news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

In 2017, artificial intelligence was first written into the Chinese government work report and became a national strategy. In 2018, Beijing Zhiyuan Artificial Intelligence Research Institute (hereinafter referred to as "Zhiyuan") was born in a small office in Lingchuang Space. At that time, the first generation of GPT model launched by OpenAI had not yet occupied the media headlines.

In 2023, Zhiyuan will upgrade my country's first ultra-large-scale intelligent model "Wudao" to 3.0, becoming one of the three most cutting-edge AI institutions in the world in the mind of Microsoft President Brad Smith. In the field of large models, Zhiyuan is the only non-enterprise research institution in the world that is not backed by a large company, and is also the earliest new research institution in China to systematically deploy large model technology research and development, open source ecosystem construction, talent training and enterprise cultivation.

"Each time in the past, equal rights in science and technology have created major industrial opportunities," said Wang Zhongyuan, the new president of Zhiyuan, who has both corporate experience and a research institute background. He has a deeper understanding of how science and technology can empower industries. The country's definition of artificial intelligence has also become more specific: artificial intelligence is an important engine for developing new quality productivity.

Work hard before the trend

Once, twice, Dark Side of the Moon CEO Yang Zhilin was surrounded by participants at least three times at the 2024 Beijing AI Conference. A year ago, when OpenAI founder Sam Altman gave a video speech at the AI ​​Conference, the restlessness at the scene was equally obvious.

From 2023 to 2024, AI scientists, CEOs of big tech companies, and founders of star startups gathered in various places to preach the big model. Whether they were technical experts or not, they had all heard of Zhiyuan, and many of them had directly participated in Zhiyuan's big model research projects. Baidu CTO Wang Haifeng was once a director of Zhiyuan, Zhipu AI founder Tang Jie was once the vice president of Zhiyuan, and Yang Zhilin participated in the research and development of Wudao.

These AI trendsetters have not only recently become connected with Zhiyuan, and Zhiyuan did not join in the big model trend until it became popular.

In 2019, Zhiyuan began to lay out the big model, and in 2020, it formed the Wudao research team and started the research and development of the big model. In March 2021, the Wudao 1.0 big model was first released, and in June, Wudao 2.0 was released. Zhiyuan used 1.75 trillion parameters to create the "world's largest" big model record at the time, which was 10 times the number of parameters of OpenAI's most advanced big model GPT-3 at the time.

Scientific research requires physical sensations, asking questions and making judgments. The judgment that "the era of big models of artificial intelligence is coming" gave Zhiyuan unreserved courage. Even the Chinese term "big model" was first proposed by Zhiyuan.

What is a big model? Huang Tiejun, chairman of the Academia Sinica, believes that it must meet three conditions: large scale, with parameters of more than 10 billion; emergent, able to generate unexpected new capabilities; and universal, not limited to specific problems or fields, and able to handle a variety of different tasks.

In 2023, big models entered the public eye. The Wudao series of models has been upgraded to version 3.0, covering basic big models such as language, vision, and multimodality, and is fully open source. At that time, in the discussion on basic models at Stanford University, Zhiyuan was listed alongside technology giants such as Google, Microsoft, and Facebook (now Meta), becoming a representative institution for the world's big model research.

"Historically, the emergence of most research results is accidental. No one can plan it. All efforts are made to increase the probability - to bring together outstanding researchers and provide them with a community environment where they can exchange ideas, discover problems, and find collaborative partners." The preface that Zhang Hongjiang, founding chairman of the Academy of Sciences, wrote to the autobiography of Turing Award winner Yann LeCun is more like the reason why Academy of Sciences came from behind and caught up quickly.

To do system engineering

In early 2018, Beijing issued the "Beijing Implementation Measures for Supporting the Construction of World-Class New R&D Institutions" to carry out a leap-forward reform of the science and technology system. In December, under the guidance and support of the Ministry of Science and Technology and the Beijing Municipal Party Committee and Government, Zhiyuan was officially established.

Previously, the scientific research management process was complicated, with a long cycle from project proposal to guideline release to funding application and review and approval, making it difficult to adapt to the ever-changing scientific research needs in a highly competitive environment. Under this system, research institutions proposed project proposals in October 2020, and large-scale model research could not be officially launched until 2022 at the earliest. The Zhiyuan model took less than 5 months from project establishment to the launch of the large model.

