news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

"Call me Zhongyuan." Compared to Dean Wang, he was more accustomed to this name.

In February 2024, Beijing Zhiyuan Artificial Intelligence Research Institute (hereinafter referred to as "Zhiyuan") announced that Wang Zhongyuan would become the second dean. In 2018, he was awarded "35 Technological Innovators Under 35" by MIT Technology Review; he also had work experience at Microsoft, Facebook (now Meta), Meituan, and Kuaishou. The technology circle is eager to hear what this "post-85" dean thinks about the present and future of artificial intelligence. In March this year, Li Qiang, member of the Standing Committee of the Political Bureau of the CPC Central Committee and Premier of the State Council, came to Zhiyuan specifically during his research in Beijing to learn more about the development of cutting-edge technologies for large models.

In the nearly one-hour interview, Wang Zhongyuan and the Beijing Business Daily reporter had an in-depth exchange on the bottlenecks and bubbles that need to be broken in the current artificial intelligence industry, the differences in Internet development between China and the United States, the direction of the domestic large-model price war, and Zhiyuan's positioning and development plan.

【Industry Reflection】

It's no surprise that the question "Which is bigger, 9.11 or 9.8" stumps the big model

Q: The question "Which is more serious, 9/11 or 9/8?" has recently stumped a number of big models. What do you think of this result?

A: I am not surprised at all. From the perspective of researchers, the amazingness of the big model comes from the fact that it solves the understanding, reasoning, and common sense problems that were difficult to solve before artificial intelligence, but it does not solve all problems, such as mathematics. We found that the big model did not do well in math problems that seemed easy. The inability to do well in problems that compare the size of numbers does not hinder the continued evolution of the big model and its industrial implementation.

Q: Big models have been popular for a year and a half, and the industry's judgment has become more rational. Do you think artificial intelligence has reached a bubble critical point?

A: It is far from the so-called bubble critical point. This is a point of view that I firmly believe in. The prelude to the big model has just begun. It took four years from the birth of the iPhone to the iteration of the iPhone 4 and iPhone 4S, which are recognized as easy to use by everyone. It took only one and a half years for the big model to go from the laboratory to the industry. However, I agree that the capabilities of the big model need to be continuously improved, and the application should be polished for different scenarios.

[Price War]

The Hundred Models Competition proves that our marketization is good

Q: Now that the Hundred Models War has been going on for a long time, do you still have reservations about price wars?

A: The iteration of large models requires continuous investment. The Hundred Models War is obviously not conducive to the development of basic large models and will cause resource dispersion. The advantage of the price war is that it can screen out some companies that are not really invested in the research and development of large models, and it is also conducive to discovering scenarios and polishing applications for scenarios. However, I hope that the price war will not cause companies to be afraid to invest in research and development. Just like the iPhone, if the comparison is always the first generation of products, it will gradually move away from the desired goal, which is what I don’t want to see.

Q: If price wars are unavoidable, what suggestions do you have for big model toB (enterprises)?

A: To establish a toB ecosystem, companies need to be willing to pay for software, big models, and knowledge. This is what I particularly want to call for. Only by forming such an ecosystem can it be beneficial to the development of big models and the development of companies that use big models.

Price wars are just one way to promote the universal use of big models. What is more important is the effect of big models. I am more looking forward to big models being able to find scenarios to improve efficiency on the B side, or for enterprises to find scenarios where big models can better solve problems. This is what big models to B need to solve. In the next one or two years, we may see some mature B-side big model applications.

Large models toC (users) may require more patience because users are very sensitive to experience. R&D personnel believe that it is normal for large models to not have a comparison of sizes when faced with the question "Which is bigger, 9.11 or 9.8?", but users cannot accept that it will not have this problem.

Q: China’s Internet has seen the Hundred Regiments War and the Thousand Broadcast War. How is this Hundred Models War likely to end?

A: In the United States, after one or two giants emerge in each industry or field, other companies will start to differentiate themselves. China has experienced full competition in both the Internet and mobile Internet. I think this shows that our marketization is good. In this case, the companies that can ultimately win have stronger competitiveness and combat effectiveness.

【Intelligent Source】

To do the research and development that universities cannot do and companies are unwilling to do

Q: Artificial intelligence is not the same as big models. What plans does Zhiyuan have for matching its power with big models?

A: There are at least three schools of thought in the development of artificial intelligence, symbolism, connectionism, and behaviorism, which correspond to knowledge graphs, neural networks, and intelligent agents, respectively. The big model is just a branch of the neural network. The big model is different because it compresses massive data through neural networks, allowing us to see its progress in common sense and reasoning. The big model is indeed not equal to artificial intelligence, but the big model is an important driving force for artificial intelligence today. Zhiyuan's vision is to become a leader in artificial intelligence innovation. We will continue to invest a lot of resources in the direction of big models to continuously solve challenging problems, but we will still deploy artificial intelligence technologies other than big models.

