Are LLMs capable of writing good code?
By "good" I mean code that is written professionally and concisely (and obviously works as intended). Apart from personal interest and understanding what the machine spits out, is there any legit reason anyone should learn advanced coding techniques? Specifically in an engineering perspective?
If not, learning how to write code seems a tad trivial now.
understanding what the machine spits out
This is exactly why people will still need to learn to code. It might write good code, but until it can write perfect code every time, people should still know enough to check and correct the mistakes.
I very much agree, thank you for indulging my question.
I used an LLM to write some code I knew I could write, but was a little lazy to do. Coding is not my trade, but I did learn Python during the pandemic. Had I not known to code, I would not have been able to direct the LLM to make the required corrections.
In the end, I got decent code that worked for the purpose I needed.
I still didn’t write any docstrings or comments.
For a very long time people will also still need to understand what they are asking the machine to do. If you tell it to write code for an impossible concept, it can't make it. If you ask it to write code to do something incredibly inefficiently, it's going to give you code that is incredibly inefficient.
I’ve even seen human engineers’ code thrown out because no one else could understand it. Back in the day, one webdev took it upon himself to whip up a mobile version of our company’s very complex website. He did it as a side project. It worked. It was complete. It looked good. It was very fast. The code was completely unreadable by anyone else. We didn’t use it.
After a certain point, learning to code (in the context of application development) becomes less about the lines of code themselves and more about structure and design. In my experience, LLMs can spit out well formatted and reasonably functional short code snippets, with the caveate that it sometimes misunderstands you or if you're writing ui code, makes very strange decisions (since it has no special/visual reasoning).
Anyone a year or two of practice can write mostly clean code like an LLM. But most codebases are longer than 100 lines long, and your job is to structure that program and introduce patterns to make it maintainable. LLMs can't do that, and only you can (and you can't skip learning to code to just get on to architecture and patterns)
I think this is the best response in this thread.
Software engineering is a lot more than just writing some lines of code and requires more thought and planning than can be realistically put into a prompt.
Very well put, thank you.
The other thing is, an LLM generally knows about all the existing libraries and what they contain. I don’t. So while I could code a pretty good program in a few days from first principles, an LLM is often able to stitch together some elegant glue code using a collection of existing library functions in seconds.
Also in my experience LLM can often propose solutions which are working but way too complex.
Story time: just yesterday, in VueJS I was trying to iterate over a list of items and render .text of reach item as HTML, but I needed to process it first. Note that in VueJS this is done by adding eg. <span v-html="item.text"></span> where content of the attribute is the JavaScript expression needed to get the text.
First I asked ChatGPT to write the function for processing the text. That worked pretty well and even used part of the JavaScript API which I was not aware about.
Next, I had a "dumb moment" when I did not realize that as I'm iterating through items I can just say <span v-html="processHtml(item.text)"></span>, that's all I really needed. Somehow I thought (or should I say, "hallucinated", ba dum tsss) for a moment that v-html is special or something (it is used differently than the most abundant type of syntax). So I went ahead and asked ChatGPT how to render processed texts while iterating.
It came with a rather contrived solution which involved creating another computed property containing a list of processed texts. I started to integrate it into the existing loop: I would have to add index and use that index to pull the code from the computed property, which already felt a little bit weird.
That's when it struck me: no, no, no, I can just f*ing use the function.
TL; DR: The point is, while ChatGPT was helpful I still needed to babysit it. And if I didn't snap from my lazy moment, or if I simply didn't know better, I would end up with code which is more complex, more surprising, which means harder to reason about for both humans and LLM's. (For humans because now it forces you to speculate about coder's intent, and for LLM's because it's less likely to be reminiscent of surrounding code in its learning data.)
Great question.
is there any legit reason anyone should learn advanced coding techniques?
Don't buy the hype. LLMs can produce all kinds of useful things but they don't know anything at all.
No LLM has ever engineered anything. And there's no sparse (concession to a good point made in response) current evidence that any AI ever will.
Current learning models are like trained animals in a circus. They can learn to do any impressive thing you an imagine, by sheer rote repetition.
