Focus:
- The development of 1AIGC did not produce a fundamental paradigm change from the earliest deep neural networks. We can't conclude from the fact that ChatGPT dialogue generates well and behaves intelligently that it really produces general intelligence, because it doesn't really know what it's talking about, it just gives people the feeling that it has understood what it generates, which is still a certain distance from human intelligence.
- 2. China does need to make further efforts in the development of AIGC. The vast majority of the work is done by a handful of research institutions in the United States, which lead the development of AI technology as a whole. Therefore, we must also work hard and strive to make more contributions to the history of AI development.
- 3ChatGPT is hot, but people have also found that it comes with a lot of negative problems. The most typical thing is that it produces false information, misinformation, it produces a bunch of content that seems right but is wrong. But if too much attention is paid to these challenges, it will more or less limit the development of technology.
- 4AIGC, and future artificial intelligence technology, will definitely bring a fundamental change to our existing production tools and productivity. These changes will inevitably lead to changes in production relations, which may have a greater impact on the future of mankind and society.
In 2022, from AI models such as DALL-E 2 and Stable Diffusion in the field of AI painting, to near-human conversational robots represented by ChatGPT, AIGC continues to flood the network, and its powerful content generation capabilities have brought people a huge shock.
When AIGC becomes a hot spot in all walks of life, people can't help but ask, will AI become a new creator? Why did AIGC suddenly erupt, does it mark that AI is ushering in the next era, and how will it go? Will AIGC models with both large and multimodal models become the new technology platform? What impact will AIGC technology and application bring to the economy and society, and how should different entities view and deal with it?
Panelists:
Yao Xin Chair Professor and Head of Department of Computer Science and Engineering, Southern University of Science and Technology
Duan Weiwen is the director of the Research Office of Philosophy of Science and Technology, Institute of Philosophy, Chinese Academy of Social Sciences
Wang Yuntao is the deputy chief engineer of the Institute of Cloud Computing and Big Data, China Academy of Information and Communications Technology
Wu Baoyuan is an associate professor at the School of Data Science, Chinese University of Hong Kong, Shenzhen
Yin Jun Director of Digital Content Technology Center, R&D Performance Department, Tencent Game CROS
Shi Shuming is the director of the natural language processing center of Tencent AI Lab, and the head of Wenyong products
Compere:
Yang Jian Vice President of Tencent Group and General Counsel of Tencent Research Institute
The power and explosion of AIGC
Yang Jian (Moderator): How do you view the sudden outbreak of AIGC in the second half of 2022? Does AIGC represent the next era of AI technology?
Yao Xin: The outbreak of AIGC is actually a comprehensive success. Although AI has a long history in content generation, it has not developed to this extent in the past due to limitations such as data, computing power and algorithm technology. An important factor in the interest in AIGC right now is that it has exceeded expectations, so people think this stuff is particularly interesting. However, it is still necessary to think about what kind of technological breakthrough AIGC has.
First, technically, it is relatively easy to achieve results that exceed people's expectations. For example, in terms of images, dialogue, and even music, it can produce pieces that have a certain composer's style, but are not exactly the same. But many people think that AIGC may be an effective path to AGI, but it is actually worth careful consideration. Because now AIGC generates content based on a large amount of data and computing power. When people generate content, sometimes through an abstract process, for example, I look at a lot of images, there is an abstract process in my mind, and then from the abstract process back to the space of the image, and finally create and generate an image. I don't know too well whether AIGC has this level of abstraction in it right now. You look at 1 million images of cats and dogs, and the machine doesn't generate a concept on its own, such as 4 legs, hair, etc.
Second, the further development of AIGC in the future, involving the field of science and technology, or the application fields that are truly related to the national economy and people's livelihood, may have additional challenges. Because AIGC generates content, it is difficult to guarantee that the content it generates meets certain constraints. To use an analogy, we can now send a lot of molecules into an AI system, and it may give you new molecules. From the generation of molecules to the generation of drugs is conceivable, but where will it really land? Where is the real distance to reach the so-called AGI? This is something we need to think about and explore.
