Who here uses AI in programming?

[nexos]

Hello everyone,
As I've been doing more programming in the last few months, I've find myself gravitating towards using AI to a somewhat limited extent in my workflow. I know it's still hotly controversial in the programming world, especially in lower level spaces but I've found that using AI to help write boilerplate code (like error cleanup branches, basic memory safety checks, function boilerplates, and those such things) has greatly increased my productivity and helped me stay more focused and motivated. With that being said, I think using AI for more difficult, non-boilerplate tasks is still not ideal and lazy. So controversial as it may, what is your position on using AI in your workflow?

[nullplan]

Using tools to deal with repetitive tasks is fine, actually, and if you have to call these tools AI, then do so. But I think it is better to think of ways to circumvent the need for the boilerplate. Can you extract the repetitive stuff into a function? Or possibly a macro (that is what they were originally for, you know)?

Problem with using tools for programming is, you already don't like the code you write enough to actually write it yourself, so who in the world is going to read it?

[Solar]

Absolutely.

On a short leash.

Explaining underdocumented APIs. Creating example uses. Generating tests on code I've written. Code review. Architectural review. Not to mention Rubberducking.

I don't tell an LLM what program I want and then let it write my code. Which is what most people think of when the subject comes up, I suppose.

Instead, I tell it what kind of feature I'd like to add, look at what the LLM suggests, and usually get an even better idea, thanks for the inspiration.

When you're in a dark corner of some esoteric API that no-one has ever actually written documentation for, much less a tutorial, having a LLM scan the code and tell you what it thinks it should be used like, for example, can be immensely helpful. And I much rather have the more egregious of my mistakes pointed out at me before I commit source of mine for review.

Treating it as a coworker, instead of you being the project lead and the AI the developer, works out really nice for me.

PS: It also helps me with my neuro-spiciness. Sometimes I know exactly what I want to do with the code, but simply cannot engage my brain to work with the code. In these cases LLMs act as "language to code" converters, which makes all the difference.

[iansjack]

Rather depends upon what you consider AI.

Lint? Valgrind? Compiler warnings?

[Octocontrabass]

Remember when cryptocurrency was the next big thing and you had to invest right now or you'd regret it later?

Remember when NFTs were the next big thing and you had to invest right now or you'd regret it later?

I'm staying away from AI right now. I don't think I'll regret it later.

[thewrongchristian]

... I've found that using AI to help write boilerplate code (like error cleanup branches, basic memory safety checks, function boilerplates, and those such things) has greatly increased my productivity and helped me stay more focused and motivated.

Working in C, all these can be a problem, so I've developed tools to make them largely non-issues:

• error cleanup branches - I've implemented an exception mechanism based on setjmp/longjmp, so all my errors are exception based with CPP wrapper for syntax sugar. As part of this, I've implemented a callback mechanism that is called when an exception is thrown, for cleaning up transient resources without relying on the destructors I'd get for free in C++. Works a treat.
• basic memory safety checks - I#ve implemented a conservative GC, so dangling pointers and use after free is not a problem. I very raely use bare arrays, instead having a map implementation that is keyed by array index.
• function boilerplates - My C code is heavily interface based, using a COM=like construct. I have interfaces for FS (VFS interface), devices, maps (with tree and array implementations). I use an IDL compiler to generate the wrapper functions, then each implementation implements the interface(s) it wants to expose.

Most of the above could be done easier in C++, but porting the exception handling to run on an unhosted runtime scares the willy's out of me. But I do like the idea of using C++'s inheritance to make my interface use easier.

But as others have said, none of these are AI, and I am probably as productive using just basic tools like code completion in editors.

I'm very on the fence in terms of AI. I'm expecting that bubble to burst any minute.

[Demindiro]

LLMs? No, I've tried them a few times but they're damn near always wrong and I'm better off just writing the code myself. It doesn't help most of them act like yesmen, which I'm sure is intended to have a certain psychological effect.

AI? Of course! Fuzzers like American Fuzzy Lop are very useful. But I'm guessing that isn't what is meant with "AI" nowadays.

