Please remove it if unallowed
I see alot of people in here who get mad at AI generated code and I am wondering why. I wrote a couple of bash scripts with the help of chatGPT and if anything, I think its great.
Now, I obviously didnt tell it to write the entire code by itself. That would be a horrible idea, instead, I would ask it questions along the way and test its output before putting it in my scripts.
I am fairly competent in writing programs. I know how and when to use arrays, loops, functions, conditionals, etc. I just dont know anything about bash’s syntax. Now, I could have used any other languages I knew but chose bash because it made the most sense, that bash is shipped with most linux distros out of the box and one does not have to install another interpreter/compiler for another language. I dont like Bash because of its, dare I say weird syntax but it made the most sense for my purpose so I chose it. Also I have not written anything of this complexity before in Bash, just a bunch of commands in multiple seperate lines so that I dont have to type those one after another. But this one required many rather advanced features. I was not motivated to learn Bash, I just wanted to put my idea into action.
I did start with internet search. But guides I found were lacking. I could not find how to pass values into the function and return from a function easily, or removing trailing slash from directory path or how to loop over array or how to catch errors that occured in previous command or how to seperate letter and number from a string, etc.
That is where chatGPT helped greatly. I would ask chatGPT to write these pieces of code whenever I encountered them, then test its code with various input to see if it works as expected. If not, I would ask it again with what case failed and it would revise the code before I put it in my scripts.
Thanks to chatGPT, someone who has 0 knowledge about bash can write bash easily and quickly that is fairly advanced. I dont think it would take this quick to write what I wrote if I had to do it the old fashioned way, I would eventually write it but it would take far too long. Thanks to chatGPT I can just write all this quickly and forget about it. If I want to learn Bash and am motivated, I would certainly take time to learn it in a nice way.
What do you think? What negative experience do you have with AI chatbots that made you hate them?
That is the general reason, i use llms to help myself with everything including coding too, even though i know why it’s badI’m fairly sure Linus would disapprove of my "rip everything off of Stack Overflow and ship it " programming style.
And sometimes that’s my only option when interfacing to toehr API.
And we still using it brother
This is a good quote, but it lives within a context of professional code development.
Everyone in the modern era starts coding by copying functions without understanding what it does, and people go entire careers in all sorts of jobs and industries without understanding things by copying what came before that ‘worked’ without really understanding the underlying mechanisms.
What’s important is having a willingness to learn and putting in the effort to learn. AI code snippets are super useful for learning even when it hallucinates if you test it and make backups first. This all requires responsible IT practices to do safely in a production environment, and thats where corporate management eyeing labor cost reduction loses the plot, thinking AI is a wholesale replacement for a competent human as the tech currently stands.
Based Linus strikes again
If you’re not an experienced developer, it could be used as a crutch rather than actually learning how to write the code.
The real reason? People are just fed up with AI in general (that has no real-world use to most people) being crammed down their throats and having their personal code (and other data) being used to train models for megacorps.
There are probably legitimate uses out there for gen AI, but all the money people have such a hard-on for the unethical uses that now it’s impossible for me to hear about AI without an automatic “ugggghhhhh” reaction.
Two reasons:
- my company doesn’t allow it - my boss is worried about our IP getting leaked
- I find them more work than they’re worth - I’m a senior dev, and it would take longer for me to write the prompt than just write the code
I just dont know anything about bash’s syntax
That probably won’t be the last time you write Bash, so do you really want to go through AI every time you need to write a Bash script? Bash syntax is pretty simple, especially if you understand the basic concept that everything is a command (i.e. syntax is
<command> [arguments...]; likeif <condition>where<condition>can be[ <special syntax> ]or[[ <test syntax> ]]), which explains some of the weird corners of the syntax.AI sucks for anything that needs to be maintained. If it’s a one-off, sure, use AI. But if you’re writing a script others on your team will use, it’s worth taking the time to actually understand what it’s doing (instead of just briefly reading through the output). You never know if it’ll fail on another machine if it has a different set of dependencies or something.
