Image source: @VisualChina
Text|Pole Business, Author|Yang Ming, Editor|Liu Shanshan
This may be the most important "stud" in the 7 years since Daniel Zhang fully helmed Alibaba, which is related to the future.
"All Alibaba products will be connected to Tongyi Qianwen and undergo a comprehensive transformation." At the Alibaba Cloud Summit at the China National Convention Center in Beijing a week ago, Daniel Zhang, who made his first public appearance as the CEO of Alibaba Cloud, stood in the center of the stage and unveiled the big model of "Tongyi Qianwen".
This is the first work to be verified after Daniel Zhang became the leader of Alibaba Cloud. According to the Daniel Zhang's announcement, 27 Ali Group's APP products, such as Tmall, DingTalk, Amap, Taobao, Youku, and Freshippo, will be connected to the "Tongyi Qianwen" model for comprehensive transformation.
Aggregating all Alibaba products under the same technical architecture, it is very similar to the famous "big middle platform, small front desk" organizational strategy path proposed when Daniel Zhang first took the helm of Alibaba in 2015 - of course, looking back now, the "big middle platform" strategy is difficult to say successful, as early as 2020, Daniel Zhang began self-reflection, and in this year's unprecedented organizational change, completely split into "1+6+N".
If the launch of "Big Middle Platform" was an innovator and leader, but in today's GPT large model arms race, Alibaba Cloud has lagged behind a lot, and even become an imitator and follower.
Not only the release time of the large model is later than the foreign OpenAI, Microsoft, Google, Meta, as well as domestic Baidu, Huawei, SenseTime, even Alibaba Cloud Intelligent Chief Technology Officer (CTO) Zhou Jingren proposed the "cloud-intelligence" strategy at the summit - this strategic direction emphasized by Alibaba Cloud Intelligent Group, in fact, Baidu officially released the "cloud-intelligence" strategy as early as 2020, and Alibaba Cloud is a full three years later.
Why did Alibaba Cloud propose the "Cloud-Intelligence Integration" strategy at this time? Is it "copying" or "industry consensus"? What are the implications for the big model arena and the cloud market landscape?
01 Cloud-intelligence integration, Alibaba Cloud "copy the concept"?
"In the era of intelligence, everyone is on the same starting line." Facing thousands of people at the venue, Daniel Zhang said that in addition to all Alibaba's products that are worth "transforming" with large models, Alibaba Cloud will open up the ability of Tongyi Qianwen and access B-end enterprises to create exclusive large models.
Through the speeches of Alibaba Cloud executives such as Daniel Zhang, we can get a glimpse of Alibaba Cloud's ambition: to bet on the future of Alibaba Cloud and even Alibaba Group through "all in" on large models.
As we all know, Ali, which occupied the first-mover advantage in the early years, relied on the explosion of e-commerce to stand at the peak of China's Internet. Its initiative, methodology, organizational ability, political commissar system, mission vision, values, etc., have influenced countless entrepreneurs and entrepreneurs.
This also includes the "big middle platform, small front desk" strategy set up in 2015 after Daniel Zhang became the CEO of Ali Group, the purpose of which is to improve the overall efficiency of Alibaba Group, minimize the behavior of recreating wheels, and reuse has the ability - although the "big middle platform" is not the first initiative of Ali, but from the experience learned after Jack Ma visited a game company in Finland, but the "big middle platform" is popular because of Ali and is still learned and imitated by many companies.
In recent years, Ali has lacked innovation, not very able to keep up with the "evolution and iteration" of the market and the general environment, and has frequently lost in various battlefields, and its market value has shrunk by nearly 70% from the peak.
For example, in the field of e-commerce, JD.com, Pinduoduo, and Douyin e-commerce have seriously shaken their basic plates; In the payment market, Alipay's share has long been left behind by WeChat Pay; Local life areas, from Ele.me to Fliggy, are powerless to stop Meituan's attack; In the field of cloud computing, Alibaba Cloud's growth rate has declined, executives have been turbulent, and Tencent Cloud, Baidu Intelligent Cloud, and HUAWEI CLOUD have brought many pressures. In the big cultural and entertainment industry, the development is even more dismal, and cases of failed acquisitions abound.
