Original title: Beijing Artificial Intelligence Data Operation Platform Release (theme)

A number of data sets collected in the public domain will be open source, and users can download them for free (subtitle)

Beijing Daily reporter Liu Suya

A few days ago, the Beijing AI data operation platform was released at the 2024 Beijing Zhiyuan Conference, and the "Industry Data Set Scenario Application Innovation Plan" was launched simultaneously. Huang Tiejun, president of Beijing Zhiyuan Artificial Intelligence Research Institute and professor of Peking University, said that to build a big model ecology, first of all, we need to establish a data circulation mechanism to promote positive feedback from data to intelligence.

Aggregated data exceeds 700 trillion bytes

"Without open source datasets, it is hard to imagine how AI can achieve such rapid iteration, however, these datasets are far from enough." Lin Yonghua, vice president and chief engineer of Zhiyuan Research Institute, said frankly, especially the data volume of Chinese Internet is far from the same, and the problem of data islands is also serious.

The Beijing AI data operation platform released this time gathers a large amount of general data and industry data, supports multiple modes such as text, image and video, and creates a full process data processing tool. More than 700 trillion bytes of general data sets can be used for general model training. At the same time, 4.33 trillion bytes of industry specific data are also very important for model training.

At present, the total amount of text data sets in the global open source industry is only about 1.2 trillion bytes. "Almost every item of this open source industry data set is far more than the total of the world's open source industry data sets, and it is the world's largest multi industry bilingual data set in Chinese and English." Lin Yonghua said that at present, the data set includes 18 categories of industry data such as medical, education, law, and news, and will be further expanded to about 30 categories in the future.

According to the characteristics of different data, the platform has identified three ways of using data. A number of data sets collected in the public domain will be open source, and users can download them for free; Subjects who are interested in exchanging high-quality data sets can participate in the construction of data pools, cooperate to build and share data sets; The high-value data set ensures that the data does not go out of the security domain during the processing and training of the model, and ensures the data security through the integrated mathematical model.

Challenge the new route of multimodal model

"In 2021, Zhiyuan Research Institute released the enlightenment 1.0 and 2.0 models, which created a series of records such as" China's first "and" the world's largest ". It can be said that Zhiyuan is closely related to the big model." Wang Zhongyuan, president of Zhiyuan Research Institute, said that with the development of the big model, artificial intelligence has gradually entered the era of general artificial intelligence, Emu3, the original multimodal world model currently being trained by the institute, has once again targeted one of the most challenging technical routes in the industry.

At present, models such as cultural map, graphic text and cultural video have their own corresponding architectures and methods, which are difficult to be compatible with different tasks. For example, the Wensheng video model Sora cannot understand images and videos. "Emu3 aimed at multimodal integration from the beginning of design, its generation and understanding ability has also been unified, and it also has more modal scalability." Wang Zhongyuan said that based on the multimodal autoregression technology path developed by Zhiyuan Research Institute, images, videos, words and other modes can be jointly trained.

This mode enables the model not only to read text, but also to "read pictures" and "watch videos". Upload a video, and the model can quickly identify the feeling of happiness and excitement of the people in the video. "If this route achieves a breakthrough, it will be a major technical contribution to the AI industry." Wang Zhongyuan said that at present, the model has the ability to generate high-quality pictures and videos, continue to write videos, understand the physical world and other multimodal capabilities, but there are still some imperfections in the middle. After continuous training and safety assessment, Emu3 will gradually open source.

Successive output of world-class research achievements

The development of embodied intelligence technology has pushed AI from the digital world to the physical world. In the exhibition hall of the conference, after the training of general grabbing model, the robot can effectively grasp objects of arbitrary shape, reflective transparency and other objects, achieve more than 95% of the success rate of grabbing on industrial level real machines, and achieve the world's leading commercial level of action execution. The robot also has the ability to think about open instructions. When the tester proposed "I'm hungry", the robot found edible oranges and bananas among more than 10 items placed on the table. "Which do you want, orange or banana?" the robot asked. After getting the answer to the need for oranges, he grabbed them and put them into the basket.

The robot arm is equipped with a probe to move back and forth on the chest of the detected person, and the image of heart beating is displayed on the bedside display screen. Wang Zhongyuan introduced that the world's first intelligent cardiac ultrasound robot jointly developed by Zhiyuan Research Institute, Tsinghua University and 301 Hospital can quickly calculate and extract cardiac features in a high-speed dynamic environment. The clinical verification results show that its detection accuracy and efficiency are basically the same as that of human doctors, but its stability and comfort are significantly higher than that of human doctors, which is of great significance for improving the popularity of ultrasonic medicine.

The world's first low carbon monomer dense trillion language model was launched, the world's first real-time twin heart computing model was built, and the whole atom biomolecular model reached the world's leading level... Wang Zhongyuan introduced that in the past year, the Zhiyuan Research Institute has made efforts in the direction of multimodal large models, physical large models, and biological computing large models, and has made many world-class achievements. In the next few years, the Institute will continue to develop in these directions.