The Future of AIGC & ChatGPT: a Chinese Experts' Roundtable at Tencent

Qiuyue Li and WANG Shuo

Source Link

I have come to realize that Pekingnology’s coverage of ChatGPT is far behind its enormous significance, so today let’s listen to seven leading Chinese scholars on the issue, via a roundtable organized by Tencent Research Institute in January. The Chinese version is available on the institute’s WeChat blog, which published quite some valuable non-Tencent content.

In 2022, AIGC (Artificial Intelligence Generated Content) took the internet by storm, from AI (Artificial Intelligence) models like DALL-E 2 and Stable Diffusion that have exploded in AI-generated art to conversational AI models like ChatGPT that exhibit human-level performance.

As AIGC becomes a hot topic across various sectors of society, people cannot help but ask: will AI become the new creator? Why has AIGC suddenly erupted? Does it signify that AI is entering a new era, and where will it lead? Will AIGC, with both large and multimodal models, become a new technology platform? How will AIGC impact the economy and society, and how should different entities perceive it?

On January 9, 2023, Tencent Research Institute held a forum titled "AIGC: Trends and Prospects of AI in the New Wave," discussing the current development and industry practices of AIGC technology and the opportunities and future challenges.

Guests:

姚新 Xin YAO, Chair Professor and Department Head of Computer Science and Engineering, Southern University of Science and Technology

段伟文 Weiwen DUAN, Director of the Department of Philosophy of Science and Technology in the Institute of Philosophy, Chinese Academy of Social Sciences

王蕴韬 Yuntao WANG, Deputy Chief Engineer of Cloud Computing and Big Data Research Institute, the China Academy of Information and Communications Technology

吴保元 Baoyuan WU, Associate Professor at the School of Data Science, the Chinese University of Hong Kong, Shenzhen

殷俊 Jun YIN, Director of Digital Content Technology Center, R&D Effectiveness Department, CROS, Tencent Games

史树明 Shuming SHI, Head of Natural Language Processing Center, Tencent AI Lab.

Moderator:

杨健 Jian YANG, Vice President of Tencent and Chief Consultant to Tencent Research Institute

AIGC's Accumulative and Explosive Growth

Jian Yang: How do you view the explosion of AIGC in the second half of 2022? Does AIGC represent the next era of AI technology?

Xin Yao: The explosion of AIGC is a comprehensive success. Although AI has a lengthy history in content generation, it has yet to advance to the extent it has today due to data, computing power, and algorithm technology limitations. One important factor contributing to the great interest in AIGC is that it has exceeded people's expectations, making it particularly interesting. However, we still need to consider what technological breakthroughs AIGC has made.

Firstly, achieving results that exceed people's expectations regarding technological advancement is relatively easy. For instance, in areas such as images, dialogues, and even music, it can generate pieces that possess a certain composer's style but are not entirely identical. However, many believe that AIGC may be an effective path to AGI (artificial general intelligence), which requires careful consideration. This is because AIGC generates content through a large amount of data and computing power, compared with the conceptual process humans undergo when creating content. For example, when I look at many images, my brain goes through an abstract process and then returns to the space of the image to create a new image. I am unsure whether AIGC currently has such abstract functions built into it. If a machine has seen a million photos of cats and dogs, it won't generate concepts such as four legs, fur, etc.

Secondly, as AIGC further develops and involves scientific and technological fields or applications that truly relate to national welfare, there may be additional challenges, as it is difficult to ensure that the content generated by AIGC meets certain constraints. Let me use an analogy. We can now send many molecules to an AI system, which may generate new molecules for you. While developing drugs from these molecules is conceivable, there may be a gap between the theoretical concept and practical application. How far are we to achieving the so-called AGI? These are questions that require contemplation and exploration.

Jun Yin: My views are consistent with Professor Yao's. Whether from the broad approach of using AI to generate content or the narrow perspective of developing deep large language models, it results from accumulated experience and research. The eruption of AIGC today is due to multiple factors, including collecting a large amount of high-quality open datasets, theoretical breakthroughs such as the diffusion model, and new computing hardware such as GBT3 with larger and more effective models. In the past, there may not have been such powerful computing devices to support training models at this scale, and even now, the cost of training such models is still very high. The combination of these factors has made generating images or text seem much more feasible than before. I think this explains why AIGC has exploded in popularity this year.

