TVP Jianfeng Dialogue with Kai-Fu Lee & Chun-Hua Shen: AI Future Progress and Technology Development

AI has gone through nearly 70 years since the Dartmouth Conference in 1956. In the past 70 years, there have been expectations for AI, panic about AI, capital's pursuit of AI one after another, and technical people's exploration and application of AI has never stopped.

We in the millennium have never imagined the prosperity of the mobile Internet. In 2022, how can we enjoy the AI development process in the next 20 years? Will AI + healthcare make us live to 100? How will AI bring the metaverse to life? Can AI help humans find happiness? Will AI deepen bias? Will AI take away human jobs? Can traditional enterprises enjoy AI dividends?

On the evening of May 8, Li Kaifu, chairman and CEO of Tencent Cloud TVP Jianfeng Dialogue with Innovation Workshop, Dean of AI Engineering Institute of Innovation Workshop, and co-author of "AI Future Progress"; Professor of Zhejiang University, Senior Consultant of Tencent Youtu Lab, Outstanding scientist Shen Chunhua; Vice President of Fourth Paradigm, Tencent Cloud TVP Zheng Zhao served as the host, invited 50 CTOs and technical experts from the AI field to participate in the forum discussion, colliding with the future spark of AI.

1. Kai-Fu Lee: "AI Future Progress" and the Trend of Technology Development

(1) The origin of the creation of "AI Future Progress"

There are two main reasons for the origin of the book "AI Future Progressive". First, I believe that AI is a particularly important technology, and everyone should understand what opportunities it can create and how it relates to themselves. Parents can help their children make study plans, young people can make career plans for themselves, and then they will see that AI may bring many new job opportunities in the future. I hope to use the way of storytelling to bring this popular science book focusing on AI technology. It is written so that everyone can understand it. At present, the response is still good, so that many people who do not understand AI at all probably understand what AI means.

On the other hand, as a technical person with a background in science and engineering, I studied machine learning during my undergraduate and doctoral studies. I deeply feel that technical people generally lack scene imagination, which is why AI is used in speech recognition, natural The fields of language and computer vision have been doing it for nearly 40 years, but they are still focusing on similar problems. Many new scenarios need to be more imaginative, so I also hope to use the way of cooperating with science fiction writers to describe the story clearly, so that those of us who do technology can see the possible application scenarios of AI in the future. Challenges, and how to resolve them, give you some inspiration and suggestions.

Therefore, the two important purposes of cooperating with science fiction author Chen Qiufan to create "AI Future Progress", one is to explain the difficult technology to everyone; the other is to hope that some technologies are strong, but the scene imagination is not Such a strong engineer, or a technical person doing AI, also allows them to have some inspiration and suggestions for this future vision and future scenarios.

(2) Investment experience in AI field and model analysis of AI + cloud

Thanks to a lot of investment experience, I also learned some cross-domain knowledge. I personally think that the greatest value created by AI must be the combination with the scene. From the relevant investment experience, three stages of development of AI entrepreneurship can be summed up: the first is in the relatively early stage, because the AI technicians are very powerful, they chose to create a company first, and then apply it; the second is in the early stage. In some fields, AI can already create great value. For example, we invested in Fourth Paradigm, Innovation Qizhi, Jifei Technology, etc., all of which have very strong commercial applications and landing scenarios. They land on the scene first, and then make the platform. Today, AI has entered the third stage. AI will intersect with other sciences, that is to say, AI + Science, AI can be used in the invention of new drugs, gene editing, new materials and new energy. The book "AI Future Progress" contains these three directions. For example, intelligent transportation and unmanned driving are an important field. For example, the AI pharmacy and the application of AI in new energy just mentioned. Books will be involved, because we also need to pay attention to and understand these industrial fields when investing, so we also try to describe these scenarios in the book, and the trends we investigate will bring some new inspiration to writing.

Most of the startups invested by Innovation Workshop use the cloud, and many companies have cooperated with some cloud vendors including Tencent Cloud to try to build an AI solution faster in the private cloud field. This is actually a very good combination with the cloud platform. point. We have also invested in some projects to help the cloud accelerate in terms of computing power, including AI of course, but also ordinary workloads. I think this also means that the cloud will improve rapidly in the future. A very interesting phenomenon is that China's AI technology is not inferior to the United States in terms of startup companies, usage rate, and value generated. Although the United States has stronger scientific research capabilities, China's landing results are not behind. In the field of cloud computing, the United States has a relatively large lead, but China has plenty of room for growth. Although the deployment of cloud computing in the United States has been completed, it has only begun to put AI capabilities into it, and China's cloud computing is already doing AI. solution, which is an interesting phenomenon.