The biggest issue is that people use the term "AI" to refer to a very narrow category of machine learning (LLMs). Fuzzers in particular are far, far more effective at testing codebases than LLMs are, yet it seems many developers apparently don't even know about them?

[rohithkumar]

I have used AI in programming and find it very helpful as it saves me a lot of time. Plus, it gives me Alternative ways on how to write the code even more efficiently. Here's the thing. You will find AI WRONG most of the times because you may have not entered the right prompt or forgot to add more details. Also, Know WHAT YOU ARE TRYING TO ACHIEVE and if it really needs COMPLEX CODE. Once you understand your PROBLEM, seek AI ASSISTANCE(NOT 100% DEPENDENCE to write the full code for you!)wherever necessary and give it the right Prompt. Believe me, it works 99 out of 100 times!! It all depends on what and you are trying to code and at what extent. I am using AI as an Assistant to check my code's efficiency and better ways to work with it. Remember, seeking help of AI contributes a lot to your productivity and depending 100% on AI will definitely ruin your project.

[nexos]

While I agree with you that AI is helpful, even after really focusing on being specific in my prompting, I've found that AI still makes tons of mistakes. It's correct an almost scary amount of the time still, but it's definitely not correct most if the time, even with good prompting.

[Solar]

That is why I said to treat it as a (slightly sycophantic) co-collaborateur instead of a source of wisdom. Bounce your ideas off it. Let it give you an overview on a subject, a piece of example code, an algorithm suggestion. Have it review your code, or your architectural idea (but beware that it's designed to please you, it's hard to make AI disagree with you even if you are wrong.)

It's far from fault-free, but even with those flaws it knows a lot more than you, and it is *faster* than you by many orders of magnitude. I didn't have a clue the EDK2 project even existed. Gemini pointed it out as a possibility, guided me through the (rather unconventional) build process, and now I have UEFI firwmare that logs all the UEFI data structures to serial for me to double-check the bootloader results against. I learned a lot about QEMU options in the process as well, all without digging through a host of documentation. I got a list of struct definitons of the special registers in the Intel Manual Volume 3. Were they all flawless? No. But I got them in seconds, and double-checking them against the documentation (and adding a few the AI might have missed) will take significantly less time and be less tedious than "typing them off" would have been.

Use it for what it's good for. Doing your job for you is not one of those things -- it's a tool, not a solution.

[sandras]

Solar makes a good point.

I don't use AI to do my work. I generally use it to discuss stuff.

[BigBuda]

On a short leash.

I just loved this part.

[WinC64]

I program exclusively in x86-64 assembly, using both GoAsm and NASM (for AVX). I frequently consult with Copilot or Gemini, and often find good answers to my questions. However...

Answers are wordy and quickly go off-topic. Also, answers are often incorrect and confused. Many times I will receive an answer with 32-bit MASM syntax, even though my questions NEVER mention that assembler OR bitness (and the AI knows what assembler I am using). So, to combat the hallucinations I have had to implement a set of rules that I must frequently remind and enforce. It can be extremely frustrating, but I forgive it by telling myself that I must always be on my toes, too.

In the case of Gemini, the AI is quite stubborn and often refuses to admit its own errors (sarcastically, too), so in this respect, it is quite human!

My overall impression is that AI is good for suggesting new approaches to solving problems and is quite adept at answering questions related to instruction usage, but often likes using obsolete or inefficient instructions. In addition, the code examples it gives will many times be a complete mess (register usage, proper stack-frame setup, etc.). It is my opinion that the answers AI gives are from poor sources.

I will finish up by saying that I will never let any current AI write the code and then just copy/paste it into a program. That is a certain recipe for failure. AI is simply not capable of writing a "good" program in assembly. In my case, I limit the use of AI to a guide-only approach and find that this limit works fairly well -- most of the time.

[bellezzasolo]

I have to vote yes here even though I'm also somewhat in the "I'd rather die" camp.

But yeah, much like Solar says, it can be useful in the right circumstances. Ask it questions about how a piece of code works, ask it to spot errors and suggest structural and architectural improvements...

For this sort of thing, it's scarily good. not accurate nearly all the time, but certainly enough to be useful.

Use it as a tool, not as the workman.

[Vaotum]

We are going to be forced to use AI i think ...