ChatGPT's road to madness

Recently, as ChatGPT has exploded out of the circle, various debates about ChatGPT have continued on the Internet; some people regard it as a more advanced chat robot, some are excited to see the opportunity for entrepreneurship, while others are not interested in it. Its ability to replace human work has revealed a lot of concerns; so is it a tool to push society forward, or is it a T-1000 that attempts to subvert human society? In this article, let’s take an in-depth look at ChatGPT.

When we talk about ChatGPT, we are discussing the joy of chatting, the worry of unemployment, and the vision of the future.

Development History

ChatGPT seems to have suddenly become popular overnight, but in fact it is not a new concept. The earliest artificial intelligence can be traced back to 1950, Allen ·Alan Turing published the landmark paper "Computing Machines and Intelligence", also known as "Can Machines Think?" ", in this paper, a very philosophical concept was proposed for the first time: the imitation game, which is also known as the famous Turing test. It was this article that won Turing the title of the father of artificial intelligence. Laurel crown.

The Turing test is to put the tester and the person being tested (a machine or a person) without face to face, and let the tester use some testing devices (such as a keyboard) to ask questions like the person being asked, if If more than 30% of the answers of the test subjects cannot make the tester confirm which answer is the human and which is the machine, then the machine has passed the test; such a machine is also called a Turing machine, and the Turing machine is just a hypothesis. , not really a machine.

Turing Test

In fact, we found that the Turing Test does not directly define artificial intelligence. What does it mean? Artificial intelligence; but does the opposite, and does not stick to complicated processes, and is result-oriented. What kind of results are achieved (deceiving testers) is considered artificial intelligence; sometimes when we deal with difficult problems You can also change your thinking and use the results as the starting point to dismantle the problem.

In the biographical film "The Imitation Game", the idea proposed by this lonely genius was also depicted. When Turing was locked up in the interrogation room by the police detective, the detective asked a meaningful question : Can machines think? Turing proposed to play a game (the Turing test); in the end Turing smiled, and he felt that the detective understood the rules of the game.

Pattern matching stage

After the Turing test was proposed, many scientists and laboratories also launched an attack on this test; in 1966, the MIT laboratory gave birth to a true The chatbot Eliza is positioned as a psychotherapist.

Eliza

Eliza's main strategy is to ask questions, rephrase what users say, and guide users to describe the problem more. For example, if you tell it: I have a headache today, it will tell you to consult a doctor for medication if you have a headache; if you say you are a little depressed today, then it will say that it heard that you are unhappy, which means it is very sad. ; If you say my mother takes care of me, it will ask who else in your family will take care of you.

Eliza decomposes the input through keyword matching rules, and then generates a reply based on the reorganization rules corresponding to the decomposition rules; in layman's terms, it means grabbing the keywords in the sentence, for example, finding that the key word mother is in the sentence words, she would say: Tell me about your family; in fact, there are a lot of if/else codes behind it.

In 1995, inspired by Eliza, Richard Wallace developed a chatbot Alice, which was open sourced in 1998; the technical principle of Alice is mainly based on natural language processing, and its purpose is Mimic natural human language to have meaningful conversations with users. It can answer questions on various topics such as weather, news, sports, etc., and can also engage in interesting chats with users.

Alice

But at this stage, whether it is Eliza or Alice, their principles are based on pattern matching (Pattern Matching) , by extracting keywords and calling preset text to reply.

Although these conversational robots can perform simple language interactions, they lack in-depth understanding and reasoning capabilities of sentences, making it difficult to compare with human conversational levels.

But this model is not without merit. It can avoid a lot of repetitive work, but it is also very common around us. For example, on common shopping websites, bank websites or apps, etc., when you enter the chat interface, you will first be greeted by an enthusiastic chatbot, which will list a bunch of keywords and ask you what you want to do.

Machine Learning Mode

As we mentioned above, no matter how much if/else code is written in the pattern matching method, the machine still cannot understand the true meaning of the sentence and just responds simply. Therefore, a new model emerged, namely: machine learning.

