Yanghua Xiao|Generative Language Model and General Artificial Intelligence: Connotation, Path and Enlightenment

Large-scale generative pre-trained language models represented by ChatGPT have led to the rapid development of a series of general artificial intelligence (AGI: Artificial General Intelligence) technologies. AGI has set off a new round of information technology revolution and has become an advanced productive force, and it is particularly urgent to deeply understand the nature of AGI. The general artificial intelligence technology represented by large-scale generative language model takes generative AI as the main form and has the ability to generate contextually, forming an intelligent refining path in three stages of knowledge, ability and value. With the development of related technologies, the intelligence level of machines is rapidly improving, which will bring blurred boundaries between man and machine and a series of social problems related to them. The development path of AGI has the characteristics of "cramming indoctrination" learning and "first pass and then specialize", which to a certain extent subverts the traditional understanding of the realization path of machine intelligence, and forces human beings to reflect on world modeling, knowledge acquisition, self-awareness and other levels. Mankind needs to be highly vigilant about the challenges brought by AGI, and actively seize the opportunities brought by it to promote the construction of a new type of human-machine harmonious relationship.

About the author:

Xiao Yanghua is a professor and doctoral supervisor at the School of Computer Science and Technology, Fudan University, and the director of Shanghai Key Laboratory of Data Science. His research interests include knowledge graph, knowledge engineering, and big data. His main works include "Graph Symmetry Theory and Its Application in Data Management", "Knowledge Graph: Concepts and Techniques" (co-authored), etc.