That means they can engineer a solution to any problem that has already been solved millions of times already. As long as the work has very little new/novel value and requires no innovation whatsoever, learning models do great work.
Horses and LLMs that solve advanced algebra don't understand algebra at all. It's a clever trick.
Understanding the problem and understanding how to politely ask the computer to do the right thing has always been the core job of a computer programmer.
The bit about "politely asking the computer to do the right thing" makes massive strides in convenience every decade or so. Learning models are another such massive stride. This is great. Hooray!
The bit about "understanding the problem" isn't within the capabilities of any current learning model or AI, and there's no current evidence that it ever will be.
Someday they will call the job "prompt engineering" and on that day it will still be the same exact job it is today, just with different bullshit to wade through to get it done.
I appreciate your candor, I had a feeling it was cock and bull but you've answered my question fully.
Wait, if you can (or anyone else chipping in), please elaborate on something you've written.
When you say
That means they can engineer a solution to any problem that has already been solved millions of times already.
Hasn't Google already made advances through its Alpha Geometry AI?? Admittedly, that's a geometry setting which may be easier to code than other parts of Math and there isn't yet a clear indication AI will ever be able to reach a certain level of creativity that the human mind has, but at the same time it might get there by sheer volume of attempts.
Isn't this still engineering a solution? Sometimes even researchers reach new results by having a machine verify many cases (see the proof of the Four Color Theorem). It's true that in the Four Color Theorem researchers narrowed down the cases to try, but maybe a similar narrowing could be done by an AI (sooner or later)?
I don't know what I'm talking about, so I should shut up, but I'm hoping someone more knowledgeable will correct me, since I'm curious about this
Isn't this still engineering a solution?
If we drop the word "engineering", we can focus on the point - geometry is another case where rote learning of repetition can do a pretty good job. Clever engineers can teach computers to do all kinds of things that look like novel engineering, but aren't.
LLMs can make computers look like they're good at something they're bad at.
And they offer hope that computers might someday not suck at what they suck at.
But history teaches us probably not. And current evidence in favor of a breakthrough in general artificial intelligence isn't actually compelling, at all.
Sometimes even researchers reach new results by having a machine verify many cases
Yes. Computers are good at that.
So far, they're no good at understanding the four color theorum, or at proposing novel approaches to solving it.
They might never be any good at that.
Stated more formally, P may equal NP, but probably not.
Edit: To be clear, I actually share a good bit of the same optimism. But I believe it'll be hard won work done by human engineers that gets us anywhere near there.
Ostensibly God created the universe in Lisp. But actually he knocked most of it together with hard-coded Perl hacks.
There's lots of exciting breakthroughs coming in computer science. But no one knows how long and what their impact will be. History teaches us it'll be less exciting than Popular Science promised us.
Edit 2: Sorry for the rambling response. Hopefully you find some of it useful.
I don't at all disagree that there's exciting stuff afoot. I also think it is being massively oversold.
Hasn’t Google already made advances through its Alpha Geometry AI?? Admittedly, that’s a geometry setting which may be easier to code than other parts of Math and there isn’t yet a clear indication AI will ever be able to reach a certain level of creativity that the human mind has, but at the same time it might get there by sheer volume of attempts.
Wanted to focus a bit on this. The thing with AlphaGeometry and AlphaProof is that they really treat doing math as a game, not unlike chess. For example, AlphaGeometry has a basic set of rules, it can apply them and it knows when it is done. And when it is done, you can be 100% sure that the solution is correct, because the rules of the game are known; the 28/42 score reported in the article is really four perfect scores and three zeros. Those systems do use LLMs, but they really are only there to suggest to the system what to try doing next. There is a very enlightening picture in the AlphaGeometry paper here: https://www.nature.com/articles/s41586-023-06747-5#Fig1
You can automatically verify correctness of code the same way. For example Lean, the language AlphaProof uses internally, can be used for general programming. In general, we call similar programming techniques formal methods. But most people don't do this, since this is more time-consuming than normal programming, and in many cases we don't even know how to define the goal of our code (how to define correct rendering in a game?). So this is only really done when the correctness of the program is critical, like famously they verified the code of the automatic metro in Paris this way. And so most people don't try to make programming AI work this way.
