The research from Purdue University, first spotted by news outlet Futurism, was presented earlier this month at the Computer-Human Interaction Conference in Hawaii and looked at 517 programming questions on Stack Overflow that were then fed to ChatGPT.
“Our analysis shows that 52% of ChatGPT answers contain incorrect information and 77% are verbose,” the new study explained. “Nonetheless, our user study participants still preferred ChatGPT answers 35% of the time due to their comprehensiveness and well-articulated language style.”
Disturbingly, programmers in the study didn’t always catch the mistakes being produced by the AI chatbot.
“However, they also overlooked the misinformation in the ChatGPT answers 39% of the time,” according to the study. “This implies the need to counter misinformation in ChatGPT answers to programming questions and raise awareness of the risks associated with seemingly correct answers.”
48% is still better than the Punxsutawney Phil.
“Major new Technology still in Infancy Needs Improvements”
– headline every fucking day
The way I see it, we’re finally sliding down the trough of disillusionment.
I’m honestly a bit jealous of you. You are going to be so amazed when you realise this stuff is just barely getting started. It’s insane what people are already building with agents. Once this stuff gets mainstream, and specialized hardware hits the market, our current paradigm is going to seem like silent black and white films compared to what will be going on. By 2030 we will feel like 2020 was half a century ago at least.
Looking forward to it, but won’t be disappointed if it takes a bit longer than expected.
Ray Kurzweil has a phenomenal record of making predictions. He’s like 90% or something and has been saying AGI by 2029 for something like 30+ years. Last I heard, he is sticking with it, but he admits he may be a year or two off in either direction. AGI is a pretty broad term, but if you take it as “better than nearly every human in every field of expertise,” then I think 2029 is quite reasonable.
That’s not very far in the future, so it’s going to be really exciting to see how that works out.
Maybe only 51% of the code it writes needs to be good before it can self-improve. In which case, we’re nearly there!
We are already past that. The 48% is from a version of chatgpt(3.5) that came out a year ago, there has been lots of progress since then.
unready technology that spews dangerous misinformation in the most convincing way possible is being massively promoted
Yeah, because no human would convincingly lie on the internet. Right, Arthur?
It’s literally built on what confidently incorrect people put on the internet. The only difference is that there are constant disclaimers on it saying it may give incorrect information.
Anyone too stupid to understand how to use it is too stupid to use the internet safely anyways. Or even books for that matter.
Holy mother of false equivalence. Google is not supposed to be a random dude on the Internet, it’s supposed to be a reference tool, and for the most part it was a good one before they started enshittifying it.
Google is a search engine. It points you to web pages that are made by people. Many times, the people who make those websites have put things on them that are knowingly or unknowingly incorrect but said in an authoritative manner. That was all I was saying, nothing controversial. That’s been a known fact for a long time. You can’t just read something on a single site and then be sure that it has to be true. I get that there are people who strangely fall in love with specific websites and think they are absolute truth, but thats always been a foolish way to use the internet.
A great example of people believing blindly is all these horribly doctored google ai images saying ridiculous things. There are so many idiots that think every time they see a screenshot of Google ai saying something absurd that it has to be true. People have even gone so far as to use ridiculous fonts just to point out how easy it is to get people to trust anything. Now there’s a bunch of idiots that think all 20 or so Google ai mistakes they’ve seen are all genuine, so much so that they think almost all Google ai responses are incorrect. Some people are very stupid. Sorry to break it to you, but LLMs are not the first thing to put incorrect information on the internet.
“Will this technology save us from ourselves, or are we just jerking off?”
in Infancy Needs Improvements
I’m just gonna go out on a limb and say that if we have to invest in new energy sources just to make these tools functionably usable… maybe we’re better off just paying people to do these jobs instead of burning the planet to a rocky dead husk to achieve AI?
Just playing devil’s advocate here, but if we could get to a future with algorithms so good they are essentially a talking version of all human knowledge, this would be a great thing for humanity.
We already had that with search engines and the world wide web.
