Overnight, the "magic capital" seemed to have become a "model capital".
Step out of the Longyao Road subway station and come to the modern, busy, artistic West Bund of Shanghai. It is a few hundred meters away from the shoreline of the Huangpu River, and you can see casual pedestrians along the road, more of them are white-collar workers in a hurry. Within a radius of one kilometer, there are tech giants and AI model startups such as Tencent, Alibaba, Microsoft, Shanghai Artificial Intelligence Lab, SenseTime, Wuwen Xinqiong, and Midu Honeynest.
Right next to this subway station is the country's first and Shanghai's only large-scale model innovation ecological community "Moldspeed Space". On September 28, 2023, after 38 days of rapid preparation, the mold speed space was officially unveiled. In less than half a year, it has attracted more than 60 large-scale model enterprises to settle in, and Xuhui District, where it is located, has gathered nearly 300 upstream and downstream enterprises in the industrial chain, and the scale is still expanding.
Artificial intelligence (AI) is a key industry in Shanghai, with an average annual growth rate of more than 29% from 134 billion yuan in 2018 to more than 380 billion yuan in 2022, and the number of enterprises above designated size has also increased from 183 in 2018 to 348 in 2022.
The industry generally believes that the large model is the core engine for general artificial intelligence, which will bring about another leap in artificial intelligence and has become the focus of global scientific and technological competition.
In 2024, Shanghai will occupy 24 of the first four batches of large models in the country that have passed the online filing. Of the 24 large-scale models for the record in Shanghai, Xuhui District occupies 15 of them, which are concentrated in the mold speed space.
It can be said that the mode velocity space is the best window to observe the development of Shanghai's AI large model industry. This area is generating more and more agglomeration, visibility and influence.
In late March, the 2024 Global Developer Pioneer Conference (GDC) opened at the West Bund of Shanghai, with the theme of "Modular Velocity Space - The Source of Modular Power for Developers", and many people braved the rain to attend. In February, Chen Jining, Secretary of the Shanghai Municipal Party Committee, investigated the Shanghai Modular Speed Space Innovation Ecological Community, presided over a symposium on artificial intelligence science and technology innovation institutions and enterprises, listened to opinions and suggestions, and conducted on-site office.
If it is said that it is now the "tuyere" of the large model - although there is a long journey in the future - at this point in time, Shanghai has seized the opportunity and ran in the forefront. The mode speed space is striving to become the "main factor" to win the initiative in global scientific and technological strategic competition on behalf of Shanghai and even the country.
Recently, the reporter took the mold speed space as the starting point to visit and interview a number of large model enterprises and relevant government leaders to explore the process and basic logic of Shanghai's layout of the AI large model track.
"Upstairs and downstairs is upstream and downstream"
From the map, on the west bank of Shanghai, where the mode speed space is located, there are well-known scientific research institutions and technology enterprises such as Shanghai Artificial Intelligence Laboratory, Qizhi Research Institute, SenseTime, Alibaba, and Tencent, forming an artificial intelligence production and research ecosystem. Modular velocity space, in which it plays the role of a professional incubation and acceleration carrier.
In late March, the reporter walked into the mold speed space, which is similar to many entrepreneurial parks, with simple and comfortable facilities and a bright environment. At present, a total of 20,000 square meters of the first and second phases of the mold speed space have been put into operation, and the third phase has been planned and constructed.
Different from other entrepreneurial parks, the mode speed space focuses on the field of AI large models. Many companies have a company profile at the door, some focus on financial and business vertical models, some are committed to improving the shopping experience through large models, some are oriented to government and enterprise services (ToB/ToG), and some directly serve individual users (ToC).
On the 5th floor of the mold speed space, Zha Xiaohui, a partner of "Jin Shengyuliang", had just finished the last meeting and quickly freed up a corner in the public area to communicate with reporters. She said that Jin Shengyuliang belongs to the financial vertical model, which can be regarded as an assistant to investment consultants, helping to answer customers' questions about individual stocks, industries and the overall market.