I worry for the future generations of people who can use chatgpt to write code but have absolutely no idea what said code is doing.
That's some 40k shit.
"What does it mean?" "I do not know, but it appeases the machine spirit. Quickly, recite the canticles."
This is directly how we're getting to a 40k future and I hate it. The bad future!
If we must I might join the Mechanicus though. I'm good at chanting and doing things by rote.
My CTO thoroughly believes that within 4-6 years we will no longer need to know how to read or write code, just how to ask an AI to do it. Coincidentally, he also doesn't code anymore and hasn't for over 15 years.
From a business perspective, no shareholder cares at how good an employee is at personally achieving a high degree of skill. They only care about selling and earning, and to a lesser degree an enduring reputation for longer term earnings.
Economics could very well drive this forward. But I don't think the craft will be lost. People will need to supervise this progress as well as collaborate with the machines to extend its capabilities and dictate its purposes.
I couldn't tell you if we're talking on a time scale of months or decades, but I do think "we" will get there.
I think he's correct and there's a ton of cope going on on lemmy right now. I also think tons of art/graphic design jobs will disappear never to return.
Don't be, there will come a time when nearly all code is AI created, and not human readable.
You need to worry for the future when big data sites are running code they literally don't know how it works and have no way to verify because of how cheap and relatively effective it is.
Then after that LLMs will get better at coding than any human can achieve, but will still be black box human unreadable code but there will be no chain of discipline left to teach new programmers.
Hardly anyone is taking this seriously because corporations stand to make a fucktonne of money and normal people are in general clueless about complex subjects that require a nuanced understanding, yet strangely arrogant about their ignorant opinions based on movies and their drinking buddies malformed opinions.
I use it to write code, but I know how to write code and it probably turns a week of work for me into a day or two. It's cool, but not automagic.
I find it better at things under 100 lines. Otherwise it starts to lose context. Any ideas how to address this?
Ask it to make a function, then do some other function, then make them work together etc. Making it write a lot in one go won't work. It's more pair programming than having it write for you.
LLMs are just computerized puppies that are really good at performing tricks for treats. They’ll still do incredibly stupid things pretty frequently.
I’m a software engineer, and I am not at all worried about my career in the long run.
In the short term… who fucking knows. The C-suite and MBA circlejerk seems to have decided they can fire all the engineers because wE CAn rEpLAcE tHeM WitH AI 🤡 and then the companies will have a couple absolutely catastrophic years because they got rid of all of their domain experts.
I'm my experience they do a decent job of whipping out mindless minutea and things that are well known patterns in very popular languages.
They do not solve problems.
I think for an "AI" product to be truly useful at writing code it would need to incorporate the LLM as a mere component, with something facilitating checks through static analysis and maybe some other technologies, maybe even mulling the result through a loop over the components until they're all satisfied before finally delivering it to the user as a proposal.
It's a decent starting point for a new language. I had to learn webdev as an embedded C coder, and using a LLM and cross-referencing the official documentation makes a new language much more approachable.
I agree, LLMs have been helpful in pointing me in the right direction and helping me rethink what questions I actually want to ask in disciplines I'm not very familiar with.
Those kinds of patterns are already emerging! That "mulling the result through a loop" step is called "reflection," and it does a great job of catching mistakes and hallucinations. Nothing is on the scale of doing the whole problem-solving and implementation from business requirements to deployed product-- probably never will be, IMO-- but this "making the LLM a component in a broader system with diverse tools" is definitely something that we're currently figuring out patterns for.
No, because that would require it being trained on good code. Which is rather rare.
If it is trained on Stack Overflow there is no chance.
The answers on stack overflow are often borderline genius.
It's training on practically all code that exists. It has plenty of good examples, and plenty of junk examples.
They can write good short bits of code. But they also often produce bad and even incorrect code. I find it more effort to read and debug its code then just writing it myself to begin with the vast majority of the time and find overall it just wastes more of my time overall.