But let’s say some company did it, a perfect AI that has read everything and doesn’t hallucinate.
A researcher is working on some experiments, if they could just route it through the AI, and it would annalyse if that experiment was even possible, maybe already done, this could speed up research.
With a truly perfect model, which the tech bros are aiming for, I can see the potential for good. I ofcourse am skeptical such a model is possible, but… I kinda see why it would be nice to have.
this would be a great thing for humanity.
That’s easy to say. Tell me how. Also tell me how to do it without it being biased about certain subjects over others. Captain Beatty would wildly disagree with this even being possible. His whole shtick in Fahrenheit 451 is that all the books disagreed with one another, so that’s why they started burning them.
There’s this series of books called the www series, about AI before AI was the new hot thing every company needed to mention at least once to get stock price to go up.
Tap for spoiler
Essentially an AI popped up on the internet, which was able to read everything. Due to this it was able to combine data in such a way that it found things like a cure for cancer by combining research papers that no one had ever combined. This is a very bad explanation, but I could see how this makes sense.
Spoiler free explanation: no human has read everything, I think there could be big implications if there’s an AI that has that can see connections that no one ever has.
they are essentially a talking version of all human knowledge
“Obama is a Muslim”
“Corporation using immature technology in productions because it’s cool”
More news at eleven
This is scary because up to now, all software released worked exactly as intended so we need to be extra special careful here.
Yes, and we never have and never will put lives in the hands of software developers before!
Tap for spoiler
/s…for this comment and the above one, for anyone who needs it
Still the same shit study that does not even name the version they used…? The one posted here 1 or 2 days ago?
I’m the footnotes they mention GPT-3.5. Their argument for not testing 4 was because it was paid, and so most users would be using 3.5 - which is already factually incorrect now because the new GPT-4o (which they don’t even mention) is now free. Finally, they didn’t mention GPT-4 Turbo either, which is even better at coding compared to 4.
4 is free for a very small number of queries, then it switches back to 3.5. Or at least that’s what happened to me the other day.
Anyone can use GPT-4 for free. Co-pilot uses GPT-4 and with a Microsoft account you can do up to 30 queries. I’ve used it a lot to create Excel VBA code for work and it’s pretty good. Much better than GPT-3.5 that’s for sure.
Still the same shit study that does not even name the version they used…?
The answer to your question would be evident to you if you had taken the time to read what you are deeming “the same shit study.” The study mentions the version used on multiple occasions:
For each of the 517 SO questions, the first two authors manually used the SO question’s title, body, and tags to form one question prompt and fed that to the free version of ChatGPT, which is based on GPT-3.5.
Additionally, this work has used the free version of ChatGPT (GPT-3.5) for acquiring the ChatGPT responses for the manual analysis.
Hence, for this study, we used the free version (GPT-3.5) so that the results benefit the majority of our target populations.
Please ensure you have read the study before making uninformed remarks.
The one posted here 1 or 2 days ago?
I have already checked for duplicates within this community before posting, and the post you are talking about is not present.
Once again, please ensure your facts are accurate before making incorrect statements.
I just use it to get ideas about how to do something or ask it to write short functions for stuff i wouldnt know that well. I tried using it to create graphical ui for script but that was constant struggle to keep it on track. It managed to create something that kind of worked but it was like trying to hold 2 magnets of opposing polarity together and I had to constantly reset the conversation after it got “corrupted”.
Its useful tool if you dont rely on it, use it correctly and dont trust it too much.
This has been true for code you pull from posts on stackoverflow since forever. There are some good ideas, but they a. Aren’t exactly what you are trying to solve and b. Some of the ideas are incomplete or just bad and it is up to you to sort the wheat from the chaff.
I couldn’t have said it better
Anyone else tired of these clickbait headlines and studies about LLM which center around fundamental misunderstandings of how LLMs work, or is it just me?
“ChatGPT didn’t get a single answer on my algebra exam correct!!” Well yes, because LLMs work on predictive generation, not traditional calculation, so of course they’re not going to do math or anything else with non-language-based patterns properly. That’s what a calculator is for.