For example, not long ago, Zhiyuan released a family of big models, including common research on language big models, as well as cutting-edge research in three directions: multimodality, biological computing, and embodied intelligence big models. Zhiyuan is also doing research on brain-like and digital hearts.

Q: Zhiyuan was established in 2018 under the guidance and support of the Ministry of Science and Technology and the Beijing Municipal Party Committee and Government. Today, artificial intelligence is an important engine for the development of new quality productivity. What role does Zhiyuan play in this?

A: Zhiyuan is an independent non-profit R&D institution that plays a unique role in China's scientific research system. The university model has been in operation for decades, and it is difficult to organize, scale, and conduct systematic R&D across teams. Enterprises will also invest in R&D, but they are facing fierce business competition and survival pressure, and are more inclined to have R&D that is strongly related to business. Zhiyuan will do R&D that universities cannot do and enterprises are unwilling to do, and research projects that will take 3 to 5 years or even longer to see results.

Q: Zhiyuan is a new type of research and development institution that maintains a keen eye for major scientific issues. What are your plans for the next 3 to 5 years?

A: The research progress and results released at the 2024 Beijing AI Conference are our answers. We have made judgments on the development trend of artificial intelligence as a whole, especially on large model technology: from language large models to unified multimodal large models, to embodied intelligence and biological computing models, and finally to world models and AGI (general artificial intelligence). This is our judgment on the development of the entire artificial intelligence technology route, mainly to overcome some technical difficulties.

The native multimodal large model trains information of different modalities, such as text, images, videos, and voice, in one model from the very beginning. Such a multimodal large model can understand and reason like the human brain.

The combination of large models and hardware, that is, embodied intelligence, is a very important development direction in the future. It enables artificial intelligence to no longer exist in the virtual world, but to enter the physical world to serve mankind.

The microscopic perspective is life science. Artificial intelligence technology can understand, model and reason about life molecules. It is of great help to AI pharmaceutical manufacturing and the entire pharmaceutical industry.

Zhiyuan hopes that artificial intelligence can model the world like humans and eventually achieve AGI.

Q: In order to catch up with the big model wave, companies are facing pressure and challenges. Does Zhiyuan also feel nervous?

A: The entire artificial intelligence industry is indeed facing many challenges at the moment, and there is also fierce competition. Everyone is always running fast, and the ultimate competition is about who can run faster without falling.

Zhiyuan focuses on original innovation in artificial intelligence, and the anxiety mainly comes from whether it can achieve the technological breakthroughs that the country urgently needs in the field of artificial intelligence technology. Therefore, Zhiyuan must not only clarify what to do, but also think clearly about whether these breakthroughs can really be achieved, and ultimately benefit the development of the entire artificial intelligence industry and even a larger industry.

【talent development】

Young people take the lead, and youth becomes the label of Zhiyuan

Q: As a dean born after 1985, you are very young and have rich experience in large Internet companies. What do you think of this label?

A: Over the past decade, my personal experience can be divided into two major periods. One was joining Microsoft Research Asia after graduation, which was a period of scientific research. Later, I moved from a pure scientific research institution to an enterprise and applied artificial intelligence technology to specific businesses. In the Internet industry, I am no longer a young person, but a veteran.

In scientific research institutions, many people think that "post-85s" are relatively young to be deans. I think this is exactly the characteristic of Zhiyuan. Zhiyuan advocates that young people take the lead, regardless of seniority. Many very important scientific research projects of Zhiyuan are undertaken by very young researchers. For example, the person in charge of our native multimodal big model is less than 30 years old, the person in charge of the biocomputing big model is just over 30 years old, and the person in charge of the embodied big model is also around 30 years old. Rather than saying that youth is a personal label, I prefer to regard it as a label of Zhiyuan. I also very much welcome all young, motivated, thoughtful and capable young scientists to join Zhiyuan to do future research together and make breakthroughs in core technologies together.

Q: Any R&D is highly dependent on talent. What are Zhiyuan’s advantages in attracting talent?

A: Talent has always been the core element of a new type of R&D institution like Zhiyuan. Zhiyuan attaches great importance to young people. As long as they have the ability, ideas and drive, they can get the top scientific research positions and projects. Many of Zhiyuan's resource allocations are centered around core and outstanding talents.

Beijing Business Daily reporter Tao Feng and Wei Wei