Yin Jun: First of all, I agree with Professor Yao's point of view, whether it is content generation based on artificial intelligence in a broad sense, or a deep-level large model in a narrow sense, in fact, it is accumulated and has a relatively long research history. The AIGC outbreak today is multifaceted, on the one hand, it has indeed accumulated a very large number of datasets, and it is a public high-quality dataset; On the one hand, there are some theoretical breakthroughs, such as represented by the Diffusion model; There is also the new computing hardware, represented by GBT3 from 2020, the model stack is large enough, the effect is good enough, there may not be such a strong computing equipment to support this scale of model training before, even now the cost of this model training is still very high. The superposition of several factors has led everyone to think that generating images or generating text may not be a very realistic thing before, but now it seems to be very close to us, and I think this year's AIGC outbreak is such a state.
The current development of AIGC has not produced a fundamental change in paradigm compared with the earliest deep neural networks. We can't conclude from the fact that ChatGPT dialogue generates well and behaves intelligently that it really produces general intelligence, because it doesn't really know what it's talking about, it just gives people the feeling that it has understood what it generates, which is still a certain distance from human intelligence.
I think the next era of AI needs to develop in several important directions. On the one hand, it can learn independently, and some of its own reasoning logic must be as explainable or understandable as ours, and it can really do cross-field. On the other hand, from the perspective of industrial landing, such as using artificial intelligence methods to assist in the content generation of games, but now whether it is AIGC generating pictures, text generating 3D models, generating character animations in games, or we use ChatGPT to do game scripts or NPC dialogues, etc., in fact, there is still a certain distance from the professional standardization of the current game.
Yang Jian (Moderator): Thank you for the views of the two teachers, generally speaking, the AIGC outbreak may still be on the verge of quantitative change to qualitative change, and it is difficult to say that it has really completed the qualitative change. It is still based on the paradigm of the past, only because other technical conditions have advanced. However, whether AIGC is in line with human intelligence or the real internal law of general artificial intelligence in the future is still uncertain, and it may be better judged according to subsequent developments.
Yang Jian (Moderator): What is the current status of AIGC's technology and industrial practice, and what are the representative applications and directions?
Wang Yuntao: First of all, from the current practice of the AIGC industry, the technical classification of AIGC can be roughly divided into text, audio, image and video and virtual space according to the mode of processing: (1) text class, mainly including article generation, text style conversion, Q&A dialogue and other AIGC technology to generate or edit text content, typical applications include writing robots, chat robots, etc.; (2) Audio category, including text-to-audio, speech conversion, voice attribute editing, etc. to generate or edit voice content, as well as music synthesis, scene sound editing and other generated or edited non-speech content, typical applications are intelligent dubbing anchors, virtual singer singing, automatic soundtrack, song generation, etc.; (3) Image video, including face generation, face replacement, character attribute editing, face manipulation, posture manipulation, etc. to generate or edit images, video content, image generation, image enhancement, image repair and other technologies are related, typical applications include face changing, pinching face, replicating or modifying image style, AI painting, etc.; (4) Virtual space class. It mainly includes 3D reconstruction, digital simulation and other generation or editing of digital characters and virtual scenes, typical applications include metaverse, digital twin, game engine, 3D modeling, VR, etc.