All of these articles are like complaining that a chainsaw is an inefficient tool for driving nails into wood. Yeah; because that’s not the job this tool was made for.
And it’s so stupid because there are ton of legitimate criticisms about AI and the AI rollout to be had; we don’t have to look for disingenuous cases of misuse for critique.
That would be fine, if people weren’t using LLMs to write code, or to do school work,
But they are. So it’s important to write these articles that say “if you keep using a chainsaw to drive nails, here are the limitations you need to be aware of.”
I see your point and I agree, except that that isn’t what these headlines are saying. Granted, perhaps that’s just the standard issue of sensationalism and clickbait rather than being specific to this issue, but the point remains that while the articles may be as you claim, the headlines are still presented instead as “A chainsaw can’t even drive a simple nail into wood without issue and that’s why you should be angry anytime you hear a chainsaw.” I dunno. I’m just so exhausted.
We have to wait a bit to have an useful assistant (but maybe something like copilot or more coded focused ai are better)
I guess it depends on the programming language… With python, I got very fast great results. But python is all about quick and dirty 😂
In Rust, it’s not great. It can’t do proper memory management in the language, which is pretty essential.
I asked ChatGPT for assistance with JavaScript doing HL7 stuff and it was a joke… After the seventh correction I gave up on it (at least for that task)
What drives me crazy about its programming responses is how awful the html it suggests is. Vast majority of its answers are inaccessible. If anything, a LLM should be able to process and reconcile the correct choices for semantic html better than a human… but it doesnt because its not trained on WIA-ARIA… its trained on random reddit and stack overflow results and packages those up in nice sounding words. And its not entirely that the training data wants to be inaccessible… a lot of it is just example code wothout any intent to be accessible anyway. Which is the problem. LLM’s dont know what the context is for something presented as a minimal example vs something presented as an ideal solution, at least, not without careful training. These generalized models dont spend a lot of time on the tuned training for a particular task because that would counteract the “generalized” capabilities.
Sure, its annoying if it doesnt give a fully formed solution of some python or js or whatever to perform a task. Sometimes it’ll go way overboard (it loves to tell you to extend js object methods with slight tweaks, rather than use built in methods, for instance, which is a really bad practice but will get the job done)
We already have a massive issue with inaccessible web sites and this tech is just pushing a bunch of people who may already be unaware of accessible html best practices to write even more inaccessible html, confidently.
But hey, thats what capitalism is good for right? Making money on half-baked promises and screwing over the disabled. they arent profitable, anyway.
If you don’t know what you are doing, and you give it a vague request hoping it will automatically solve your problem, then you will just have to spend even more time to debug its given code.
However, if you know exactly what needs do do, and give it a good prompt, then it will reward you with a very well written code, clean implementation and comments. Consider it an intern or junior developer.
Example of bad prompt: My code won’t work [paste the code], I keep having this error [paste the error log], please help me
Example of (reasonably) good prompt: This code introduces deep recursion and can sometimes cause a “maximum stack size exceeded” error in certain cases. Please help me convert it to use a
whileloop instead.I wouldn’t trust an LLM to produce any kind of programming answer. If you’re skilled enough to know it’s wrong, then you should do it yourself, if you’re not, then you shouldn’t be using it.
I’ve seen plenty of examples of specific, clear, simple prompts that an LLM absolutely butchered by using libraries, functions, classes, and APIs that don’t exist. Likewise with code analysis where it invented bugs that literally did not exist in the actual code.
LLMs don’t have a holistic understanding of anything—they’re your non-programming, but over-confident, friend that’s trying to convey the results of a Google search on low-level memory management in C++.
If you’re skilled enough to know it’s wrong, then you should do it yourself, if you’re not, then you shouldn’t be using it.
Oh I strongly disagree. I’ve been building software for 30 years. I use copilot in vscode and it writes so much of the tedious code and comments for me. Really saves me a lot of time, allowing me to spend more time on the complicated bits.