As the name suggests, it does not artificially prescribe questions and answers, but gives the machine a bunch of ready-made cases for the machine to learn. This method is also more in line with the cognitive laws of human learning.

The more famous one at this stage is SmarterChild (smarter child), which was developed by ActiveBuddy Co., Ltd. in 2001. This chatbot is used in the old instant messaging AIM, which can conduct interesting conversations and quickly access Data from other services, like weather, stocks, and movie data; you can even scold it and make things difficult for it, and it always seems to know how to deal with it.

SmarterChild was also one of the first chatbots integrated into instant messaging platforms. It quickly caused a sensation at the time, with more than 30 million users using it. It was later acquired by Microsoft in 2006 and was also applied to MSN Messenger. superior.

SmarterChild

However, after SmarterChild, the chatbot was abandoned for several years, and Microsoft also closed SmarterChild’s high-tech department.

In 2010, with the explosion of Artificial Neural Networks (ANNs), a field in machine learning, artificial intelligence ushered in unprecedented development.

It is inspired by biology. In junior high school biology textbooks, we have learned that the neural network of animals can process a large amount of complex information, which is to perform functions through about 100 billion neurons connected to each other; artificial What a neural network has to do is simulate the basic unit in the brain: neurons.

Artificial neural network looks very powerful, but in fact it is essentially a process of constantly extracting features. It is very similar to how we learned to recognize things when we were young, which is to find the characteristics of different things; artificial neural network is getting a string of After the sample data is collected, the features of each part of the observed thing are also extracted through learning, the features are associated, and then after repeated training, the correct answer is finally output.

Through this process, we can also find that artificial neural networks require massive data training and powerful computing power support; with the rapid development of the Internet, large data sets are no longer a problem.

Siri was founded in 2007 and was acquired by Apple for US$200 million in 2010. Initially, it focused on text chat. Later, it cooperated with Nuance, the world's largest speech recognition manufacturer, to realize the speech recognition function and It debuted on the iPhone 4S in 2011, causing a sensation at the time, and the iPhone 4S also became a miracle machine.

Siri uses a deep convolutional network to recognize your voice, which is also a type of artificial neural network; the iPhone has a dedicated low-power processor to run this neural network. When the similarity reaches a certain level, threshold, Siri will be activated.

❝

The launch of Siri also marks that chat robot technology has entered a new era,

❞

Siri promoted on the official website

The time came in 2017, Google released a new machine learning model in the article "Attention is all you need": Transformer model, which is mainly used to overcome the problem that the traditional network training time in machine translation is too long and it is difficult to achieve parallel computing well.

Traditional natural language processing (such as speech recognition, language modeling, machine translation) relies on Recurrent Neural Network (RNN), which uses loops to perform sequential operations, that is, learning word by word , has the disadvantages of too long training time and difficulty in parallel computing.

The Transformer model abandons the traditional temporal structure, processes all words or symbols in the sequence in parallel, and uses the self-attention mechanism to combine context with distant words; this is equivalent to the scumbag still learning When reading a book word by word, the top student has read ten lines at a glance and several paragraphs. This naturally improves learning efficiency.

Microsoft’s GPT model and Google’s BERT model, the T in them both stands for Transformer model.

We have discovered a rule that the development and progress of science and technology are always accompanied by the proposals and breakthroughs in theoretical research.

Now that the model is available, there must be companies to commercialize it. Let’s first talk about ChatGPT’s parent company OpenAI. It was established in 2015 by a group of Silicon Valley technology tycoons, including the well-known Te Tesla founder Elon Musk and others confirmed the company's main goals at the beginning of its establishment:

Including creating "universal" robots and chatbots that use natural language.

In 2018, OpenAI released a generative pre-trained language model (Generative Pre-trained Transformer, GPT-1) based on the Transformer model; however, its old rival Google's BERT soon appeared , and completely crushed GTP in terms of performance.