This is a new type of R&D institution between universities and enterprises. It is new in that it does not use papers or products as the final evaluation indicators, but aims to build an innovative system; it is new in that it brings together scholars from different institutions and enterprises such as Tsinghua University, Peking University, Facebook AI Lab, Baidu, etc. to do big things; it is new in that it aims at big problems, keeps a keen eye on major scientific issues, and makes forward-looking arrangements.

"The university model has been running for decades, and it is difficult to carry out systematic research and development in an organized, large-scale, and cross-team manner. Companies will also invest in research and development, but they prefer research and development that is strongly related to business. Zhiyuan will do research projects that will take 3 to 5 years or even longer to see results," said Wang Zhongyuan in an exclusive interview with a Beijing Business Daily reporter.

At present, Zhiyuan's confidence also lies in the country's firm belief in artificial intelligence. In March 2024, Li Qiang, member of the Standing Committee of the Political Bureau of the CPC Central Committee and Premier of the State Council, made it clear during a research trip to Beijing that artificial intelligence is an important engine for the development of new productivity.

New quality productivity is the advanced productivity state in which innovation plays a leading role, breaking away from the traditional economic growth mode and productivity development path, with high-tech, high-efficiency and high-quality characteristics, and conforming to the new development concept. It is spawned by revolutionary technological breakthroughs, innovative allocation of production factors, and deep transformation and upgrading of industries. "Every time technological equality was achieved in the past, it was able to generate major industrial opportunities, and the big model can bring new technological equality," Wang Zhongyuan firmly believes.

For example, the big model is the carrier of "intelligence", and the AI-centered wave is the operation of intelligence. The bottom layer is the technical software and hardware system, and the top layer is the AI ​​application. The big model is in the middle of the two, playing the role of "trunk". The significance of the big model is to turn "intelligence" into a public service like water, electricity, and the Internet, and provide AI services to a large number of enterprises or individuals through cloud computing.

This is a systematic project, "which requires concentrated investment of resources and manpower. It cannot be done by many people doing their own thing, but a technical system must be formed." Huang Tiejun gave an example, "The development of artificial intelligence is like steelmaking and power generation. It requires a complete technical system to ensure the production of high-quality steel, stable power generation at a relatively low cost, etc." Huang Tiejun said.

To be ahead of the industry

The technical system built by Zhiyuan includes: a large model family bucket, a large model operating system, a data set, a training framework, an operator library, etc. These achievements and Zhiyuan’s vision and goals are hung on the wall on the first floor of Zhiyuan Building.

The vision and goals are divided into five major sections, including mechanism and system, industrial development, etc., which can be summarized as innovation. "Zhiyuan will do the research and development of the most cutting-edge artificial intelligence technology, to lead and predict the development of artificial intelligence, and to be ahead of the industry," Wang Zhongyuan explained in detail, "Zhiyuan will do research and development 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."

In his opinion, when the technical capabilities of large models reach a certain level, they will be divided into two major directions. "One direction is to combine with products, promote applications, and exert commercial value. On the other hand, there are a small number of institutions that continue to iterate and optimize the most advanced large models. Whether it is enterprises or research institutions, they should continue to tackle the technology."

For example, in the case of multimodality, most domestic companies choose the DiT architecture for research and development. "This is because DiT is a proven route. Zhiyuan hopes to train information of different modalities, such as text, images, videos, and voice, in one model from the beginning." Wang Zhongyuan used the human brain as an analogy, "This multimodal big model can see the world, understand and reason. In the future, the big model will be combined with hardware, that is, embodied intelligence, and will be able to enter the physical world to serve humans."

Enterprises are close to the market and look for scenarios upwards, while R&D institutions focus on core technology breakthroughs and provide support downwards. Zhiyuan belongs to the second type, which is far away from applications and scenarios, but uses open source methods to support the industry.

"Artificial intelligence is not equivalent to a big model, it is just a school of artificial intelligence," Wang Zhongyuan explained to the Beijing Business Daily reporter. At present, Zhiyuan is also going all out on other artificial intelligence technologies such as brain-like research and digital heart, which means that there is no upper limit to the imagination space of artificial intelligence in other industries.

Beijing Business Daily reporter Wei Wei