I. Preface

Since the release of ChatGPT in December 2022, large-scale generative pre-trained language models have caused an uproar in academia and industry, driving the rapid development of a series of general artificial intelligence technologies (AGI: Artificial General Intelligence), including graphic generation models, such as Midjourney's high-precision, highly simulated graphic and text generation; Embodied multimodal language models, such as Google's successive PaLM-E (D. Driess et al., 2023) and PaLM 2 (A. Rohan et al., 2023). AGI has rapidly evolved from simulating the thinking ability of the human brain (represented by language models) to embodied models that "manipulate the body" (represented by embodied large models). AGI has comprehensively invaded all territories of human intelligence, from artistic creation to code generation, from problem solving to scientific discovery, from question and answer chat to assisted decision-making, and almost all fields that human intelligence can involve have traces of AGI. A new round of information technology revolution driven by AGI has swept through. Mankind has ushered in a technological revolution about "intelligence" itself. As an advanced productive force, AGI brings both exciting opportunities and worrying challenges to the whole society. The excitement and apprehension boil down to the fact that our understanding of AGI is still far from keeping pace. Specifically, human thinking about AGI technical principles, intelligent forms, and upper limits of capabilities, and the evaluation of its impact on society and individuals, lag behind the development speed of AGI. It can be said that the rapid development of AGI and the significant lag of human cognition constitute a pair of sharp contradictions, and grasping this contradiction is the key to understanding the current development law of AGI and its social impact. It is precisely based on the understanding of the above contradictions that many scientists and AI business leaders have issued a call to suspend giant model experiments, calling for accelerating the development of safe and provable AI systems. Admittedly, understanding AGI is difficult. The three words in the term AGI express the challenges of understanding AGI from different perspectives. From its core word "Intelligence", there have always been different views on what intelligence is, such as traditional computer science believes that the ability to "acquire and apply knowledge and skills" is intelligence, but it is necessary to consider whether this definition still applies to AGI represented by large-scale generative language models today. The term "general" exacerbates the difficulty of understanding AGI. Compared with traditional AI for specific functions, AGI aims to simulate human mental capacity, and the uniqueness of human intelligence is clearly reflected in its ability to adapt to different environments and be able to perform different types of tasks that have never been seen before. What is the connection and difference between dedicated AI and general AI, and should general AI be implemented first or private AI first? The word General will lead to a lot of thinking like that. The term "artificial" speaks to the nature of AGI's artificial creations, rather than intelligence that spontaneously evolved from its natural environment. This naturally raises a series of questions about the similarities and differences between instrumental intelligence and natural intelligence. Despite the challenges, this article attempts to analyze some aspects of AGI. This paper focuses on generative artificial intelligence, especially general artificial intelligence technologies represented by large-scale generative language models. The "intelligence" mentioned in this article is not limited to human intelligence, but also includes machine intelligence, and will use machine intelligence and human intelligence as a reference for comparative analysis. This paper will analyze in detail the connotation of "intelligence" and the evolution path of "intelligence" caused by the development of generative language models, and reflect on many aspects of human intelligence on this basis, including creativity, world modeling, knowledge acquisition, and self-cognition. The author believes that the thinking of this article can eliminate people's concerns about the rapid progress of machine intelligence on the one hand, and on the other hand, it can also remove obstacles to the further development of machine intelligence and help establish a new type of human-machine harmonious relationship. It should be noted here that some of the thoughts and conclusions of this paper are beyond the scope of current engineering practice, and still need to be strictly demonstrated and tested. 2. What is intelligence? How did ChatGPT succeed? Generative VS Discriminant. ChatGPT is the representative of generative artificial intelligence. Generative AI has achieved good results in the fields of text generation, text and image generation, and image generation. Traditional artificial intelligence is mostly discriminative artificial intelligence. Why is generative AI rather than discriminative AI the main form of AGI? This is a question worth pondering. Discriminant AI, through the training of labeled data, guides the model to learn the ability to correctly give answers to questions. Generative AI is often trained on self-supervised learning tasks based on masking content restoration for unlabeled data, and guides the model to generate content that meets the context context. Generative models not only have the ability to generate results, but also generate processes and explanations. Therefore, the generative task can be regarded as a more intellectually challenging task than the discriminant task, which can effectively guide the model to acquire a high level of intelligence. Specifically, for judgment questions, discriminative AI only needs to give a right or wrong answer, and even if it is randomly guessed, there is still a 50% probability that it is correct. However, generative AI not only needs to generate answers, but may also need to generate the problem-solving process at the same time, which makes it difficult to muddle through. Therefore, compared with discrimination, generation can be said to be a type of task closer to the essence of intelligence. Intelligence and contextualized generation. What is the essence of intelligence? The development of large models has brought a lot of new inspiration to human thinking about this problem. The intelligence of a large model is essentially the ability to generate contextualized generation, that is, the ability to generate relevant text based on contextual prompts. Therefore, the application effect of large models depends to a certain extent on whether the prompt is effective or not. If we can give a valid and reasonable hint, then large models such as ChatGPT can often produce satisfactory answers. This contextualized generation capability (the ability to "prompt + generate") is not only applicable to text, but also to a wide range of complex data of various types, such as images, speech, protein sequences, etc. Different data contexts are different, for example, for images, the context is peripheral images. The contextualization generation ability of large models is formed through in-context learning in the training stage (Q. Dong et al., 2022）。 