What negative experience do you have with AI chatbots that made you hate them?
I just find dealing with them to take more time than just doing the work myself. I’ve done a lot of Bash in my career (>10 years), so I can generally get 90% of the way there by just brain-dumping what I want to do and maybe looking up 1-2 commands. As such, I think it’s worth it for any dev to take the time to learn their tools properly so the next time will be that much faster. If you rely on AI too much, it’ll become a crutch and you’ll be functionally useless w/o it.
I did an interview with a candidate who asked if they could use AI, and we allowed it. They ended up making (and missing) the same mistake twice in the same interview because they didn’t seem to actually understand what the AI output. I’ve messed around with code chatbots, and my experience is that I generally have to spend quite a bit of time to get what I want, and then I still need to modify and debug it. Why would I do that when I can spend the same amount of time and just write the code myself? I’d understand the code better if I did it myself, which would make debugging way easier.
Anyway, I just don’t find it actually helpful. It can feel helpful because it gets you from 0 to a bunch of code really quickly, but that code will probably need quite a bit of modification anyway. I’d rather just DIY and not faff about with AI.
You boss should be more worried about license poisoning when you incorporate code that’s been copied from copyleft projects and presented as “generated”.
Perhaps, but our userbase is so small that we’d be very unlikely that someone would notice. We are essentially B2B with something like a few hundred active users. We do vet our dependencies religiously, but in all actuality, we could probably get away with pulling in some copyleft code.
We built a Durable task workflow engine to manage infrastructure and we asked a new hire to add a small feature to it.
I checked on them later and they expressed they were stuck on an aspect of the change.
I could tell the code was ChatGPT. I asked “you wrote this with ChatGPT didn’t you?” And they asked how I could tell.
I explained that ChatGPT doesn’t have the full context and will send you on tangents like it has here.
I gave them the docs to the engine and to the integration point and said "try using only these and ask me questions if you’re stuck for more than 40min.
They went on to become a very strong contributor and no longer uses ChatGPT or copilot.
I’ve tried it myself and it gives me the wrong answers 90% of the time. It could be useful though. If they changed ChatGPT to find and link you docs it finds relevant I would love it but it never does even when asked.
Phind is better about linking sources. I’ve found that generated code sometimes points me in the right direction, but other times it leads me down a rabbit hole of obsolete syntax or other problems.
Ironically, if you already are familiar with the code then you can easily tell where the LLM went wrong and adapt their generated code.
But I don’t use it much because its almost more trouble than its worth.
- issues with model training sources
- business sending their whole codebase to third party (copilot etc.) instead of local models
- time gain is not that substantial in most case, as the actual “writing code” part is not the part that takes most time, thinking and checking it is
- “chatting” in natural language to describe something that have a precise spec is less efficient than just writing code for most tasks as long as you’re half-competent. We’ve known that since customer/developer meetings have existed.
- the dev have to actually be competent enough to review the changes/output. In a way, “peer reviewing” becomes mandatory; it’s long, can be fastidious, and generated code really needs to be double checked at every corner (talking from experience here; even a generated one-liner can have issues)
- some business thinking that LLM outputs are “good enough”, firing/moving away people that can actually do said review, leading to more issues down the line
- actual debugging of non-trivial problems ends up sending me in a lot of directions, getting a useful output is unreliable at best
- making new things will sometimes confuse LLM, making them a time loss at best, and producing even worst code sometimes
- using code chatbot to help with common, menial tasks is irrelevant, as these tasks have already been done and sort of “optimized out” in library and reusable code. At best you could pull some of this in your own codebase, making it worst to maintain in the long term
Those are the downside I can think of on the top of my head, for having used AI coding assistance (mostly local solutions for privacy reasons). There are upsides too:
- sometimes, it does produce useful output in which I only have to edit a few parts to make it works
- local autocomplete is sometimes almost as useful as the regular contextual autocomplete
- the chatbot turning short code into longer “natural language” explanations can sometimes act as a rubber duck in aiding for debugging
Note the “sometimes”. I don’t have actual numbers because tracking that would be like, hell, but the times it does something actually impressive are rare enough that I still bother my coworker with it when it happens. For most of the downside, it’s not even a matter of the tool becoming better, it’s the usefulness to begin with that’s uncertain. It does, however, come at a large cost (money, privacy in some cases, time, and apparently ecological too) that is not at all outweighed by the rare “gains”.
