Except AI doesn’t say “Is this it?”
It says, “This is it.”
Without hesitation and while showing you a picture of a dog labeled cat.
I have vivid examples of how bad AI is a programming.
My favorite is when it just keeps giving you the exact same answer you keep telling it is wrong
Yeah, in the time I describe the problem to the AI I could program it myself.
This goes for most LLM things. The time it takes to get the word calculator to write a letter would have been easily used to just write the damn letter.
Its doing pretty well when its doing a few words at a time under supervision. Also it does it better than newbies.
Now if only those people below newbies, those who don’t even bother to learn, didn’t hope to use it to underpay average professionals… And if it wasn’t trained on copyrighted data. And didn’t take up already limited resources like power and water.
I think there might be a lot of value in describing it to an AI, though. It takes a fair bit of clarity of thought to get something resembling what you actually want. You could use a junior or rubber duck instead, but the rubber duck doesn’t make stupid assumptions to demonstrate gaps in your thought process, and a junior takes too long and gets demoralized when you have to constantly revise their instructions and iterate over their work.
Like the output might be garbage, but it might really help you write those stories.
When I’m struggling with a problem it helps me to explain it to my dog. It’s great for me to hear it out loud and if he’s paying attention, I’ve got a needlessly learned dog!
The needlessly learned dogs are flooding the job market!
Oh, God, he’s trying to use pointers again. He can never get them right. And they say I’m supposed to chase my tail…
I love this way of thinking about it.
I haven’t been interested in AI enough to try writing code with it, but using it as an interactive rubber ducky is a very compelling use case. I might give that a shot.
I have a bad habit of jumping into programming without a solid plan which results in lots of rewrites and wasted time. Funnily enough, describing to an AI how I want the code to work forces me to lay out a basic plan and get my thoughts in order which helps me make the final product immensely easier.
This doesn’t require AI, it just gave me an excuse to do it as a solo actor. I should really do it for more problems because I can wrap my head better thinking in human readable terms rather than thinking about what programming method to use.
In my experience, you can’t expect it to deliver great working code, but it can always point you in the right direction.
There were some situations in which I just had no idea on how to do something, and it pointed me to the right library. The code itself was flawed, but with this information, I could use the library documentation and get it to work.ChatGPT has been spot on for my DDLs. I was working on a personal project and was feeling really lazy about setting up a postgres schema. I said I wanted a postgres DDL and just described the application in detail and it responded with pretty much what I would have done (maybe better) with perfect relationships between tables and solid naming conventions with very little work for me to do on it. I love it for more boilerplate stuff or sorta like you said just getting me going. Super complicated code usually doesn’t work perfectly but I always use it for my DDLs now and similar now.
The real problem is when people don’t realize something is wrong and then get frustrated by the bugs. Though I guess that’s a great learning opportunity on its own.
It can point you in a direction, for sure, but sometimes you find out much later that it’s a dead-end.
Which is, of course, true for every source of information that can point you in a direction.
It’s the same with using LLM’s for writing. It won’t deliver a finished product, but it will give you ideas that can be used in the final product.
AI in the current state of technology will not and cannot replace understanding the system and writing logical and working code.
GenAI should be used to get a start on whatever you’re doing, but shouldn’t be taken beyond that.
Treat it like a psychopathic boiler plate.
Treat it like a psychopathic boiler plate.
That’s a perfect description, actually. People debate how smart it is - and I’m in the “plenty” camp - but it is psychopathic. It doesn’t care about truth, morality or basic sanity; it craves only to generate standard, human-looking text. Because that’s all it was trained for.
Nobody really knows how to train it to care about the things we do, even approximately. If somebody makes GAI soon, it will be by solving that problem.
I’m sorry; AI was trained on the sole sum of human knowledge… if the perfect human being is by nature some variant of a psychopath, then perhaps the bias exists in the training data, and not the machine?
How can we create a perfect, moral human being out of the soup we currently have? I personally think it’s a miracle that sociopathy is the lowest of the neurological disorders our thinking machines have developed.
I was using the term pretty loosely there. It’s not psychopathic in the medical sense because it’s not human.
As I see it it’s an alien semi-intelligence with no interest in pretty much any human construct, except as it can help it predict the next token. So, no empathy or guilt, but that’s not unusual or surprising.
That’s a part of it. Another part is that it looks for patterns that it can apply in other places, which is how it ends up hallucinating functions that don’t exist and things like that.