Maybe in a couple of years they might be good enough. But it looks like their growth is starting to flatten off so it is up for debate as to if they will get there in that time.
No, a large part of what "good code" means is correctness. LLMs cannot properly understand a problem so while they can produce grunt code they can't assemble a solution to a complex problem and, IMO, it is impossible for them to overtake humans unless we get really lazy about code expressiveness. And, on that point, I think most companies are underinvesting into code infrastructure right now and developers are wasting too much time on unexpressive code.
The majority of work that senior developers do is understanding a problem and crafting a solution appropriate to it - when I'm working my typing speed usually isn't particularly high and the main bottleneck is my brain. LLMs will always require more brain time while delivering a savings on typing.
At the moment I'd also emphasize that they're excellent at popping out algorithms I could write in my sleep but require me to spend enough time double checking their code that it's cheaper for me to just write it by hand to begin with.
A broken clock is right twice a day.
Yes ... and it doesn't know when it is on time.
Also, machines are getting better and they can help us with inspiration.
This question is basically the same as asking "Are 2d6 capable of rolling a 9?"
I have no knowledge of coding, my bad for asking a stupid question in NSQ.
Sorry, I wasn't trying to berate you. Just trying to illustrate the underlying assumption of your question
Yes, two six-sided dice (2d6) are capable of rolling a sum of 9. Here are the possible combinations that would give a total of 9:
So, there are four different combinations that result in a roll of 9.
…
See? LLMs can do everything!
This seems like the most sane take.
A computer can do a lot. But if you give the computer to a regular fish instead of a regular human, that’s just a regular fish next to a computer. Not very useful.
Of course it can. It can also spit out trash. AI, as it exists today, isn't meant to be autonomous, simply ask it for something and it spits it out. They're meant to work with a human on a task. Assuming you have an understanding of what you're trying to do, an AI can probably provide you with a pretty decent starting point. It tends to be good at analyzing existing code, as well, so pasting your code into gpt and asking it why it's doing a thing usually works pretty well.
AI is another tool. Professionals will get more use out of it than laymen. Professionals know enough to phrase requests that are within the scope of the AI. They tend to know how the language works, and thus can review what the AI outputs. A layman can use AI to great effect, but will run into problems as they start butting up against their own limited knowledge.
So yeah, I think AI can make some good code, supervised by a human who understands the code. As it exists now, AI requires human steering to be useful.
For basic boiler plate like routes for an API, an etl script from sample data to DB tables, or other similar basics, yeah, it's perfectly acceptable. You'll need to swap out dummy addresses, and maybe change a choice or two, but it's fine.
But when you're trying to organize more complicated business logic or debug complicated dependencies it falls over
For small boilerplate or very common small pieces of code, for instance a famous algorithm implementation. Yes. As they are just probably giving you the top stack overflow answer for a classic question.
Anything that the LLM would need to mix or refactor would be terrible.
That all depends on where the data set comes from. The code you'll get out of an LLM is the average code of the data set. If it's scraped from the internet (which is very likely) the code you'll get will be an amalgam of concise examples from one website, incorrect examples from another, bits from blogs with all the typos and all the gunk and garbage that's out there.
Getting LLM code to work well takes an understanding of what the code it gives you actually does and why it's bad. It will always be bad because it cannot be better than the dataset and in order for a dataset to be big enough to train an LLM it'll have to have everything they can get including all the trash. But it can be good for providing you a framework to start with. It is however never going to replace actual programming and understanding of programming. The talk of LLMs completely replacing programers is mostly coming from people who do not understand coding or LLMs at all.
Can't LLM's eventually gain some form of "sentience", and be able to self correct? A sort of thinking before speaking kind of situation.
This question right here perfectly encapsulates everything wrong with LLMs right now. They could be good tools but the people pushing them have no idea what they even are. LLMs do not make decisions. All the decisions an LLM appears to make were made in the dataset. All those things that an LLM does that make it seem intelligent were done or said by a human somewhere on the internet. It is a statistical model that determines what output is mostly likely to come next. That is it. It is nothing else. It is not smart. It does not and cannot make decisions. It is an algorithm that searches a dataset and when it can't find something it'll provide convincing-looking gibberish instead.