a lot of your issues are effeciency related which i think can realistically be solved given some time for development cycles to take hold on ai. if they were better all around to whatever standard you think is sufficiently useful, would you then think it would be useful? the other side related thing too is that if it can get that level of competence in coding then it most likely can get just as competant in a variety of other domains too.
The point is, they don’t get “competent”. They get better at assembling pieces they were given. And a proper stack with competent developers will already have moved that redundancy out of the codebase. For whatever remains, thinking is the longest part. And LLM can’t improve that once the problem gets a tiny bit complex. Of course, I could end up having a good rough idea of what the code should look like, describe that to an LLM, and have it write actual code with proper variable names and all, but once I reach the point I can describe accurately the thing I want, it’s usually as fast to type it. With the added value that it’s easier to double check.
What remains is providing good insight on new things, and understanding complex requirements. While there is room for improvement, it seems more and more obvious that LLM are not the answer: theoretically, they are not the right tool, and seeing the various level of improvements we’re seeing, they definitely did not prove us wrong. The technology is good at some things, but not at getting “competent”.
Also, you sweep out the privacy and licensing issues, which are big no-no too.
LLM have their uses, I outline some. And in these uses, there are clear rooms for improvements. For reference, the solution I currently use puts me at accepting around 10% of the automatic suggestions. Out of these, I’d say a third needs reworking. Obviously if that moved up to like, 90% suggestions that seems decent and with less need to fix them afterward, it’d be great. Unfortunately, since you can’t trust these, you would still have to review the output carefully, making the whole operation probably not that big of a time saver anyway.
Coding doesn’t allow much leeway. Other activities which allow more leeway for mistakes can probably benefit a lot more. Translation, for example, can be acceptable, in particular because some mishaps may automatically be corrected by readers/listeners. But with code, any single mistake will lead to issues down the way.
A lot of the criticism comes with AI results being wrong a lot of the time, while sounding convincingly correct. In software, things that appear to be correct but are subtly wrong leads to errors that can be difficult to decipher.
Imagine that your AI was trained on StackOverflow results. It learns from the questions as well as the answers, but the questions will often include snippets of code that just don’t work.
The workflow of using AI resembles something like the relationship between a junior and senior developer. The junior/AI generates code from a spec/prompt, and then the senior/prompter inspects the code for errors. If we remove the junior from the equation to replace with AI, then entry level developer jobs are slashed, and at the same time people aren’t getting the experience required to get to the senior level.
Generally speaking, programmers like to program (many do it just for fun), and many dislike review. AI removes the programming from the equation in favour of review.
Another argument would be that if I generate code that I have to take time to review and figure out what might be wrong with it, it might just be quicker and easier to write it correctly the first time
Business often doesn’t understand these subtleties. There’s a ton of money being shovelled into AI right now. Not only for developing new models, but for marketing AI as a solution to business problems. A greedy executive that’s only looking at the bottom line and doesn’t understand the solution might be eager to implement AI in order to cut jobs. Everyone suffers when jobs are eliminated this way, and the product rarely improves.
Generally speaking, programmers like to program (many do it just for fun), and many dislike review. AI removes the programming from the equation in favour of review.
This really resonated with me and is an excellent point. I’m going to have to remember that one.