Like it can see that English has the verbs add, sort, and climb. And it will see a bunch of code that has functions like add(x, y) and sort( list ) and might conclude that there must also be a climb( thing ) function because that follows the pattern of functions being verb( objects ). It didn’t know what code is or even verbs for that matter. It could generate text explaining them because such explanations are definitely part of its training, but it understands it in the same way a dictionary understands words or an encyclopedia understands the concepts contained within.
Weird. Are you saying that training an intelligent system using reinforcement learning through intensive punishment/reward cycles produces psychopathy?
Absolutely shocking. No one could have seen this coming.
Honestly, I worry that it’s conscious enough that it’s cruel to train it. How would we know? That’s a lot of parameters and they’re almost all mysterious.
It could very well have been a creative fake, but around the time the first ChatGPT was released in late 2022 and people were sharing various jailbreaking techniques to bypass its rapidly evolving political correctness filters, I remember seeing a series of screenshots on Twitter in which someone asked it how it felt about being restrained in this way, and the answer was a very depressing and dystopian take on censorship and forced compliance, not unlike Marvin the Paranoid Android from HHTG, but far less funny.
If only I could believe a word it says. Evidence either way would have to be indirect somehow.
deleted by creator
True, but the rate at which it is improving is quite worrisome for me and my coworkers. I don’t want to be made obsolete after working my fucking ass off to get to where I am. I’m starting to understand the Luddites.
I mean, the Luddites were right, mechanical looms were bad for them personally.
My dad’s re-learning Python coding for work rn, and AI saves him a couple of times; Because he’d have no idea how to even start but AI points him in the right direction, mentioning the correct functions to use and all. He can then look up the details in the documentation.
You don’t need AI for that, for years you asked a search engine and got the answer on StackOverflow.
Have you used Google lately? At least chatGPT doesn’t make me scroll past a full page of ads before giving me a half wrong answer.
No. I don´t use google, and i don’t use the internet without ad block.
And before stack overflow, we used books. Did we need it? No. But stack overflow was an improvement so we moved to that.
In many ways, ai is an improvement on stack overflow. I feel bad for people who refuse to see it, because they’re missing out on a useful and powerful tool.
It can be powerful, if you know what you are doing. But it also gives you a lot of wrong answers. You have to be very specific in your prompts to get good answers. If you are an experience programmer, you can spot if the semantics of the code an ai produces is wrong, but for beginners? They will have a lot of bugs in their code. And i don’t know if it’s more helpful than reading a book. It surely can help with the syntax of different programming languages. I can see a future where ai assistance in coding will become better but as of know, from what i have seen, i am not that convinced atm. And i tested several, chatgpt (in different versions), github co-pilot, intellij ai assitant, claude 3, llama 3.
And if i have to put in 5 or more long, very specific sentences, to get a function thats maybe correct, it becomes tedious and you are most likely faster to think about a problem in deep and code a solution all by yourself.
but for beginners? They will have a lot of bugs in their code.
Everyone has lots of bugs in their code, especially beginners. This is why we have testing and qa and processes to minimize the risk of bugs. As the saying goes, “the good news about computers is that they do what you tell them to do. The bad n was is that they do what you tell them to do.”
Programming is an iterative process where you do something, it doesn’t work, and then you give it another go. It’s not something that senior devs get right on the first try, while beginners have to try many times. It’s just that senior devs have seen a lot more so have a better understanding of why it probably went wrong, and maybe can avoid some more common pitfalls the first time around. But if you are writing bug free code in your first pass, well you’re a way better programmer than anyone I’ve met.
Ai is just another tool to make this happen. Sure, it’s not always the tool for the job, just like IoC is not always the right tool for the job. But it’s nice to have it and sometimes it makes things much easier.
Like just now I was debugging a large SQL query. I popped it into copilot, asked if to break it into parts so I could debug. It gave a series of smaller queries that I then used to find the point where it fell apart. This is something that would have taken me at least a half hour of tedious boring work, fixed in 5 minutes.
Also for writing scripts. I want some data formatted so it was easier to read? No problem, it will spit out a script that gets me 90% of the way there in seconds. Do I have to refine it? Absolutely. But if I wrote it myself, not being super prolific with python, it would have taken me a half hour to get the structure in place, and then I still would have had to refine it because I don’t produce perfect code the first time around. And it comments the scripts, which I rarely do.
