this post was submitted on 07 Oct 2023
74 points (100.0% liked)
Science
374 readers
1 users here now
Subscribe to see new publications and popular science coverage of current research on your homepage
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
a test for creativity seriously that work? also after scraping the entire of internet of course someone could think that, ask any programmer and they gonna explain that the IA don't create anything, it can't even do basic msth because it don't gave logic in that,maybe one day, but not with chatgpt of today
Yes, a test for creativity. If you're going to say something "doesn't have the power of creativity" then it behooves you to accept the notion that creativity is measurable.
This is the key - it does not create, it can only copy. Which is good enough to fool us - there's enough stuff to copy out there that you can spend your whole life copying other people and nobody will ever notice you're not actually creating anything new. What's more, you'll probably come across as pretty clever. But you're not creating anything new.
For me, this poses an existential threat to academia. It might halt development in the field without researchers even noticing: Their words look fine, as if they had thought it through, and they of course read it to make sure it's logically consistent. However, the creative force is gone. Nothing new will come under the sun - the kind of new thoughts that can only be made by creative humans thinking new thoughts that have never been put on paper before.
If we give up that, what's even the point of doing science in the first place.
There's a difference between:
Using ChatGPT to help write parts of the text in the same way you'd use a grammar- or spell-checker (e.g. if English isn't your first language) after you've finished the experiments
Using ChatGPT to write a paper without even doing any experiments
Clearly the second is academic misconduct. The first one is a lot more defensible.
Yes, absolutely. But I still think it has its dangers.
Using it to write the introduction doesn't change the substance of the paper, yet it does provide the framework for how the reader interprets it, and also often decides whether it'll be read at all.
Maybe worse, I find that it's oftem in the painful writing and rewriting of the introduction and conclusion that I truly understand my own contribution - I've done the analysis and all that, but in forcing myself to think about the relevance for the field and the reader I also bring myself to better understand what the paper means in a deeper sense. I believe this kind of deep thinking at the end of the process is incredibly valuable, and it's what I'm afraid we might be losing with AI.
I have asked ChatGPT to write poetry on subjects that I know with great certainty have never had poems written about them.
You can of course shuffle around the meanings of "create" and "copy" to try to accommodate that, but eventually you end up with a "copying" process that's so flexible and malleable that it might as well be creativity. It's not like what comes out of human brains isn't based on stuff that went into them earlier either.
@Vilian @FaceDeer I agree. I’m no programmer but do a fair bit of Linux/powershell/bash scripting. Virtually all the code that ChatGPT gives me is wrong. You tell it the errors, and it gives you a modified script with errors, point out those errors and it’s go back to its first answer. The only thing it is useful for is writing lots of basic code, really quickly. I can just copy/paste then start debugging.
I am a programmer and I've found ChatGPT to be able to produce plenty of good, useful code. I haven't encountered the problems you're describing in correcting its errors, perhaps you're not prompting it well.
@FaceDeer @floofloof @henfredemars @PoisonedPrisonPanda @sab @Vilian nah, it was trained in 2021 and parrots 10 year old stack overflow pages. That may have worked a decade ago, but stuff has moved on since then. It still spits out code using AzureAD cmdlets as it doesn’t know MSGraph replace a lot of it in last couple of years. I guess it could be ok if you’re on a legacy tech stack though.
So you're telling me that the code it generated for me wasn't good and useful, and that when I told it to correct errors it actually did introduce new errors and restore old ones, contrary to what I just said? Guess all that stuff I got done using its help didn't actually get done after all and I'm descending ever deeper into a world of delusion, thinking my projects are finished and working when in fact they aren't.
Obviously if you're trying to get it to use APIs from after 2021 that's not going to work. It also won't bake you a cake if you ask it to. Use tools for the tasks they're good for, don't use them for things you know they can't do.
@FaceDeer @floofloof @henfredemars @PoisonedPrisonPanda @sab @Vilian the 500 char limit made me pick the first failure that spring to mind. Maybe you forget, but AI wasn’t trained on “good” data. It was trained “all” data and large amounts of that is plain wrong. People with problems pasting blocks of code and responses correcting a single line. ChatGPT isn’t smart enough to merge those into a single block of working code.
You have to tell ChatGPT that you want good code, then.
I'm actually serious. If you just ask for something generic, it'll assume you want something generic. If you ask it for something that's "high efficiency, well commented and maintainable" then it's going to know you wanted that and give you something more along those lines. Just like if you asked it for something "that looks crappy and sloppy, like an amateur wrote it."
Very often when people complain about ChatGPT's "style" or say they can immediately spot something that "sounds like" ChatGPT it's because they're not giving it good directions. It can't read your mind. Yet.
@FaceDeer @floofloof @henfredemars @PoisonedPrisonPanda @sab @Vilian nope. I ask for highly precise stuff. When I say, “I’m no coder,” it’s coz I use interpreters and don’t compile “real” code, plus it only accounts for 10% of my day job. ChatGPT maybe useful for Hello World, Towers of Babel or other stuff it scraped from udemy, but when you ask it to assist in automating complex production systems, it really falls down.
Maybe ChatGPT just hates you personally, then.
You're saying "it can't work for anyone because it doesn't work for me!" And I'm saying "well, it worked for me, so maybe you're using it wrong."
You can't insist it's not working for me because it did. I'm not disputing that it didn't work for you, all I can suggest is reading up a bit on prompt engineering to see if you can find out what you're doing differently.
@FaceDeer @floofloof @henfredemars @PoisonedPrisonPanda @sab @Vilian I’m not saying it didn’t work for you. I’m saying it’s only good for entry level stuff… It’s not coming for my job yet, but I can seriously see it stealing work from the overseas, underpaid, code shops in China/India.
I am not an "entry level" coder.
I also am not concerned with it "coming for my job." It's a collaborator with humans, not a replacement for them.