Welcome to the WeChat subscription account of “Sina Technology”: techsina
Source: Xinzhiyuan
[Introduction to Xinzhiyuan] The popular ChatGPT is used by netizens to generate massive answers. And Stack Overflow has been overwhelmed by the disturbance, and launched a “chasing and killing”: it should be blocked!
OpenAI’s new model ChatGPT has only been born for a few days, and it has become a “powerful weapon” for the majority of netizens.
Its answer does not say whether it is correct or not, but at first glance it is a fortress.
Coupled with the fact that it is very convenient to generate, these “looks good” answers flocked like a tide to the legendary question-and-answer website with codes from all over the world, the real teacher of computer students-Stack Overflow.
However, in order to judge whether these answers are right or wrong, a large number of people with professional knowledge need to carefully examine them.
In desperation, Stack Overflow had to urgently issue a statement: ChatGPT, banned!
There are too many mistakes, they should be sealed
Seeing that the discussion is in full swing, the official position of Stack Overflow, the world’s second largest programmer dating site, is also clear:
Don’t think about using AI robots to confuse the audience, just send the answer well, and it is forbidden to use ChatGPT to directly generate garbage answers, otherwise the account will be banned.
Stack Overflow stated in the official notice: This is a temporary regulation, the purpose is to deal with the current flood of content generated by ChatGPT on the forum.
The announcement pointed out that the main reason for this regulation is that the quality of the answers automatically generated by ChatGPT is too low, there are too many mistakes, and it looks like that, even people who don’t understand at all can generate answers at will.
As a community where the quality of Q&A is the standard, these low-quality answers can cause serious distress and inconvenience to those who come here for help, and are also harmful to Stack Overflow itself.
Stack Overflow said it will continue to weigh new policies around ChatGPT “and other similar tools” in the future, but the date of the decision and what the final restrictions will be is unclear.
In short, the attitude of Stack Overflow is very clear. It can be summed up as follows: “Fun, play, make trouble, don’t joke about Q&A”.
Netizens applaud
As for this decision, the onlookers below expressed their complete approval.
– Well done! Really hope this decision is not ad hoc and not limited to ChatGPT but should be extended to all AI generated answers. Relying on AI will never be able to give high-quality programming answers, and it will be impossible in another 100 years.
— This ban should not be a temporary measure.
ChatGPT can be used to score points by flooding and automatically generating answers. Even if these answers are correct, they don’t know or care. The harm to the ecology of the Stack Overflow forum is obvious.
However, it is actually not easy to identify such answers, and how to ensure “should be sealed” is a big problem.
Many professionals can see the mistakes in this kind of “false answer”, but all they can do is click on it, and there is no guarantee that these contents will be cleaned up.
Banning content like this is great, but what can moderators do? A post above said that these AI-generated answers may look like high-quality answers, which means that for a well-trained human, errors may be found and clicked.
The answers of these AI impersonated users may have some likes and some downvotes, because these answers look like high-quality answers.
I don’t know how to tell if an answer was generated by ChatGPT, and with the custom tagging now taking months to resolve, these users could continue to damage Stack Overflow until it’s dealt with in a timely manner.
So, to what extent is ChatGPT flooded?
Some netizens tried to post the questions they raised in the ChatGPT dialog box, and compared them with the answers given by a “real person” later to see if the generated answers could be used. It turned out that this “real person” was obviously wrong.
Here is his own answer generated using ChatGPT:
Here’s the answer given by so-called “real users”:
It is not difficult to see that the two answer routines are exactly the same, the tone of the language is also very similar, and the code examples are almost the same.
This is clearly a fake answer generated by ChatGPT. Currently this “user” has been banned from Stack Overflow.
In fact, from the view of OpenAI, the publisher of ChatGPT, one of the important uses of this AI tool is to help programmers check the code, and as a coding aid, answering questions on Stack Overflow is originally one of the legitimate uses of ChatGPT.