AI + cloud is not described in the book "AI Future Progress", not because cloud computing is not important, but because cloud is too important, it has become a must-have platform-type (technical service), just like operation Like systems and databases, there is no need to stress its importance. For example, in the field of unmanned driving, we assume that the cloud and bandwidth are much larger and faster than the current one. After the development of 5G and 6G, the challenges of data mobilization on the cloud and the integration of edge computing are described in the book.

(3) Development and application of NLP technology

In the book "AI Future Progress", there is a story chapter of "Double Sparrow". In this story, AI becomes the companion assistant for children's growth. Under the leadership of human teachers, AI has transformed into a long-term accompanying teaching assistant role, combining children's learning and interests to become more efficient and proactive. Although in the AI era, I think AI cannot replace the educational work of human teachers, but AI can do many useful supplements, so in such a scenario, NLP (Natural Language Processing) technology is the key point.

During my sophomore year, the first AI technology I was just getting started with was NLP (Natural Language Processing). Compared with the development of early NLP, many technologies have been developed in recent years, and deep learning is an important core of them. We see a big breakthrough in the field of natural language processing, which is using self-supervised learning technology to label massive data, which solves a huge bottleneck. Since this model was made, it has brought many technological breakthroughs in the field of NLP including transformer and GPT3.

I think NLP will have a lot of development in the next three to five years. On the one hand, there will be more breakthroughs in existing applications in the past, such as speech recognition, machine translation, etc.; on the other hand, in scenarios that have not yet happened Get a lot of attempts, such as the ultimate search engine of voice dialogue and so on. The future development of NLP will not only change existing applications from unusable to usable, and from usable to useful, but also make applications that were impossible in the past possible. This is our major investment direction now. Very optimistic about this field.

2. Shen Chunhua: Industrial Layout of Youtu Lab and Tencent AI

(1) Technical input of Youtu Laboratory

Youtu is one of Tencent's top artificial intelligence laboratories. It has been focusing on basic research and landing exploration in the field of computer vision for many years. Youtu's AI capabilities are representative of the well-known WeChat face payment, automatic AI beauty and other consumer Internet applications; at the same time, the laboratory is also widely used in industrial quality inspection, finance, education and other industries. landing.

Because Tencent started out as a consumer Internet product, it is very good at understanding the needs of users, and this ability is now being verified in the industrial Internet field. I personally understand that C to B is definitely a very important part of the industrial Internet. This may also be a unique advantage of Tencent, because Tencent is good at helping B-side partners to improve their business through connector-type products such as official accounts, mini programs, corporate WeChat, etc. Serve downstream customers well.

In the process of the industrial Internet developing in depth, more tests are the ability of comprehensive solutions. Tencent has been deeply cultivating in these vertical industries for many years and has many application scenarios of technology landing, and has also sorted out some vertical fields. Standardized solutions can be quickly replicated, which is Tencent's biggest advantage over other companies in AI applications.

(2) How Tencent uses AI technology to help the development of the real economy

On the one hand, the deep integration of cloud and AI encapsulates basic AI capabilities such as speech recognition, image recognition, and NLP into APIs or tool software suitable for different scenarios, making AI a tool for optimizing production and improving efficiency in all walks of life. For example, in the field of industrial manufacturing, Tencent Industrial Cloud combined with the classic technology of computer vision, coupled with the powerful computing power of the cloud GPU, can provide the factory with an ultra-accurate AI-based quality inspection solution system, which takes 20 minutes to complete manually. The quality inspection work is compressed to a few seconds, bringing a qualitative leap and saving tens of millions of costs every year.

On the other hand, the deep integration of cloud and AI also transforms very sophisticated and complex statistical machine learning/deep learning algorithms and theories into APIs that ordinary developers can easily call, which greatly reduces the threshold for AI applications, allowing the AI has become a tool that small and medium-sized enterprises and ordinary developers can easily use, enabling talents from all walks of life to bring AI to more industrial landing scenarios, thereby generating greater economic and social value.

Through Tencent Cloud, Tencent has provided hundreds of AI atomic capabilities for industry partners and developers in all walks of life, covering many fields such as machine vision, NLP, pattern recognition, etc., and created AI solutions for different industry scenarios. More small and medium-sized enterprises can quickly deploy and apply AI, so that AI can generate value and help the industry to further develop digitally and transform and upgrade.