Under various difficulties, on March 28 this year, Ali ushered in the largest organizational structure adjustment in history. For Alibaba, which adjusts its organizational structure every year, it is easy to adjust to "1+6+N", and the most difficult thing is how to find a new development path.
When it is considered to be the cornerstone of the next generation of the Internet, the AI model represented by ChatGPT is competed by all the world's technology giants, Ali will naturally not miss it, just as Daniel Zhang predicted at the summit: the current technology field is the juncture of forming a major historical opportunity in the next 10 or 20 years.
But if compared with OpenAI, Microsoft, Google's large-model products, and even Baidu, "Tongyi Qianwen" is regarded by many people as a "Great Leap Forward" in the outside world.
According to Tongyi Qianwen's own introduction, this product was developed by Ali's DAMO Academy and was born in November 2022. In February this year, a senior technical expert of Ali broke the news that Damo Academy was developing Ali's version of the chatbot ChatGPT, and Ali later confirmed that it was in internal testing. On April 7, Alibaba Cloud suddenly officially announced that the large model Tongyi Qianwen invited the test, after which the content was shown to have been deleted by the publisher. On April 11, Tongyi Qianwen came out with high specifications in the debut of Alibaba Cloud Daniel Zhang.
Strictly speaking, there is not much innovation in the technical route and product form of "Tongyi Qianqian". On the technical route, the products that have been launched are similar to ChatGPT mode, multi-modal hybrid mode, and Tongyi Qianwen is similar to the Wen Xin Yiyan released by Baidu in March (digression: the ancient well-known literary theory works have Wen Xin Diao Long, White Tiger Tongyi), is a large-scale pre-trained language model similar to ChatGPT, with extensive knowledge reserve and universality, learning a large amount of text data in the training process, so as to have cross-domain knowledge and language understanding capabilities, suitable for the needs of different scenarios.
Even if Daniel Zhang proposed to help enterprises build a common platform for exclusive GPT, in fact, similar "open source" also has cloud enterprises in the forefront. At the end of March this year, Baidu Intelligent Cloud officially launched the "Wenxin Qianfan" large model platform to provide customers with enterprise-level large language model services.
It is worth noting that Alibaba Cloud also clearly stated for the first time that "cloud-intelligence integration" is its future strategy. Including Zhou Jingren, Chief Technology Officer of Alibaba Cloud Intelligence, who delivered a speech on "Cloud-Intelligence Integration, Making Intelligent Innovation Accessible", and Cai Yinghua, Chief Business Officer of Alibaba Cloud Intelligence, also said that only cloud-intelligence integration can make the comprehensive industry intelligent.
The concept of "cloud-intelligence integration" was first proposed by Baidu. In May 2020, Baidu Intelligent Cloud announced that CTO Wang Haifeng would unify the cloud, AI and basic technology systems, and clearly put forward the "cloud-intelligence integration" strategy. In the following two years, Baidu Intelligent Cloud completed the evolution of the cloud-intelligence integrated architecture from 1.0 to 3.0 at a one-year frequency.
The integration of "cloud + AI" is actually not uncommon in cloud computing companies at home and abroad. In addition to Baidu's "cloud intelligence integration", Microsoft proposed "office + Teams + Azure", Amazon proposed "cloud, data, intelligence trinity" in 2022, Tencent Cloud proposed "cloud intelligence integration" last year, Jingdong called "digital intelligence supply chain", Alibaba Cloud previously had "cloud nail integration" - even if everyone's technical route and concept meaning are similar, they will avoid using the same concept.
In the past three years, Baidu has been strongly bound to "cloud-intelligence integration" in various occasions and financial reports. This also puzzles the industry, why did Alibaba Cloud re-propose the concept of "cloud-intelligence integration" three years later, is it "copying homework" or "industry consensus"?
