Zhang Yaqin: In the field of artificial intelligence, China can catch up with AI·20 people

When Zhang Yaqin, dean of the Intelligent Industry Research Institute (AIR) at Tsinghua University, received the invitation to sign the invitation from the Center for AI Safety, he chose to sign his name without much hesitation.

The letter, which later attracted global attention, had only one paragraph: "Mitigating the risk of extinction posed by AI should be a global priority, along with other social-scale risks such as epidemics and nuclear war." ”

More than 350 industry executives, experts and professors in the field of artificial intelligence (AI) around the world, including Yaqin Zhang, signed the letter, including OpenAI CEO Sam Altman, DeepMind CEO Demis Hassabis to Microsoft founder Bill Gates, and many more — almost all of the names that are crucial to the leap forward in human AI technology are signed in the joint letter.

Zhang Yaqin is not only a professor and dean, but also a senior executive at Microsoft and Baidu. Before that, he was a world-class scientist in the field of artificial intelligence, and he was also a genius prodigy. At the age of 12, Zhang Yaqin was admitted to the junior class of the University of Science and Technology of China, becoming the youngest college student in China at that time, and then he studied at George Washington University in the United States, and the world-renowned distinguished scientist Professor Pick Holtz repeatedly boasted of Zhang Yaqin in public: "He is a treasure in the world." At the age of 31, Zhang became the youngest member of the Institute of Electrical and Electronics Engineers (IEEE) in its centennial history.

In December 2020, the Intelligent Industry Research Institute (AIR) of Tsinghua University was established, with Zhang Yaqin as the dean, and he and his team studied AI big models and also promoted the implementation of the industry.

Compared with the worry about the fate of mankind, Zhang Yaqin is now more focused on domestic technological progress and industrial application. At the end of April, he and his team released BioMedGPT, an open source model for biomedicine. "I hope that the research we do is useful and can be used in industry." Zhang Yaqin told the Economic Observer reporter.

His Tsinghua University's Intelligent Industry Research Institute (AIR) is located in the Wudaokou Tus Technology Building, which has recently become the "central brain" of domestic AI practitioners. The heads of AI rookie companies, AI R&D leaders, and top AI investors who have been familiar to us have recently visited here. The common topic they chatted about was AI, or more accurately, large models.

More communication with Zhang Yaqin will reveal that although he signed an open letter warning about AI risks, during the interview, he repeatedly mentioned that he is an AI optimist - of course, this optimism is based on the premise that AI can be controlled.

When listening to Zhang Yaqin talk about artificial intelligence, you can feel his excitement. He stood up and pointed to the PowerPoint made on the screen in June this year and said that ChatGPT was the first product to pass the Turing test, "It's amazing." "People who used to study AI tried their best, but they were always stuck in a bottleneck. After the appearance of the big model, a new direction is lit, and although it is not perfect at the moment, the ending is clear.

When people have doubts about the current domestic nearly 100 companies launching large models in a short period of time, Zhang Yaqin is also optimistic about this. Zhang Yaqin feels that the number is a bit too much, but he believes that working hard in a fully competitive market will eventually have a good result. That's where long-standing Chinese entrepreneurs excel, and he doesn't doubt it.

That optimism led him to disagree with another earlier open letter that also shocked overseas. Tesla founder Elon Musk said in the risk warning letter that "calls for a pause in the development of advanced AI" that AI has the potential to destroy humanity. Zhang Yaqin said that he is not worried about this, "Historically, every technological revolution has brought risks, but every time human beings have managed the risk well, and this time, I think it will be no exception." ”

|Dialogue┃

Technical risks

Economic Observer: You signed the May 30 open letter warning about AI risks, why did you sign it?

Zhang Yaqin: Actually, as early as 20 years ago, I have been calling attention to the possible risks and ethical challenges of AI. In fact, not only AI, all technologies are facing similar crises, when technology is getting stronger and stronger, it will appear in two ways: one is to use well, can serve human beings, for society; The other is that it is not used well, which brings negative effects.

I'm actually optimistic, I think we can control where it's going, but we have to have a sense of risk. We humans have two kinds of intelligence, one is the ability to invent technology, and the other is the ability to control the direction of technology. Now it is clear that wisdom two is behind wisdom one, and this is the time to step up. This kind of intensification is not only in terms of policies and regulations, but also in terms of technology.

