A stronger "push-back feeling" than the previous AI wave.Author Sukhoi.
Edited by Chestnut.
Do you think you know more about strategic direction than I do?At the annual meeting at the end of 2018, the CEO of A Technology Company asked the young man in front of him.
I don't think so, but the security field winter is coming, we should shrink the business, continue to invest in full will only lose the whole market, I can't watch the company jump into the fire pit!”
The young man who said this is Jiang Jun, who is now the CEO of Ant Cloud Technology Group.
Then in 2019, the first class of security equipment really collapsed. Everything is just as Jiang Jun calculated, the original product of 10,000 yuan per camera software authorization has been reduced to hundreds or even dozens of yuan. The market size of 100 billion yuan has also rapidly collapsed to more than one billion or hundreds of millions of yuan, and the entire AI industry has been dragged into a cold winter trough. At that time, the "AI Four Little Tigers" lost tens of billions of yuan and billions of yuan.
Time has passed, and Jiang Jun, who has faded from his edge, is heartbroken every time he recalls this scene.
Now, four years later, Shanghai is cooling off a cliff. In stark contrast to the cold weather, Shanghai has a fiery and strong AI atmosphere. Especially in the Xuhui Binjiang sector, which is an important carrier of Shanghai's innovation center, there are benchmarking new R&D institutions such as Shanghai Artificial Intelligence Laboratory and Qizhi Research Institute, with 72 invention patents per 10,000 people2 pcs.
Ant Cloud Technology Group's AI large model research institute and operation center are located in Xuhui Binjiang.
Ant Cloud Technology Group is an artificial intelligence company focusing on the research and development of AI large models, upstream and downstream application development, and industrial ecological expansion. On December 18, 2023, Jiang Jun and the others ushered in the joy of housewarming. From starting to form a team again by one person, to being full of seats in the office at this moment, Jiang Jun felt a lot of feelings in his heart: "Everything has finally returned to the starting point. ”
Along the way, there are peaks and valleys. Now, Jiang Jun is once again standing at the starting line of the large model. He firmly believes:There must be nothing wrong with sticking to what is difficult and right.
As a veteran of the AI industry, this year is Jiang Jun's 10th year of deep cultivation. He has always adhered to one principle of action, which is first principles:Go back to the essence of things and rethink what to do
Jiang Jun personally participated in the first phase of the AI wave, and his feelings can be summed up in one sentence: I saw him rise a tall building, I watched him feast guests, and I saw his building collapse.
In this wave, Jiang Jun insisted on making first-principles decisions rather than emotion and general knowledge, and many choices were not understood at the time.
In 2012, he gave up as Huawei's sales director to become a clerk in Huawei's sales management departmentIn 2016, he gave up the position of general manager of the business department of 70 people in the management of a listed company and went to a small company with dozens of people to set up a business department from scratchIn 2018, when AI was in full swing, the boss of the company in the position of "AI Four Tigers" put forward a proposal of "comprehensive contraction of the company's business".In 2020, when the "AI trough" was launched, the "5 billion AI plan" was proposed to break through the current barriers.
The relationship with AI began in 2013 when Jiang Jun was in charge of public security monitoring business in a security company in Suzhou.
At that time, although the company ranked fourth in the security industry, life was not easy. The pressure mainly comes from the security giants Haikang and Dahua. Their hardware is inexpensive, and their sales team is large and well-known.
Jiang Jun believes that from the perspective of core technology, the security business can be divided into three stages:
At the heart of the first phase is the hardware camera. The most striking pain point at this stage is that users need to manually identify content across hundreds of cameras. For example, when investigating suspicious persons, many police officers need to spend a lot of time, energy and manpower to conduct a "carpet" investigation.
At the heart of the second phase is the image application software. With relay tracking of security imaging software, map GIS systems and camera locations, criminals can be caught in a matter of hours. A major breakthrough has been achieved compared to the first stage.
At the heart of the third phase is artificial intelligence. Using AI for image search, users can find the identity and whereabouts of a thief in minutes or even seconds.
At that time, Jiang Jun knew that the company was lagging behind Haikang, Dahua, etc. in the first stage. Therefore, he proposed to the upper management to strive for the first place in the second stage and establish the company's "advantage moat".
So between 2014 and 2015, Jiang Jun planned to use Huawei's group army as the combat mode, with the application of security image software as the focus of publicity and sales, and 180 public security exhibitions in three months, quickly becoming the first public security graphic detection software in China.