"Our customers have a lot of financial influencers, who used to need to record short videos every day or regularly, and they would lose fans after a long time of stopping. Now with a digital clone, you can solve the problem of the lack of skills of the big V clone. Hot spots, content, text, and portrait actions can be generated with one click. Zha Xiaohui said that Jin Shengyuliang's corpus is updated 24 hours a day, and many financial investment hotspots can be answered in real time.
Many entrepreneurs in the mold speed space are as busy and capable as Zha Xiaohui, and reporters can hear and see R&D personnel holding offline or online meetings in Chinese and English almost at any time. Many Chinese AI technicians are working overtime every day in the hope of surpassing OpenAI.
Zhang Ning, director of the Science and Technology Commission of Xuhui District, said in an interview with reporters that in this round of artificial intelligence enterprise entrepreneurship, the founding team has a strong industrial background: 90% of the founding team of the mold speed space has a background of continuous entrepreneurship or large factories, 80% of the founders come from the head of the famous school, and 50% of the team has overseas experience.
There is a saying here: "Going up and down the stairs is going up and down." Zhang Ning explained: "In one building, different enterprises in the industrial chain such as the underlying technology of the large model, application research and development, scene design, computing power support, and product marketing are gathered. ”
The government, as the "shopkeeper", hopes that within a very short physical space, researchers can find anyone they want to see, connect with any resources they want to contact, and obtain any life services they want, which is the "agglomeration" that the modular space should achieve.
This kind of cluster effect makes the enterprises settled in it feel it. "A lot of our clients are in this building, and they can walk into the next office and ask questions, or learn about the latest developments in the industry at an event organized by the incubator. Zhang Jiaqing, the founder of "Open Expressive", attributed the company's accelerated development to the spatial pattern of one building gathering the entire industrial chain.
Liu Yidong, president of Mido Honey Nest and CTO of Mido, told reporters that the company is based on the general model of scholars developed by Shanghai Artificial Intelligence Laboratory, and makes efforts in the application of government affairs verticals, and the self-developed government affairs model of Honey Nest can provide knowledge management, data mining and other services for enterprises and governments, and has released version 2.0.
"As a pedestal model, 'Scholar' also needs the participation of vertical applications, and we hope to become a strong ecological builder in it. Liu Yidong said. The assistance is mutual, and Mido Honeynest has previously provided corpus to Shusheng Puyu and joined the "Corpus Data Alliance" led by the Shanghai Artificial Intelligence Laboratory. High-quality, large-scale, secure and trustworthy corpus data is essential for the training and performance improvement of large language models.
Zhang Ning said that while bringing together more than 60 large-scale model enterprises, the mold speed space has formed a comprehensive innovation and entrepreneurship guarantee such as computing power scheduling, open data, evaluation services, financial services, and comprehensive services.
For example, computing power is the biggest bottleneck for large model companies, and without the support of infrastructure with computing power as the core, large models cannot run at all. The computing power scheduling platform of Modespeed Space has connected with 9 suppliers including SenseTime, Huawei, Tencent, Alibaba, and the three major telecom operators, with more than 50,000 GPU cards, to ensure that the computing power of large-scale enterprises is available, sufficient, and easy to use.
"Half-step advantage"
Artificial intelligence is not a new thing, if you count from the 1950s when Turing asked "can machines have human intelligence", AI has gone through more than half a century of development. It's just that it wasn't until the past two years, with the emergence of ChatGPT and Sora, that the AI model finally detonated the entire industry, allowing many people to see the "business value".
2023 is known as the first year of large models, and domestic and foreign technology and AI companies have tested the waters, and more than 200 large models have emerged in China, setting off a "100-model war" for a while. General model, industry model, vertical model...... All kinds of large models appeared, and all walks of life wanted to get a piece of the pie.
Although the industry is very lively and entrepreneurs are very excited, many people seem to be confused about how to put the large model into commercial use. Some industry insiders found that unlike the investment boom brought by ChatGPT in 2023, after the emergence of Sora in 2024, many large manufacturers appear to be much more "calm".
Also in 2023, when many large factories and enterprises finally change their vision of AI from "technology" to "industry", Shanghai Xuhui has quietly created an "AI ecological community" to bring together the upstream, midstream and downstream of the industrial chain to empower applications in an all-round way.
In other words, when everyone is rolling models, Shanghai has created a new "way to play".
