Reference sources:
书名：一本书读懂ChatGPT、AIGC和元宇宙
作者：王喜文
出版社：电子工业出版社
出版时间：2023年5月
ISBN：9787121453571
Scott said that ChatGPT will subvert the world;
Microsoft has invested tens of billions of dollars
in ChatGPT and plans to integrate it into Office software and Bing search engine;
in some universities and academic institutions, there is a rise in the legal compliance of using ChatGPT to write papers. There has been a big discussion;
even
some consulting companies have begun to worry that their jobs will be taken away...
In 2023, the enthusiasm for applying ChatGPT has been ignited, and the application scenarios of ChatGPT continue to expand rapidly.
The future is here, whether humanity welcomes it or not.
In this era, everyone feels that the world is changing too fast and knowledge is updating too fast, and it is difficult for us to keep up with this change. How do we respond to massive changes psychologically, behaviorally, physically, and strategically? Every job, every type of work, every step, every process is empowered and changed by technology.
For example, many people now believe that artificial intelligence (AI) will replace humans. With ChatGPT, human jobs will be replaced. In fact, everyone has ignored the most basic point. Human beings, animals called "Homo sapiens", have been on the verge of extinction many times in the long history of history, but humans have not become extinct because humans have subjective initiative. Human beings are extremely intelligent creatures that can adapt to complex changes. Why does the emergence of AI bring so much controversy? Some people know what to do with AI but decide not to do it because they think it's bad.
Of course, there are still many people who know that AI is good, but don’t know how to do it specifically. Whether or not to use AI to replace human workflow depends on a person’s world view, values, knowledge, experience, and abilities, as well as the civilization, culture, relationships, and characteristics of the ethnic group they belong to. Only when these factors are superimposed, it is possible to draw a conclusion— —Whether humans are willing to embrace the latest technology, learn, and understand. However, will AI replace humans and benefit mankind, or is this enough? These situations may occur. The door of AI has been opened. Whether you are willing to embrace it or not, whether you think it is good or bad, you must study hard, embrace it, accept it critically, and let humans become its masters. Today, between human and machine education, human education is actually more important. There is no clear conclusion yet on how AI will evolve, so we must embrace and learn first.
Human beings used to think that their knowledge was valuable, but in the future, knowledge needs to be expressed with human wisdom. Although knowledge is dead, human beings can use knowledge through subjective initiative, generate wisdom after practice, and then form new knowledge. Today, new knowledge that can be learned by machines and provided by human wisdom is constantly emerging in society.
Therefore, knowledge must be utilized, and machines must continue to learn, refine, and precipitate. In this way, society is likely to develop benignly. Dr. Wang Xiwen's book shares the latest achievements of cutting-edge science and technology with readers. It not only provides research basis for scientific and technological workers, government employees, scholars, and corporate practitioners, but also brings rich classroom content to students, teachers, and other groups. case. This book is logically clear and easy to understand. It is refreshing and full of longing for the future of AI...
ChatGPT is the embodiment of the technological progress of AIGC (AI Generated Content, artificial intelligence generated content). In our lives, artificial intelligence has long been popularized. For example, industrial robots replace humans for painting and welding, navigation apps automatically plan paths, and face recognition technology is widely used in life... Although the above-mentioned artificial intelligence can already replace humans to complete certain tasks. The work may have some characteristics of humans, but artificial intelligence is still far from real humans. To make artificial intelligence closer to humans, artificial intelligence must have the creative capabilities that humans have. This is the meaning of the existence of AIGC. AIGC has opened a new chapter in AI painting, AI composition, AI writing, and even AI-generated film and television works. It can be said to be a leapfrog upgrade in the history of artificial intelligence. The emergence of ChatGPT provides everyone with new ways and tools for text generation in the form of natural language conversations. It will significantly lower the threshold for building the metaverse, allowing us to build the metaverse in a completely different way from how we built the Internet in the past. A large number of non-professionals can describe their needs through language, and ChatGPT will automatically generate design drawings and codes based on the received needs, greatly improving the efficiency of building the metaverse and greatly reducing the cost. The content has also been greatly enriched.
ChatGPT will completely change ChatGPT is the embodiment of the technological progress of AIGC (AI Generated Content, artificial intelligence generated content). In our lives, artificial intelligence has long been popular, such as it is widely used in industrial machinery... Although the above-mentioned artificial intelligence can replace humans to complete certain tasks or possess certain characteristics of humans, artificial intelligence is still far away from real humans. Very far. To make artificial intelligence closer to humans, artificial intelligence must have the creative capabilities that humans have. This is the meaning of the existence of AIGC. AIGC has opened up new methods and tools for AI painting, AI composition, and AI writing, which will significantly lower the threshold for building the metaverse, allowing us to build the metaverse in a completely different way from how we built the Internet in the past.