The precautionary action brought Jiang Jun and them a huge sales profit. However, the competitive advantage of application software is usually only one year, and in order to consolidate the first place and promote hardware sales to catch up with Hikvision Dahua through software sales, Jiang Jun's plan was interrupted by a black swan event in 2016 when he was preparing a larger-scale sales plan.
At the beginning of 2016, the founder of A Technology Company went to Suzhou and brought a test report to Jiang Jun, who has been deeply involved in the security field for more than ten years. According to the report, in a face recognition competition in Guangzhou, 4,600 faces were randomly screened from 8.6 million citizens, and more than 2,100 of them were accurately identified by A Technology Company, with a first-place hit rate of 47%, while the best record of such competitions in the past was only 59%。
AI image recognition technology, which has not progressed in a decade, has improved by orders of magnitude. Jiang Jun keenly captures the trend of "the industry is about to reshuffle".
In his opinion, the security application system will be reconstructed based on AI, and the security field will enter the AI stage unstoppable. The second phase of the company's newly formed moat will quickly disappear because of the development of AI.
Therefore, Jiang Jun immediately urgently proposed to the company to acquire an AI company. But it was rejected by the boss who insisted on independent research and development.
He didn't want to miss the opportunity of the times, so he decisively resigned from his well-paid job and decided to join A Technology Company, which was less than 100 people at the time, as the general manager of the security division, and set up a security team from scratch.
Jiang Jun analyzes according to first principlesThe main role of AI in the security industry is to search for pictures with images。The key is that the clarity of the source and the size of the comparison library have a decisive impact on the accuracy of the system by matching the accuracy of the comparison source and the comparison library. Therefore, improving the quality of the source and controlling the size of the comparison library is the key to improving the accuracy of the AI system. Only after achieving a recognition accuracy rate of more than 30% to 40% can it become a usable system for the actual combat business of the public security organs.
Based on this concept, Jiang Jun redesigned the AI security software and launched the "Fire Eye" system in early 2017. Using "artificial intelligence + big data technology", the whereabouts of the target person can be grasped in real time, and the time and space range of the target person can be predicted, which expands the breadth and depth of face applications. It has comprehensively improved the ability of public security organs to accurately handle incidents and crack down on crimes, and achieved the goal of rapid handling and detection of cases.
2016 is widely regarded as the first year of AI.
Artificial intelligence has gone all the way through the noise and doubts, and finally ushered in a key breakthrough in deep learning in that year. The most well-known example is Deep Mind's AlphaGo defeating Chinese player Ke Jie, known as "Mankind's Last Hope", with a score of 3:0 in a game of Go.
By the 2017 China International Public Security Expo, hundreds of security companies have invested in the development of AI systems, and dozens of them have also advertised that "self-developed AI algorithms are the best". All the ** are talking about AI, and capital is also pouring AI like crazy.
In this "arms race" of AI algorithms, dozens of AI vendors compete in hundreds or thousands of POC tests, and the algorithm evolves at a rate of % improvement every month. In the end, four algorithm masters were born - the "AI Four Little Dragons".
The A technology company that Jiang Jun jumped to is one of the "AI Four Little Dragons".
At the time, the company had rapidly expanded from a few dozen to thousands. The performance of the department in charge of Jiang Jun is growing at a rate of about 600% per year.
Everything looks so good!
Measured by the rhythm of the life cycle of traditional security products, this is an important juncture to expand market share and strive to become a market position.
In 2018, the Internet and even the entire Chinese industry have entered the "second half", and how artificial intelligence develops, how it is implemented in various industries, and how it affects people's lives and work has become the focus of attention in the industry.
Just when the industry entered a crazy sprint, Jiang Jun rethought the follow-up development of the security market. Behind the fiery appearance, he captured the slightest "chill".
At that time, the technology and data of the security industry had fallen into a bottleneck. A technology companies and their peers have billions of capital support, regardless of the cost, at the expense of "self-harm" to fight the battlefield. Jiang Jun believes that such a trending product will go all the way until it completely collapses.
on this, he put forward the strategy of "building high walls, slowly becoming kings, and connecting vertical and horizontal".
Jiang Jun believes that first of all, we should give up the obsession with "security brother". Open source and reduce expenditure, intensive cultivation, and shrink the business to several nodes of major public security customers and provincial customers;Secondly, change their cognitive role, from signing directly with customers to providing "ammunition" as an integrator;In addition, it integrates various industries within the company, such as security, finance, transportation, and commerce, and aims to be a "smart city" and "city brain".