Compared with the earliest deep neural networks, the development of AIGC has yet to undergo a fundamental paradigm shift. Therefore, we cannot conclude that ChatGPT has achieved artificial general intelligence (AGI) simply because it performs well in generating conversations and appears intelligent. ChatGPT does not truly understand what it is saying and only creates the impression that it comprehends the generated content, which is still far from human intelligence.

The next era of AI needs to focus on several essential directions. On the one hand, it should be able to learn autonomously. Like humans, some of its reasoning logic should be explainable or understandable, and be able to perform tasks across domains. On the other hand, from the perspective of industrial implementation, such as using AI methods to assist in game content generation, it is still far from the professional norms of the gaming industry, whether in generating images, text, 3D models, or character animations, or using ChatGPT to create game scripts or NPC (Non-Player Character) dialogues.

Jian Yang: Thank you for sharing your insights. In summary, AIGC may be on the edge of quantitative to qualitative change, and it is difficult to say whether it has truly undergone a qualitative change. It is still based on the past paradigm, only progressing due to advancements in other technological conditions. Therefore, it remains to be seen whether AIGC truly conforms to the inherent principles of human intelligence or general AGI in the future.

Jian Yang: What is the current state and industrial practice of AIGC? What are some representative applications and directions?

Yuntao Wang: Regarding the current industrial practice of AIGC, the technology can be classified into text, audio, image and video, and virtual space based on the processing modes.

(1) Text. That mainly includes generating or editing text content, such as article generation, text style conversion, and question-and-answer dialogue. Typical applications include writing robots, chatbots, and so on.

(2) Audio. That includes text-to-speech (TTS) conversion, speech conversion, and speech editing, as well as non-speech content such as music generation and scene sound editing, with typical applications including intelligent dubbing, virtual performances, automatic music composition, and song generation.

(3) Image and video. That includes generating or editing image and video content such as facial generation, facial replacement, character attribute editing, face manipulation, and posture manipulation. Technologies such as image generation, enhancement, and restoration are also relevant. Representative applications include beauty filters, face swapping, image replays, style modification, and AI painting.

(4) Virtual space. That mainly includes generating or editing digital characters and virtual scenes such as 3D reconstruction and digital simulation. Typical applications include metaverse, digital twins, game engines, 3D modeling, and VR (virtual reality).

In terms of AIGC applications, there are significant advantages in providing more diverse, dynamic, and interactive content. AIGC has achieved significant innovative improvements in industries with a high level of digital culture and diverse content demands, such as media, e-commerce, film and television, and entertainment. AIGC+media promotes media convergence through human-machine collaboration. The time it takes for a robot to generate an in-depth report has been reduced from the initial 30 seconds to less than two seconds. In addition, AIGC+e-commerce focuses on developing 3D models of products to enhance the online shopping experience through virtual product displays and trials. AIGC is also used to create virtual anchors to boost live-streaming sales. AIGC+film & television expand the creative space for film and television to improve the quality of productions. Currently, products are already providing new ideas for script creation, such as intelligent writing services for script selection, including the film "Hi, Mom" and "The Wandering Earth." AIGC also provides image processing services based on AI for film and television editing and post-production, including the movie "Amazing China" and the remastering of the 1937-released film "Street Angel." AIGC+entertainment generates fun images and audiovisual content. There are also some examples of developing consumer-focused applications to explore the metaverse. However, while AIGC has some practical applications in the medical and industrial sectors, it is still in the exploratory phase of deeper industry integration and business logic implementation.

Shuming Shi: AIGC has made significant advances in overall technological progress. Only TTS generation was considered usable in the AIGC field five years ago. Three years ago, it was unimaginable that AI could generate high-quality and relevant images according to text. Previously, most text generation was based on specific models, which had a limited scope of applicability. With the emergence of large language models and the continuous improvement of language models, AIGC has become increasingly impressive. Both Stable Diffusion and ChatGPT amaze people with their robust text comprehension and content generation.