Listen think of it like this; a man decides to take exams to become a doctor in France, but for some reason he doesn't learn either french or medicine. No, no instead he studies every former exam and all the answers to them. He gets very good at regurgitating those answers so much so that he can even pass the exam. But at no point does he understand what any of it means and when asked new and novel questions he provides utter nonsense answers. No matter how good he gets at memorising those answers he will never get any better at medicine. LLMs are as likely to gain sentience as my excel spreadsheets are.
No LLM is trust worthy.
Unless you understand the code and can double check what it’s doing I wouldn’t risk running it.
And if you do understand it any benefit of time saved is likely going to be offset by debugging and verifying what it actually does.
In my experience, not at all. But sometimes they help with creativity when you hit a wall or challenge you can't resolve.
They have been trained off internet examples where everyone has a different style/method of coding, like writing style. It's all very messy and very unreliable. It will be years for LLMs to code "good" and will require a lot of training that isn't scraping.
AI can only really complete tasks that are both simple and routine. I'd compare the output skill to that of a late-first-year University student, but with the added risk of halucination. Anything too unique or too compex tends to result in significant mistakes.
In terms of replacing programmers, I'd put it more in the ballpark of predictive text and/or autocorrect for a writer. It can help speed up the process a little bit, and point out simple mistakes but if you want to make a career out of it, you'll need to actually learn the skill.
I've tried Copilot and to be honest, most of the time it's a coin toss, even for short snippets. In one scenario it might try to autocomplete a unit test I'm writing and get it pretty much spot on, but it's also equally likely to spit out complete garbage that won't even compile, never mind being semantically correct.
To have any chance of producing decent output, even for quite simple tasks, you will need to give an LLM an extremely specific prompt, detailing the precise behaviour you want and what the code should do in each scenario, including failure cases (hmm...there used to be a term for this...)
Even then, there are no guarantees it won't just spit out hallucinated nonsense. And for larger, enterprise scale applications? Forget it.
Yes, in small bits, after several tries, with human supervision. For now.
No in large amounts, too hard to human review, they're still doing it anyway.
Dunno. I'd expect to have to make several attempts to coax a working snippet from the ai, then spending the rest of the time trying to figure out what it's done and debugging the result. Faster to do it myself.
E.g. I once coded Tetris on a whim (45 min) and thought it'd be a good test for ui/ game developer, given the multi disciplinary nature of the game (user interaction, real time engine, data structures, etc) Asked copilot to give it a shot and while the basic framework was there, the code simply didn't work as intended. I figured if we went into each of the elements separately, it would have taken me longer than if i'd done it from scratch anyway.
Theoretically, I would say yes it’s possible, insofar as we could break down most subtasks of the development process into training parameters. But we are a long way from that currently.
ETA: I suspect LLM’s best use-case in this hypothetical would not be in architecting or implementation, but rather limited to tasks with human interfaces (requirements gathering, project planning and logistics, test scaffolding, feedback collection/distribution, etc).
If the unironic goal is to develop things without any engineering oversight (mistake) then there’s no point to using programming languages at all. The machine might as well just output assembly or bin code.
What’s more likely in the short term are software LLMs generating partial solutions that human engineers then are asked to “finish” (fix) and maintain. The effort and hours required to do so will, at a guess, balloon terribly and will often be at best proportional to the resources saved by the use of the automatic spaghetti generator.
I eagerly await these post mortems.
Technically it's possible, but it's neither probable nor likely, and it's especially not effective. From what I understand, a lot of devs who do try to use something like ChatGPT to write code end up spending as much or more time debugging it, and just generally trying to get it to work, than they would have if they'd just written it themselves. Additionally, you have to know how to code to be able to figure out why it's not working, and even when all of that is done, it's almost impossible to get it to integrate with a larger project without just rewriting the whole thing anyway.
So to answer the question you intend to ask, no, LLMs will not be replacing programmers any time soon. They may serve as a tool of dubious value, but the idea that programmers will be replaced is only taken seriously by by people who manage programmers, and not the programmers themselves.