A large number of non-professionals can build technology for digital humans: from the appearance, clothing, and shape design of digital humans, the actions and postures of digital humans, and the natural language communication of digital humans, to the execution of tasks by digital humans, and the relationship between digital humans and the environment. Perception and interaction, etc. ChatGPT will make the construction of digital people more convenient and simple, and make the functions of digital people richer. 2023 is a milestone year for artificial intelligence. With ChatGPT and AIGC becoming popular around the world, artificial intelligence technology has once again emerged in the past 10 years and has come to the forefront and entered the public eye.
In the past few years, technology giants have successively established artificial intelligence laboratories and invested more and more resources to seize the artificial intelligence market. Some companies have even transformed into artificial intelligence companies as a whole and stepped up planning for the future layout of artificial intelligence. Our country and other governments regard artificial intelligence as a strategic leader in the future, and have issued strategic development plans to promote overall development at the national level to welcome the coming era of artificial intelligence. This time the rise of artificial intelligence is not limited to laboratory research. The research and commercial application of relevant theories and key common technologies are advancing at the same time, which has led to the emergence of more product solutions and service-oriented application cases in the field of artificial intelligence, allowing the public to truly feel its impact and influence. . Especially in fields based on deep learning algorithm applications such as large language models (LLM), reinforcement learning based on human feedback, multimodal models, and natural language processing, artificial intelligence is rapidly being industrialized, and the track for industrial competition will follow. becomes more crowded.
Digital human construction technology: from digital human appearance, clothing, morphological design, digital human actions and postures, and digital human natural language communication, to digital human task execution, digital human perception and interaction with the environment, etc. ChatGPT will make the construction of digital people more convenient and simple, and make the functions of digital people richer.
2023 is a milestone year for artificial intelligence. With ChatGPT and AIGC becoming popular around the world, artificial intelligence technology has once again emerged in the past 10 years and has come to the forefront and entered the public eye. In the past few years, technology giants have successively established artificial intelligence laboratories and invested more and more resources to seize the artificial intelligence market. Some companies have even transformed into artificial intelligence companies as a whole and stepped up planning for the future layout of artificial intelligence.
Our country and other governments regard artificial intelligence as a strategic leader in the future, and have issued strategic development plans to promote overall development at the national level to welcome the coming era of artificial intelligence. This time the rise of artificial intelligence is not limited to laboratory research. The research and commercial application of relevant theories and key common technologies are advancing at the same time, which has led to the emergence of more product solutions and service-oriented application cases in the field of artificial intelligence, allowing the public to truly feel its impact and influence. . Especially in fields based on deep learning algorithm applications such as large language models (LLM), reinforcement learning based on human feedback, multimodal models, and natural language processing, artificial intelligence is rapidly being industrialized, and the track for industrial competition will follow. becomes more crowded.
In December 2022, ChatGPT was just a social network media application with chat function. But at the beginning of 2023, ChatGPT has been recognized as a technological product that has brought the third "revolution" to mankind after the Internet and smartphones. The Internet has opened up a "space revolution", making real-time connections with the whole world a reality. We do not have to travel thousands of miles to the scene, but we can communicate, teach, and video conference through the Internet, causing chain changes in politics, society, and business; the advent of smartphones The emergence of the Internet has brought about the "time revolution". Through various Apps (application software) that can be expanded and installed, we can achieve the fastest transactions and fast delivery, bringing about huge changes in life, work and consumption; and the emergence of ChatGPT , is expected to set off a "thinking revolution". ChatGPT can replace humans in creative creation, consulting and answering, translation services, customer service... changing the way humans think and deal with problems, and thus reshape the ecology of various industries, and even reshape the entire world.