After a round of "burning money" in 2023, many companies have reached a new consensus: the general model is just a game for a few people, and in the next stage, the opportunity is in the industry model and vertical model. Not long ago, Robin Li, the founder of Baidu, also said: "Don't roll the model, roll the application, only the application directly creates value." ”
More than 80% of the more than 60 large-scale model companies settled in the mold speed space are application companies from the vertical field. "Vertical may be the only choice for most companies in the process of commercial landing. From the perspective of the industry, Zha Xiaohui analyzed that the cost of the general model is too high, and the product is easy to be "modest", "It seems that everything can be done, and it seems that nothing can be done well." ”
In the industrial outlet, it is commendable that the mold speed space can quickly gather a number of high-level large-scale model enterprises. In Zhang Ning's view, on the one hand, this is based on the accumulation of artificial intelligence industry in Shanghai and Xuhui District, "we have long grasped the transformation trend and industrial outlet of artificial intelligence".
Since 2017, Shanghai has taken the lead in the development of the artificial intelligence industry in the country, and in 2018, the first World Artificial Intelligence Conference was held, and Xuhui has been the core carrier for 6 consecutive years.
Many people think that the mold speed space was "born", and Zhang Ning knows the accumulation and layout behind it. A few years ago, when people saw AI as a "technology" that didn't make money, Shanghai started to treat it as an industrial track. A typical example is the cultivation and development of SenseTime, an artificial intelligence unicorn company.
On the other hand, Zhang Ning said that the rapid birth and growth of the mold speed space, the relevant departments have done a lot of research and analysis, peeling off the cocoon, "we are to see the large model industry and then attract investment." Zhang Ning said that after many years of dealing with AI scientists and entrepreneurs, he found that good policies are important for large model entrepreneurs, but enterprises often prefer and need a government that can understand the large model industry and know how to help large model enterprises.
Based on this, as the first large-scale model innovation ecological community in China and the only one in Shanghai, Mode Speed Space was born in Xuhui. Zhang Ning humbly describes the process as a "half-step advantage."
However, even if there is a "half-step advantage", in the increasingly fierce competition of large models, no city can relax.
In November 2023, Shanghai issued the "Several Measures for Promoting the Innovation and Development of Large Artificial Intelligence Models in Shanghai (2023-2025)", introducing a series of policies and measures to support the innovation and development of large models, promoting the substantive operation of the Large Model Corpus Alliance, and open-source a series of high-quality corpus data, so as to build a better ecological environment for the innovation and creation of global AI developers.
Xuhui District has gone one step further and launched a number of incremental initiatives. This platform lowers the threshold for enterprise large model training and drives the vertical application of large models in typical fields such as finance, education, medical care, and manufacturing.
While attracting high-quality enterprises, Shanghai actively attracts young talents. At the beginning of March 2024, Cao Liqiang, Secretary of the Xuhui District Party Committee, personally led a team to organize more than 30 enterprises and institutions in the high-tech field to recruit at Tsinghua University and Peking University, and more than 1/3 of the participating enterprises belong to the field of artificial intelligence.
Innovations are emerging
At present, four base basic large models have been born on the west bank of Shanghai, namely the general model of scholars in Shanghai Artificial Intelligence Laboratory, the "daily new" senseNova large model system of SenseTime, the MINIMAX-ABAB large model, and the Jieyue Xingchen large model. There are more vertical applications, and there are more than 100 related products.
On March 23, at the 2024 Global Developer Pioneer Conference, Chen Jie, vice mayor of Shanghai, said that Shanghai's large-scale model innovations are constantly emerging, the "Shushengpuyu" large-scale model developed by the Shanghai Artificial Intelligence Laboratory has been released open-source, and a variety of humanoid robots will also be released soon. He also mentioned that the large-scale model industry carrier continues to expand, and the mold speed space has attracted more than 60 large-scale model enterprises to settle in.