The development history of ChatGPT

Naturally, OpenAI was not willing to be surpassed by its opponents, so it spent money like crazy to increase its size. The training data set was obtained, and the GPT-2 and GPT-3 models were released one after another. The number of parameters of the model also rapidly "spiked" from 125 million in GPT-1 to 175 billion in GPT-3; the number of parameters was even larger. This means that the model has more powerful expressive capabilities and smaller prediction errors, which means it can generate longer and more natural text.

The massive model parameters allow GPT-3 to perform well on some more complex problems, such as writing some papers on behalf of humans, or even writing SQL statements, JavaScript codes, etc.

In 2021, OpenAI modified and improved based on the GPT-3 model, adjusted the model parameters, added training data, and released GPT-3.5 at the end of the year, which is the original model of the currently popular ChatGPT.

In November 2022, ChatGPT gained 1 million users within just a few days of its launch. Within two months of its launch, its monthly active users reached an astonishing 100 million, making it the consumer application with the fastest growing user base in history.

How many months does it take for each App to have 100 million monthly users?

Behind the influx of a large number of users is the need for huge computing power Cost and server investment; according to calculations by a research institution, the number of GPU chips required to run such a complex GPT-3.5 model is as high as 20,000, and professional-grade graphics cards generally use A100; according to data from a shopping website, 10 The price of an A100 graphics card is 10,000 yuan, which is still priceless in China; therefore, according to a rough calculation, the investment in the graphics card alone is at least 2 billion; there are other soft costs such as data collection, manual annotation, and model training, which are even more difficult. statistics.

Registration Guide

After talking so much about the development history of ChatGPT, I believe that everyone must be impatient and eager to try this "advanced chat robot" with tens of billions of investments, so here is the author. Share the registration strategy; due to some well-known reasons, domestic users cannot directly access smoothly, and the registration process will have some twists and turns, so we need to make the following preparations:

• Be able Access Google's environment (this is a prerequisite, everyone knows it)
• An email, preferably Gmail or Outlook
• A foreign mobile phone number (if not, sms-activate.org is recommended)

First we prepare a mobile phone number to receive the SMS verification code, open sms-activate.org and register an account, adjust the language to Chinese, click the recharge button in the upper right corner, and select Alipay as the payment method.

Recharge

The minimum recharge amount for the account is 1 US dollar, which is about more than 7 RMB when converted according to the current exchange rate; we You can recharge $1 first. Select the country on the left, then search for OpenAI. The lease period defaults to 4 hours, and click Rent. The prices in different countries here are also different. Generally, prices in Southeast Asian countries such as Indonesia, Thailand, and Indonesia are cheaper, but more people are using them now. There is a chance that it will be blocked by OpenAI or there will be no number; other European and American countries are relatively expensive and may not be worth 1 dollar. For example, Spain, the author’s choice, costs a full 2 ​​dollars.

❝

You can try different country numbers. If you don’t receive a text message within four hours, you can get a refund.

❞

Rental number

After the mobile phone number is prepared, we can register. Open the signup page. After entering the email address, password and verification email, we come to the page to verify the mobile phone number. We select the corresponding country for renting the mobile phone number above. , then paste the number and click Send.

When pasting the mobile phone number, you need to delete the international area code in front.

Verify mobile phone number

Return to the sms-activate page, we can see the received OpenAI code in the message , just paste it.

Receive the verification code

After the registration is completed, we can open chat.openai.com/chat and use it.

Main interface

It is also very simple to use. Enter your question in the input box below and click the send button , you can mention a chatty robot that knows everything about astronomy and geography, and is omnipotent.

Application Scenario

So, so many technology giants spent tens of billions to develop ChatGPT, is it just for us to chat? To understand this problem, we might as well ask ChatGPT itself.