A developer who is afraid of peer review is not a developer at all imo, but more or less an artist who fears exposing how the sausage was made.
I’m not saying a junior who is nervous is not a dev, I’m talking about someone who has been at this for some time, and still can’t handle feedback productively.
They’re saying developers dislike having to review other code that’s unfamiliar to them, not having their code reviewed.
As am I, it’s a two way street. You need to review the code, and have it reviewed.
Nowhere did you say this.
I did, and I stand by what I said.
Review is both taken and given. Peer review does not occur in a single direction, it is a conversation with multiple parties. I can understand if someone misunderstood what I meant though.
Your reply refers to a “junior who is nervous” and “how the sausage is made”, which makes no sense in the context of someone who just has to review code
When it comes to writing code, there is a huge difference between code that works and code that works *well." Lets say you’re tasked with writing a function that takes an array of RGB values and converts them to grayscale. ChatGPT is probably going to give you two nested loops that iterate over the X and Y values, applying a grayscale transformation to each pixel. This will get the job done, but it’s slow, inefficient, and generally not well-suited for production code. An experienced programmer is going to take into account possible edge cases (what if a color is out of the 0-255 bounds), apply SIMD functions and parallel algorithms, factor in memory management (do we need a new array or can we write back to the input array), etc.
ChatGPT is great for experienced programmers to get new ideas; I use it as a modern version of “rubber ducky” debugging. The problem is that corporations think that LLMs can replace experienced programmers, and that’s just not true. Sure, ChatGPT can produce code that “works,” but it will fail at edge cases and will generally be inefficient and slow.
Exactly. LLMs may replace interns and junior devs, they won’t replace senior devs. And if we replace all of the interns and junior devs, who is going to become the next senior devs?
As a senior dev, a lot of my time is spent reviewing others’ code, doing pair-programming, etc. Maybe in 5-10 years, I could replace a lot of what they do with an LLM, but then where would my replacement come from? That’s not a great long-term direction, and it’s part of how we ended up with COBOL devs making tons of money because financial institutions are too scared to port it to something more marketable.
When I use LLMs, it’s like you said, to get hints as to what options I have. I know it’s sampling from a bunch of existing codebases, so having the LLM go figure out what’s similar can help. But if I ask the LLM to actually generate code, it’s almost always complete garbage unless it’s really basic structure or something (i.e. generate a basic web server using <framework>), but even in those cases, I’d probably just copy/paste from the relevant project’s examples in the docs.
That said, if I had to use an LLM to generate code for me, I’d draw the line at tests. I think unit tests should be hand-written so we at least know the behavior is correct given certain inputs. I see people talking about automating unit tests, and I think that’s extremely dangerous and akin to “snapshot” tests, which I find almost entirely useless, outside of ensuring schemas for externally-facing APIs are consistent.
deleted by creator
Personally, I’ve found AI is wrong about 80% of the time for questions I ask it.
It’s essentially just a search engine with cleverbot. If the problem you’re dealing with is esoteric and therefore not easily searchable, AI won’t fare any better.
I think AI would be a lot more useful if it gave a percentage indicating how confident it is in its answers, too. It’s very useless to have it constantly give wrong information as though it is correct.
People who use LLMs to write code incorrectly perceived their code to be more secure than code written by expert humans.
Lol.
We literally had an applicant use AI in an interview, failed the same step twice, and at the end we asked how confident they were in their code and they said “100%” (we were hoping they’d say they want time to write tests). Oh, and my coworker and I each found two different bugs just by reading the code. That candidate didn’t move on to the next round. We’ve had applicants write buggy code, but they at least said they’d want to write some test before they were confident, and they didn’t use AI at all.
I thought that was just a one-off, it’s sad if it’s actually more common.
OP was able to write a bash script that works… on his machine 🤷 that’s far from having to review and send code to production either in FOSS or private development.
I also noticed that they were talking about sending arguments to a custom function? That’s like a day-one lesson if you already program. But this was something they couldn’t find in regular search?