Of course, China still needs to make more effort to advance AIGC. The vast majority of work is being done by a few research institutions in the United States, which have led the entire AI technology field. Therefore, we must also strive to contribute more to AI development.

As for commercial applications, a prominent direction is assisting humans, such as AI-assisted creation, with AIGC playing an auxiliary role. It may not be meaningful for AIGC to generate a lot of pictures independently, but when people need it, AIGC can produce exquisite images based on some prompts, which are a combination of hints. Then, through repeated testing and interacting with them, AIGC can assist most people who are not good at drawing to create. The same goes for text generation. With the assistance of AIGC, the efficiency in text continuation and text rewriting will be higher, and it can also inspire our thinking. Therefore, objectively speaking, AIGC improves our productivity and work efficiency. As for commercial applications, the most direct one is AI-assisted creation, and other aspects still need further exploration. Of course, some people ask if ChatGPT can replace search engines, but it seems unlikely at the moment. It may only complete a part of the functions in search engines, but it cannot replace them entirely.

In summary, technological progress is very fast and beyond expectations; secondly, there is much room for imagination in commercialization, but we have not yet tapped into the most critical aspect.

Jian Yang: Although AIGC is still in the exploration stage, it is already very exciting. In the past, those who studied design and fine arts had to start with sketches, but now these techniques seem to be less and less useful. Therefore, there may be a greater need for breakthroughs in creativity.

Jian Yang: Why is AIGC so important? What kind of value and significance does it have? In which areas can it bring change? Apart from the material aspect, what kind of impact can it bring about at the spiritual and values levels?

Weiwen Duan: I will mainly talk about it from four aspects:

First, AIGC provides a new way of content creation, but it should also be a new way of cognitive perception. Now, when AIGC generates content, individual ideas become more important.

Second, AIGC is a new learning and research tool because it empowers individuals with higher-level creative abilities. For example, recently, there has been controversy over whether many undergraduate theses can already be written using AIGC, and people think it's cheating. If we carefully examine most undergraduate theses today, they already copy and paste [from somewhere else], but they have done it skillfully. However, using AIGC can make it easier to finish the literature search and processing and improve learning efficiency. In research work, AIGC may become a routine and beneficial tool in the collaborative process of human-machine cognition.

Third, it may involve the metaverse, which is the creation of a virtual world. AI art combines all possibilities, similar to Saul Kripke's possible worlds theory. In the past, the individual's brain had limited access to these potential data resources. But AIGC can fully combine these resources according to your imagination and become a machine for creating possible worlds. Therefore, the potential of the metaverse has significantly expanded, and we can combine all of humanity's spiritual wealth, creative ideas, and cultural heritage to create.

[According to American philosopher Saul Kripke, modal facts are construed as facts about possible worlds, where the actual world is just one in which as many aspects of the world as possible as similar to ours.]

Fourth, there has been some interesting applications of AIGC, one of which is to prevent autism. Many people have social anxiety, so they can create a digital person to chat with. One artist used her childhood diary to train AI and eventually conducted a conversation with her younger self. This allowed her to understand what she worried about during her adolescence and achieve a therapeutic effect. Therefore, AIGC can also contribute to spiritual self-awareness and self-healing and even become our good companion that enables us to self-help and take on new life through AI to gain greater spiritual strength.

Jun Yin: I think that AIGC can significantly lower the threshold for virtual content production, including the metaverse. For example, an AI-generated artwork titled "Space Opera Theater" won first place in a competition. After people have access to AIGC, they see a new possibility where mass creation could become a reality, and all one needs is a good idea to start creating. For instance, I could use ChatGPT to generate a complete script based on my thoughts. While previous production tools and methods may not meet the demand for massive content, AIGC can convince people that it could be the next-generation production tool.

Jian Yang: Professor Duan and Professor Yin discussed how new technologies could help improve and outperform ourselves. However, there may be another aspect worth considering, which is the potential impact on existing ways of survival, lifestyle, and production. How shall we deal with these possible changes and conflicts?