Application scenarios

Chat robot

Indeed, using ChatGPT as a chat robot is indeed a bit of a killing. It's a bull's-eye knife; but let's take a look at it first. Is it fast enough to kill a chicken? If we want to do something, we need to call the API, and OpenAI has just recently provided access to API Keys, which can integrate ChatGPT into our applications and services. Visit platform.openai.com, click the API Keys => Create button, and copy the keys in the pop-up box that appears.

❝

After the pop-up box is hidden, the api key cannot be seen and needs to be regenerated.

❞

Generate API keys

After successful copying, we recommend the wechatbot project to connect personal WeChat to ChatGPT; there are many ways to run the project, which can be based on the source code You can also run it based on docker, but you need to use the key copied above. For the specific running method, you can check the project description, which will not be expanded here; if you don’t know about docker, you can check this article.

After the project is running, just use WeChat to scan the code to log in, and then our WeChat ID will automatically connect to the ChatGPT chat; the method of use is also very simple, the WeChat ID will reply directly to the private chat, group chat You need @this WeChat ID.

❝

Important reminder: Abuse may lead to the risk of being banned from WeChat. Try to use a smaller account. This article does not bear any responsibility.

❞

WeChat Chatbot

However, it should be noted that the API calls for each account are also limited, currently it is 5 US dollars, and there is also an expiration date. Just have fun with it. Friends who have the conditions can recharge.

Work and study assistant

In addition to being used for chatting, it is also highly sought after by many students and professionals during work and study. When a philosophy professor at a foreign university was grading, I was very surprised to read a paper that was "the best in the class." The paper discussed a philosophical issue with concise paragraphs, appropriate examples and rigorous arguments; however, after the professor's questioning Next, the student "admitted that the paper was written using ChatGPT."

At work, ChatGPT has also saved the hairline of many professionals. It has been used to generate documents including but not limited to: leadership speeches, media releases, group profiles, award speeches, toasts, publicity It can even take care of trivial matters such as weekly and monthly reports and reasons for asking for leave.

Reasons for leave

In terms of text polishing, ChatGPT does not lose to professional editors at all. In weekly, monthly or even annual reports, this is quite practical and can be understood by everyone.

Weekly Report

For some simple tool functions, we can easily let ChatGPT generate them directly; for example, I need A function that hides a mobile phone number can just describe the function of the function; it can even be connected to the context, which is something that previous artificial intelligence has not achieved.

Generate javascript function

ChatGPT not only helped us solve the problem, but also explained the logic behind the problem in a well-founded way; For example, the author encountered a problem when using the KeepAlive component in Vue3 last week. After searching Baidu, although there were many answers, we still need to conduct secondary screening on a large number of web pages. In the end, the solutions that may be screened out are all They are all the same (copying each other); and they are full of errors and various error reports. The user experience is very unfriendly. It probably took more than half a day to solve the problem.

Error case

For example, in this article found by Baidu, if you use the writing method in the red line frame, vue- The router will display various strange errors, and the error messages that appear cannot be searched at all.

But if the author uses ChatGPT to ask questions, we see that its logic is clear, and there are specific cases and tips to pay attention to. We only need to put the name of the component under the include attribute; with its help , I believe that I was able to jump out of the wrong logic at that time and greatly shorten the time to solve the problem.

ChatGPT’s answer

ChatGPT is also a good hand in copywriting generation; when we are faced with blank documents, we are struggling When thinking, we might as well open ChatGPT, describe our needs, easily generate a rough copy, and edit it again based on this to save time and stress.

We introduced ONLYOFFICE in the previous article on office software. Download the desktop version for free on the ONLYOFFICE official website or register for the online personal version for free. Use ChatGPT in the office software to quickly generate copywriting.

First we need to install the plug-in. After cloning the code on github, we find /sdkjs-plugins/content/openai/, select all files to add to the ZIP file, and then change the file format to plugin; open the document interface and find Plug-in => Setting => Add plugin, select our plugin file, and the plugin will be installed successfully.

Add plug-in

Then enter the API Key obtained above, and our plug-in is activated; in the text field Describe the copy content we want to generate and click the submit button; ChatGPT will process the request, return a response within a few seconds, and insert it into the document as plain text.