In terms of mathematical essence, large models learn the joint probability distribution between tokens or corpus basic units in the training stage. Contextualized generation can be thought of as a conditional probability estimate, i.e., given context or cue (i.e., given evidence), the probability of remaining text appearing based on the joint distribution. The traditional understanding of intelligence is more or less related to "knowledge" (such as defining intelligence as "the ability to discover and apply knowledge") or to people (such as defining intelligence as "the ability to think and act like a person"), and its essence is still human-centered and understands intelligence from an epistemological perspective. The situational generation ability presented by the big model has nothing to do with "knowledge", "knowledge" is ultimately a man-made invention made by human beings to understand the world. The existence of the world does not depend on "knowledge", does not depend on human beings, situational generation gets rid of the "knowledge" defined by human beings, and returns to the world itself - as long as the world can be reasonably generated, it is intelligence. Intelligence is reduced to a generative ability, which can be non-human-centric or dependent on human civilization, which is an important revelation brought to us by AGI. Intelligent analysis and restoration. The large model training and optimization process can provide useful inspiration for us to better understand the formation process of intelligence. The "release" of the general large model basically goes through three stages (W. X. Zhao et al., 2023): The first stage is the training of large models with bases; The second stage is task-oriented instruction learning, also known as instruction fine-tuning; The third stage is value alignment. The training of the first stage of the base large model is essentially to let the large model acquire the knowledge contained in the corpus or data. But the knowledge here is a parametric, probabilistic knowledge (essentially modeling a joint distribution between words in the corpus), which makes contextualized generation possible. Therefore, the essence of the first stage is knowledge acquisition (or knowledge acquisition), the second stage of instruction learning aims to allow the large model to acquire the ability to complete the task, and the last stage is the acquisition of values. The intelligence of large models is broken down into three stages: knowledge, ability and value, which is a noteworthy feature. Knowledge is the foundation of ability and value, so the "refining" of the base model is particularly critical. ChatGPT has undergone nearly four years of training and optimization from the first version of GPT-1 in 2018 to GPT-3.5 in 2022. The deeper and broader the knowledge base of the large model, the more complex and diverse the skills that can be learned in the future, the more accurate the value judgment and the more agile the value alignment. The big model separates the three core elements of intelligence from each other, and human knowledge, ability and value acquisition are often mixed together. It is difficult to define whether an article in an elementary school textbook is imparting knowledge, training skills, or shaping values. This discrete intelligent development of the large model can be compared to the higher education of human society. Undergraduate education in human society aims to cultivate the ability to learn to acquire knowledge, master's education aims to cultivate problem-solving ability to solve problems, and doctoral education aims to cultivate the ability to make value judgments to discover problems. The separation of knowledge, capability and value has positive enlightenment significance for the future intelligent system architecture, the establishment of a new human-machine collaboration relationship, and the design of human-machine hybrid intelligent system architecture. With the gradual development of machine intelligence, the things that humans are good at relative to machines will gradually decrease. However, there is still some room for human intervention in certain scenarios. The key to the future development of human-machine hybrid systems is still to answer what work is most worthy of human work. Only after the seemingly complete task can the subtasks that the man and machine are good at and suitable can be disassembled. For example, separating knowledge and capabilities is extremely valuable for protecting private domain knowledge: large models are responsible for core tasks such as language understanding, while confidential data and knowledge are still managed by traditional databases or knowledge bases. Such a system architecture not only makes full use of the core capabilities of the large model, but also fully takes into account the privacy of knowledge. Intelligent testing and human-machine differentiation. The development of general artificial intelligence technology has significantly improved the level of machine intelligence, especially the level of language understanding, and machines have reached the level of ordinary humans and even language experts in text processing, language understanding and other related tasks. A key problem is the increasingly blurred boundaries between man and machine. It's hard to tell whether you're communicating behind a window or a machine just by a few rounds of conversation. In other words, the traditional Turing test is no longer up to the task of human-machine differentiation. Anyone who has used ChatGPT knows that what ChatGPT is best at is chatting, and even if we chat with it for a long time, we may not find it boring. The blurring of the human-machine boundary will bring many social problems. First, ordinary people, especially teenagers, may indulge in ChatGPT-like conversational models out of trust in technology. As ChatGPT becomes smarter, we get used to asking it questions, getting used to receiving its answers, and over time, the questioning spirit on which humans rely will gradually lose. In the face of the increasingly powerful AGI, how to avoid the degradation of the spiritual essence of man? These are questions that we need to think about and answer. Second, when it is difficult to distinguish between true and false information and false information, fraud will emerge one after another. Recently, more and more criminals have successfully carried out many fraud cases through AI face swapping and AI video generation. How to deal with the social deception caused by the blurring of human-machine boundaries will become a very important AI governance issue. Finally, it's worth noting captcha, an app that we use widely in our daily lives but can quickly become a problem. The verification code is our weapon for man-machine differentiation, but with the development of AGI, especially after its increasing ability to control various tools, the human-machine differentiation function of the verification code will face increasingly severe challenges. As humanoid robotics matures, how to prove that you are a human and not a machine, or vice versa, how to prove that a machine is a machine and not a human will become an increasingly difficult problem in the future. The blurring of the human-machine boundary essentially comes down to the problem of human-machine intelligence testing. We need to portray territories that are unique to human