I’m closing in on 30 years too, started just around '95, and I have yet to see an LLM spit out anything useful that I would actually feel comfortable committing to a project. Usually you end up having to spend as much time—if not more—double-checking and correcting the LLM’s output as you would writing the code yourself. (Full disclosure: I haven’t tried Copilot, so it’s possible that it’s different from Bard/Gemini, ChatGPT and what-have-you, but I’d be surprised if it was that different.)
Here’s a good example of how an LLM doesn’t really understand code in context and thus finds a “bug” that’s literally mitigated in the line before the one where it spots the potential bug: https://daniel.haxx.se/blog/2024/01/02/the-i-in-llm-stands-for-intelligence/ (see “Exhibit B”, which links to: https://hackerone.com/reports/2298307, which is the actual HackerOne report).
LLMs don’t understand code. It’s literally your “helpful”, non-programmer friend—on stereoid—cobbling together bits and pieces from searches on SO, Reddit, DevShed, etc. and hoping the answer will make you impressed with him. Reading the study from TFA (https://dl.acm.org/doi/pdf/10.1145/3613904.3642596, §§5.1-5.2 in particular) only cements this position further for me.
And that’s not even touching upon the other issues (like copyright, licensing, etc.) with LLM-generated code that led to NetBSD simply forbidding it in their commit guidelines: https://mastodon.sdf.org/@netbsd/112446618914747900
I’m very familiar with what LLMs do.
You’re misunderstanding what copilot does. It just completes a line or section of code. It doesn’t answer questions - it just continues a pattern. Sometimes quite intelligently.
Shoot me a message on discord and I’ll do a screenshare for you. #locuester
It has improved my quality and speed significantly. More so than any other feature since intellisense was introduced (which many back then also frowned upon).
deleted by creator
Example of (reasonably) good prompt: This code introduces deep recursion and can sometimes cause a “maximum stack size exceeded” error in certain cases. Please help me convert it to use a
whileloop instead.That sounds like those cases on YouTube where the correction to the code was shorter than the prompt hahaha
I will resort to ChatGPT for coding help every so often. I’m a fairly experienced programmer, so my questions usually tend to be somewhat complex. I’ve found that’s it’s extremely useful for those problems that fall into the category of “I could solve this myself in 2 hours, or I could ask AI to solve it for me in seconds.” Usually, I’ll get a working solution, but almost every single time, it’s not a good solution. It provides a great starting-off point to write my own code.
Some of the issues I’ve found (speaking as a C++ developer) are: Variables not declared “const,” extremely inefficient use of data structures, ignoring modern language features, ignoring parallelism, using an improper data type, etc.
ChatGPT is great for generating ideas, but it’s going to be a while before it can actually replace a human developer. Producing code that works isn’t hard; producing code that’s good requires experience.
This has been my experience as well. If you already know what you are doing, LLMs can be a great tool. If you are inexperienced, you cannot assess the quality nor the accuracy of the answers, and are using the LLM to replace your own learning.
I like to draw the parallel to people that have learnt to paint only using digital tools. They often show a particular colouring that shows a lack of understanding of colour theory. Because pipette tools mean that you never have to mix colours, you never have to learn to do so. Painting with physical paint isn’t superior, but it presents a hurdle (mixing paint) that is crucial to learn to overcome. Many digital-only artists will still have learnt on traditional media. Once you have the knowledge, the pipette and colour pickers are just a tool, no longer inhibiting anything.
Actually the 4o version feels worse than the 4. Im getting tons of wrong answers now…
Yeah, it’s not supposed to be better than 4 for logic/reason/coding, etc… its strong points are it’s natural voice interaction, ability to react to streaming video, and its fast and efficient inference. The good voice and video are not available to many people yet. It is so efficient that it is going to be available to free users. If you want good reasoning, then you need to stick with 4 for now, or better yet, switch to something like Claude Opus. If you really want strong reasoning abilities, then at this point, you need a setup using agents, but that requires some research and understanding.
Removed by mod
Well, I do it 99% of the times
Hey! I can keep my job for at least a few more years
Removed by mod