Generate copy

Fungibility

Many friends will definitely be amazed when they see this: This ChatGPT is really awesome! Many of my programmer friends are also lamenting: Are we going to be unemployed, but it seems that we won’t be unemployed yet. After all, its training cost is indeed high, and it is not easy to implement the application.

Replace workers

But some positions, such as customer service positions, are more susceptible to impact; I believe everyone should I have received various types of robot customer service calls, all of which identify specific keywords, follow fixed procedures, and reply sentence by sentence; with the emergence of ChatGPT, I believe that robot customer service will be able to deal with more complex scenarios in the future and understand customers more accurately. needs to respond flexibly.

❝

Will ChatGPT completely replace programmers’ work?

❞

In terms of programming, I feel that ChatGPT cannot completely replace programmers at the current stage; although it can be well implemented in generating code snippets, it is somewhat similar to the previous Copilot; but in complex projects, it is necessary to ChatGPT cannot replace the ability to understand the meaning of different file modules and make calls; it cannot replace manual labor in work that requires organization and coordination, such as code bug repair, front-end and back-end joint debugging, cross-department coordination, and business trip docking.

The author believes that ChatGPT brings more improvements in programming efficiency; for example, with its assistance, we can generate some repetitive and simple modules directly by throwing them to him. , and ultimately it may be completed in less than half a day.

❝

Chairman of SenseTime: 80% of future software code may be generated by AI.

❞

Although it does not seem to be a complete replacement at present, with the continuous reduction of hardware costs and the improvement of machine learning capabilities; when the cost of machine training is lower than the salary of programmers, do you think capitalists need to pay five insurances and one housing fund, and pay repayments from time to time? How to choose between you who want to fish and a machine that costs 30 to 40 cents per kilowatt hour and runs 24 hours a day? I believe that in the future, the R&D department that may originally need dozens of people will eventually only need three or four core engineers to maintain the normal operation of the business.

❝

How do we choose employment in the future?

❞

Although ChatGPT will replace some low-end jobs in the future, But new positions will certainly continue to be created; just as the emergence of computers eliminated typists but created a large number of programmer positions, artificial intelligence has created many emerging positions, such as prompt word consultants (the author's own conjecture), Professionals are responsible for generating corresponding copywriting or materials for ChatGPT questions; this is an inevitable trend of social progress, and it also forces us, who are forced by the historical trend to move forward, to continue learning and progress.

Replacing programmers

In terms of front-end, many junior programmers will definitely be eliminated, and the threshold for programmers continues to lower; Therefore, we need to continue learning and progress before artificial intelligence replaces us; the author believes that in terms of WebGL, artificial intelligence is still unable to replace manual labor for complex graphics and aesthetic understanding, so front-end partners can try to advance in this field. At the same time, senior architect is also a good direction.

Summary and Outlook

I believe that in the near future, ChatGPT can help us greatly improve productivity and learning efficiency; when facing new technologies or new articles, there is no need to read them in full, so that ChatGPT generates the outline and main content of the article to help us learn quickly; when writing documents, it can also be used to quickly generate beautiful content.

We are at a moment when we are witnessing history. No one can stand alone in the face of the historical torrent of artificial intelligence. Maintaining an arrogant and disdainful attitude will only accelerate elimination, as it is said in the novel "The Three-Body Problem":

❝

Weakness and ignorance are not obstacles to survival, arrogance is.

❞

In the movie The Wandering Earth 2 , Liu Peiqiang asked Moss, can humans survive? Moss said that the fate of human beings depends on their own choices; just like we are now asking whether ChatGPT will replace programmers, although it tells us that it cannot be completely replaced, how to make choices is related to the fate of each of us; and When the wheel of history slowly passes by, the only thing we have to do is to try to run ahead of it.

The above is the detailed content of ChatGPT's road to madness. For more information, please follow other related articles on the PHP Chinese website!