However, this strategy is completely contrary to the direction of the CEO's plan. The CEO hopes to continue to fully enter the security industry and strive for the first market share.
Fierce clashes of opinion are inevitable. Eventually, Jiang Jun opted out of A Technology Company in 2019. As described at the beginning of this article, the "cold winter" has destroyed the invincible security market and also crushed capital's confidence in AI commercialization.
At that time, many investment institutions thought that this group of 'AI' people were bragging about cheating money. Jiang Jun said helplessly to Jiazi Lightyear.
When it comes to deep learning, Jiang Jun describes himself as a believer in it. His affection for artificial intelligence is far greater than other pursuits.
After leaving A Technology Company, Jiang Jun joined Alibaba in 2020 as the head of the ** and AI industry, hoping to make a career that truly changes the status quo of the AI industry with the support of Alibaba's potential.
In Jiang Jun's view, the development of AI technology had begun to slow down or even stagnate.
From the perspective of the three elements of AI, the training data we can obtain has reached the upper limit. How can I get more data?Second, in the case of the collapse of the industry market and the capital winter, it is impossible for enterprises to spend huge amounts of money to buy computing power or rebuild models that are several times more complex than before. Jiang Jun explained," he explainedTo overcome this problem, more resources are needed to increase the scale of computing power, data, and algorithms. In this year, when the 'AI Four Tigers' used hundreds of millions of AI R&D funds, but the progress of AI algorithms stagnated and slowed down, it would take 1 or even 2 orders of magnitude to break through, which would require at least billions of investment. ”
As a result, Jiang Jun proposed the plan of "AI City". He independently designed a 5 billion yuan AI large model platform, planned a massive data collection training pool and computing power at the city-wide level, and led the Alibaba DAMO Academy to work with more than 50 AI companies to promote the breakthrough of China's AI large model.
He hopes that with the joint efforts of ** and industrial capital, the digitalization of AI in cities will be increased from 5% to 70% to 80%. He plans to make AI digitalization cover the entire city through at least a decade of "mandatory" promotion, integrated into **, audio and text-aware neurons.
At the same time, it will open up industry applications in smart cities, and invite dozens of AI companies to develop AI algorithm training and industry applications based on data.
Jiang Jun knows that the larger the project and the more links involved, the risk of failure will increase by several levels. But in order to achieve a breakthrough, such resources and scale are indispensable.
At first, Jiang Jun visited many cities, and the interviewees felt that he was describing a "fantasy", and even a certain district chief, after listening to Jiang Jun's idea, bluntly expressed to the recommender: "He's crazy!".”
But the hard work paid off, and in the end, the leadership team of a certain city agreed to carry out this plan locally after repeated research and research.
It's just that at the end of November 2020, due to some force majeure factors of Alibaba, they were unable to cooperate with ** to complete the super project, and Jiang Jun also left the scene.
Until this year, everyone saw that OpenAI spent three years and invested billions of dollars to gradually upgrade the model parameters from 100 million parameters to a 100 billion-level super large model, and finally achieved a breakthrough in the AI industry.
Time will prove that Jiang Jun's madness at that time was correct. His heart was full of regrets, three years ago, this advanced plan, even if he knew that the opportunity was slim, he was like Don Quixote rushing to the windmill alone, fearlessly pursuing his faith.
Just like Jiang Jun's nickname in Alibaba, Ruiwu, a determined general. He has always called himself an "AI veteran", and when asked why, Jiang Jun explained: "Although my name is General (Jiang Jun), I have been doing it from zero to zero for nearly ten years.
First, pioneering and innovative work. This makes me have to lead by example and lead everyone forward on the front line. ”
After leaving Alibaba, Jiang Jun hibernated for three years, and he finally restarted his journey in September this year and joined Ant Cloud Technology.
The ups and downs of the AI industry in the past 10 years have given Jiang Jun the confidence to command more exciting battles from a higher dimension in this AI tide. From only caring about the business itself, to the direction that needs to be considered, the strength of joint business partners, etc., the responsibility on Jiang Jun's shoulders is also becoming more and more important. He understood that he had to take on the heavy responsibility of being a "general" and take charge of the overall situation.
In July this year, Jiang Jun met Feng Qiaoyu, chairman of Ant Cloud Technology Group.