02 The logic still needs to be self-consistent, and the gap between actual measurement and publicity is a bit large
In any case, in the eyes of the industry, Alibaba Cloud's follow-up means that the opening and integration of "cloud-intelligence integration" has become an industry consensus, and also indicates the arrival of the big industry opportunity of cloud-intelligence integration.
Alibaba Cloud executives did not elaborate on the specific meaning of cloud-intelligence integration. According to Baidu's previous statement, "cloud" refers to cloud computing as it is based on, enabling enterprises to access AI capabilities as conveniently and quickly as using "water, electricity, and coal"; "Intelligence" refers to artificial intelligence, which meets the needs of enterprises for AI technology in complex and changeable scenarios through full-stack technical capabilities such as flying, vision, language, and knowledge.
Simply put, it is "cloud suitable for running AI" superimposed on "AI that understands the scene" to jointly build infrastructure in the intelligent era.
In the concept of "cloud-intelligence integration", the core is intelligence. After years of layout development, the cloud has become the infrastructure of all walks of life. Whether it can provide and deliver comprehensive, convenient, efficient, stable and reliable AI intelligent services has become the key to the competition of major technology enterprises.
Alibaba Cloud, which has gradually reached the ceiling in the government, enterprise and personal cloud markets, may see the same opportunity: with the differentiated play of "cloud-intelligence integration", Baidu has formed an intelligent closed-loop path of "chip-frame-big model-industry application", helping Baidu find the second growth curve for the next decade. It can be said that "cloud-intelligence integration" is quite important in the process of reshaping Baidu.
In addition, with the acceleration of the digital transformation of China's economy and society and the advent of the era of the Internet of Everything, "going deep into the industry and promoting industrial intelligent upgrading" has clearly become an important part of each cloud computing strategy.
This is even more inseparable from the "cloud-intelligence integration" to carry large model computing power and unlock AI productivity - in fact, whether it is an AI large model, or other clouds, AI product matrices, although the concept is different, it is actually based on the "cloud-intelligence integration" base. This is true of Wen Xin Yiyan, Tongyi Qianqian, 360 Wisdom Brain, SenseTime, and many other ChatGPT products.
Judging from the actual measurement of "Extreme Business", the answer given by Tongyi Qianwen is not ideal overall.
At present, Tongyi Qianwen only provides text generation function, and cannot generate pictures, audio and video content for the time being. In terms of natural language processing and Chinese and English dialogue, the performance of Tongyi Qianwen is acceptable, but in terms of basic facts, time, calculation, reasoning, etc., which require thinking cognition and logical understanding dialogue, it is not a small gap compared with other large models.
"How to make shredded fish" is the same, and the question that is widely mentioned in various models, Tongyi Qianqing really answered well. However, "Pole Business" uses the hot topic "Romance of the Three Kingdoms" to focus only on portraying the chaotic era in the early stage, and downplaying the final unification process? When asked a question, Tongyi Qianwen replied, "I'm sorry, I don't have the answering skills at the moment."
In contrast, Wen Xin pointed out, saying that "the main plot of the novel is to describe the struggle for hegemony during the Three Kingdoms period, and the final unification process is only one of the small plots." This treatment makes the novel pay more attention to describing the historical background and character of the chaotic era in the early stage, so as to better present the history and characters of the Three Kingdoms period. ”
On the issue of "how to view the Measures for the Administration of Generative Artificial Intelligence Services (Draft for Comments)", the pseudo-ChatGPT provider should be responsible for the legality of data sources", another AI model affirmed, pointing out that it is an important measure to strengthen the supervision of artificial intelligence. And Tongyi Qianqing is still "not learning how to answer".
When it comes to basic facts and timeline combing, Tong Yi Qianwen may not understand how to answer. For example, in the question "What negative events Alibaba has experienced in the past few years", Tongyi Qianqian's answer is not only full of errors, but even many fabrications of Alibaba's facts and time.
In the dialogue "What are the advantages and disadvantages of Tongyi Qianwen compared with Wen Xin's words", Tongyi Qianwen made another factual low-level mistake: saying that Tongyi Qianwen and Wen Xin Yiyan are both AI pre-training models launched by Alibaba.