3. The value of AI and the double-edged sword of technology

(1) How to view the social value and significance of AI

Kai-Fu Lee: In the book "AI Future Progress", I took a constructive attitude towards the value of AI. There have been many discussions on the negative effects of AI in the industry, but I personally think that the technology itself is neutral, and whether the application of the technology is correct or not is the reason for the different effects. We can see that AI will certainly continue to create enormous economic value in the long run, while being able to get things done more accurately, cheaply, and efficiently.

At the same time, AI is also a double-edged sword. It can save us a lot of repetitive tasks and work, but in the process, some people's jobs will inevitably be replaced by AI. But on the positive side, when these job replacements occur, some new jobs are bound to be created as well. I think 20 years from now, we will reach a very good state. This state is that humans will have more time to do things that only humans can do, things that suit individual interests and abilities, and make everyone's work more interesting. more satisfied.

At the same time, with the maturity of AI and various other technologies, we may reduce the cost of production to a very low level in the future, which may have the opportunity to completely eliminate poverty and famine, which is also a major milestone in human history. Of course, there are still many challenges to reach these goals, but at least from a technical perspective, there are opportunities.

Although the current application of AI still has problems such as the information cocoon effect and privacy protection of the recommendation algorithm, these are some common social phenomena after the application of AI technology. However, I think that when a new technology is launched, it will definitely have an impact on society, but the final solution is that technicians invent new technologies to solve these negative effects. The group of people who came to listen to the live broadcast today may be the solution in the future. AI workers for most problems.

Shen Chunhua: There is no unified answer to what the AI ideal country is like, but I believe that "science and technology for good" is one of the interpretations. The greatest value of AI is to serve people and society.

On the one hand, AI has been able to make society more convenient, such as the Siri voice assistant and WeChat face payment, which can be seen everywhere in real life. On the other hand, AI is also tackling some previously unsolvable problems in the social field. Youtu previously used AI technology to assist in finding people, helping many children who have been lost for many years to find their parents. Last year, Youtu teamed up with the National Astronomical Observatory to release a star exploration program, using Youtu's computer vision technology to help China Sky Eye FAST to greatly improve the search efficiency of pulsars. The learned technology may be processed in a few days, which is an improvement of several orders of magnitude, which can greatly accelerate the efficiency of scientific exploration.

In addition, I have just started to get in touch with AI For Science research, such as the use of artificial intelligence algorithms to analyze protein sequences. As far as I know, Youtu Laboratory, Shanghai Institute for Advanced Study of Zhejiang University and other institutions are in these directions. All have achieved relatively good stage results, which are all very good examples, and I believe we will see more and more such examples.

(2) The wonderful combination of AI and medical care

Kai-Fu Lee: We think that from now to the next ten years, or even twenty years, medical care will be a very good investment area. There are several specific reasons. First, the traditional medical industry is now being digitized in an all-round way, including the informatization of medical processes, the recording of whole-process health data by wearable devices, and the massive amounts of data generated by new technologies. "Nutrition" generates valuable algorithms to assist doctors in diagnosis and treatment in disease early warning, diagnosis, treatment, monitoring, and long-term management. Second, medical care is not only practice and science with massive data. For example, many medical schools may not have too many samples of each type of cancer in a year. Medical data is used for teaching, and our data is used to teach AI. When teaching When the data has only a few samples, the AI capability leadership after training with massive data is produced. Third, there is too much medical knowledge, and it is impossible for a doctor to be omniscient and omnipotent. He has limited time to see a patient to understand the patient's background and cases, and even more limited time to see a doctor.

Therefore, I think that the optimization iteration of AI in the medical industry can indeed do better, but various interests, morals, and laws must be considered comprehensively. Therefore, more running-in and training will be needed in the actual diagnosis. . I think the human-machine collaboration model is more acceptable to the industry in terms of diagnosis: the doctor is the mainstay, he is the boss of AI, and AI assists him in making a diagnosis. In this case, on the basis of a doctor, AI will only improve his correct diagnosis. The probability.

From an investment perspective, we believe that in the medical industry, research and development of new drugs is a very good field, which is also highly consistent with the goal of AI. Because in such a scenario, people's ideas and AI's ideas are the same, both are how to develop the most effective new drug with the lowest cost and the shortest time. In this direction, the technical crowd of AI and the real business side's demands All are consistent and can work together to get things done. In this regard, we also have some investment cases to follow.

Finally, what I want to say is that there are many people in the fields of biology, chemistry, and pharmaceuticals behind the medical industry. They are in the laboratory work stage after computer simulation and before clinical practice, which is actually very suitable for AI to do. The value it brings is not to replace labor costs, but more importantly, the machine itself can do experiments 24 hours a day, and the process of research and development is exponentially improved. In the end, the ideal effect of AI is to make new drug research and development faster and cheaper, make many rare and incurable diseases treatable, and allow people to enjoy longer-term health. This is what we can expect, and what AI can do. largest and uncontested contribution ever made.