The thinking revolution triggered by ChatGPT
In ChatGPT, GPT is the Generative Pre-training Transformer (pre-training is linked to the world in real time, allowing us to communicate, teach, and video conference through the Internet without having to travel thousands of miles to the scene, creating a chain of events in politics, society, and business) Transformation service...changes the way humans think and deal with problems, and thus reshapes the ecology of various industries. In ChatGPT, GPT is Generative Pre-training Transformer (pre-training generation model). OpenAI's language model can help the education field, Virtual therapists, writing aids, role-playing games, etc. In these fields, the existence of social bias, misinformation and toxic information is more troublesome, and only circumventing these system flaws can make it more useful. ChatGPT can answer continuous questions Questions, generate text summaries, translate documents, classify information, write code, etc. It will also admit mistakes, question incorrect premises and reject inappropriate requests. In just two months, people have continuously unearthed updates to ChatGPT. Multiple skills, including writing code, assignments, papers, speeches, event planning, advertising copy, movie scripts and other types of texts, drawing, translating, writing poems based on descriptions, and even playing interviewers, roles in movies, chatting, etc. A worry-free storyteller, and even gives advice on home decoration design, programming and debugging, life planning, etc. As long as you train it carefully, ChatGPT can even quickly evolve from a "consulting master" who is good at communication to an efficient learning tool. After continuous questioning, it can Lists a large number of book lists and material links to assist learning, helps you refine the key points of an article, the knowledge map and core context of a field, and even helps you open up your creative mind when your inspiration is exhausted. ChatGPT seems to have everything Understand, like an encyclopedia.
His fluent answering method and rich knowledge reserve have greatly shocked users. A report from UBS shows that just two months after ChatGPT was launched (end of January 2023), its number of active users has exceeded 100 million, breaking Douyin’s 9-month record and becoming the fastest growing user in history. One of the fastest apps. To reach 100 million users, it took App Store 2 years, Instagram 2.5 years, WhatApp 3.5 years, and Twitter 5 years (see Figure 1-2). Over the past few months, a large number of people have flooded the site with various requests for ChatGPT. A software engineer asked it to debug code, and it did it; a food blogger asked it to write a recipe for healthy chocolate chip cookies, and it did it; and a user asked it to write drawing prompts for input into another human. Smart painting app Midjourney does just that. Midjourney successfully creates works of art based on its text descriptions.
It is said that when Roxana Daneshjou, a dermatologist at Stanford University School of Medicine, was studying the application of AI in medicine, she asked it many medical questions and received sufficient answers... Many intelligent chatbots have appeared before, but none of them ChatGPT is so amazing. ChatGPT can hold long, fluent conversations, answer people's questions, and produce almost any type of written material people ask for, including business plans, advertising campaign proposals, poetry, jokes, computer code, and movie scripts. ChatGPT's response time is very short, it generates answers within a few seconds, users don't have to wait, and a lot of the content it generates is of good quality.
The main features of ChatGPT ChatGPT
suddenly aroused public opinion this time, which is quite incredible in the eyes of industry insiders. Even OpenAI, which developed ChatGPT, did not expect that they could successfully obtain a US$10 billion investment from Microsoft. Microsoft has invested heavily in OpenAI and announced that in addition to Office, the search engine Bing will also fully integrate ChatGPT, which is bound to break Google's 20-year monopoly on search engines. This move forced Google to invest in competitors and completely change its business organization. OpenAI is an American AI laboratory and a non-profit organization whose function is to promote and develop friendly artificial intelligence to benefit mankind as a whole.
OpenAI was founded at the end of 2015 by Elon Musk and former YC President Sam Altman.
Wikipedia information shows that from a timeline perspective, OpenAI was established at the end of 2015. The organizational goal is to open patents and research results to the public through free cooperation with other institutions and researchers. In 2016, OpenAI announced that it would build a universal robot, hoping to prevent artificial intelligence 21. The founders are Elon Musk and former YC president Sam Altman (Sam Altman). The goal is to freely collaborate with other institutions and researchers Cooperate to open up patents and research to the public and promote artificial intelligence to play a positive role. On March 1, 2019, the OpenAI LP subsidiary was established with the goal of profitability and commercialization. On July 22, 2019, Microsoft invested in OpenAI 10 billion, the two parties cooperated to develop artificial intelligence technology for Azure (Microsoft's cloud service). On June 11, 2020, OpenAI announced the launch of the GPT-3 language model, and Microsoft obtained an exclusive license on September 22, 2020. November 30, 2022 On the same day, OpenAI released a natural language generative model called ChatGPT, which interacts in a conversational manner. In January 2023, Microsoft and OpenAI negotiated an investment of US$10 billion, and hoped to integrate OpenAI's artificial intelligence technology into Word, Outlook, and PowerPoint and other applications.