Yuntao Wang: As someone who frequently interacts with the industry, the first thing that comes to mind when I hear AIGC is "supercomputing." AIGC will likely present new challenges to our computing architecture, including the computing system. When we actually adopt heterogeneous AI systems effectively, we discover many deficiencies in computing devices, data storage, and hardware-software coordination, all of which require us to find solutions.

Regarding applications, the biggest challenge AIGC poses to traditional industries is content technology. Content creation has shifted from being centralized on a platform to being more decentralized with user-generated content. In this process, AI technology is playing an increasingly disruptive role, whether in content generation, distribution, or moderation.

The most crucial aspect is the metaverse, which is undoubtedly a testing ground for all kinds of content. Unlike traditional games, the metaverse appears to lack a specific goal everyone needs to complete. This means that the game must continue indefinitely. So, how can we create an infinite set of game rules in the metaverse? AIGC will have a crucial role in helping humans design the content system for the future metaverse that offers a never-ending digital environment.

Jian Yang: What are the potential challenges for AIGC now? These challenges can be divided into two aspects: what are the technical and industrial difficulties, what legal, ethical, and social issues AIGC may bring, and how should we deal with them?

Baoyuan Wu: First, recently, ChatGPT has become very popular, but people have also found that it can bring many negative issues. The most typical one is that it produces false and erroneous information and generates a lot of seemingly correct but wrong content. However, focusing too much on these challenges will inevitably impose restrictions on technology development. For example, when the academic community is researching Deepfake, ethical implications and potential technical risks need to be declared in generating attacks but not for defense and detection, which leads to a greater focus on researching defense and detection. However, attacks inspire defense. Second, in the digital economy, AIGC can serve as a tool for generating data, protecting privacy, significantly reducing the cost of data collection, and creating new data. In conclusion, AIGC needs more application scenarios to drive its positive development.

Weiwen Duan: The legal, ethical, and social issues of AIGC have already been discussed in many forums. For example, copyright issues have already been raised in art creation. During the era of search engines and platform economy, we were turning the world into data, that is, datafication of the world, and the corresponding privacy, ethical, and legal issues were also constantly being addressed. After entering the AIGC technology stage, we are generating content based on the datafication of the world, which is the second-order datafication era. Therefore, its legal and ethical issues may differ from those in the past, and we need new social agreements and consensus on acceptable and unacceptable behaviors.

If we talk about the content generated by AIGC in the general sense of knowledge production, it produces new content based on the datafied world. It is like when Euclid invented Euclidean geometry, which did not exist before. There was only measurement, and Euclidean geometry developed from it. Similarly, AIGC is a new form of cognition or knowledge production. Therefore, we need to provide AIGC with an innovative protection space regarding legal and ethical governance. Why talk about the protection of innovation? It is not only about protecting your economic interests but also because society can only accept technology when ethical and legal issues are considered from the beginning. Therefore, regulators, managers, legal and ethical scholars, and the industry should work together to build a predictable governance model to enhance AIGC’s development by exploring legal and ethical issues.

For example, the toxicity of data is often mentioned nowadays, which is the toxicity of real life. There are two sides to this issue. On the one hand, AIGC can expose data toxicity, biases, and discrimination in society and purify our social life in return. However, this purification cannot be absolute because absolute purification violates some of our primary intentions in modern life. There is no definite standard for what is clean or not, and it is a process that we must all accept together. Therefore, we need to recognize the complexity of things under such circumstances. Only by recognizing this complexity can we make progress. We could then learn what is acceptable and not yet acceptable in the exploration process.

In my opinion, legal, ethical, and social issues should be incorporated into the new cognitive paradigm brought by AIGC. We need to respond dynamically to the impact of new cognitive methods on society's legal and ethical issues and how to address them.

Jian Yang: Many technical problems often come down to finding the right balance. AIGC needs to be subject to certain legal and ethical constraints as a new technology, but we also can't stifle its development. So, it's crucial to figure out how we can develop AIGC applications in a safe, trustworthy, and responsible way. What conditions should we have to do this well?

Xin Yao: The development of safety, trustworthiness, and responsibility has fallen behind a bit in terms of content generation. The first problem is that most of the data currently comes from the internet, and a significant proportion is either incorrect or inaccurate. This data is then used to train large AI models, which in turn generate new data used by the next generation of AI models for training. So, as errors can accumulate when doing calculations, some mistakes in large models may become permanent and difficult to resolve.