From the perspective of AIGC applications, it currently has great advantages in providing richer and diverse, dynamic and interactive content, and has made some significant innovative developments in industries with a relatively high degree of digital culture and rich content demand, such as media, e-commerce, film and television, and entertainment. More representative include AIGC+ Media, mainly in human-machine collaborative production to promote media integration, in which the writing robot has shortened the time to generate an in-depth report from the initial 30 seconds to less than two seconds; There is also AIGC+ e-commerce, the core of which is to generate 3D models of goods, use them for product display and virtual trials, and improve the online shopping experience. There are also virtual digital humans to create virtual anchors and empower live broadcasts; The other piece is AIGC+ film and television, mainly to expand the space of film and television creation and improve the quality of works. At present, there are already products that provide new ideas for script creation, such as intelligent writing services for novel selection, including "Hello, Li Huanying" and "The Wandering Earth". There are also AIGC-enabled upgrade services for film and television editing and post-production, including "Awesome My Country" and "Road Angel" and many other film and television dramas have used AI-based image processing services; AIGC+ entertainment, mainly interesting images, audio and video generation, etc. Then at the same time, at present, there is also this kind of digital avatar to develop this kind of C-end application to lay out related application cases related to the meta-universe, which may still be seen by everyone. In addition, AIGC also has some practices in the medical and industrial fields, but it may only be used in virtual interaction, and it is still being explored in terms of in-depth industry and covering industry business logic.
Shi Shuming: From the perspective of overall technological progress, AIGC has indeed made great progress. 5 years ago, only text-generated speech (TTS) was considered available in the AIGC field. Three years ago, if AI generated images based on text, the quality of the generated pictures was high and relevance, which was unimaginable. But now that's all a reality. In addition, in the past, text generation was mostly based on templates, and this pattern was very less versatile and only suitable for very narrow fields. Now with the advent of large models, and the language model itself is also improving, AIGC is impressive. Whether it's Stable Diffusion or ChatGPT, it's amazing how powerful text understanding and content generation they are.
Of course, China does need to make further efforts in the development of AIGC. The vast majority of the work is done by a handful of research institutions in the United States, which lead the development of AI technology as a whole. Therefore, we must also work hard and strive to make more contributions to the history of AI development.
In terms of commercial application, a significant direction that can be seen at present is to assist people, such as AI-assisted creation, and AIGC plays an auxiliary role. It doesn't necessarily make sense for AIGC to generate many images on its own. But when people need it, AIGC can use some prompts, that is, enter some prompt words and their combinations, and after continuous testing and interaction, they can finally produce beautiful pictures and apply them in specific work or life. AIGC can really assist most people who are very bad at drawing to create. The same is true for text generation, for example, with the assistance of AIGC, we will be more efficient in text continuation and text rewriting, and it can also inspire our thinking. Therefore, objectively, AIGC has improved our labor production efficiency and work efficiency. From the perspective of commercial application, the most direct is AI-assisted creation, and other aspects need to be further explored. Of course, some people ask whether ChatGPT can replace the search engine, but now it seems unlikely, it may only complete some of the functions of the search engine, but it cannot replace it yet.
To sum up, first, the technological progress is very fast, exceeding expectations; Second, commercialization has a lot of room for imagination, but at the moment we don't seem to have tapped into the most critical things.
Yang Jian (Moderator): Although AIGC is still in the exploration stage, I think it is already very fascinating. In the past, those who learned design and art started from sketching, but now these techniques seem to be becoming more and more useless with the assistance of AI technology, so they may need some breakthroughs in creativity, which is also related to the relationship between art and Tao in technology.
AIGC Value Leadership
Yang Jian (Moderator): Let's go back to the value level, why is AIGC so important? What kind of value and significance does it have? In which areas can it trigger change, in addition to the material level, what kind of more impact and change can it bring about from the spiritual level and value level?
Duan Weiwen: I will mainly talk about four aspects:
First, AIGC brings a new kind of content creation, but it should also be a new way of knowing. Now when it comes to AIGC generating content, personal ideas become even more important.
Second, AIGC is a new learning and research tool because it empowers each individual with a higher level of creative ability. For example, it has been controversial recently that many undergraduate papers can already be written in AIGC, which is considered cheating. In fact, if we look carefully at most undergraduate theses now, it is indeed copying a paragraph here and there, but he will copy it better, and he summarizes it more reasonably. But in fact, if you use a technology such as AIGC, it can make it easier for you to complete the work of literature collection and processing, and improve the learning efficiency. In research work, the application of AIGC may be normalized and become a favorable tool in the collaborative process of human-machine cognition.
