Asklemmy

I've been messing around with running my own LLMs at home using LM Studio and I've got so say it really helps me write code. I'm using Code Llama 13b, and it works pretty well as a programmer assistant. What I like about using a chatbot is that I go from writing code to reviewing it, and for some reason this keeps me incredibly mentally engaged. This tech has been wonderful for undoing some of my professional burnout.

If what keeps you mentally engaged does not include a bot, then I don't think you need any other reason to not use one. As much as I really like the tech, anyone that uses it is still going to need to know the language and enough about the libraries to fix the inevitable issues that come up. I can definitely see this tech getting better to the point of being unavoidable, though. You hear that Microsoft is planning on adding an AI button to their upcoming keyboards? Like that kind of unavoidable.

This might cheer you up: https://visualstudiomagazine.com/articles/2024/01/25/copilot-research.aspx

I don't think we have anything to worry about just yet. LLMs are nothing but well-trained parrots. They can't analyse problems or have intuitions about what will work for your particular situation. They'll either give you something general copied and pasted from elsewhere or spin you a yarn that sounds plausible but doesn't stand up to scrutiny.

Getting an AI to produce functional large-scale software requires someone to explain precisely the problem domain: each requirement, business rule, edge case, etc. At which point that person is basically a developer, because I've never met a project manager who thinks that granularly.

They could be good for generating boilerplate, inserting well-known algorithms, generating models from metadata, that sort of grunt work. I certainly wouldn't trust them with business logic.

Yeh

I'm gonna sum up my feelings on this with a (probably bad) analogy.

AI taking software developer jobs is the same thinking as microwaves taking chefs jobs.

They're both just tools to help you achieve the same goal easier/faster. And sometimes the experts will decide to forego the tool and do it by hand for better quality control or high complexity that the tool can't do a good job at.

As a welder, I've been hearing for 20 years that "robots are going to replace you" and "automation is going to put you out of a job" yadda yadda. None of you code monkies gave a fuck about me and my job, but now it's a problem because it affects you and your paycheck? Fuck you lmao good riddance to bad garbage.

I’m a 50+ year old IT guy who started out as a c/c++ programmer in the 90’s and I’m not that worried.

The thing is, all this talk about AI isn’t very accurate. There is a huge difference in the LLM stuff that ChatGPT etc. are built on and true AI. These LLM’s are only as good as the data fed into them. The adage “garbage in, garbage out” comes to mind. Anybody that blindly relies on them is a fool. Just ask the lawyer that used ChatGPT to write a legal brief. The “AI” made up references to non-existent cases that looked and sounded legitimate, and the lawyer didn’t bother to check for accuracy. He filed the brief and it was the judge that discovered the brief was a work of fiction.

Now I know there’s a huge difference between programming and the law, but there are still a lot of similarities here. An AI generated program is only going to be as good as the samples provided to it, and you’re probably want a human to review that code to ensure it’s truly doing what you want, at the very least.

I also have concern that programming LLMs could be targeted by scammers and the like. Train the LLM to harvest sensitive information and obfuscate the code that does it so that it’s difficult for a human to spot the malicious code without a highly detailed analysis of the generated code. That’s another reason to want to know exactly what the LLM is trained on.

Yes of course you are.

I use GitHub Copilot from work. I generally use Python. It doesn't take away anything at least for me. It's big thing is tab completion; it saves me from finishing some lines and adding else clauses. Like I'll start writing a docstring and it'll finish it.

Once in a while I can't think of exactly what I want so I write a comment describing it and Copilot tries to figure out what I'm asking for. It's literally a Copilot.

Now if I go and describe a big system or interfacing with existing code, it quickly gets confused and tends to get in the weeds. But man if I need someone to describe a regex, it's awesome.

Anyways I think there are free alternatives out there that probably work as well. At the end of the day, it's up to you. Though I'd so don't knock it till you try it. If you don't like it, stop using it.

It doesn't matter what you think about AI. It's very clear that this technology is here to stay and will only improve. From this point on AI will become deeply integrated into human culture and technology, after all we've been fetishizing it for almost 100 years now. Your only logical option as a developer is to learn how to use it and abuse it. Choosing not to do so is career suicide, possibly even societal suicide depending on how quickly adoption happens.

You're probably right, in the near future people that can't use it will be fired. To that point they should be fired. Why the fuck would I allow my accounts to do their financal work on paper when Excel exists?