intelligence and that cannot or at least are difficult to infringe upon by machine intelligence. From the history of machine intelligence, the scope of this territory will become narrower and narrower. We used to think that it was difficult for machines to surpass humans in highly intelligence-intensive activities such as playing Go, and we used to think that it was difficult for machines to surpass humans in high-quality conversations, and we once thought that scientific discoveries such as protein structure prediction were difficult for machines to surpass humans... The list of tasks for these machines to struggle with humans, once long, is now getting shorter. The Turing test has failed, but humans have not yet had time to come up with new effective alternative test schemes. It has been suggested that only human beings can make mistakes and the uncertainty of their actions are unique to human beings. This view is not worth refuting, because machines can easily implant errors and uncertainties to disguise their intelligence. How we will prove in the future that a machine is trying to jailbreak, and whether a machine is disguising its abilities, these are issues that require high attention from AI security. Third, the evolution route of intelligence, how to develop and progress in general artificial intelligence? "Feedback evolution" and "cramming indoctrination". Human intelligence is a typical biological intelligence that has been formed after a long period of evolutionary development. Human beings continue to practice, receive feedback, and continue to experiment in the natural and social environment, forming a high degree of intelligence. The intelligence of all kinds of animals can be classified as evolutionary intelligence. The evolution of evolutionary intelligence takes a long time, in other words, given enough time, the natural environment may be able to shape any level of intelligence. Lower animals may also develop advanced intelligence over a long period of time. However, the current machine intelligence takes a "cramming" path, which is a shortcut to achieve advanced intelligence. All the corpus, books, and documents that human society has accumulated are "instilled" into the big model, and after careful "refining", the large model can learn the achievements of civilization accumulated by human beings for thousands of years. Although it also takes days and months to "refine" large models, the long evolution process relative to human intelligence is almost instantaneous. It is a miracle in itself that machines can acquire thousands of years of accumulated knowledge in such a short period of time. Human society mostly regards "cramming indoctrination" as a mechanical and inefficient way of transferring knowledge, but this has become an efficient way for humans to transfer knowledge to machines. If students are evaluated solely by test scores, crude cramming and indoctrination education is very effective. However, this kind of education produces students who often have high scores and low energy, and it is difficult to flexibly apply knowledge to solve practical problems. So our students also need to receive a lot of practical education, learn from feedback, and eventually become experts and integrate knowledge. The cultivation process of human experts is very instructive for understanding the development of large models. At present, the cramming learning stage of the large model has been basically completed, and soon the large model will control various tools and carry out practical learning, thus entering a new stage of knowledge learned from practice. "First through and then specialized" or "first specialized and then through". Another enlightenment brought to us by the development of general artificial intelligence is that machine intelligence has embarked on a development path of "first through and then specialized". From the perspective of the application of large-scale language models, it is first necessary to "refine" the general large language model, generally speaking, the more extensive and diverse the training corpus, the stronger the ability of the general large model. But such a general-purpose large model still does not work well when completing the task. Therefore, it is generally necessary to go through domain data fine-tuning and task instruction learning to make them understand the domain text and be competent for specific tasks, which shows that the intelligence of large models is first universal and then professional. The General Intelligence stage focuses on general learning, including language understanding and reasoning skills and a wide range of general knowledge; The professional intelligence stage allows the large model to understand various task instructions and be competent for various specific tasks. Such an intelligent evolution path is similar to the human learning process. Basic education for human beings focuses on general studies, while higher education focuses on specialized learning; Kung Fu masters in martial arts novels often practice internal strength first and then moves. These are similar to the development path of the large model "first through and then specialized". The development path of "first through and then specialized" of large models subverts the mainstream development path of artificial intelligence in the past. Before the birth of ChatGPT, the main position of AI research was dedicated AI or functional AI, which mainly aimed to make machines capable of specific scenarios and tasks, such as chess, computing, speech recognition, image recognition, and so on. The traditional concept is that several specialized intelligence can be stacked together to approach general intelligence; Or if professional intelligence cannot be achieved, it is even less likely to achieve general intelligence. It can be seen that "first specialize and then pass" is the basic consensus of the development of traditional artificial intelligence. However, the large-scale generative language model represented by ChatGPT basically subverts this traditional understanding and shows that machine intelligence, like human intelligence, requires general knowledge before it can develop professional cognition. Under the new understanding, we need to re-understand domain-specific AI. The domain is relative to the general. In fact, without universal cognitive abilities, there is no domain cognitive ability. For example, healthcare is a typical vertical field, and the traditional concept is that intelligent systems for diagnosing certain types of diseases can be built at a low cost. For example, for tinnitus diseases, traditional methods generally instill relevant professional knowledge, text, and data into machines in order to achieve intelligent diagnosis of tinnitus, a very subdivided disease. But in practice, the idea never really succeeded. At the root of this, doctors need to understand health in order to understand disease, and health does not belong to the category of disease. An otologist spends most of his or her visit checking for health conditions that do not require treatment. That is, to truly understand a domain, you need to recognize concepts outside the domain. It can be seen that domain cognition is based on general knowledge. These new understandings bring new inspiration to our redevelopment of domain cognitive intelligence, and it can be said that with the support of the general large model of ChatGPT, cognitive intelligence in various fields will usher in new development opportunities. Symbol before experience, form before content. Large-scale language models are trained by using a corpus expressed by text or symbols. Human natural language is a symbolic expression, and language models express statistical associations between linguistic symbols. However, symbols are only forms, and statistical learning based on symbols alone is not enough for machines to understand the meaning of symbols or language. Intelligent systems with purely formal symbols are bound to be criticized by John Searle's "Chinese house" ideas. Therefore, AGI does not stay in the stage of simple language model, but actively integrates multimodal data for hybrid training. All kinds of multimodal data, such as images, voice, and video, can express the rich human experience of the world (X. Zhu et al., 2022）。 