Ethan

I tried for a good 20-30 minutes to get it to generate something that worked. It just gave me a sequence of garbage. Each one looked plausible but as soon as I really read the code it was clear there were issues. Different issues each time. I wasted more time trying to get it to produce correct code than it took to write it myself, once I gave up on the AI.

For me the most productive way to use AI is to ask it to suggest a strategy and to implement that myself. If the code is simple enough for the AI to get it right, I can often write it myself with less effort. If it’s something that requires multiple reprompts to get right, I can almost always do it faster myself.

Or something more complicated than it can handle. Ask it to implement an ASN.1 floating point parser. It looks good but the closer you look at the code the more trash it is.

A good teacher will do that. And even the best teacher can’t force you to learn if you don’t want to. LLMs can be a powerful tool for learning if you’re sufficiently motivated.

AI is an extremely useful tool. I’m on the other end of the career track, but it seems to be that it’s almost like having a personal tutor. And as with any other teacher, if you use it as an aid to figure things out yourself I imagine it would help immensely, but at the same time if you use it as a crutch to do your work for you, you skills will be as weak as someone who cheats off their friends in school. I attribute a large part of my skills to spending lots of time reading other people’s code and understanding why they wrote it the way they did (usually because some library didn’t do what I wanted so I figured out how to beat it into submission out of pure stubbornness). If you use the AI as an aid and spend the time to really understand the code it’s producing (and the flaws in that code), I think you’ll build up your skills well.

My rant about code monkeys was inspired by people I’ve interviewed and worked with who had to be told exactly how to solve a problem since they apparently had zero problem solving skill themselves. The “programming is just writing code” attitude drives me up the fucking wall and “LLMs are going to make programmers obsolete” is just the latest iteration of that bullshit.

Knowing how to write code has only ever been half (or less) of the job. A real programmer solves problems with code, especially problems that aren’t like any they’ve seen before. Someone who can write code but can’t solve problems or can only ‘solve’ problems they’ve seen before is just a code monkey. AI can regurgitate code it’s seen before (that is, code it was trained on) and it can do some problem solving but it falls on its face quickly if you ask it to do anything complex (at least for my metric of what is complex).