this post was submitted on 25 Jul 2023
346 points (99.7% liked)
Technology
37757 readers
572 users here now
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
Subcommunities on Beehaw:
This community's icon was made by Aaron Schneider, under the CC-BY-NC-SA 4.0 license.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
ChatGPT has no knowledge of the answers it gives. It is simply a text completion algorithm. It is fundamentally the same as the thing above your phone keyboard that suggests words as you type, just with much more training data.
Who cares? It still gives me the answers i am looking for.
Yeah it gives you the answers you ask it to give you. It doesn't matter if they are true or not, only if they look like the thing you're looking for.
An incorrect answer can still be valuable. It can give some hint of where to look next.
@magic_lobster_party I can't believe someone wrote that. Incorrect answers do more harm than being useful. If the person asks and don't know, how should he or she know it's incorrect and look for a hint?
In the context of coding it can be valuable. I produced two tables in a database and asked it to write a query and it did 90% of the job. It was using an incorrect column for a join. If you are doing it for coding you should notice very quickly what is wrong at least if you have experience.
I don't know about others' experiences, but I've been completely stuck on problems I only figured out how to solve with chatGPT. It's very forgiving when I don't know the name of something I'm trying to do or don't know how to phrase it well, so even if the actual answer is wrong it gives me somewhere to start and clues me in to the terminology to use.
Google the provided solution for additional sources. Often when I search for solutions to problems I don’t get the right answer directly. Often the provided solution may not even work for me.
But I might find other clues of the problem which can aid me in further research. In the end I finally have all the clues I need to find the answer to my question.
How do you Google anything when all the results are AI generated crap for generating ad revenue?
Well then I guess I have to survive with ChatGPT if the internet is so riddled with search engine optimized garbage. We’re thankfully not there yet, at least not with computer tech questions.
In my experience, with both coding and natural sciences, a slightly incorrect answer that you attempt to apply, realize is wrong in some way during initial testing/analysis, then you tweak until it's correct, is very useful, especially compared to not receiving any answer or being ridiculed by internet randos.
Well if they refer to coding solution they’re right : sometimes non-working code can lead to a working solution. if you know what you’re doing ofc
Even if you don't know what you're doing ChatGPT can still do well if you tell it what went wrong with the suggestion it gave you. It can debug its code or realize that it made wrong assumptions about what you were asking from further context.
How is that practically different from a user perspective than answers on SO? Either way, I still have to try the suggested solutions to see if they work in my particular situation.
At least with those, you can be reasonably confident that a single person at some point believed in their answer as a coherent solution
That doesn't exactly inspire confidence.
Better than knowing there's some possibility that the answer was generated purely because the sequence of characters had the highest probability of convincing the reader that it seems correct based on the sequence of characters it was given as input (+/- a decent amount of RNG)
Still debatable, IMO. Human belief is stubborn and self-justifying whereas an RNG can be rerolled as many times as needed.
Yeah but if you keep rerolling the RNG, how do you know when a right answer gets randomly generated?
Also, my point above was that if a human believed the solution was true, it probably was true at some point. With generative language models, there's no guarantee that there's any logic to what it tells you.
You know when the code compiles and does what you want it to do. What's the point in asking for code if you're not going to run it? You'd be doing that with anything you got off of Stack Overflow too, presumably.
What point are you trying to make? LLMs are incredibly useful tools
Yeah for generating prose, not for solving technical problems.
You’ve never actually used them properly then.
One example is writing complex regex. A simple well written prompt can get you 90% the way there. It's a huge time saver.
It's great a writing boilerplate code so I can spend more of my time architecturing solutions instead of typing.
the good thing if it gives you the answer in a programming language is that its quite simple tontestvif the output is what you expect, also a lot of humans hive wrong answers...
There was a story once that said if you put an infinite number of monkeys in front of an infinite number of typewriters, they would eventually produce the works of William Shakespeare.
So far, the Internet has not shown that to be true. Example: Twitter.
Now we have an artificial monkey remixing all of that, at our request, and we're trying to find something resembling Hamlet's Soliloquy in what it tells us. What it gives you is meaningless unless you interpret it in a way that works for you -- how do you know the answer is correct if you don't test it? In other words, you have to ensure the answers it gives are what you are looking for.
In that scenario, it's just a big expensive rubber duck you are using to debug your work.
There's a bunch of people telling you "ChatGPT helps me when I have coding problems." And you're responding "No it doesn't."
Your analogy is eloquent and easy to grasp and also wrong.
Fair point, and thank you. Let me clarify a bit.
It wasn't my intention to say ChatGPT isn't helpful. I've heard stories of people using it to great effect, but I've also heard stories of people who had it return the same non-solutions they had already found and dismissed. Just like any tool, actually...
I was just pointing out that it is functionally similar to scanning SO, tech docs, Slashdot, Reddit, and other sources looking for an answer to our question. ChatGPT doesn't have a magical source of knowledge that we collectively also do not have -- it just has speed and a lot processing power. We all still have to verify the answers it gives, just like we would anything from SO.
My last sentence was rushed, not 100% accurate, and shows some of my prejudices about ChatGPT. I think ChatGPT works best when it is treated like a rubber duck -- give it your problem, ask it for input, but then use that as a prompt to spur your own learning and further discovery. Don't use it to replace your own thinking and learning.
Even if ChatGPT is giving exactly the same quality of answer as you can get out of Stack Overflow, it gives it to you much more quickly and pieces together multiple answers into a script you can copy and work with immediately. And it's polite when doing so, and will amend and refine its answers immediately for you if you engage it in some back-and-forth dialogue. That makes it better than Stack Overflow and not functionally similar.
I've done plenty of rubber duck programming before, and it's nothing like working with ChatGPT. The rubber duck never writes code for me. It never gives me new information that I didn't already know. Even though sometimes the information ChatGPT gives me is wrong, that's still far better than just mutely staring back at me like a rubber duck does. A rubber duck teaches me nothing.
"Verifying" the answer given by ChatGPT can be as simple as just going ahead and running it. I can't think of anything simpler than that, you're going to have to run the code eventually anyway. Even if I was the world's greatest expert on something, if I wrote some code to do a thing I would then run it to see if it worked rather than just pushing it to master and expecting everything to be fine.
This doesn't "replace your own thinking and learning" any more than copying and pasting a bit of code out of Stack Overflow does. Indeed, it's much easier to learn from ChatGPT because you can ask it "what does that line with the angle brackets do?" or "Could you add some comments to the loop explaining all the steps" or whatever and it'll immediately comply.