For example, people's understanding of the symbol "horse" depends to some extent on people's experience and understanding of the horse as an animal, such as high-pitched neighing (voice), robust image (image), galloping action (video). Human experience supports people's understanding of the concept of "horse", just as people's sad experience of Wan Ma Qi is based on the experience of a healthy and positive image of horses. Therefore, AGI has embarked on a development path of first symbol and then experience, from form to content. This is the opposite of the development of human intelligence, where human beings first have rich experience or experience before abstracting into symbols, words and concepts. "Brain before body" and "body before brain". At present, the development trend of AGI is to first develop language models to simulate the cognitive ability of the human brain, and then drive various tools and body parts based on the cognitive ability of the machine brain. The brain's complex planning and reasoning abilities are indispensable for the body's interaction and movement with tools in the real world. AGI has embarked on a development route of "first realizing the cognitive ability of the brain, and then realizing the ability of the body to interact with the physical world". It is clear that this route of development of AGI is significantly different from the evolution of human intelligence. To a certain extent, humans first acquire physical abilities and shape and develop the cognitive abilities of the brain during the continuous interaction between the body and the world. The traditional artificial intelligence technology route also tends to realize the basic functions of various organs or components of the body first, and then realize the complex cognitive ability of the brain, and tends to accept the view that the ability of the mechanical body to interact with the real world is easier to achieve than the complex cognitive ability of the brain. However, the current AI development path subverts our traditional understanding of the machine intelligence implementation path to a certain extent. The generalization of human self-examination and revelation combinations triggered by general artificial intelligence is a kind of creation, but it may be a low-level form of creation. A very important reason why AGI has attracted great attention from the industry is that it shows a certain creative ability. We found that ChatGPT or GPT-4 already has a relatively powerful combinatorial generalization ability: large models can be qualified for some new combinatorial tasks after learning instructions for sufficient common tasks. Specifically, when the large model learns to complete two types of tasks, A and B, it can complete new tasks such as A+B to a certain extent. GPT-4, for example, was able to write mathematical theorem proofs using the style of Shakespeare's poetry. In fact, this is due to GPT-4's acquisition of two abilities, mathematical proofs and writing Shakespearean poems, respectively, and then combined to generalize the new abilities. First, we must recognize this combined innovation ability of large models. In contrast, many innovations in human society are essentially combined innovations, and this form of innovation even accounts for the vast majority. For example, in the technological innovation in the field of engineering, many graduate students are good at applying the B method proposed for scene A to scene X and have achieved good results; The mediocre plot creation in popcorn movies mostly borrows the framework of story A, the characters of story B, the plot of story C, the bridge of story D, and so on. Second, AGI's ability to combectively innovate far exceeds the level of human cognition. AGI can combine the abilities of any two disciplines, many of which may never have been imagined, such as writing code annotations in the style of Li Qingzhao's poetry. This novel combination of innovation capabilities may be a valuable asset brought to us by AGI and will greatly stimulate human imagination. Third, AGI's ability to combine innovation basically declares that the collage content innovation of human society will lose its meaning. Because AGI can combine innovative materials, and its generation efficiency is far beyond that of humans. The integrated innovation that we once were proud of will also lose its aura, and the original innovation will be even more valuable in the face of AGI. Fourth, AGI's combinatorial innovation will force humanity to rethink the nature of innovation. The innovations that humans can make and AGI cannot achieve will be even more valuable. AGI will encourage humans to no longer indulge in random splicing or simple assembly creation, but pay more attention to content creation with rich connotation, unique perspective and novel viewpoint. Self-supervised learning is an effective way to model the world. Self-supervised learning can be thought of as a fill-in-the-blank game, that is, filling in gaps based on context. For example, we mask a word in a complete sentence beforehand, and then have the machine restore the covered word according to the context of the sentence. Similarly, in terms of images, we can block part of the image area, so that the large model restores the content of the occluded image according to the surrounding background image. Why such a self-supervised learning paradigm can achieve large-scale pre-trained language models such as ChatGPT is a question worth pondering. The "mask+reduction" style self-supervised learning task aims to acquire the world model. For example, people know that throwing heavy objects at height, the object will definitely fall, but it will not float upwards and cannot be suspended in the air. Recently, many scholars, including Turing Award winner Yann LeCun, have pointed out that the world model (Y. Lecun, 2022) for AGI. The data that human society has accumulated reflects human understanding of the real world, and through the learning of this data, machines will have the opportunity to build a model of the world. When the data is enough, precise enough, and rich enough, it can express the complete cognition of the complex real world to a certain extent, and based on the self-supervised learning mechanism of "masking + reduction", the machine can realistically establish a model of the world. On the other hand, the human world model comes to a large extent from experience and civilization inheritance. On the one hand, we form experiences in the process of