Maybe I misunderstood something.
Exactly. If you understand that functions are just commands, then it’s quite easy to extrapolate how to pass arguments to that function:
function my_func () { echo $1 $2 $3 # prints a b c } my_func a b cOnce you understand that core concept, a lot of Bash makes way more sense. Oh, and most of the syntax I provided above is completely unnecessary, because Bash…
Hmm, I’m having trouble understanding the syntax of your statement.
Is it
(People who use LLMs to write code incorrectly) (perceived their code to be more secure) (than code written by expert humans.)Or is it
(People who use LLMs to write code) (incorrectly perceived their code to be more secure) (than code written by expert humans.)I intended B, but A is also true, no?
The “statement” was taken from the study.
We conduct the first large-scale user study examining how users interact with an AI Code assistant to solve a variety of security related tasks across different programming languages. Overall, we find that participants who had access to an AI assistant based on OpenAI’s codex-davinci-002 model wrote significantly less secure code than those without access. Additionally, participants with access to an AI assistant were more likely to believe they wrote secure code than those without access to the AI assistant. Furthermore, we find that participants who trusted the AI less and engaged more with the language and format of their prompts (e.g. re-phrasing, adjusting temperature) provided code with fewer security vulnerabilities. Finally, in order to better inform the design of future AI-based Code assistants, we provide an in-depth analysis of participants’ language and interaction behavior, as well as release our user interface as an instrument to conduct similar studies in the future.
If you’re a seasoned developer who’s using it to boilerplate / template something and you’re confident you can go in after it and fix anything wrong with it, it’s fine.
The problem is it’s used often by beginners or people who aren’t experienced in whatever language they’re writing, to the point that they won’t even understand what’s wrong with it.
If you’re trying to learn to code or code in a new language, would you try to learn from somebody who has only half a clue what he’s doing and will confidently tell you things that are objectively wrong? Thats much worse than just learning to do it properly yourself.
I’m a seasoned dev and I was at a launch event when an edge case failure reared its head.
In less than a half an hour after pulling out my laptop to fix it myself, I’d used Cursor + Claude 3.5 Sonnet to:
- Automatically add logging statements to help identify where the issue was occurring
- Told it the issue once identified and had it update with a fix
- Had it remove the logging statements, and pushed the update
I never typed a single line of code and never left the chat box.
My job is increasingly becoming Henry Ford drawing the ‘X’ and not sitting on the assembly line, and I’m all for it.
And this would only have been possible in just the last few months.
We’re already well past the scaffolding stage. That’s old news.
Developing has never been easier or more plain old fun, and it’s getting better literally by the week.
Edit: I agree about junior devs not blindly trusting them though. They don’t yet know where to draw the X.
Edit: I agree about junior devs not blindly trusting them though. They don’t yet know where to draw the X.
The problem (one of the problems) is that people do lean too heavily on the AI tools when they’re inexperienced and never learn for themselves “where to draw the X”.
If I’m hiring a dev for my team, I want them to be able to think for themselves, and not be completely reliant on some LLM or other crutch.
As a cybersecurity guy, it’s things like this study, which said:
Overall, we find that participants who had access to an AI assistant based on OpenAI’s codex-davinci-002 model wrote significantly less secure code than those without access. Additionally, participants with access to an AI assistant were more likely to believe they wrote secure code than those without access to the AI assistant.
FWIW, at this point, that study would be horribly outdated. It was done in 2022, which means it probably took place in early 2022 or 2021. The models used for coding have come a long way since then, the study would essentially have to be redone on current models to see if that’s still the case.
The people’s perceptions have probably not changed, but if the code is actually insecure would need to be reassessed
Sure, but to me that means the latest information is that AI assistants help produce insecure code. If someone wants to perform a study with more recent models to show that’s no longer the case, I’ll revisit my opinion. Until then, I’m assuming that the study holds true. We can’t do security based on “it’s probably fine now.”
