Wen | Han Jiang, a weekly magazine of Finance and Economics
Editor| Zhang Ze
<h1 class="pgc-h-arrow-right" data-track="5" > when AI becomes coal-water power</h1>
In 2015, the Go battle between South Korean chess player Lee Sedol and Alpha Dog brought AI into the public eye. But for a long time after that, in people's impressions, AI is just a bunch of high-tech endorsements in the laboratory, very hot but rarely landed. Nowadays, AI technology is gradually crossing the various barriers of commercialization from the door gates and gates of sensorless passage and the beauty filters in mobile phones to urban infrastructure like hydropower and coal, gradually crossing the various barriers of commercialization and creating huge social value.
In Shanghai, which has an average daily management population of up to 30 million, there is a large number of grid members active in the grassroots frontline, whether it is to check safety hazards or mediate neighborhood disputes, when the masses need help, they can always be seen. However, on the streets of Shanghai, where there are many alleys and small shops, in the face of a huge number of and delicate practical problems, it is obvious that manual grid members alone cannot be efficiently disposed of. Today, behind the ignorance of many citizens, AI is becoming the invisible assistant of urban management.
Starting in 2020, Shanghai, which is at the forefront of digital transformation, has launched the "one network unified management" platform, and more than 15 million components such as manhole covers, street lamps, and fire hydrants in the city, and more than 26,000 kilometers of underground pipe networks have been connected.
(Photo: SenseTime Ark City Open Platform)
As a member of the system construction, SenseTime and Shanghai Changning District jointly piloted the "AI + one network unified management", through AI scene segmentation and other technologies to transform 1,000 cameras in the region into intelligent perception neurons, the camera can identify where there are problems such as garbage overflow, bicycles parked and placed randomly in a few seconds, and push these situations to the government Affairs WeChat used by the grid members in real time to file a case. Data show that since the operation of the system, more than half of the events have been disposed of within 4 hours, and the fastest can be completed in 20 minutes from identification to disposal, which greatly improves the management efficiency.
Recently, electric vehicle explosion news has been frequent, and how to manage electric vehicles entering elevators is a headache for many jurisdictions. In Tianlin Street, Xuhui District, Shanghai, SenseTime's AI technology not only senses electric vehicles entering the elevator, but also checks whether life passages are blocked. Data show that with the help of AI, the complaint rate related to life channel incidents has dropped by 90% month-on-month.
In Shanghai's public spaces such as ports, construction sites, and historic former residences, SenseTime's AI technology is escorting the safety of the city. These are based on the SenseTime Ark City Open Platform. SenseTime's prospectus disclosed that the platform had been deployed in 119 cities at home and abroad during the reporting period.
Not only is it urban auxiliary management, but also in the digital transformation of enterprises, SenseTime also has such an ark of wisdom - SenseTime Ark Enterprise Open Platform, which has landed in the fields of manufacturing, transportation, and residential property management. For example, in the production workshop of FAW Group, the SenseTime Ark Enterprise Open Platform has been integrated into the existing production line, which can help quality inspectors automatically detect more than 34 defects in 6-12 seconds, avoiding losses caused by product defects in the production process.
(Photo: SenseTime Ark Enterprise Open Platform)
SenseTime's AI has also long expanded to autonomous driving. At this year's World Artificial Intelligence Conference, SenseTime announced its independent new brand of smart car solutions, SenseAuto Smart Car Platform, which covers L2-L4 level autonomous driving. In fact, as early as 2017, SenseTime has begun to lay out this field, and honda has jointly developed L4 level autonomous driving.
(Photo: SenseTime Smart Car Platform)
As smart cars begin to land, the seeds sown 5 years ago have begun to bear fruit. According to the data disclosure, SenseTime has cooperated with more than 30 domestic and foreign head partners and customers, and has been selected as a supplier of more than 50 models, covering a total of more than 20 million vehicles (within the life cycle) of fixed-point mass production projects in the next few years, and will apply intelligent technology in the automotive industry on a large scale.
From urban management, corporate services to personal life, from smart business, smart city, smart life to smart cars, SenseTime's AI technology is empowering more and more industries, and AI is becoming an infrastructure like water, coal and electricity.