interacting with the world to build a model of the world; On the other hand, cultural transmission and educational inheritance shape our perception of the world. So the way humans model the world is fundamentally different from the way machines model the world. Tacit knowledge learned by large models. Large-scale pre-trained language models are developed with the help of Transformer(A. Vaswani et al., 2017), a deep neural network architecture that learns statistical correlations between language elements and has the ability to contextualize generation. The large size of the large model is mainly reflected in the huge number of parameters. Such a complex deep cyberspace encodes the various knowledge contained in the corpus, which has two distinctive characteristics of parametric expression and distributed organization. The so-called distributed organization means that a certain knowledge does not correspond to a specific neuron, but is dispersed and expressed as weight parameters of different neurons and the interconnection structure between them. With specific inputs, knowledge is obtained computationally by activating certain neurons. Therefore, a large model can be thought of as a container for tacit knowledge. The tacit knowledge encoded by the big model is significantly beyond the explicit knowledge that humans have expressed. In a sense, the knowledge that humans can express in natural language is exhaustible and limited. The common sense used by humans subconsciously, the meaning of words in texts, the experience that is difficult for domain experts to express, etc., all exist in the form of tacit knowledge. Large models provide us with more possibilities for understanding this tacit knowledge. The big model is a generalist, which is trained and generated by using the corpus of all humans and disciplines, and some implicit associations or statistical patterns it has learned may correspond to unspeakable tacit knowledge of human beings. For example, in the diplomatic scene, the choice of words and sentences often has special meanings, and the emergence of large models has brought new opportunities to interpret such meanings and unique connotations. Much of the knowledge encoded by the big model has never been deciphered by humans, especially the implicit correlation between interdisciplinary knowledge points. This is also a major opportunity brought by the big model to the development of our entire human civilization. As the interpretation of tacit knowledge by large models deepens, human knowledge will show explosive growth. We have to ponder a profound question: whether excessive knowledge will become an unbearable burden for the development of human civilization. In fact, when knowledge accumulates to a certain extent, simple knowledge acquisition has deviated from the main channel of human civilization development. In a future of rapid growth in knowledge, discovering "wisdom" is more important than acquiring "knowledge." Many times, we don't need much knowledge, just the ability to get knowledge from large models. Theoretically, the amount of knowledge that each individual of human beings (even the most outstanding elites of mankind) can know must be much lower than that of intelligent machines. The value of each of us is not reflected in how much knowledge we have, but knowing how to use knowledge, and the wisdom of using knowledge will be the core value of individual human beings. The development of AGI has forced the development of human society to move from the pursuit of knowledge to a new stage of pursuing wisdom. The big model forces humans to rediscover themselves. AGI technology will be deeply integrated with the process of human social development, bringing unprecedented major opportunities and severe challenges to human society. With the rapid development of artificial intelligence technology, the risks brought by AGI have gradually become prominent. First of all, AGI brings challenges to AI technology governance and social governance. Loss of control of AGI will have more catastrophic consequences than current artificial intelligence. At present, the risk of AGI technology "getting out of control" is increasing, and timely intervention is necessary. For example, AGI has lowered the threshold for content generation, leading to the proliferation of false information, which has become a serious problem. For example, if AGI, as an advanced productive force, cannot be grasped by the majority but in the hands of a few people or institutions, technological hegemony will have a negative impact on social development. Second, AGI technology will bring challenges to the development of human individuals. In the future, social production seems to be completed by a few elites plus intelligent machines, and the 2/8 rule of the industrial age may become the 2/98 rule in the AGI era. In other words, more and more jobs and tasks may lose their meaning in the face of a strong AGI, and the value and meaning of individual existence need to be redefined. Our lifespan may be greatly extended, but the quality of life is gradually weakening. How can we help the vast majority of us find meaning in life? How to pass your leisure time elegantly? These are all questions that require deep thought. Finally, advances in AGI may carry the risk of overall human regression. When humans developed poultry technology, hunting technology regressed significantly; As textile machines matured, embroidery techniques became unnecessary. Our various intangible cultural heritages and sports are essentially preventing human regression. Just because a machine is good at a certain job or task for a human cannot be allowed to gradually deteriorate. If the progress of various technologies in the past has only gradually distanced human beings from the primitive state of nature, the regression of the limb ability developed by human beings in the struggle against the harsh natural environment is a sacrifice that must be made in the development of human civilization; So, will AGI, which aims to replace human brain power, cause a regression of human intelligence? The regression of intelligence will inevitably lead to the loss of human subjectivity and the collapse of civilization. How to prevent the regression of our brain power or intelligence is a problem that must be seriously considered. Despite the challenges, AGI is undoubtedly an advanced productive force, and its momentum is unstoppable. In addition to the specific technological empowerment mentioned above, the new opportunities brought by AGI should be re-emphasized from the height of the development of human civilization. First of all, AGI is of great significance for accelerating the process of human knowledge discovery. As discussed earlier, the interpretation of tacit knowledge encoded by large language models will accelerate human knowledge discovery, but it will also lead to the depreciation of knowledge. In the future, we will witness the "uselessness of knowledge" brought about by the explosion of knowledge. Second, the greatest significance of AGI development may be to force human progress. Mediocre creation loses meaning, combinatorial innovation loses meaning, exhaustive exploration loses meaning... This list is destined to grow longer and longer. But human existence cannot lose its meaning, we must rediscover our own value, and rethink the philosophical proposition of why people are human.