Unexpectedly, less than a week after it was released, it was ruthlessly banned. I don’t know if I blame myself for being “not good at learning” and the answers I gave were not helpful, or if I blame the netizens for being too cunning and eventually being overwhelmed and ruined.
We reached out to OpenAI to ask how it plans to handle the accuracy of these auto-generated answers, but have yet to hear back.
Stack Overflow can’t be replaced yet
As a developer, you must know the Stack Overflow that allows hundreds of millions of companies to run programs and enable hundreds of millions of computer students to obtain diplomas.
Here, you can find best practices for implementing certain features, or find the solution to that obscure bug that only occurs at 4 am on a full moon night.
However, after being acquired by Prosus for US$1.8 billion in 2021, everyone was generally worried about whether to start the paid VIP mode. And the arrival of that day will probably be the “doomsday” of programmers.
Fortunately, so far, the user experience of Stack Overflow is still excellent.
As for OpenAI’s ChatGPT, not only can it answer questions in any field, but every time the answer is very precise and detailed.
For example, when you asked “How to iterate over an keys and values array in Javascript”, this is the answer:
The icing on the cake is that ChatGPT can also tell you the best solution to this common problem in JavaScript with a supporting example and explanation:
So the question is, since ChatGPT can give the answer within seconds of entering the question, and also provides the code that can be copied directly, why don’t we use it systematically?
Because, ChatGPT has three “fatal” flaws:
1. Input cannot be wrong
When using ChatGPT, you must be precise, even a change of pronoun (a/an) can make the answer completely different.
For example, the previous question “How to iterate over an keys and values array in Javascript”, this time we remove an “an” and become “How to iterate over keys and values array in Javascript”
Now the solution given by ChatGPT is only valid for simple arrays consisting of a uniform list of elements.
2. Interpretation is impersonal
For example, the following question with the most votes on Stack Overflow:
“Why are operations on sorted arrays faster than operations on unsorted arrays?”
The answer given by ChatGPT is undoubtedly correct, but nothing more.
Because, if you can understand the answer, you don’t need to ask the question.
And if you need to ask this question, it means that you can’t understand the answer given by ChatGPT…
In contrast, the answerers on Stack Overflow first provide general background in non-technical terms, and then gradually pull the topic back to the original and final questions to deepen the thinking.
Of course, not all answers on Stack Overflow are (and need not be) so qualitative and detailed.
But this example fully demonstrates that the biggest difference between humans and AI is that the latter cannot determine the level of understanding of the other party, so as to adjust the answer.
To put it simply, AI will explain the “Theory of Relativity” in the same way and terminology, no matter you are a professor in the physics department, an ordinary college student, or a novice who knows nothing about it.
This is not important for AI.
3. The role of the community
Like it or not, Stack Overflow’s greatest strength is its community. This platform was created and operated in such a way that it encourages and encourages contributions from the largest number of people.
And it is this wide variety of answers and opinions that allows users to position themselves and make choices by considering the advantages and limitations of each solution.
In addition, another advantage of Stack OverFlow is peer validation. A solution that has been proven and tested by thousands of developers provides great assurance of its correctness. (Of course this doesn’t mean it’s 100% correct)
PS can still be used
Needless to say, Stack Overflow will never be replaced. It, and GitHub, will always be a good place to go to solve problems that require personal explanations or difficult problems.
Like Copilot, Intellisense, and advanced IDEs, ChatGPT (even if it’s not limited to code) is just one tool in a developer’s increasingly rich palette of tools. Like any self-learning model, it will learn and self-correct and improve over time.
But next time, instead of waiting a few days on Stack OverFlow for a relatively simple question (how to trigger a rebuild of a HookConsumerWidget in Flutter), why not just ask it on ChatGPT.
After all, there is a good chance that you will have a starting point that you can use to solve your problem:
Why does ChatGPT make so many low-level mistakes?
In the past few days, all netizens have played ChatGPT crazy. Many people have realized that this new AI is indeed a “powerful weapon”-without any effort, it can generate an awesome answer for you.