In the uncertain global economic environment, the birth of ChatGPT seemed to be a breath of fresh air. ChatGPT was entrusted with the good wishes of using smart technology to improve the world economy and promote social progress. Since the emergence of ChatGPT, suddenly everyone is talking about how artificial intelligence affects their work, study and life. The reason why ChatGPT shocks everyone is that its user experience greatly exceeds previous human-computer dialogue products. Ordinary users feel that they are no longer talking to an "artificial retarded person". ChatGPT has a deep understanding of the problem and the text it generates is also very smooth. It really seems like a "person" is replying. There are even engineers trying to use ChatGPT to improve the smart home experience. It is said that a senior web developer used SiriShortcuts to create a smarter voice assistant in less than an hour by interacting with the GPT-3 large model behind ChatGPT. This voice assistant can not only control the entire Apple HomeKit smart home system, but also can easily answer various questions with ultra-low latency. He gave ChatGPT a very high rating, saying that after trying this product, all "smart" assistants, including Apple's Siri, Amazon Alexa, and Google Home, seemed so stupid and useless. In addition, there are sensational claims that ChatGPT will replace some human workers: software developers, web developers, programmers, advertisers, journalists and other content creators, as well as lawyers, market research analysts, Teachers, financial analysts, and entrepreneurs. Now Microsoft is just back, trying to use OpenAI’s ChatGPT to shatter the halo Google has accumulated through its investments in DeepMind, Boston Dynamics, and Waymo. Microsoft CEO believes that its user experience greatly exceeds previous human-computer dialogue products. Ordinary users feel that ChatGPT improves their smart home experience. It is said that a senior web developer can use all "smart" assistants including Siri, Amazon Alexa, and Google Home in less than 1 hour. They are: software developers, network developers, programmers, and advertisers. , content creators such as journalists, as well as lawyers, market research analysts, teachers, financial analysts, financial advisors, traders, graphic designers, accountants, customer service, etc. This replacement process seems a bit cruel. In the future, from a product and investment perspective, the current customer service, translators, clerks, junior programmers, copywriters, tutors and other text-based workers will be affected by the first wave, reaching tens of millions of people. For example, India will be greatly affected. As the unemployed population increases and the industry changes drastically, a large number of language processing-related companies will lose value, and no one will care about the voice assistants that have been popular for a while...
Compared with the previous InstructGPT, ChatGPT has a slightly different training process. The previous InstructGPT model gave an input and an output, and then compared it with the training data. If it was right, there would be a reward, and if it was wrong, there would be a penalty; now ChatGPT only has one input, and the model gives multiple outputs, and then "people" give The output results are sorted, and then the model is asked to sort these results from "understanding" to "unknowing", so that the model can learn the way humans sort. This strategy is called supervised learning. In summary, the difference between ChatGPT and InstructGPT lies in how the data is set up and used for training (and collection). In early 2023, OpenAI proposed that it would release a more powerful GPT-4 in the near future. It is said that GPT-4 will be released in 2024. It will be able to pass the Turing test and be as advanced as humans. In addition, the cost for enterprises to access and use GPT-4 will also drop significantly. Some experts speculate that GPT-4 may make progress in multi-modality, that is, it will introduce video, audio, etc.
Human-computer interaction system
This kind of industrial change and model innovation will be reflected in at least the following aspects:
1. Change the existing human-computer interaction model. Users will be able to interact with smart products using natural dialogue. Because ChatGPT can more accurately understand the user's intentions, subsequent software and service calls can be more in line with the user's needs, thereby improving interaction efficiency and task success rate. This change in human-computer interaction mode will change the way current apps are used. For example, more functions will be integrated into Apps, and there may even be super-universal Apps with "unified" capabilities.
2. Change the information distribution and acquisition model. Based on cognitive intelligence technology, more efficient information integration and knowledge recommendation can be achieved. Take search as an example. Traditional search engines match content based on keywords, and users need to filter out useful information from massive search results. However, the Bing search engine powered by ChatGPT can directly integrate service calls, which can better meet the needs of users. Improve interaction efficiency and task efficiency. Taking search as an example, traditional search engines match content based on keywords, and users need to give answers online. This improves the matching between questions and answers, greatly improving the user experience. The changes in the information distribution and acquisition model caused by ChatGPT affect the distribution of traffic and change the business model of traffic monetization.
Core technologies of deep learning
The emergence of the GPT pre-training model represents a technological leap in the field of natural language processing, both from an academic research perspective and from a scene application perspective, and has brought about a transformation of the research paradigm in the entire field.
Under the framework of large models, the number of parameters of each generation of the GPT model used by ChatGPT is rapidly expanding, and the data volume requirements and costs of pre-training are also rapidly increasing. ChatGPT official website attracted as many as 25 million daily visitors from January 27, 2023 to February 3, 2023.