The second problem is that if AIGC is to be truly applied in industry or applications closely related to people, when should the issues of safety and trustworthiness be considered? It cannot be after AI generation and then finding ways to determine whether it is safe and feasible. It must be thought out throughout the entire process of building and training the model.

The third problem is that some students copy from here and there for their graduation thesis, and perhaps what they eventually produce is not as good as ChatGPT. So why not let them use this tool? On this, what should education teach students? How should students be taught? This is an important issue because relying on a single AI model to generate knowledge could lead to a loss of diversity in knowledge production. If we lose diversity, what impact will it have on our society? All of these issues should be considered at the beginning of the development of AIGC. Otherwise, we may end up going down the same path as recommendation algorithms, where our worldviews are closed off by one recommendation algorithm after another, and in the future, we may be trapped by a large model.

Jian Yang: Thank you, Professor Yao. The three problems raised by Professor Yao are all significant. Firstly, the issue of data source pollution is very frightening. Secondly, at what stage and to what degree should we intervene in technology? Thirdly, will AI models become a shackle for humans, turning from serving as human assistants to becoming restraints that confine human progress?

Baoyuan Wu: Let me discuss from the perspective of trustworthy AI, which is my research field. The definition of trustworthy AI is already very clear, including robustness, fairness, privacy, and interpretability. However, these only apply to discriminative and decision-making AI, and there is still relatively little research on AIGC. First, as Professor Yao said, the security problems of AIGC are more likely to be created at the source with more significant harm. In addition to the old problems, new challenges, such as copyright issues and accountability tracing, should also be focused on. Thus, the problems need to be defined before exploring the technical solutions.

Second, a feature of AIGC is that its harmful effects appear less direct. As technical personnel may not think so clearly about the derivative problems of AIGC, the governance of AIGC requires more interdisciplinary participation earlier to jointly define the issues and proactively control them at their source, which is conducive to the healthy development of AIGC.

The Promising Future of AIGC

Jian Yang: What are your expectations and prospects for the future development of AIGC and AI? What could be the potential impact on the future of human society?

Jun Yin: The entire AIGC and the future AI technology will undoubtedly bring about a fundamental transformation to our existing means of production and productivity. These reforms will inevitably cause changes in the relationship of production, which could significantly impact the future of humanity and society.

Yuntao Wang: AIGC may be a significant opportunity for the future digital native world, but it is also a new challenge. Compared to the digital transformation of the physical world, the future digital native world is likely to be a metaverse where humans can create many new applications, new formats, and business models out of thin air, and AIGC is an indispensable part of it.

However, it also poses many challenges, including the challenge to traditional economic theory, as AIGC may change the future cost structure of human production and life. The cost of intelligence is expected to decrease significantly, resulting in a much higher marginal return on using smart technologies. As a result, humanity may face a more complex and diverse new world.

Xin Yao: Firstly, it is undoubtedly necessary to embrace AIGC technology. Secondly, we must be aware of its potential challenges while adopting AIGC. But, of course, we don't necessarily have to address these challenges before advocating for the use of AIGC.

Shuming Shi: First, AIGC and the entire field of AI will continue to develop rapidly. Second, I look forward to this advancement improving the quality of human life and making our lives more comfortable and convenient.

Weiwen Duan: AIGC mainly brings about content automation, and this automation will fundamentally change the cognitive collaboration process between humans and machines. AIGC is an engine for content or knowledge production, and the real challenge is whether we are prepared for the content, including legal and ethical rules.

Baoyuan Wu: AIGC should be another wave of enthusiasm for AI. There is also a potential impact here: current AI teaching and materials need significant modifications. Our teaching focus was on discriminative networks in the past, but now we may need to add content on generative AI.

Jian Yang: Thank you to all the guests for the wonderful sharing! We are currently experiencing the generative revolution as a result of AIGC, which is a technological advancement and a trend that society must face. We must approach it with an open mind and acknowledge it with an optimistic and cautious attitude, and only then can we truly understand and benefit from this wave of change. (Enditem)