But I don't think the way to close the gap is to stop frontier exploration. I understand their concerns about a previous letter suggesting that you suspend the development of GPT for 6 months, but I don't think this is possible. Because human curiosity and the pace of human innovation cannot be stopped. People's intentions are good, but it is likely that this company stops, that company does not stop, this country stops, that country does not stop, so this kind of call is actually very difficult to achieve.

The Economic Observer: How should we deal with AI risks?

Zhang Yaqin: I believe that humans have the ability to manage AI risks, because the second intelligence of humans has always existed. For example, launching rockets also has many unknown risks. And the way it defuses risk is to do various tests in different, scientific, and technical ways, such as rocket safety research, and there is a branch of rocket that is like this. AI can also have a branch that does this to minimize all risks.

We believe that this is also the reason why human beings can continue to develop, survive and progress rapidly. I am an optimist, I believe that technological innovation is the biggest driving force for social progress, the industrial revolution was a technological revolution before, and this time it is the same, artificial intelligence is the technological driving force of the fourth industrial revolution.

In fact, in every technological revolution, humans have managed risk well, not perfectly, but all managed. Our human life expectancy has been extended, social productive forces have developed, and our lives have become happier and more abundant.

Large model evolution

Economic Observer: What is the difference between the emergence of large models this time and the previous iteration of AI technology?

Zhang Yaqin: A major challenge in artificial intelligence in the past was the lack of common sense, a model can only do a certain thing, and each specific task must be specially made a set of models. Now that AI common sense capabilities are getting stronger and stronger, it is developing towards AGI, which is really amazing.

Moreover, AI imitates not one person, it imitates all humans. It accumulates the best things in human beings, so it's called superintelligence. In this way, it will include what we humans have learned, human common sense.

In the past, we joked that artificial intelligence, there are more people doing labeling than intelligence. But now there is no need for human annotation. For example, in Go, you don't need to tell it the rules, just tell it what is winning or losing, and it can learn it itself and apply it to other chess games.

Economic Observer: Are there also shortcomings in large models?

Zhang Yaqin: Its efficiency is still very low. The current large model is much less efficient than the human brain, perhaps at least 1,000 times worse. From the perspective of scale and energy consumption, the human brain is the most efficient agent, it has 86 billion neurons, each neuron has tens of thousands of synapses, but only weighs less than 3 pounds, consumes less than 20 watts, the storage and efficiency of the human brain is unmatched by any current large model.

So now it is necessary to improve the efficiency of large models and make energy consumption smaller. And in the human brain, when you ask a sentence, it does not mobilize all neurons, only a small part, and the smarter the person, the less it mobilizes. But the big model is different, and any question to the big model may have to mobilize all the resources, which is undoubtedly a huge waste.

However, resolving these issues will take time. This is also the direction of our future research, to learn human brain perception, decision-making, thinking.

But you must know that to do research and products, time is not the same. When do large models consume less energy? It doesn't get smaller in a day. Some studies have reduced it by 20%, some studies have reduced it by 30%, and it may reach a certain point after many years and it will be successful. In fact, now the large model can be used, but the energy consumption is very high, so Microsoft, Google and other companies in the large model business has not yet been profitable.

Economic Observer: What can the Intelligent Industry Research Institute (AIR) of Tsinghua University bring to the development of AI technology?

Zhang Yaqin: We are a basic research institute with more than 20 professors and more than 100 postdoctoral fellows, doctoral students and interns. We are a little different from the research of the faculties of ordinary schools, and I hope that the research we do will be useful and can be applied to industry. This study may be 3 years, maybe 5 years, maybe longer, it doesn't matter, because we are oriented to solving real problems, not pure curiosity research.

There are three directions of our scientific research, and they are also the three directions in which artificial intelligence will have a huge impact in the next 5 or 10 years. The first is robotics and driverless; The second is intelligent IoT, including green computing, small model deployment, etc.; The third is smart healthcare, including AI-driven new drug research and development.