As a front-line representative of scientific research, Qiao Yu, assistant director and leading scientist of Shanghai Artificial Intelligence Laboratory, said: "Large models are currently evolving in the direction of multimodality, and we have established a new multi-modal large model evaluation system with three levels: knowledge, alignment, and value, which will provide new reference and guidance for the development and evaluation of large models in the future." ”
At this conference, StepStar released the company's Step series of general large models, including the Step-100 billion parameter language model, the Step-1V 100 billion parameter multimodal large model, and the Step-2 trillion parameter MoE language large model preview - which is also the first trillion parameter large model among domestic startups.
Jiang Daxin, CEO of Shanghai Stepleap Xingchen Intelligent Technology Co., Ltd., a former global vice president of Microsoft, said that the parameters of the model have increased by an order of magnitude from 100 billion to trillions, and the difficulty has also posed challenges to computing power, systems, data, and algorithms.
The reporter noticed that whether it is a basic large-scale model enterprise such as Shanghai Artificial Intelligence Laboratory, SenseTime, and Jieyue Xingchen, or a vertical model start-up such as Midu Honeynest and Jinsheng Yuliang, the current development speed and update iteration are very fast, and they have begun to reap commercial success.
Sheng Shiwei, managing director of the board secretary office of SenseTime, revealed in an interview with reporters that the generative AI business has become the core business of SenseTime from scratch with a revenue volume of more than 1 billion yuan. According to the financial report, in 2023, SenseTime will achieve a total revenue of 3.4 billion yuan, and its generative AI business will generate 1.18 billion yuan in revenue, with a revenue growth rate of 200%, accounting for 35% of the group's total revenue.
In February 2024, SenseTime launched version 4.0 of the "RiRixin" large model, which has achieved the ability to match GPT-4 in multiple scenarios such as code writing, data analysis, and medical Q&A. And just 2 months later, the new version will be iteratively updated.
"In the coming late April, we will have a Tech Day where we will demonstrate the capabilities of our updated 'Everyday New' 5.0 model. Sheng Shiwei revealed that the performance of the "Ririxin 5.0" version is fully benchmarked against GPT4-Turbo.
For start-up large-scale model enterprises, attracting investment and cooperation from large companies is an important opportunity for development and growth. For many large enterprises, "Al+" is the core of this wave of large models, which are expected to empower thousands of industries and become a new engine for the development of the real economy.
"The AI knowledge management that Honeynest has just introduced will be of great help to us. On the afternoon of March 29, Lei Hongjie, general manager of the airborne aviation industry of the central enterprise, said when visiting the mode speed space.
"Will you consider cooperating with Shanghai large-scale model enterprises to apply such new technologies to enterprise management?" the reporter asked.
"I think this is necessary, to improve the efficiency of enterprises, such as the knowledge management, data mining and other models just mentioned, now the big model may be the best solution. "But first of all, we have to ensure security, which is very important for enterprises." ”
"This time I came to Shanghai to see (the large model industry), and we gained a lot. Lei Hongjie added.
With the help of the industrial outlet, vertical large-scale model enterprises have ushered in unprecedented opportunities. Zha Xiaohui also deeply experienced this change. When the company talked about its first brokerage client, it took a full year. But in recent months, new customers have been coming to the door on their own. "The financial industry is increasingly aware of the need for AI. She said.
On the 5th floor of the mold speed space, Best Smart is a large model enterprise focusing on precision marketing of goods, similar to an e-commerce shopping guide, and its co-founder Qiu Yunwei is a post-90s generation who previously worked in Silicon Valley in the United States, and began to take root in Shanghai mold speed space in November 2023.
At present, there are only more than 10 people in this entrepreneurial team, and the main members have previously worked in IBM, Facebook and other technology companies. Qiu Yunwei turned on the computer to show the company's first product launched in China in January 2024, "Sesame AI", which was on the hot search on Weibo on the first day of launch, with more than 100,000 users.
"The future may be full of uncertainties, and the company is currently focusing on verifying the performance of our products in the relevant vertical markets within six months, three to six months, collecting more data, and improving product performance. This young entrepreneur speaks with vigor and vigor.
Although it is a small start-up, if the e-commerce vertical model like BestSmart develops well, it is likely to completely subvert the existing e-commerce model in the future.
Xuhui District Rong Media Center, The Paper Reporter: Li Jiawei, Liu Kun, Hu Hanyan, Wang Shiyao (intern)
Editor: Ye Fangfang