Text: Read Tron/Shenzhen Business Daily reporter Chen Shu Correspondent Tang Yuting
Photo: Read Tron/Shenzhen Business Daily reporter Zhong Huadeng
Big models are making waves around the world, and Shenzhen AI companies are exploring how to make big models penetrate into thousands of industries and empower the development of the digital economy.
In the administrative service hall of Shenzhen Longgang Political and Numerical Bureau, the large model has been applied to the government consulting service system. Through the structured combing and analysis of a large number of policies, regulations, procedures and other texts, an easy-to-read knowledge base is formed. Help service personnel quickly query and obtain accurate information, reduce reference time, and improve work efficiency. Behind it is the large model called "Tianshu" of Yuntian Lifei Technology Co., Ltd. The company revealed that in the future, the "Book of Heaven" plan to gradually expand to other government agencies and public services, such as legal, finance, education, medical care, transportation, etc.
Founded in Shenzhen in 2014 and listed on the Shanghai Stock Exchange in April this year, Yuntian Lifei has grown from a small team of only a few returnee doctors to a leading enterprise in the industry and embarked on a differentiated artificial intelligence development path.
The "10,000-model war" began, and the "Yuntianshu" responded
Chips and large models have become two hot topics in the artificial intelligence industry. Chips are the bearers of computing power and are the key core technologies that need to be tackled urgently; Pre-trained large models are regarded as the prototype of general artificial intelligence, and it is the current boom in artificial intelligence development. Since the advent of ChatGPT, the wave of pre-trained large models has swept the artificial intelligence industry, and a boom in large models has also been set off in China.
At the "2023 World Artificial Intelligence Conference" that just concluded in Shanghai, Yuntian Lifei exhibited the "Yuntiantianshu" for the first time. According to incomplete statistics, more than 80 large models have been released in China, many of which are from Shenzhen, such as Huawei's "Pangu", Tencent's "Mixed Yuan", Pengcheng Lab's "Pengcheng Mind", IDEA Research Institute's "Jiang Ziya", and Yuntian Lifei's "Tianshu".
Chen Ning, chairman and CEO of Yuntian Lifei, believes that Shenzhen has become a fertile soil for the cultivation of large models. At present, the domestic large model is in the stage of "a hundred flowers blooming and a hundred schools of thought contending", and the research and development of the large model is not only related to the future development of artificial intelligence enterprises, but also related to the improvement of the future productivity level of a city. Therefore, it is important to create large-model products that are deployed locally and can be controlled independently.
Luo Yi, vice president of Yuntian Lifei, told reporters that from the perspective of model pendants, Yuntian Lifei will launch NLP, CV, multi-modal and other large models. From the perspective of model hierarchy, the basic large model of "Tianshu" includes the general large model of the L1 layer, the industry large model of the L2 layer, and the scene large model of the L3 layer. Based on the algorithm development platform and algorithm chip platform, carry out the research and development of basic large models, and produce general large models through pre-training; On the basis of the general large model, combined with the industry know-how (technical know-how), the industry large model training is carried out; Based on the industry large model, fine-tune the subdivision scene dataset to produce the scene large model.
From the perspective of parameter magnitude, Yuntian plans according to the R&D path from small to large parameter magnitude, including 10 billion, 100 billion and other large models of different magnitude. According to the application requirements of different scenarios, comprehensively consider various factors such as timeliness, accuracy, and deployment cost, and match large models with appropriate parameter magnitude. In addition, the efficient basic self-developed framework, according to the needs of different industries and scenarios, adopts the appropriate scale of parameters to achieve the desired effect, which is helpful to expand the edge side application in combination with domestic self-developed chips.
Let the big model go out of the office and into the industry
Yuntian Lifei is one of the few artificial intelligence companies in the industry with full-stack capabilities of chips, algorithms and big data. Chen Ning himself is a chip expert, the company's founding team is mostly from the field of integrated circuits, the company has long built a "algorithm + chip + big data" full-stack capability platform as the goal of future development, "core manufacturing" is ahead of many peers.
In 2015, Yuntian Lifei started the research and development of neural network processors closely related to artificial intelligence algorithms. Li Aijun, vice president and head of the chip product line, told reporters that since 2015, Yuntian Lifei has completed the research and development of 3 generations of instruction set architecture and 4th generation neural network processor architecture, and has been commercialized one after another. Through years of investment, the company has established a core chip team, with an average design experience of more than 14 years.