Third, it will involve the metaverse, that is, the creation of possible worlds. AI painting actually combines various possibilities together, similar to Kripke's theory of possible worlds. In the past, the ability of the individual brain to call up these possible data resources was very limited. With AIGC, it can be combined exactly as you imagine and become a manufacturing machine for possible worlds. Therefore, in this way, the vision of the meta-universe is broader, but we can completely combine all the spiritual wealth, thought creation, and cultural heritage of human beings through combination, plus people's inspiration choices to create.
Fourth, AIGC has some interesting applications, one of which is that it can be used to prevent autism. There are a lot of people with social phobia these days, so he can engage in a digital person of his own and talk to himself. Another artist used her childhood diary to train the AI, and finally realized the dialogue with her childhood self, she was able to understand what she worried about when she was a teenager, and achieve the role of psychotherapy. Therefore, I think AIGC also plays a role in spiritual self-awareness and self-healing, and even in the end, AIGC will become our good companion and good companionship, allowing us to help ourselves through AI to obtain new life and generate greater spiritual strength.
Yin Jun: I think AIGC can lower the threshold of content production for the entire virtual content production, including the metaverse. For example, the AI-generated Space Opera House-themed painting won first place. After people have AIGC, they see a new possibility, will it become a mass creation in the future, and they only need to have a good idea to create. For example, using ChatGPT to interpret a whole script according to my ideas. Returning to the topic of the metaverse, the previous production tools and production methods may not be able to meet the production needs of the massive content of the metaverse. But now AIGC makes people think it could be the next generation of production tools.
Yang Jian (Moderator): Just now, Teacher Duan and Teacher Yin are more about how human beings can improve themselves, break through themselves, and use new technologies to achieve this goal. But inevitably there may also be another side, it may have a relatively large impact on our existing way of life, lifestyle, and production mode, how do we digest these possible changes and conflicts?
Wang Yuntao: We have more contact with the industry, and when I first saw AIGC, the first word that came to mind was "supercomputing". AIGC is likely to pose new challenges to our current computing systems, including computing systems. In fact, when we really use a computing system to make a good parallel deployment of heterogeneous AI systems, we will find that there are many shortcomings in computing devices, data reserves, and software and hardware collaboration. For these deficiencies, they may be such a challenge that we need to meet.
In addition, from the perspective of application, the biggest challenge of AIGC to traditional industries is the challenge of content technology. In fact, the creation of the entire content has moved from the centralized creation of the platform to the decentralized user creation, so AI technology has also played a more and more subversive role in this process. Whether it's content generation, content dissemination, content moderation? The disruptive role of AI is getting stronger and stronger.
In fact, one of the most core here is the metaverse, which must be a testing ground full of various content. Unlike traditional games, the metaverse does not seem to have a goal that everyone needs to achieve or complete, which means that the metaverse must continue non-stop. So how can you create an infinitely continuous game rule in the metaverse scenario? AIGC will play a very important role in assisting humans to achieve infinite scrolling when designing the content system of the future metaverse.
Build a trusted AIGC
Yang Jian (Moderator): What are the potential challenges for AIGC now? These challenges are divided into two levels, one is what are the technical and industrial difficulties, and the other is what legal and ethical and social problems it may bring, and how should we deal with them?
Wu Baoyuan: First of all, ChatGPT has been very popular recently, but people have also found that it will bring a lot of negative problems. The most typical thing is that it produces false information, misinformation, it produces a bunch of content that seems right but is wrong. But if too much attention is paid to these challenges, it will more or less limit the development of technology. For example, when academia studies deepfakes, it is necessary to make a statement of ethical impact and potential risks of technology to generate attacks, while defense detection is not required, resulting in people being more inclined to do research on defense detection. But offensive techniques often inspire defensive techniques. Second, in the digital economy, AIGC can be used as a tool for data generation, which can protect privacy, greatly reduce the cost of data collection, and create new data. In summary, the social problems brought by AIGC and the challenges it faces, and it needs more application scenarios to drive its positive development.