Assuming that at the current stable state, each user asks about 10 questions every day, there will be approximately 250 million consultations per day. According to the report "How much computing power does ChatGPT require" published by Guosheng Securities computer analysts Liu Gaochang and Yang Ran on February 12, 2023, an average of about 13 million unique visitors used ChatGPT every day in January 2023, and its corresponding chip demand For more than 30,000 NVIDIA A100 GPU graphics processors, the initial investment cost is about US$800 million, and the daily electricity bill is about US$50,000. The cost of GPT-3 training once is about US$1.4 million. For some larger LLMs, Training costs range from $2 million to $12 million. This is not cheap for global technology giants, but it is still acceptable. For ChatGPT technology, leading Internet giants such as Google and Meta will master it sooner or later, but the arms race between technology companies will still unfold.
On the one hand, the reasoning cost of ChatGPT is high, and the computing power consumed is highly positively correlated with user experience; on the other hand, search engine user stickiness is low, and users will always use the one with the best experience, which will greatly increase the efficiency of the search algorithm. The cost makes it impossible for the search advertising business that Google relies on to survive to make money. Microsoft's revenue will be more diversified, so the impact will be smaller. It can enjoy visitors using ChatGPT. The corresponding chip demand is more than 30,000 NVIDIA A100 GPUs. The cost of graphics graphics is high, and the computing power consumed is different from users. Experience is highly positively correlated; on the other hand, search engine user viscosity is small, and users will always use the one with the best experience. This will significantly increase the benefits of searching for the Edge browser and even increasing the market share of Windows systems. To put it bluntly, Microsoft is now equivalent to starting a large-scale arms race and wants to "kill" its competitors.
In the early years, companies met the power needs of their operations by generating their own electricity, which not only cost a lot of money but also required some special capabilities that were not closely related to the company's business. The operation of the power grid infrastructure has made power supply a public utility, which also enables enterprises to purchase electricity instead of generating electricity themselves. In essence, enterprises turn their own power generation into purchasing power generation services. Centralized power generation can make the use of electricity more efficient, which means that more businesses and even individuals can buy electricity according to their own needs without having to pay anything for other electricity. The publicization of electricity supply has increased the productivity of various departments, improved the quality of social life, and created development opportunities for emerging industries. The information and communications technology industry is undergoing a similar evolution. For decades, institutions and individuals have purchased information and communications technology like a commodity by investing in computer software and hardware.
In the past decade, the supply of information and communication services has been rapidly updated. With the popularization of high-speed broadband infrastructure, information and communication technology can be purchased as a service through the Internet. Today, computing power is like electricity, which is widely used by everyone, and is potentially destructive and transformative at the same time. If AI is compared to electricity, then the computing power of large models is equivalent to a generator, which can popularize intelligent applications on a larger scale. The intelligent capabilities of large models will become a public basic resource in the future, as readily available as electricity or running water. In the future, every smart terminal, every App, and every smart service platform can be connected to the computing network composed of IT infrastructure just like the power grid, allowing AI algorithms and technology to be more widely used in all walks of life. Industry. If users from all walks of life want to use services but do not want to purchase, install and run expensive computer hardware, they can use ubiquitous wired or wireless networks to obtain computing power from the cloud, which is no different from using other public services.
Research has estimated the huge computing power requirements and capital consumption
. Training GPT-3, a large language model with 175 billion parameters, requires tens of thousands of CPUs/GPUs to input data 24 hours a day, and the energy consumption required is equivalent to driving back and forth. The Earth and the Moon, and one operation costs $4.5 million.
ChatGPT demonstrates three powerful capabilities: ● Language generation: following the prompt word (Prompt), and then generating sentences that complete the prompt word. This is also the most common way for humans to interact with language models. ● In-context Learning: Follow several examples of a given task and then generate solutions for new test cases. It is worth mentioning that although GPT-3 is a language model, its focus is not language modeling (Language Modeling), but context learning. ● World knowledge learning: including factual knowledge (Factual Knowledge) and common sense (Commonsense). The above three capabilities all come from large-scale pre-training: a model with 175 billion parameters is pre-trained on a corpus of 300 billion words (60% of the training corpus comes from the Common Crawl corpus from 2016 to 2019, and 22% comes from the WebText corpus , 16% from books and newspapers, 2% from Wikipedia).
Having the ability to create has the potential to surpass artificial intelligence in terms of professional capabilities and popularization. It also means that it begins to possess human thinking abilities and may replace humans in more and more aspects.
A picture of infinite possible futures.
