Use of ChatGPT generated text for posts on Stack Overflow is temporarily banned
meta.stackoverflow.com
meta.stackoverflow.com
— ChatGPT arguing its own case
— Human who watches SciFi movies
And the fact that ChatGPT answers are easy to produce means that more people can contribute.
Actually, that means that fewer people and one robot will contribute.
But as the SO administrators say, these answer are close enough to true to require a lot of effort on the part administrators - and this stuff may promise a rocky road for answer and discussion sites in the near future. And that would not mean that more people are "participating" either.
This pattern...
<1-3 paragraphs of ChatGPT generated text here>
<Summary saying I used ChatGPT to write this>
But now the novelty is wearing thin and all I see are the fallacies inherent in trying to put this forward as some sort of argument/comment in favour of a viewpoint... because I'm still giving the commenter the benefit of the doubt and assuming they are trying to make a substantive contribution to the discussion.
- False Attribution (https://www.logicalfallacies.org/false-cause-and-false-attri...) almost universally at play since before the bait and switch they are attributing AI/ML blather for that of a genuine human's reasoning and opinion, revealing it was from the AI doesn't remove that. By virtue of posting it on your own account, there is the inherent presumption that its coming from the human who registered the account, and thus the false attribution.
- Written by a real human, I promise
Can't agree. When writing code I actually pretty sure that 1 + 1 = 2, and 100 + 100 = 200. Of course, of course it could be not true for all languages and all environments, but in one language and in one environment 1 + 1 = 2 is true is always, and not with 99.9999999% chance.
Thats my issue with ChatGPT answers - I can't trust it. ChatGPT shifts complexity of reading, analyzing, verifying correctness of answer to human.
After the abuse: intelligent experts using their time to check the vaticinations of a delirious toy.
In it (I'm paraphrasing here) the open internet is destroyed by media companies who flood the internet with "good garbage", that is stuff that is _almost_ correct.
It forced people to pay for curated accurate content (that is, pay the same people who generated the good garbage in the first place).
I know it’s not actually meant to do that in the first place, but it’s quite excellent at it despite that shortcoming. I’m not looking forward to AIs learning to become even more convincing and authoritative.
a) Wow such a nice tool that generates nice prose for emails and communication! b) Wow, such a nice tool that filters and summarizes the mountains of prose coming into my communication channels!
It's as if we're just making the channel broader with less information for no reason.
It's the usual: all that effort is generally wasted, because it averages out to zero, but the money is made on high-frequency fluctuations, the momentary imbalances. It's just another hamster wheel for the economy to work, and hopefully produce something useful every now and then as an accidental spinoff.
I hadn't considered that the success of large language models might break the current system in a way that can only be fixed by moving it in the direction I'd like to see (but maybe not the micropayments). What an exciting thought :-D
I must read Anathem, maybe it'll change my mind. Neal Stephenson's always great anyway.
https://www.engadget.com/2017-10-24-adblock-plus-relaunches-...
The scary part was just how confidently incorrect it was. The text looked very good, but there were big errors in there.
I also tried to ask it a scientific question on a pretty niche field, and it produced a very reasonably looking answer. Most of the text was actually correct, if a bit generic. But there was one big factual mistake in there, it gave a value range that is simply wrong for this question.
Ask ChatGPT to do any engineering work, and it falls flat. It's poor mathematical skills and lack of ability to create complex plans mean that it is relying purely on it's coherent-sounding dialog.
For a factual claim, it could look at each sentence and see if the equivalent information exists on Wikipedia, or another online source, or even a book. And then link to that source. Google has already digested all that information, so it's just a matter of finding clever ways to connect them.
I don't think it would ever be flawless, but it could be turned into an extremely capable research assistant.
I can do the complex plans myself, but automation is quite helpful when looking up, for example, what is the idiomatic way to do datetime manipulation in some language which I haven't used in years.
This will be crucial for preventing unfavorable training loops among a myriad of other human-led control efforts.
- obviously generated by ChatGPT
- my opinion, not generated by ChatGPT.
- same as you
> Write a Tweet tht answers the question, "how can you tell if a comment was written by ChatGPT and not a human?"
The first three replies were:
> "If a comment seems overly robotic or generic, it's likely written by ChatGPT rather than a human!" #AI #ChatGPT
> If a comment reads like it was written by a robot and lacks the nuance and creativity of human language, it's likely written by ChatGPT. #AI #chatbot
> "If the comment seems too perfect or robotic in its language and tone, it's likely written by ChatGPT. Look for a lack of personality and human error in the writing." #ChatGPT #AIwriting
When I tried asking it to "Provide a Reddit style answer to the question, ...", it gave me a one-page essay with 5-item bullet point list, which ironically argued - among other things - that one indicator is an answer that's longer / more verbose than what you'd expect.