I don't know how to program, but to a very limited extent can sorta kinda almost understand the logic of very short and simplistic code that's been written for me by someone who can actually code. I tried to get to get chat GPT to write a shell script for me to work as part of an Apple shortcut. It has no idea. It was useless and ridiculously inconsistent and forgetful. It was the first and only time I used chat GPT. Not very impressed.
Given how it is smart enough to produce output that's kind of in the area of correct, albeit still wrong and logically flawed, I would guess it could eventually be carefully prodded into making one small snippet of something someone might call "good" but at that point I feel like that's much more an accident in the same way that someone who has memorised a lot of French vocabulary but never actually learned French might accidentally produce a coherent sentence once in a while by trying and failing 50 times before succeeding and then failing again immediately after without ever having even known.
Its the most ok’est coder with the attention span of a 5 year old.
For repetitive tasks, it can almost automatically get a first template you write by hand, and extrapolate with multiple variations.
Beyond that… not really. Anything beyond single line completion quickly devolves into either something messy, non working, or worse, working but not as intended. For extremely common cases it will work fine; but extremely common cases are either moved out in shared code, or take less time to write than to "generate" and check.
I've been using code completion/suggestion on the regular, and it had times where I was pleasantly surprised by what it produced, but even for these I had to look after it and fix some things. And while I can't quantify how often it happened, there are a lot of times where it's convincing gibberish.
I've also had some decent luck when using a new/unfamiliar language by asking it to make the code I wrote more idiomatic.
It's been a nice way to learn some tricks I probably wouldn't have bothered with before
The LLM can type the Code, but you need to know what you want / how you want to solve it.
Absolutely, but they need a lot of guidance. GitHub CoPilot often writes cleaner code than I do. I'll write the code and then ask it to clean it up for me and DRYify it.
Yes and no. GPT usually gives me clever solutions I wouldn’t have thought of. Very often GPT also screws up, and I need to fine tune variable names, function parameters and such.
I think the best thing about GPTis that it knows the documentation of every function, so I can ask technical questions. For example, can this function really handle dataframes, or will it internally convert the variable into a matrix and then spit out a dataframe as if nothing happened? Such conversions tend to screw up the data, which explains some strange errors I bump into. You could read all of the documentation to find out, or you could just ask GPT about it. Alternatively, you could show how badly the data got screwed up after a particular function, and GPT would tell that it’s because this function uses matrices internally, even though it looks like it works with dataframes.
I think of GPT as an assistant painter some famous artists had. The artist tells the assistant to paint the boring trees in the background and the rough shape of the main subject. Once that’s done, the artist can work on the fine details, sign the painting, send it to the local king and charge a thousand gold coins.
Ai is excellent at completing low effort ai generated Pearson programming homework while I spend all the time I saved on real projects that actually matter. My hugging face model is probably trained on the same dataset as their bot. It gets it correct about half the time and another 25% of the time, I just have to change a few numbers or brackets around. It takes me longer to read the instructions than it takes the ai bot to spit out the correct answer.
None of it is "good" code but it enables me to have time to write good code somewhere else.
my dad uses this LLM python code generation quite routinely, he says the output's mostly fine.
For snippets yes, ask him to tell it to make a complete terminal service and see what happens
I use LLMs for C code - most often when I know full well how to code something but I don't want to spent half a day expressing it and debugging it.
ChatGPT or Copilot will spit out a function or snippet that's usually pretty close to what I want. I patch it up and move on to the tougher problems LLMs can't do.
Writing code is probably one of the few things LLMs actually excell at. Few people want to program something nobody has ever done before. Most people are just reimplimenting the same things over and over with small modifications for their use case. If imports of generic code someone else wrote make up 90% of your project, what's the difference in getting an LLM to write 90% of your code?
IMO this perspective that we're all just "reimplementing basic CRUD" applications is the reason why so many software projects fail.
I see where you're coming from, sort of like the phrase "don't reinvent the wheel". However, considering ethics, that doesn't sound far off from plagiarism.