So many low-level mistakes are not only inconsistent with Alibaba Cloud's first layout of cloud computing and market position in China, but also difficult to undertake the ambition of redoing all of Alibaba's product access and creating an exclusive large model for B-end enterprises - after all, understanding the basic logic and following the facts first, if the measured results are not ideal, it will not only affect the digital layout of B-side enterprises, but also bring users such as DingTalk and Tmall Genie a bad experience.
According to Alibaba Cloud's layout, intelligent customer service is its key application scenario - it can help Alibaba break through the limitations of traditional e-commerce "people find goods" and iteratively upgrade to "goods find people". Moreover, in terms of wine tourism and local life, AI large models can also be accurately marketed to meet the real needs of more consumers
However, from the actual measurement, the performance of Tongyi Qianwen in intelligent customer service and intelligent recommendation is still far from commercialization. For example, "from Beijing, go to Chongqing Nantianhu May Day for three days" and "budget 6,000 yuan, what range hood and gas stove are better to buy", the answer given by Tongyi Qianwen is of very little practical reference significance to consumers
It is worth mentioning that these questions are raised casually by "Pole Business", not intentionally. At the previous press conference, Daniel Zhang called Tongyi Qianwen "the world's first AI large model that has exceeded 10 trillion parameters". In addition, it is reported that Ali has also led the construction of the largest AI model service community "Magic Match" in China. So, why is the logical ability of generative AI so weak?
03 It's not easy to follow the trend, and Alibaba Cloud is still missing
This involves the essence of "cloud-intelligence integration" - the cloud must be better integrated with wisdom in order to give users a better experience and make the industry move towards intelligence.
The answer may be simple, even if Alibaba Cloud has strong cloud computing and AI capabilities, the two have not been well integrated. In addition, the gaps in data parameters, scale and quality of pre-training data, and data cleaning are also directly related to the answer results of different large models, and Alibaba Cloud does not have advantages in this regard
Some insiders said that in the past, within some cloud vendors, AI and cloud were generally in two different systems. The sale of resource-based products such as computing and storage has become the main business model, resulting in AI not being able to exert maximum value in the process of industrial digital transformation and intelligent upgrading.
From the perspective of "extreme business" understanding, Alibaba Cloud is such a typical case.
For a long time, Alibaba's artificial intelligence and cloud computing have been in two separate systems - within Alibaba, the research department of artificial intelligence was once led by Ali Artificial Intelligence Laboratory (AI Labs) and DAMO Academy, which was a parallel department to Alibaba Cloud.
AI Labs is a laboratory within the Alibaba system, researching speech, language, and machine vision, and later changed to Tmall Genie Business Unit, focusing on intelligent hardware. It was not until 2021 that the AI Labs team was integrated into Alibaba Cloud Intelligence in the midst of architectural changes.
The DAMO Academy, which develops Tongyi Qianqianqian, is an independent research institute located outside Alibaba's huge business system, and its research scope includes cutting-edge technologies such as data intelligence and human-computer natural interaction. In the 2021 organizational structure upgrade, the "cloud and technology sector" in charge of former Alibaba Cloud President Zhang Jianfeng includes five lines: Alibaba Cloud, DAMO Academy, DingTalk, Intelligent Interconnection (Tmall Genie), and Pingtou Ge. At that time, although the AI technology of DAMO Academy was exported through Alibaba Cloud, it was still a parallel "brother department".
In addition to the parallel departments, what has a greater impact on Alibaba Cloud is the successive years of executive turmoil. In May last year, there were a series of executive changes between DAMO Academy and Alibaba Cloud, Hua Xiansheng (head of DAMO Academy City Brain Lab), Jin Rong (senior researcher of City Brain), Zhang Lei (deputy director of DAMO Academy AI Center) left, and Xiao Lihua (president of Alibaba Cloud Research Institute) left to start a business. Overnight, the star project that has been cultivated for many years since Wang Jian's period, the Damo Academy Urban Brain Laboratory, almost all the core personnel were lost.