EDIT: OK, you can say that "it's likely written by..." is a hedge. So I now tried this:
> Write a Tweet that confidently answers the question, "how can you tell if a comment was written by ChatGPT and not a human?"
The third result I got was:
> "Just look for the tell-tale signs of a ChatGPT comment: robotic language, repetitive phrases, and a lack of genuine emotion or personal touch. Trust me, you'll know it when you see it!" #ChatGPT #artificialintelligence
- Written by ChatGPT under prompt Write a Hacker News comment that answers the question, "how can you tell if a comment was written by ChatGPT and not a human?"
I input several emails I've received in private and it correctly answered those. This is not a huge library ecosystem or even a large language .
It then translated several Haskell code samples I have into c++.
I'm not sure if I'm in awe or just in shock to be honest.
It's not a place to ask questions and receive help from an AI. That would be a different product. It might in the future even be a better product, who knows? But it is a different product and there needs to be a clear line between them.
That product is called "Google Cloud Support".
I agree OpenAI have something going on here, I'd assumed it was that they wanted to know the strategies we came up with for bypassing their filters, and to get into a cat and mouse game with us so that they could improve them to the point of being productized. Also, marketing, both sales, recruitment, and mindshare.
ETA: I think I underestimated before the extent to which the output is often just your own words reflected back to you, in which case, it may well be adding more signal then noise.
https://news.ycombinator.com/item?id=33863413
Still, some prompts return a walk of text from a single sentence, and I have to imagine those are worse then useless for training.
And it seems that AI has much more than 50% answers wrong.
Better job offers because you have high rep?
Would like to learn more about this. Tons of SO posts are wrong, it's almost always the case that you have to scroll past the accepted answer to find one that's actually right. Is ChatGPT much worse?
Yes, there are terrible and incorrect answers, but adding ChatGPT allows this to happen at a much higher rate than previously. And it allows folks to put together coherent or correct sounding answers without knowing anything about the subject.
But it also routinely provides very useful answers, and you can ask it to fix bugs you find.
I assume that within the next two years, Stack Overflow will be overtaken by a similar site that defaults to Chat-GPT-like AI answers, and has incentives for human programmers to train it by checking or improving its answers.
Unless the Stack Overflow people decide to do something like that themselves.
But even this version of ChatGPT often has amazing code answers, and if you could put it in a loop with a compiler/JavaScript runtime that could go pretty far.
I imagine a couple more years of larger/better models (maybe with some visual understanding of UIs) etc., and the website will just be "write, test and deploy this entire program for me, here is a list of requirements" -- taking out a lot of not only Stack Overflow's business but also Upwork etc.
https://meta.stackoverflow.com/a/421832
I guess that cat won't go back in its bottle [0].
[0] Attempt at humor was not AI generated. [1]
[1] And neither is this footnote.
If answering a question was all about generating the correct subsequent/pattern of words based on prior context or prompts, then we can argue that sales consultant, marketing personnel, journalists etc can write and design software or come up with scientific theories? After all, they also can use buzz words they have learnt contextually very well and hence they sell the product, isn't it?
(Also don't let ChatGPT pose as a human - give it it's own user name and automate it.)
But there is an answer suggesting using it: https://meta.stackoverflow.com/a/421836/7884305.
Yeah I know, humans will team up with bots and praise the new chat overlords.
partial answer which some faulty reasoning about precision:
First, the Celsius scale is based on a more intuitive and logical reference point. The Celsius scale sets the freezing point of water at 0 degrees and the boiling point of water at 100 degrees, whereas the Fahrenheit scale sets the freezing point of water at 32 degrees and the boiling point at 212 degrees. This means that the Celsius scale has a smaller unit interval, which allows for more precise temperature measurement.
Or any post on Internet.
It is to reduce abuse of their API because these requests are very expensive to serve as it requires 100s of gigabytes of VRAM to serve a request, requests take a few seconds at a time, and they are handling a ton of free requests. If you were paying for their API you could be paying up to about $0.10 per text generation.
This is expected and it may not look like it now, but banning an advancement in technology (even temporary) is usually the first step to irrelevance of those pushing back.
In a few years from now when ChatGPT and perhaps the much awaited GPT-4 are released, get better, (or even an implementation is open sourced) it will further plummet the usage for SO and Google.
The genie is out of the bottle and there is no going back, I'll give it less than 15 years.
You could also imagine inserting adverts into the generated response.
Stable diffusion is powerful because there's no "right answer". Just better/worse variations of a prompt.