Duan Weiwen: The legal, ethical and social issues of AIGC have been discussed in many discussions, for example, in artistic creation, artists have raised the issue of copyright. In the past, in the period of search engine and platform economy, we were actually turning the world into data, that is, the world was dataized, and the corresponding privacy, ethical and legal issues were also deepening and governing. Then, after entering the AIGC technology stage, it is generated on the basis of world dataization, that is, the second-order world data era. Then, its legal and ethical issues should be different from those that existed before, so we need to have some new social contract and consensus, that is, which behaviors are acceptable to us and which behaviors are unacceptable.
AIGC generated content If in the general sense of knowledge production, the content it generates is new content produced on top of the world's dataization. This is like Euclid invented Euclidean geometry, Euclidean geometry is not in the original world, the world originally only measurement, it is developed on the basis of measurement. So now AIGC is also like this, it is a new form of cognition or knowledge production. Therefore, in this new form, I think that in terms of legal and ethical governance, AIGC should be given an innovative protection space. Why talk about the protection space of innovation, it is not only to protect your economic interests, but only when technology is accepted by society, and technology attaches importance to ethical and legal issues from the beginning, can it be stable and far-reaching. Therefore, regulators, managers, scholars of legal ethics, and industry should work together to build a predictable governance model, and through legal and ethical exploration, AIGC can have better development. For example, data toxicity is often said nowadays, it is actually toxicity in real life. There are two sides to this, the first is that AIGC can expose data toxicity, some prejudice discrimination in society, etc., which in turn can purify our social life. But this purification cannot be absolute purification, because absolute purification actually goes against some of the most basic intentions of our modern life. Because there is no absolute standard for what is clean or not. In the end, there is a process that everyone needs to accept. So in this case, we have to recognize the complexity of things, and only by recognizing such a complexity can we forge ahead. In the process of development, we can know what is acceptable and what is not acceptable now.
In my opinion, legal, ethical and social issues should be incorporated into the context of the new cognitive paradigm brought by AIGC, the dynamic impact of new cognitive methods on the legal ethics of the whole society, and how we respond dynamically.
Yang Jian (Moderator): Many times technical problems are indeed a matter of grasping degrees. As a new technology, AIGC must be legally and ethically restricted, but it cannot inhibit its development. So, how to develop and apply AIGC safely, reliably and responsibly in the future has become very important, what conditions should we have in these aspects to be able to do it well?
Yao Xin: I think in terms of content generation, safe, credible and responsible development is indeed a little behind. The first problem is that most of the data now comes from the Internet, and a large proportion of the data on the Internet is wrong or inaccurate, but this data is used to train AI large models, and then use AI large models to generate new data, and finally these generated data will also be used by a new generation of AI large models to train. So it can be imagined that just like the errors will be superimposed when doing calculations, there are some errors that will be solidified in the large model, and once solidified, it will be more difficult to solve.
The second question is, if AIGC really has some industrial applications, or the application is closely related to people, at what stage should the security and credibility issues be considered? It is impossible to find a way to judge whether it is safe and feasible after AI is generated. It must be considered throughout the model building and training process.
The third question is, for example, some students write their graduation thesis and copy it here, copy a little there, maybe what he wrote in the end is not as good as ChatGPT, so why not let him use this tool? There is a deeper problem here. What should education teach students? How should students be taught? This is an important question. Because relying on a certain AI model to generate knowledge, will it lose the diversity of knowledge, and if I lose the diversity of knowledge, what impact will it have on my society? This is what should be considered at the beginning of the development of AIGC, otherwise it is possible to follow the old road of recommendation system, as if our world vision is closed by a recommendation system, will it be closed by a large model in the future.