In just one month, the two orders of the two sides reached 100 million yuan. The smooth cooperation made Feng Qiaoyu particularly optimistic about Jiang Jun, and the two had a very unified view on the large model, and he invited Jiang Jun to join Ant Cloud Technology: "I will fully support your plan for the large model business." ”
Jiang Jun himself believes that it is time to start "all in AI" again. The "blueprint" that was not completed three years ago may be reopened.
Unlike the "1+N" model (1 basic model + n vertical field small models) of most other players in the industry, Ant Cloud Technology Shanghai AI Large Model Research Institute is mainly based on the development of open source large models on the market, guided by business needs, and provides professional industry large models for all walks of life through innovation and research and development, while the development of basic large models only uses open source or docking existing large models.
It's not that you don't do it, the time has not yet come. Jiang Jun believes that although the start of the large-scale model track is slow, there is a lot to be done in "corner overtaking".
It took OpenAI three years to develop ChatGPT, while many players in China only started to lay out in 2023. He estimated that perhaps in 2024, we will have large-scale computing power demand.
I'm skeptical of claims that some companies claim to complete trillions of models autonomously. In the next 2 to 3 years, the pattern of the domestic large model field will be clearly differentiated, and then we will consider mining general large model talents and carrying out deeper development. According to my judgment, after 3 or 4 years, there is a high probability that the general large model will have the same major collapse as the visual AI in 2019. Jiang Jun told Jiazi Lightyear, "So we want to be a supporting service provider for the large-scale model gold rush wave, just like the merchants who sell jeans, shovels and gold jewelry processing." ”
Jiang Jun believes that at present, it is more important to establish a data training and annotation platform, cut the data of various industries into forms that can be used for AI training, and combine them with its own computing power platform. While providing services to other AI companies, it also "feeds" Ant Cloud's industry model.
Under his plan, Ant Cloud Group's core business can be briefly summarized as "one platform and two centers". The platform refers to the AI computing platform project that undertakes the construction and operation, and the two centers include the data annotation center and the model development center.
Ant Cloud Technology's core business architecture, **Respondent.
The AI computing platform is built and operated by Ant Cloud Technology. Different from the builders of traditional computing power centers, the core of Ant Cloud Technology is to help local governments introduce a large number of AI large model ecosystems, and at the same time promise that the AI ecology will land on enterprise revenue, and make a bet on performance.
The data annotation center is built on top of the local computing power center, which has become the core grasp of Ant Cloud Technology's traction of ecological model companies. It not only provides auxiliary annotation tools and resource manpower management for many annotation outsourcing companies, reduces costs and improves efficiency, but also provides data collection and annotation services for large-scale model enterprises and vertical industry model enterprises.
The establishment of "one platform and two centers" will provide a strong backup power for Ant Cloud's model development center.
The Model Development Center is dedicated to the development of open-source-based general models and industry-specific vertical models. In addition to the Shanghai AI Large Model Research Institute, Ant Cloud Technology has also established the Hangzhou Digital Human Research Institute. At present, the important components of its AI agent have been developed.
Jiang Jun believes that education, civil affairs and other industries are the most likely to give birth to disruptive applications, so Ant Cloud Technology will develop independent AI models and applications in the above fields. The business development of other industry conferences will be completed together with the leading players in the industry. At present, Ant Cloud Technology has signed strategic cooperation with more than 10 industry leaders, such as 4Paradigm, MindMinds, Computing, Inspur, Global Digital, Tianchuang, etc., to make every effort to deepen and thoroughly develop the industry model.
The rapid progress of the business exceeded Jiang Jun's expectations. With his own feelings, he has a stronger "push back feeling" than the previous wave of AI.
At the relocation ceremony, Jiang Jun looked at the team and said with some excitement: "With the development of the team and the expansion of the business, the original office space can no longer meet the demand in just one month, so we decided to move to a larger office space immediately. In about 3 months, our large model research institute will break through a team of 100 people, and the location of the headquarters will be clear, and it will be relocated to an office space that can accommodate 300-500 people. ”
In 2024, Jiang Jun plans to find suitable areas in the Yangtze River Delta and the Greater Bay Area, and successively establish a wholly-owned subsidiary of the group in the AI industry, a wholly-owned operating subsidiary with AI computing power construction as its core business, and a joint venture industry model company jointly established with one or two listed companies. The Group will move its IPO headquarters to a new address, and Xi'an will continue to serve as the Group's supporting business center.
Cover image**: Ant Cloud Technology Group).