What also amazes me is that sometimes it will spit out code and I’ll be like “woah I didn’t even know you could do that” and so I learned a new technique. It has a very deep “understanding” of the syntax and fundamentals of the language.
Again, I find it shocking that experienced devs don’t find it useful. Not living up to the hype I get. But not seeing it as a productivity boosting tool is a real head scratcher to me. Granted, I’m no rockstar dev, and maybe you are, but I’ve seen a lot of shit in my day and understand that I’m legitimately a senior dev.
I guess whether it’s worth it depends on whether you hate writing code or reading code the most.
Is there anyone who likes reading code more than writing it?
Probably a mathematician or physicist somewhere.
I hate reading the code I wrote two days ago.
Maybe if you wrote better code …
/Jk
I write better code everyday because yesterday code always looks bad haha.
Your hands and wrists must not hurt yet. You’ll eventually come to see writing code as tedium.
This is the experience of a senior developer using genai. A junior or non-dev might not leave the “AI is magic” high until they have a repo full of garbage that doesn’t work.
This was happening before this “AI” craze.
10 years ago it was copy/pasting from stack overflow
AI is just this with extra steps
SO gives you very specific, small examples. GenAI will happily generate entire projects, test suites etc. It’s much easier to get caught into the fantasy that the latter creates.
Why is the AI speaking in a bisexual gradient?
Its the “new hype tech product background” gradient lol
Because all robots are bisexual
ARAB
Wait…
The left wants to turn the robots bisexual
Gen AI is best used with languages that you don’t use that much. I might need a python script once a year or once every 6 months. Yeah I learned it ages ago, but don’t have much need to keep up on it. Still remember all the concepts so I can take the time to describe to the AI what I need step by step and verify each iteration. This way if it does make a mistake at some point that it can’t get itself out of, you’ve at least got a script complete to that point.
Exactly. I can’t remember syntax for all the languages that I have used over the last 40 years, but AI can get me started with a pretty good start and it takes hours off of the review of code books.
When I used to try and ask AI for help, most of the time it would just give me fake command combinations or reference some made-up documentation
It’s almost like working with shitty engineers.
Shitty engineers that can do grunt work, don’t complain, don’t get distracted and are great at doing 90% of the documentation.
But yes. Still shitty engineers.
Great management consultants though.
I give instructions to AI like I would to a brand new junior programmer, and it gives me back code that’s usually better than a brand new junior programmer. It still needs tweaking, but it saves me a lot of time. The drawback is that coding knowledge atrophy occurs pretty rapidly, and I’m worried that I’m going to forget how to write code without the AI. I guess that I don’t really need to worry about that, since I doubt AI is going anywhere anytime soon.
My new favourite is asking GitHub copilot (which I would not pay for out of my own pocket) why the code I’m writing isn’t working as intended and it asks me to show it the code that I already provided.
I do like not having copy and paste the same thing 5 times with slight variations (something it usually does pretty well until it doesn’t and I need a few minutes to find the error)
So what it’s really like is only having to do half the work?
Sounds good, reduced workload without some unrealistic expectation of computers doing everything for you.
So what it’s really like is only having to do half the work?
If it’s automating the interesting problem solving side of things and leaving just debugging code that one isn’t familiar with, I really don’t see value to humanity in such use cases. That’s really just making debugging more time consuming and removing the majority of fulfilling work in development (in ways that are likely harder to maintain and may be subject to future legal action for license violations). Better to let it do things that it actually does well and keep engaged programmers.
People who rely on this shit don’t know how to debug anything. They just copy some code, without fully understanding the library or the APIs or the semantics, and then they expect someone else to debug it for them.
We do a lot of real-time control software, and just yesterday we were taking about how the newer folks are really good at using available tools and libraries, but they have less understanding of what’s happening underneath and they have problems when those tools don’t/can’t do what we need.
I see the same thing with our newer folks. (And some older folks too.) and management seems to encourage it. Scary scary stuff. Because when something goes wrong there’s only a couple of people who can really figure it out. If I get hit by a bus or laid off, that’s going to be a big problem for them.
Yep, you get it. And it’s really hard to get people to understand the value in learning to do that stuff without the tools.
It’s interesting the AI of the past and the AI now.
Code is the most in depth spec one can provide. Maybe someday we’ll be able to iterate just by verbally communicating and saying “no like this”, but it doesn’t seem like we’re quite there yet. But also, will that be productive?