In 2018, Yuntian Lifei's first chip "Shenmu" 1000 tape-out that can be directly sold to the outside world was successful, and it was officially sold and commercialized in 2019. At the end of last year, the new generation of edge computing chip DeepEdge10 independently designed and developed by Yuntian Lifei was successfully tape-out. At the "2023 World Artificial Intelligence Conference", DeepEdge10 was also displayed, which can be applied to AIoT edge video, mobile robots and other scenarios, and is expected to be put into mass production this year.
The core capability of Yuntian Lifei is "algorithm chip". The so-called "algorithm chip" is not "algorithm + chip", but an AI chip design process that integrates the concepts, ideas and algorithms of chip designers based on a deep understanding of the scenario and quantitative analysis of the key computing tasks of the algorithm in the application scenario. This concept is reflected in both the Deep Edge10 chip and the "Book of Heaven" large model. DeepEdge 10 can provide sufficient computing power for large models in edge inference scenarios, solve various challenges in the application and deployment of large models, including new neural network computing paradigms, etc., which can bridge large models and practical application scenarios, enable large model technology to be deployed to terminals efficiently and at low cost, and help scenarios explore more innovative applications.
"We hope that we can implement the capabilities of artificial intelligence large models into more scenarios through chips, so that the big model is not only a product that can only be used in the office, but can bring real cost reduction and efficiency value to more industries and scenarios." Chen Ning said.
Shenzhen installs "accelerator" for artificial intelligence development
Yuntian Lifei was established in Shenzhen in 2014 and officially landed on the science and technology innovation board on April 4 this year, becoming the "first stock of artificial intelligence in Shenzhen". From the initial entrepreneurial team of a few people, it has developed into an enterprise with nearly 1,000 people today; The business scope has expanded from smart policing to urban governance, smart transportation, smart parks and other fields; The world's first city-level dynamic portrait recognition system, the world's first 5AIOT chip DeepEye1000, the first concept of self-evolving urban intelligent twin... Many "firsts" in the growth of Yuntian Lifei are inseparable from Shenzhen.
Chen Ning has praised Shenzhen's business and innovation environment in more than one media interview, and said: "Shenzhen is the most suitable city for overseas returnees to start a business. Returning to China to innovate and start a business in Shenzhen is the most correct decision in this life, and there is no one." He also said: "Shenzhen's biggest support for Yuntian Lifei is to open the artificial intelligence market to us, and Shenzhen's courage and trust have made us." ”
Last year, the mainland's first special legislation for the artificial intelligence industry, the Regulations on the Promotion of the Artificial Intelligence Industry in the Shenzhen Special Economic Zone, was officially implemented; On May 31 this year, the "Shenzhen Action Plan for Accelerating the High-quality Development of Artificial Intelligence and High-level Application (2023-2024)" was released, and the first list of "city + AI" application scenarios was released at the same time, thereby building a policy system for the high-quality development and high-level application of artificial intelligence of "one case, one plan, one list, and one fund group".
Chen Ning believes that the advantage and feature of Shenzhen's artificial intelligence development lies in having an active market, so it should give full play to this advantage and promote the continuous progress and iteration of artificial intelligence technology in accordance with the idea of "application generates data, data training algorithms, algorithms define chips, and chips enable applications".
"Shenzhen's soil has given birth to Yuntian Lifei, and the various policies issued by Shenzhen have undoubtedly installed an accelerator for the development of the artificial intelligence industry," Chen Ning said, in the future, Yuntian Lifei will also further strengthen technology exploration combined with existing downstream scenarios to develop a more versatile and intelligent large model, so that artificial intelligence can better empower the industry.
From 2021 to 2023, the work report of the Shenzhen Municipal Government mentioned the construction of "Pengcheng Self-evolving Intelligent Twin" for three consecutive years, and was written into the "Shenzhen Digital Twin Pioneer City Construction Action Plan (2023)" released in June this year, which became an important guide for the development of artificial intelligence and smart city in Shenzhen. Chen Ning said that it will give full play to the generalization ability of large models, rapidly expand new scenarios, and further feed back technology through application landing, and accelerate the self-learning and self-evolution of urban intelligent bodies.
Reviewed: Sun Shijian