This caused Aliyun and DAMO Academy to "greatly hurt their vitality" and accelerated the "division" of the two. According to a previous report by "Lei Feng.com", Alibaba Cloud and DAMO Academy were divided at the end of 2021, and some product technologies incubated by DAMO Academy were carved out, while DAMO Academy operated in a more independent form and took on the challenge of 10-figure revenue figures - the specific way is that DAMO Academy sells products externally through Alibaba Cloud business lines, and the two implement a dual computing model.
At the beginning of January this year, Alibaba Cloud's senior management changed personnel again. Daniel Zhang personally succeeded Zhang Jianfeng as president of Alibaba Cloud Intelligence, and Zhou Jingren served as CTO of Alibaba Cloud Intelligence. However, with the failure of the "big middle platform", whether the cloud and AI will be better integrated in the future has become a question.
After the "largest" organizational structure change on March 28, DAMO Academy and Alibaba Cloud were both placed under the cloud intelligence group headed by Daniel Zhang, but they are still divided into different business segments and have a certain degree of business independence. At the same time, executives are still lost, Ali AI bull Jia Yangqing, Ali M6 model former leader Yang Hongxia has also left.
The core person in charge of the big model or AI continued to be in turmoil for two years, and the result was that the launch of Alibaba's big model was late, and the actual performance of product capabilities was more like "catching ducks on the shelf".
Some artificial intelligence practitioners revealed that a stable management team is the key to the continuation of product and technical routes. The person gave an example, in Baidu Intelligent Cloud, the combination of Baidu CTO Wang Haifeng + Baidu Technical Committee Chairman Wu Hua + Baidu Group Vice President Wu Tian has remained stable since the first day of the launch of the Wenxin large model ERNIE 1.0 series in 2019, and only then has the continuous iteration of the Wenxin series.
Unfortunately, for Alibaba's cloud and AI businesses, Baidu has never had the stable advantage of frequent organizational structure adjustments and continuous personnel turmoil.
This is also reflected in large models. Ali is more than two years slower than Baidu and Huawei in model iteration. In 2019, when Baidu and Huawei launched large models, Ali began research and development. In 2021, other companies began to continue to iterate on the big model, and Alibaba released the language and multimodal big model, which was merged into the Ali Tongyi big model in September last year.
The result is the difference in data quality and scale of large models. "There is a clear gap between Alibaba's model and the industry's or domestic leading level, in terms of data parameters, pre-training data quality and data cleaning." ICT industry service platform "Jiwei" reported.
Alibaba Cloud CTO Zhou Jingren also previously admitted that the development of large models often requires more than 100 billion parameters, and its difficulty far exceeds the simple stacking of a single algorithm or GPU chip.
In fact, this gap is expected by the industry. Baidu has massive C-end search data, and Tencent also has more than one billion users WeChat ecological data precipitation, but Ali has the most TOB merchant customer data in vertical fields. From the perspective of general large model training, TOB merchant data that is too commercial is not suitable for mass addition.
Before the release of Tongyi Qianwen, a senior expert of Alibaba Cloud Research Institute admitted that in terms of parameters, Tongyi Qianwen's PLUG-based model has parameters of 20 billion to 30 billion, while GPT-3 parameters are 175 billion, and it is mainly biased towards text and does not involve images and videos; In terms of pre-trained corpus datasets, they are not comparable to OpenAI and Baidu in terms of quality and scale.
This may be one of the reasons why Alibaba Cloud "copied homework" and implanted the concept of "cloud-intelligence integration". After all, as Zhou Jingren said, the current AI big model competition is essentially a competition in computing power. It is a systematic project that includes complex technologies such as underlying computing power, network, storage, big data, AI framework, and AI model, and requires full-stack technical capabilities of AI cloud computing.
In any case, with the opening of Pandora's box of AI large models, the cloud market GameChanger has arrived. For Alibaba Cloud, the former big brother, from a "leader" to a "follower", how many opportunities are there to reshape the rules of the industry?