For example, our Professor Liu Yang (male) has been making calls between models and interactions between models, which is a study of large models. Professor Liu Yunxin is making lightweight models, which is to miniaturize the capabilities of large models. Professor Liu Jingjing makes multimodal large models, and jointly trains sounds, images, and words between modalities, and she made large models at Microsoft before. Professor Nie Zaiqing made big data models, and his team recently released BioMedGPT, a large model of biomedicine. Associate Professor Liu Yang (female) does federated learning and privacy computing. Associate Professor Li Peng has also been doing general large models, Professor Chen Yilun has also been doing transportation models, and Assistant Professor Zhan Xianyuan is also doing reinforcement learning, etc., and their work is top in the world.

We recently released BioMedGPT, a biomedical open source model, which is a model for the bio-vertical. Our team from the Collaborative Vision and Robotics Lab (DISCOVER Lab) participated in ICRA 2023 PUB. R International Robot Competition and won the championship, at the same time, this DISCOVER laboratory is the research group with the most manuscript papers in China, the director of the laboratory, Associate Professor Zhou Guyue is the Chinese scholar and doctoral student with the most manuscript papers in the world, and Li Pengfei is the student author with the most manuscript papers in the world.

Long-termism

Economic Observer: Incomplete statistics have shown that 79 companies in China are already making large models. Is this homogeneous competition?

Zhang Yaqin: Although they all make large models, they do not necessarily make horizontal models, and some are doing vertical models. I have discussed with many entrepreneurs, most of them are very smart people who have ideas, want to do things, and they will adjust according to the industrial development plan, and there will be a few winners later. While most of them will disappear, how do you know who will exist and who will disappear without going through this process? Internet companies are like this, such as the hundred regiment war, video war, search war, etc.

From the perspective of Internet history, PC Internet and mobile Internet in the later stage, Chinese companies have a lot of scale innovation and platform innovation. In the era of the PC, we almost copied the United States, and in the era of mobile Internet, China's mobile Internet products such as payment, short video, WeChat, and O2O are better than the United States.

Therefore, I think in the field of artificial intelligence, China can also do well in the future, and we can catch up.

The Economic Observer: Does this mean that new entrepreneurial opportunities are emerging?

Zhang Yaqin: The effect of each change of industrial platform is an order of magnitude increase. From the past, the industrial opportunity in the mobile Internet era is at least 10 times larger than the PC era, and the industrial opportunity in the artificial intelligence era is at least 100 times larger than the PC era and 10 times or higher than the mobile Internet era.

If we look back at the history of the IT industry, we will find that every time there are small companies that disrupt large companies, but different companies do it differently. Now 79 companies, more than 100 companies, there is always someone to find his new positioning. There are a lot of things that can be done, such as vertical models, autonomous driving, biomedicine, and so on, and there are many, many opportunities now. You have to believe these people who came out of the Internet age, they absolutely know how to compete.

Economic Observer: There are concerns about whether this wave of generative AI will be similar to many previous gusts of technology.

Zhang Yaqin: To do research, we must still have a long-term approach, be quiet, and not be in a hurry. You see, OpenAI is not in a hurry, and DeepMind is not in a hurry.

But it must be seen that for a long time, once a certain field in China is hot, everyone will fall, including the investors will also urge entrepreneurs to go down to do the company, thus forming a short-term feverish atmosphere. But after a while, there seemed to be no new progress, so everyone's attention shifted to a new place. A real technology, in the short term we tend to overestimate its capabilities, in the long term we often underestimate its impact. I think it's similar to the situation on the internet around 1998, and then around 2000 the bubble started to burst, and you see, so many VCs are out of nowhere. At the time, many people had pessimistic expectations for Internet companies, but now, the Internet is still one of our most important capabilities.

As a big platform, good companies will get bigger and bigger, but in two years everyone may forget about it. It's like a few years ago when autonomous driving was hot, and it was discussed all day long. Now that autonomous driving is the best time for development, everyone has not responded.

Of course, what can be seen is that after the previous bubble, the atmosphere of long-termism is slowly forming. In the past, people had little money, but now some people make a lot of money, and in the future there will be some NGOs, foundations, or very wealthy individuals. So I think we might as well have some patience and time to wait for this shift.