V. Conclusion

The exploration and thinking about AGI has just begun, and we still have a long way to go. We must be highly alert to the problems posed by AGI and pay full attention to the opportunities created by AGI. More than 2,000 years ago, Socrates said "know thyself", and today, under the pressure of the development of AGI technology, mankind needs to "know yourself again".

Author: Xiao YanghuaSource: Public Number [Academic Frontier Magazine]

Illustration by IconScout Store from IconScout

-The End-

Scan the code to watch!

New this week!

"AI Technology Stream" original submission plan

TechBeat is an AI Learning Community (www.techbeat.net) established by Jiangmen Ventures. The community has launched 480+ talk videos and 2400+ technical dry goods articles, covering CV/NLP/ML/ROBOTIS, etc.; Hold top meetings and other online communication activities on a regular basis every month, and hold offline gatherings and exchange activities for technicians from time to time. We are striving to become a high-quality, knowledge-based communication platform that AI talents love, hoping to create more professional services and experiences for AI talents, and accelerate and accompany their growth.

Contents

Latest Technology Interpretation/Systematic Knowledge Sharing //

Cutting-edge information commentary/experience narration //

Instructions for submission

Manuscripts need to be original articles and indicate author information.

We will select some directions in in-depth technical analysis and scientific research experience, inspire users with more inspirational articles, and do original content rewards

Submission method

Send mail to

[email protected]

Or add staff WeChat (chemn493) to submit articles to communicate the details of submissions; You can also pay attention to the "Jiangmen Venture Capital" public account, and reply to the word "submission" in the background to get submission instructions.

>>> Add WeChat!

About me "door" ▼

Jiangmen is a new venture capital institution focusing on the core technology field of digital intelligence, and is also a benchmark incubator in Beijing. The company is committed to discovering and cultivating scientific and technological innovation enterprises with global influence by connecting technology and business, and promoting enterprise innovation and development and industrial upgrading.

Founded at the end of 2015, the founding team was built by the founding team of Microsoft Venture Capital in China, and has selected and deeply incubated 126 innovative technology-based startups for Microsoft.

If you are a start-up in the technology field and want not only investment, but also a series of continuous and valuable post-investment services, please send or recommend a project to my "door":