<h1 class="pgc-h-arrow-right" data-track="93" > "AI device" focused on infrastructure</h1>
The realization of such a large cross-border is mainly due to the "AI big device" that SenseTime has been continuously polishing for 7 years. What is an "AI Device"? It is a new artificial intelligence infrastructure built by SenseTime - by integrating a strong computing power foundation and algorithm capabilities, it can disassemble and collide massive amounts of data, dig deep into potential value, and break the boundaries of cognition and application.
As early as 2016, when AI technology went to the industry, most industry participants tended to brush their faces to pay for such a large and frequent application, and among these companies, SenseTime was the first AI company to integrate biometric technology into the subway ticketing system of more than 10 million people in 2019. With the in-depth layout of the industry, SenseTime also realized earlier that in urban management, there are still a large number of long-tail scenes such as garbage dumping, vehicle illegal parking, and occupied road operations, although it is just needed, but it has not been intelligently transformed, and a large number of AI companies are competing for limited head scenes. Xu Li, co-founder and CEO of SenseTime, also pointed out that "the core must be to make a breakthrough in the performance of long-tail applications to truly form a value closed loop, which is the most critical point for artificial intelligence to penetrate into the industry." ”
The reason why the long-tail scene cannot be broken is that before the AI landed, everyone could only rely on project customization and pile of heads to make a single breakthrough. In this way, most of the cost of production factors cannot be reduced, and the willingness of users to pay for it will be reduced.
However, in the past few years, driven by the need for general problems, some general language models, general vision models, and even general multimodal models have begun to make breakthroughs. In the process of switching new solutions, the more versatile the model, the larger the training data, the larger the model parameters, and the higher the requirements for computing power.
Why not make a particle collider similar to that in physics? You know, the results of particle collisions in many processes are unpredictable, and through continuous exploration attempts, in the midst of randomness and mutation, it is possible to find some real laws that can explain the physical world.
(Photo: SenseTime AI device)
As a result, the "AI big device" came into being. Its essence is to get AI out of the manpower-intensive state. The new paradigm provided by ai devices, exploring the unknown, can not rely on the input that humans already understand as input, can try to put bigger and more data in, find a law that is closer to the essence.
This is a huge challenge: the larger the amount of training data, the more parameters of the model and the higher the requirements for hash rate. Sullivan's report reveals that with the rapid increase in the volume of data and the increasing complexity of algorithms, the computing power required to train the most advanced large-scale AI models has increased by a cumulative 1 million-fold over the past 10 years. For example, in 2020, OpenAI launched the strongest GPT-3 language model in history, and the data training volume behind it was as high as 570GB of text, 355 GPU years, and cost about $12 million.
But it's a more rewarding adventure. From the beginning of its inception to the present, after 7 years of continuous polishing, the current "AI large device" as a whole includes three pieces: computing power layer (AI chip and processing card + AIDC + AI sensor), platform layer (model production + training platform + data platform) and algorithm layer (algorithm toolbox + open source framework).
The prospectus discloses that the number of commercial models produced through "AI large devices" is increasing rapidly, from 1152 in 2019 to 9673 in 2020, and as of June this year, more than 22,000 artificial intelligence models have been produced. Compared to the industry average of weeks, SenseTime's R&D and engineering teams can reduce development time to hours. In 2019, 2020 and the first half of 2021, senseTime's R&D personnel increased the average number of commercial models produced per person per year from 0.44 to 3.45 and continued to increase to 5.24, greatly improving production efficiency. In the long run, this is a key step in driving the reduction of AI production factors.
In order to further enhance the production capacity of the "AI device", SenseTime's Shanghai Lingang Artificial Intelligence Computing Center (AIDC), which will invest 5.6 billion yuan in 2020, will be put into use in early 2022, which will become one of the largest supercomputing centers in Asia and will also provide stronger computing power support for the "AI large device".
(Photo: SenseTime Shanghai Lingang Artificial Intelligence Computing Center)
From the prospectus submitted by SenseTime, it is not difficult to see that "60% of the funds raised in the public offering will be used for research and development", from the beginning of its establishment to the present and beyond in the time period, SenseTime will continue to invest most of its resources in the creation of original technology long boards.
<h1 class="pgc-h-arrow-right" data-track="95" > Shang Tang's "long-termism" chess game</h1>
"AI Device" is one of the core competitiveness of SenseTime's strategic design and continuous polishing based on the pursuit of long-term value.