And if you take a closer look at this layer of gorgeous coat, you will find that its answers are often full of mistakes and omissions.
For example: I am 37 years old in 2022, so how old was I in 1985?
ChatGPT: 1985-2022=-37, since the age cannot be negative, there is no solution to this question.
Um, is there a possibility that the author of the question was born in 1985?
For another example, this netizen asked ChatGPT, does Hobbs advocate the separation of powers?
ChatGPT gave the answer confidently, and did not forget to quote the classics.
However, its answer is wrong.
Hobbes, a proponent of despotism, believed that the only viable alternative to anarchy was the empowerment of a monarch.
The idea of power distribution and checks and balances between the executive branch and the legislative branch was put forward by John Locke, a contemporary philosopher with Hobbes.
So, why does ChatGPT make such low-level mistakes?
Obviously, this is because in political philosophy, Hobbes and Locke are almost always mentioned together. When ChatGPT went online to find information about Hobbes, it naturally found Locke’s explanation of the separation of powers, and then mistakenly pinned this view on hobbes 
.
From this example, we can also see the limitations of the GPT-3 language model supporting ChatGPT-it can only grasp, but cannot reason or think.
So, it’s actually a matter of probability
Recently, many people have been discussing: which jobs will be disrupted by AI first?
This is an open question. But at least we can get a little inspiration from the homework assigned by the teacher.
When a teacher assigns a student a political philosophy paper, the student’s output is usually, to the world, only a reflection on what has already been written a million times.
Interestingly, the text of AIGC is not like a math problem, it does not have a single path to the correct answer.
Yes, the output of AI is probabilistic: ChatGPT does not have any internal records marked right or wrong. It is actually a statistical model. In this model, it is a combination of languages in different contexts.
The basis of the context is the overall data set used for GPT-3 training, the additional context of RLHF training from ChatGPT, prompt and previous conversations, and feedback that will soon be available.
The results of these combinations are undoubtedly exciting.
ChatGPT will run the entire virtual machine and write code.
Of course, ChatGPT is not running python, the answer is actually a probability result collected from the Internet data corpus that makes up GPT-3 –
ChatGPT made the best guess of the result within 10 seconds, which is likely to be correct, and it feels as if a real computer is executing the relevant code.
This possibility is enough to surprise human beings.
Although for coding work that requires accuracy, it has quite a few gray areas.
But for other fields of AIGC, such as generating text and images, this multiple possibility will undoubtedly help humans expand the boundaries of imagination.
ChatGPT’s business model promotes the revolution of artificial intelligence
Why has ChatGPT exploded these days?
GPT-3, as the underlying model, has actually been born for two years. The difference is that ChatGPT is free and easy to use.
It’s one thing to read examples of AI output, and quite another to generate the output yourself.
The AIGC revolution has already started. When Midjourney made AI-generated art easy and free, human interest and awareness exploded.
ChatGPT is free, and that’s the key.
You know, on OpenAI’s API, the biggest limiting factor is cost.
Using OpenAI’s most powerful language model, Davinci, to generate 750 words costs 2 cents; using RLHF or other methods to fine-tune the model will also cost a lot of money, generating 750 words from the fine-tuned model costs 12 cents.
What’s fascinating about ChatGPT is that it makes OpenAI the same leader in consumer AI products as MidJourney.
The midjourney subscription business model, for something that has a marginal cost in terms of GPU time, makes a lot of sense.
In this regard, former GitHub CEO Nat Friedman once put forward such an interesting point: In the real world, AI applications are actually scarce.
Now, researchers are delivering tons of new features to the world every day at a rapid pace, and businesses and product people are just beginning to digest these new features.
Thinking one step further, will humans be replaced by AI when AI is no longer scarce in the real world?
Let’s take a look at AI’s answer to this question.
(Disclaimer: This article only represents the author’s point of view, not the position of Sina.com.)
This article is reproduced from: https://finance.sina.com.cn/tech/csj/2022-12-06/doc-imqmmthc7209236.shtml
This site is only for collection, and the copyright belongs to the original author.