Shen Chunhua: In the past few years, when I was working in an Australian university, I also did some research on medical images. I very much agree with what Mr. Kaifu mentioned just now, that is, for medical image practitioners, the goal of AI, at least at this stage, cannot replace doctors to see a doctor. Therefore, we should make tools to help doctors see X-rays and CT images, so as to assist doctors in making diagnosis better and faster.

If AI itself is to completely replace doctors, I think there are still many problems to be solved. For example, the interpretability of the current Deep Learning algorithm. If this interpretability problem is not solved, no one dares to directly give the diagnosis predicted by the Deep model to the patient, because you do not know how the algorithm's decision is made.

Tencent has a medical imaging team that has developed a system called "Tencent Miying", which is a solution for diagnosing new coronary pneumonia through AI. The system uses the patient's CT images to provide doctors with an auxiliary diagnosis result in one minute or less, thereby helping doctors to more accurately judge the severity and development of a patient's pneumonia.

We all know that a CT scan of the chest can often produce hundreds of images. If it is completely dependent on the human eye, it will take ten to twenty minutes. Now, the use of AI algorithms can improve the doctor's inspection efficiency by an order of magnitude. , so that patients can receive more timely treatment.

So I personally think that AI is an assistant used to help doctors make rapid diagnosis and improve efficiency, which may be a development direction in the next few years.

(3) How to do a good job of science and technology and avoid the double-edged sword of data hidden dangers?

Kai-Fu Lee: These problems with data hidden dangers have caused many people's perception of AI to be more negative than positive. This is a very unfortunate thing. As mentioned earlier, AI does bring some problems, but it will also be resolved with time and new technologies, including the improvement and improvement of some relevant laws and regulations, Web 3.0 returning data to individuals, and so on. These are all directions that can be explored, and the solution is not as difficult as we thought.

For example, the issue of privacy protection, how to ensure that data is not abused, involves many algorithms for privacy computing, such as federated learning. The privacy algorithm represented by federated learning allows us to have the best of both worlds. It can not only authorize data for training, but also ensure that the data in the training model is desensitized without causing privacy issues.

From big data to recommendation algorithms, there are indeed many bad social phenomena, which leads to a very interesting topic in the field of AI, how can we make an objective function measure a relatively long-term and difficult thing. The commercialized AI algorithm maximizes the interests of large companies, but we should also consider the individual needs of AI for thousands of people. Is it possible to refer to personal needs as the objective function to train AI models, taking into account the needs of enterprises and individuals, this is a technical direction worth thinking about and breaking through.

On the other hand, the problem of prejudice actually comes from the imbalance of data. These problems are the problems that AI technicians should be vigilant about. Before we make products, we must ensure that the data has reasonable coverage and balance. Need to have some tools for reminder correction. The interpretability involved behind this is quite difficult. On the one hand, we can make some interpretable machine learning models. On the other hand, we can try to explain the existing standard models, but do not worry about the accuracy and fineness of the results. Too picky.

The last thing I want to say is that if we criticize AI so much, will people do better? Will humans be more unbiased than AI? In fact, Israel has done an experiment. The judge's judgment before lunch will be harsher than after lunch, which means that he will ignore the fairness of the work when he is upset. Human prejudice is very serious, and people hide and refuse to admit their prejudice. AI is an objective, fair, transparent, and data-based field, so we have a greater hope of making AI a low-bias decision maker, far lower than human bias. Let's not magnify some individual cases because of external forces, but in fact, I am confident that I can do much better than now in these four points, and do better than people.

(4) For technical people, how do you view the two sides of AI technology?

Shen Chunhua: When a technology like Deepfake was first open-sourced, some people used it to replace and generate face pictures. Now the fake face pictures are becoming more and more realistic, to the point where it is difficult for the human eye to distinguish the real and the fake. This does come with a lot of risk. The dissemination of these fake pictures and video information on the Internet will bring challenges. Deepfake is actually based on the GAN technology of the past few years.

As far as I know, at least for the images generated by GAN, even if the human eye cannot see the difference between the real and the fake, it is still possible for the algorithm to capture the subtle difference, because there is a distribution between the generated data and the real data. a domain gap. That is to say, you use a large amount of data to train a deep learning model to discriminate between real and fake images. At present, it is easy to judge whether the image is generated or real. At least for now.