In fact, long-term value is something that SenseTime's management continues to think deeply about. What is the value of SenseTime?' Xu Li proposed that SenseTime's mission is to "let artificial intelligence lead human progress". Then, Professor Tang Xiaoou, the founder of SenseTime, deliberately added four words before this sentence, "Adhere to the original".
"Sticking to the original" - it is easier to understand doing things in the laboratory, but this time it is necessary to use the industry, which is a little different. The application of original technology and original mode exists in the market exploration and education period, and the follower has ready-made practical guidance on the maturity of technology application and the development of business model, so many model innovations are not the first original in the end, but the technical scale advantages and research and development accumulation of the original are difficult to be easily copied by the followers.
If a company chooses to adhere to the original, to endure in the complex and changeable market environment, and unswervingly go on, the test is the company's faith and determination. From SenseTime's research and development and investment in talents, it can be confirmed that it does adhere to the mission of originality.
According to SenseTime's prospectus, in 2018, 2019, 2020 and the first half of 2021, SenseTime's R&D expenditure was RMB850 million, RMB1.92 billion, RMB2.45 billion and RMB1.77 billion, respectively, with a cumulative total of about RMB7 billion in three and a half years. According to the disclosed financial data, in addition to convertible, redeemable changes in the fair value of preferred shares, and share payment expenses arising from equity incentives, huge R&D investment has become a reason for losses, but it has also created SenseTime's moat in the longer-term industry competition.
At the same time, originating from the multimedia laboratory of hong Kong Chinese University founded in 2001, the founding team of SenseTime established the company in 2014, almost recruited as many corresponding graduate students and doctoral students as possible in the circle, built its own top talent pool, and gradually formed a hematopoietic mechanism including mentors and interns.
(Photo: SenseTime and Tsinghua University "Perceptual Computing" Industry-University-Research Deep Integration Project)
Nowadays, SenseTime has jointly established joint training laboratories with Chinese University of Hong Kong, Zhejiang University and other universities, and has also formed a guarantee for the supply of Talents in SenseTime with the technology landing research of the industrial community. At present, SenseTime has 40 professors and more than 5,000 employees, of which about two-thirds are scientists and engineers, including more than 250 doctoral and doctoral candidates; in terms of academic research, the company has won more than 70 global competition championships, published more than 600 top academic papers, and has more than 8,000 artificial intelligence patents and patent applications, ranking leading the industry.
Without these early strategic investments and the accumulation of talents, SenseTime's AI technology will not be able to move from the laboratory to the industry, and it is impossible to talk about landing in such a complex large project as "one network unified management" in Shanghai.
(Photo: SenseTime ARTIFICIAL INTELLIGENCE Education Textbook)
In addition to senior talents, SenseTime has also established an education sub-brand to promote the basic education of AI technology to basic schools such as primary schools, junior high schools and high schools, and cultivate talents for China's future AI industry.
Continuous investment in talent and underlying technology has brought rewards to SenseTime. According to the prospectus, in 2018, 2019 and 2020, the company's revenue was 1.85 billion yuan, 3.03 billion yuan and 3.45 billion yuan respectively; in the first half of 2021, SenseTime's revenue was 1.65 billion yuan, an increase of 91.8% over the same period last year. On the indicator of gross profit margin reflecting the profitability of the company's main business, SenseTime also performed well. From 2018 to 2020, SenseTime's gross profit margin was 56.5%, 56.8% and 70.6%, respectively, and in the first half of 2021, the gross profit margin further increased to 73%. According to Sullivan's industry report, SenseTime has become asia's largest AI software company based on its 2020 revenue.
Since its establishment 7 years ago, SenseTime's business map has become more and more abundant, becoming a leader in AI enterprises, SenseTime is currently involved in the formulation of more than 80 industries, national and international standards, and has worked closely with a number of domestic and multilateral institutions on the sustainable and ethical development of artificial intelligence, and assumed the social responsibility of promoting the sustainable development of AI.
If we want to really let artificial intelligence lead human progress, what awaits SenseTime is a long marathon, but only by unswervingly investing can we stabilize the technical foundation and form our own moat. In this IPO, SenseTime said that 60% of the public offering funds will also be used for research and development. The development of hard technology is a long-term practice, which requires AI companies to have patience and perseverance, and also needs a more tolerant vision of the market.