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.
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.
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...
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)))
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.
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?
You could also imagine inserting adverts into the generated response.