If we say that one day we invented an algorithm, no matter what kind of generative model it is based on, it can be GAN or other generative models. If we say that the synthetic data it generates, the algorithm can’t tell the truth from the fake. Then there will be major good news for deep learning researchers. Why? Because we are now training deep learning models, such as classifiers for recognition models, often require a large amount of manually labeled data. If there is no domain gap between the synthetic data generated by the model and the real data, and the data distribution is exactly the same, then we can use this technology to obtain a large amount of training data without manual annotation. This will greatly reduce data costs.

4. TVP big coffee asks: those things about AI

(1) How do traditional enterprises use AI technology?

Kai-Fu Lee: According to my observation, many traditional enterprises in China have not done a good job in digitalization. When they started to use AI, the problem they encountered was that they needed a lot of resources and time to integrate the data first, and then they could enter the implementation of AI. Entrepreneurs of many traditional enterprises either do not realize such challenges, or the digital teams below do not synchronize information well, and finally find that AI still cannot work after a lot of efforts. In fact, data storage, sorting, and analysis are the most difficult parts. Once done, the implementation of AI is a relatively small problem.

There are many other challenges. For example, the needs of each industry are different, and it is impossible to achieve a platform to solve the problems of all industries. Many of the companies we have invested in are still working on solutions. In the long run, we certainly hope to solve all problems, but this requires a lot of time and technical exploration. Overall, most domestic AI applications in the next five years still need a solution provider, which can be a cloud or a vertical AI enterprise. At the same time, we still need to continue to explore how to create a more standardized AI capability or platform that can cover the needs of users in most industry scenarios.

Shen Chunhua: Yes, I think this is something that all cloud computing providers are doing. The purpose is to greatly lower the threshold for applying AI technology, and then it can be quickly rolled out, but there is still a long way to go. way to go. I believe that with the development of technology, we will reach that step one day, and the process may be longer.

(2) The distinction between AI theory and practical applications

Kai-Fu Lee: I think there are actually great opportunities on the deep learning platform, especially considering the application scenarios. In fact, the problem is not that we haven't invented enough good algorithms, but that we haven't been able to come up with a way to implement it. The application is introduced faster, so I think there is still a lot of room for dividends in the application.

It has also been suggested that deep learning itself is a black box-based technical direction, and the theoretical basis behind it is not strong enough, and it is understandable that it is hoped that it can be advanced. Some people have also put forward some ideas, thinking that whether deep learning itself can be developed in an interpretable and analyzable way like humans, there needs to be a way to combine the human brain and human thinking characteristics with deep learning to produce 1+1>2 development. , which is also a direction. I think it is a good thing to do more theoretical participation in AI, but at least from the perspective of application, we are not encountering a bottleneck in application because of lack of theoretical foundation.

Shen Chunhua: Let me add one more thing. I have been doing computer vision related algorithms. I personally understand that whether it is computer vision or NLP, in the past few years, some people feel that the development of deep learning technology has reached a bottleneck, and it will be difficult to develop below. However, what we have seen in the past few years is that suddenly there is another breakthrough algorithm that drives the development of the entire field. There have been many such examples in the past few years. It may take 10 years from the development of artificial intelligence this time to today. If it is counted from Hinton's paper on ImageNet image classification in 2012, it is exactly 10 years.

You can see that its development process started from AlexNet, from various convolutional neural networks to a bottleneck period, everyone thinks that it seems that there is no need to do it, and it develops to a bottleneck. Suddenly Transformer was invented, and Transformer again Greatly improved performance on many tasks. Then the GAN technology just mentioned, GAN has not been around for a few years. The effects that a generative model such as GAN can achieve today was unimaginable just a few years ago: how can a generative model get such beautiful results?

GPT3 technology was invented three years ago, and GPT3 overturned the research results of NLP in the past few decades. I personally feel that AI is far from reaching the ceiling, or to the extent of the bottleneck period. The field of AI is developing very fast.

At the end of the sharing, the host Zheng Zhao expressed a lot of emotion. He said that Mr. Kai-Fu Li’s long-term research and exploration in the field of AI gave us a lot of fresh perspectives and perspectives; Mr. Chunhua Shen, as a top scholar in the field of AI, also led us to appreciate The deep charm of artificial intelligence technology.

V. Conclusion

The road to the future is never smooth. This requires that the pioneers who create the future can strengthen their beliefs and apply the technology itself, which is neither good nor bad. What will AI development look like in 20 years? This discussion outlines just the tip of the iceberg. There are more scenarios waiting for every practitioner to imagine and make breakthroughs.

From the very beginning of its establishment, TVP hopes to "influence the world with technology", so that technology can benefit everyone, and practice the original intention and original intention of science and technology for good. On the road to the future, I hope we can join hands and move forward together.