GPT is going to be amazing for telling stories. Particularly, for "too literal" software engineers, like me. However, it doesn't seem to offer much in terms of hard, factual information that technical fields rely on.
Which makes sense considering that these large scale language models represent the diluted sum of human generated content and most humans are by definition relatively mundane, as are the works that they produce.
Now I won't argue that the stories that these large scale language models can produce are highly cohesive, they're just not very interesting or compelling in any form.
But so is modern TV (that includes streaming services for the nitpickers) and look how much money they're making...
Given Google's resources and scale, I'd wager that whatever they have is on par with, if not better than, OpenAI's models.
All this being said, I disagree with your hypothesis. Google does have certain benefits that pretrained models won't easily be able to replicate. For instance, keeping up-to-date and accurate data about restaurant hours, locations, etc. Additionally, it's pretty difficult to know if the LLMs are giving you correct information or not. I think these models still need more innovations before they replace Google. Not to say the current gen isn't useful, but the current Venn diagram of Google use cases and ChatGPT use cases is barely overlapping IMO.
they are swarmmed with a massive influx of incorrect answers to questions that 'look somewhat correct at first glance'
sbcl --script example.lsp --output example.c
As far as AI answers go, I would give it two thumbs up. However, the information is wrong, and so I must save my thumbs for another time.
And it'd get things like the sbcl command line above, because that was the right answer in the particular universe where the question could be answered.
(defun num-atoms(list)
(cond
((null list) 0)
((atom list) 1)
(t (+ (num-atoms (car list)) (num-atoms (cdr list))))))
#include <stdio.h>
#include <stdlib.h>
typedef struct list_node {
int data;
struct list_node *next;
} list_node;
int num_atoms(list_node *list) {
if (list == NULL) {
return 0;
} else if (list->next == NULL) {
return 1;
} else {
return num_atoms(list->next) + num_atoms(list->next->next);
}
}
int main() {
list_node \*list = NULL;
printf("%d\n", num_atoms(list));
return 0;
}
"Note that this implementation of the num_atoms function is not equivalent to the original Lisp function, as it does not correctly handle the (atom list) case. In C, it is not possible to directly check whether a value is an atom, as atoms are not a fundamental concept in C. To properly implement the num_atoms function in C, a more detailed specification of what is meant by an "atom" would be needed."
I then asked ChatGPT: "Make a function in C that duplicates (atom list) from Lisp" since that was the recommendation.
I like to think it got impatient with me badgering it about this subject, because the response starts with "As mentioned earlier, atoms are not a fundamental concept in C, so it is not possible to directly implement a function that duplicates the behavior of (atom list) in Lisp." But it then followed up with C code that approximates it (by simply checking if something is a list or not), and then explains why it is still not exactly the same as Lisp's implementation, and that it won't work correctly in every case as it would under Lisp.
It then ended with "The definition of an atom and the way it is identified may vary depending on the specific requirements of the problem being solved."
This could be quite educational. I really expected it to "make shit up," but it was pretty spot on.
(+ (num-atoms (car list)) (num-atoms (cdr list)))
return num_atoms(list->next) + num_atoms(list->next->next);
(+ (num-atoms (cdr list)) (num-atoms (cddr list)))
Answers given by chatgpt have a high chance of being incorrect, how is this a ban on advancement in technology? How is banning chatgpt in it's current state unreasonable for a website that tries to provide correct answers?
Depends on the reason for banning a technology. In this case it isn't because SO's business model is threatened, really; right now it's largely because the ChatGPT answers are usually wrong, and getting spammed to high hell by people who think they're clever by using ChatGPT to answer questions.
It's similar to image boards banning or requiring labels for AI-generated art. It's not that artists feel threatened... it's that the people looking at art don't want to have to wade through a ton of similar-looking, mediocre AI-generated crap.
Or to put it another way; if I set up some bots to start spamming every HN post with dozens of ChatGPT generated comments based on the submission title[0], and I (rightfully) got banned for such, is that because HN feels threatened by ChatGPT?
[0] not using text from the article as a prompt to more accurately simulate the average HN comment
Where I fear you might be right though, is that the web in general (and so, Google) and sites like SO with their internal content, will inevitably lose value as people post ChatGPT-generated content which superficially looks reasonable, but is subtly wrong. Wading through all those subtly-wrong things looking for the actual answer will be a far worse experience than today.
Which is unfortunate. I don't see a good way to avoid it, sadly. SO's attempt to ban ChatGPT-generated content is useless: while people have the incentive to try to get kudos from SO, and ChatGPT requires skilled, detailed analysis to detect its errors, the ban is unenforceable.
A version of ChatGPT that has access to the web (can crawl, can link to content) will be a big blow for google.