitting next to me (yes, I refuse to do cloud-based AI, so my company bought me an expensive standalone AI-ready machine): "In Python, in OpenCV2, write a piece of code that takes a image as a numpy array of uint8's and detects contours with 4 segments of equal lengh to within 10%". The thing will piss 10 / 15 lines of code in seconds with the right OpenCV calls that would take me a hour to figure out myself.
In other words, I use AI as a supercharged search engine. Because I figured what a coding model really is is a condensate of technical stuff from the internet, both mainstream and obscure, with an ability to understand what I'm talking about.
It's the opposite of "vibe coding", in which somebody with little to zero coding skills asks a powerful AI to do their job (as in "Write an application that does xyz"). That's bound to generate terrible buggy code, and it's way beyond the capabilities of the models I can run locally on my little AI-ready machine. What I do is code exactly like I used to, but use the AI model to do the boilerplate and figure out the boring-ass stuff piecemeal, in tiny enough chunks that I can verify the code it wrote from start to finish and make sure it didn't hallucinate stuff it didn't actually know - which it does on a regular basis.
And it turns out, I don't actually use AI that much when I do that, because it doesn't take much for a prompt to take longer to write than writing the code yourself when you're prompting simple, short things. The OpenCV example above is worth writing a prompt for because it's quicker than perusing the extensive OpenCV documentation. I might even ask it to take some really long declaration and rework the format in some way that might take too long in vi. But for the rest? I do it myself really, because it's plain quicker.
And so for me, the art of AI coding is figuring out when it's quicker to do it myself and when it's quicker to let the model do it and me reviewing and adapting what it did. And at least the way I do it, 95% of the time, it's quicker when I do it myself 🙂
all 9 comments