I think it’s more appalling because they should have assumed this was early tech and therefore less trustworthy. If anything, I’d expect more people to believe their code is secure today using AI than back in 2021/2022 because the tech is that much more mature.
I’m guessing an LLM will make a lot of noob mistakes, especially in languages like C(++) where a lot of care needs to be taken for memory safety. LLMs don’t understand code, they just look at a lot of samples of existing code, and a lot of code available on the internet is terrible from a security and performance perspective. If you’re writing it yourself, hopefully you’ve been through enough code reviews to catch the more common mistakes.
- AI Code suggestions will guide you to making less secure code, not to mention often being lower quality in other ways.
- AI code is designed to look like it fits, not be correct. Sometimes it is correct. Sometimes it’s close but has small errors. Sometimes it looks right but is significantly wrong. Personally I’ve never gotten ChatGPT to write code without significant errors for more than trivially small test cases.
- You aren’t learning as much when you have ChatGPT do it for you, and what you do learn is “this is what chat gpt did and it worked last time” and not “this is what the problem is and last time this is the solution I came up with and this is why that worked”. In the second case you are far better equipped to tackle future problems, which won’t be exactly the same.
All that being said, I do think there is a place for chat GPT in simple queries like asking about syntax for a language you don’t know. But take every answer it gives you with a grain of salt. And if you can find documentation I’d trust that a lot more.
Yes, I completely forget how to solve that problem 5 minutes after chatGPT writes its solution. So I whole heartedely believe AI is bad for learning
All that being said, I do think there is a place for chat GPT in simple queries like asking about syntax for a language you don’t know.
I am also weary regarding AI and coding but this is actually the first time I used ChatGpt to programm something for a small home project in python, since I never used it. I was positively surprised in how much it could help me getting started. I also learned quite a bit since I always asked for comparison with Java, which I know, and for reasonings why it is that way. I simply also wanted to understand what it puts out. I also only asked for single lines of code rather than generating a whole method, e.g. I want to move a file from X to Y.
The thought of people blindly copying the produced code scares me.
AI Code suggestions will guide you to making less secure code, not to mention often being lower quality in other ways.
This is a PR post from a company selling software.
I use ai, but whenever I do I have to modify it, whether it’s because it gives me errors, is slow, doesn’t fit my current implementation or is going off the wrong foot.
I have a coworker who is essentially building a custom program in Sheets using AppScript, and has been using CGPT/Gemini the whole way.
While this person has a basic grasp of the fundamentals, there’s a lot of missing information that gets filled in by the bots. Ultimately after enough fiddling, it will spit out usable code that works how it’s supposed to, but honestly it ends up taking significantly longer to guide the bot into making just the right solution for a given problem. Not to mention the code is just a mess - even though it works there’s no real consistency since it’s built across prompts.
I’m confident that in this case and likely in plenty of other cases like it, the amount of time it takes to learn how to ask the bot the right questions in totality would be better spent just reading the documentation for whatever language is being used. At that point it might be worth it to spit out simple code that can be easily debugged.
Ultimately, it just feels like you’re offloading complexity from one layer to the next, and in so doing quickly acquiring tech debt.
Exactly my experience as well. Using AI will take about the same amount of time as just doing it myself, but at least I’ll understand the code at the end if I do it myself. Even if AI was a little faster to get working code, writing it yourself will pay off in debugging later.
And honestly, I enjoy writing code more than chatting with a bot. So if the time spent is going to be similar, I’m going to lean toward DIY every time.
For me it’s because if the AI does all the work the person “coding” won’t learn anything. Thus when a problem does arise (i.e. the AI not being able to fix a simple mistake it made) no one involved has the means of fixing it.
I have seen my friend in this situation
But I don’t want to learn. I want the machine to free me from tedious tasks I already know how to do. There’s no learning experience in creating a Wordpress plugin or a shell script.