Welcome to the future.

i'm still in uni so i can't really comment about how's the job market reacting or is going to react to generative AI, what i can tell you is it has never been easier to half ass a degree. any code, report or essay written has almost certainly came from a LLM model, and none of it makes sense or barely works. the only people not using AI are the ones not having access to it.

i feel like it was always like this and everyone slacked as much as they could but i just can't believe it, it's shocking. lack of fundamental and basic knowledge has made working with anyone on anything such a pain in the ass. group assignments are dead. almost everyone else's work comes from a chatgpt prompt that didn't describe their part of the assignment correctly, as a result not only it's buggy as hell but when you actually decide to debug it you realize it doesn't even do what its supposed to do and now you have to spend two full days implementing every single part of the assignment yourself because "we've done our part".

everyone's excuse is "oh well university doesn't teach anything useful why should i bother when i'm learning ?" and then you look at their project and it's just another boilerplate react calculator app in which you guessed it most of the code is generated by AI. i'm not saying everything in college is useful and you are a sinner for using somebody else's code, indeed be my guest and dodge classes and copy paste stuff when you don't feel like doing it, but at least give a damn on the degree you are putting your time into and don't dump your work on somebody else.

i hope no one carries this kind of sentiment towards their work into the job market. if most members of a team are using AI as their primary tool to generate code, i don't know how anyone can trust anyone else in that team, which means more and longer code reviews and meetings and thus slower production. with this, bootcamps getting more scammy and most companies giving up on junior devs, i really don't think software industry is going towards a good direction.

AI allows us to do more with less just like any other tool. It's no different than an electric drill or a powered saw. Perhaps in the future we will see more immersive environment games because much of the immersive environment can be made with AI doing the grunt work.

I'm a composer. My facebook is filled with ads like "Never pay for music again!". Its fucking depressing.

As someone with deep knowledge of the field, quite frankly, you should now that AI isn't going to replace programmers. Whoever says that is either selling a snake oil product or their expertise as a "futurologist".

Your fear is in so far justified as that some employers will definitely aim to reduce their workforce by implementing AI workflow.

When you have worked for the same employer all this time, perhaps you don't know, but a lot of employers do not give two shits about code quality. They want cheap and fast labour and having less people churning out more is a good thing in their eyes, regardless of (long-term) quality. May sound cynical, but that is my experience.

My prediction is that the income gap will increase dramatically because good pay will be reserved for the truly exceptional few. While the rest will be confronted with yet another tool capitalists will use to increase profits.

Maybe very far down the line there is blissful utopia where no one has to work anymore. But between then and now, AI would have to get a lot better. Until then it will be mainly used by corporations to justify hiring less people.

I love llms! I'm using them to answer all sorts of bullshit to become a manager....like here's a bunch of notes make me a managers review of Brian. LOL.

I think Google is struggling to control the bullshit flood from the Internet and so AI is about to eat their lunch. Like I already decided that all AIs are just bullshit and the only really useful AIs are the ones that can actually search the Internet live. Perplexity AI was doing this for a while but someone chopped off it's balls. I've been looking for a replacement ever since.

I also use it for help with python, with Linux, with docker, with solid works and stuff around the house like taxes, kombucha, identifying plants and stupid stuff like that.

But I can definitely see the future when the police are replaced with robo dogs with lases heads that can run at 120mph and shoot holes through cars. The only benefit being that the hole doesn't get infected and there's no pool of blood. That future is coming. I'm going to start wearing aluminum reflective shield armor.

I wish your fear were justified! I'll praise anything that can kill work.

Hallas, we're not here yet. Current AI is a glorified search engine. The problem it will have is that most code today is unmaintainable garbage. So AI can only do this for now : unmaintainable garbage.

First the software industry needs to properly industrialise itself. Then there will be code to copy and reuse.

🙄 no I'm sure you're the only one

Programming is the most automated career in history. Functions / subroutines allow one to just reference the function instead of repeating it. Grace Hopper wrote the first compiler in 1951; compilers, assemblers, and linkers automate creating machine code. Macros, higher level languages, garbage collectors, type checkers, linters, editors, IDEs, debuggers, code generators, build systems, CI systems, test suite runners, deployment and orchestration tools etc... all automate programming and programming-adjacent tasks, and this has been going on for at least 70 years.

Programming today would be very different if we still had to wire up ROM or something like that, and even if the entire world population worked as programmers without any automation, we still wouldn't achieve as much as we do with the current programmer population + automation. So it is fair to say automation is widely used in software engineering, and greatly decreases the market for programmers relative to what it would take to achieve the same thing without automation. Programming is also far easier than if there was no automation.

However, there are more programmers than ever. It is because programming is getting easier, and automation decreases the cost of doing things and makes new things feasible. The world's demand for software functionality constantly grows.

Now, LLMs are driving the next wave of automation to the world's most automated profession. However, progress is still slow - without building massive very energy expensive models, outputs often need a lot of manual human-in-the-loop work; they are great as a typing assist to predict the next few tokens, and sometimes to spit out a common function that you might otherwise have been able to get from a library. They can often answer questions about code, quickly find things, and help you find the name of a function you know exists but can't remember the exact name for. And they can do simple tasks that involve translating from well-specified natural language into code. But in practice, trying to use them for big complicated tasks is currently often slower than just doing it without LLM assistance.

LLMs might improve, but probably not so fast that it is a step change; it will be a continuation of the same trends that have been going for 70+ years. Programming will get easier, there will be more programmers (even if they aren't called that) using tools including LLMs, and software will continue to get more advanced, as demand for more advanced features increases.