Yang Jian (Moderator): Thank you, Teacher Yao. The three questions raised by Teacher Yao are very important. The first is the problem of data source pollution, which is very scary to enter the entire content gene; The second is what stage of technological intervention is and what degree to grasp; The third is whether the AI big model will change from a human assistant to a shackle of human beings and become a cage that restricts human beings.
Wu Baoyuan: Let me discuss it from the field of trusted AI that I am studying. The definition of trusted AI is already clear, robustness, fairness, privacy, and explainability. However, these are limited to previous discriminant and decision-making AI, and there are still relatively few related studies on AIGC. First, as Yao said, AIGC's security problems are more likely to be created at the source and may be more harmful. Therefore, in response to the credibility of AIGC, in addition to the old problems, we should also pay attention to new challenges, such as copyright issues and liability retrospective issues. Therefore, it is necessary to define its problems clearly, and the technical solutions can be further explored later.
Second, AIGC has a characteristic that its harm seems to be less direct, that is, the derivative problem of AIGC, as a technical person, he may not think so clearly, so AIGC governance needs more interdisciplinary participation earlier, jointly define the problem clearly, and control it from the source, which will help the healthy development of AIGC.
AIGC has a promising future
Yang Jian (Moderator): What are your expectations and prospects for the future development of AIGC and artificial intelligence? What is its likely future impact on human society?
Yin Jun: The entire AIGC, as well as the future artificial intelligence technology, will definitely bring a fundamental change to our existing production tools and productivity. These changes will inevitably lead to changes in production relations, which may have a greater impact on the future of mankind and society.
Wang Yuntao: AIGC may be a major opportunity for the future digital native world, but also a new challenge. Compared with the digital transformation of the physical world, the future digital native world is likely to be the metaverse world, and humans can create many new applications, new formats, and new business models out of thin air, and AIGC is an indispensable part.
Then it also has many challenges, including the challenge to traditional economic theory, that is, AIGC may change the cost structure of human production and life in the future, the cost of future intelligent capabilities will fall a lot, that is, the marginal benefits of the use of wisdom will increase a lot, so human beings will face a more complex and more diversified new world.
Yao Xin: First, we should embrace AIGC technology, there is no doubt about that. Second, in the process of embracing AIGC, it is necessary to identify its potential challenges, and of course it is not necessary to solve these challenges in order to promote AIGC applications.
Shi Shuming: First, I believe that AIGC and the entire artificial intelligence technology will continue to develop at a high speed; Second, I am looking forward to this development that will help improve the quality of life of the entire human race and make our lives more comfortable and convenient.
Duan Weiwen: AIGC mainly brings a kind of content production automation, then this automation will actually completely change the cognitive collaboration process between man and machine. The real challenge is that AIGC as an engine for content production or knowledge production, are we prepared for the content itself, including legal and ethical rules.
Wu Baoyuan: AIGC should be another wave of enthusiasm for artificial intelligence. There is also a potential impact here, that is, the current artificial intelligence teaching and teaching materials need to be greatly updated. In the past, our teaching focus was on discriminative networks, but now we may need to add content on generative AI.
Yang Jian (Moderator): Thank you for your wonderful sharing! It can be said that we are experiencing such a generation wave led by AIGC, it is not only the progress of the technology industry, but also a trend that the whole society has to face, we must use a more open mind to recognize it, with optimism and caution to accept it, may be able to see clearly, and benefit from this wave.
On January 9, 2023, Tencent Research Institute held a thematic forum entitled "AIGC: Trends and Prospects under the New Wave of AI", which specifically discussed the technical status and industrial practice, development opportunities and future challenges of AIGC, a cutting-edge technology trend. We publish the proceedings of the conference here for the reference of thinkers in various fields.