Programming

I honestly feel like I'm going crazy. Everybody is constantly telling me that LLMs are so good and capable, and a lot of people even consider them superior and talk about that like it's a known fact. But I just don't see it.

I tried to use them lot of times, but the results were simply not good and I always end up just wasting my time. And everybody is always like "well you have to write the right prompts, they have to be like this and this" or "well it might be that you're not using the right model, try this and this model" or whatever.

One guy even suggested that I might not be advanced enough to recognize it as good code, because again, LLMs are just so good and that's a fact. Which, to be fair, I fully admit to not being advanced or very experienced, especially in C++ which I am currently using and trying to learn. So I'm not rejecting this notion outright, it might actually be true. Although, I am experienced and I'd even say advanced in Python, and there I can say for sure that the code written by LLMs is just bad, convoluted, and unmaintainable.

A lot of times that I talk to people about this and explain my experience, and manage to convince them that I tried different prompts, models and everything, they always land on the same conclusion - that the code doesn't actually have to be readable and maintainable. LLMs are a new interface for coding, thus the code only has to be good enough for the LLM to be able to work with it.

And I honestly can't even say that this is such a ridiculous idea. That may very well be the path we're headed towards. And if that is the case, then my insisting on not using LLMs just puts me at a disadvantage since I'm much less productive (assuming that people who use LLMs are actually more productive - which I haven't seen conclusive studies about). I've read a little about this topic on reddit and stuff, and a lot of people really do agree with this - if you don't use LLMs, you're just too slow and it will only get worse. And this is coming from people who are (claiming to be) experienced software devs.

My situation is a little bit different though - I'm writing research code and I definitely DO have to be able to read and debug the code. If I don't, my results might be completely false and come from bugs or wrong functionalities. So I strongly reject using LLMs in my code, at least in the 95% of the code that actually implements my methodology.

But even still, I haven't had success with getting good code out of LLMs for the 5% that doesn't deal with the research, so what the hell is going on? I seriously feel like I'm going crazy since it seems to me like I'm the only one who isn't satisfied with the "capabilities" of LLMs. I can admit that my prompting needs work and that I don't exactly put much effort into it, but at some point, where do you draw the line? How many paragraphs do I have to write to ensure that the LLM does exactly what I want, before it becomes more work than actually doing it myself? I swear, some of the tips I get from people on prompting are crazy - writing .md files with instructions, writing out every single detail, explaining exactly how to do it and on and on and on. Its honestly easier and quicker just doing it myself.

And another thing - when did it happen that people stopped being embarrassed about heavily using LLMs to code?? Like this just flipped at some point and now people proudly talk about not having written a single line of code in a project (and I mean this literally, I've had multiple people tell me this). I cringe so hard every time in these conversations.

So I really have no idea what to think anymore. In any case, I won't be using LLMs in my research because the main goal is for me to actually learn something, as well as to produce reliable code and results. I'm still early in my research, so there's no problem with me doing that right now. But there will probably come a time where I will have learned enough about the topic to allow for "blackbox" code because I will be able to verify the functionalities and results, and I certainly expect higher pressure to use LLMs then. I have no idea what I'll do then.