2023 is undoubtedly a year for large models. AIGC has opened up the public's imagination of artificial intelligence application scenarios and set off a new round of artificial intelligence entrepreneurship.
In the context of the era of large models competing to emerge and empower thousands of industries, Wu Xiaobo led the author team to record and analyze nearly 50 representative cases from popular industries and fields such as computers and the Internet from the four main dimensions of AIGC innovative enterprises through field research and in-depth interviews, and condensed them into this book "China on the Cloud 3: The AI Era in Dramatic Changes", which presents the application and future of large models in the era of intelligent native in a panoramic way.
Today's article is selected from the second chapter of the book, "AIGC Innovative Enterprises: From Large Model R&D to Popular Applications", to share with you.
01
Thousands of troops crossing the single-plank bridge:
Involution, because I don't see the second way
On August 24, 2022, on the shores of Lake Washington in Seattle, USA, OpenAI's two founders, Sam Altman and Greg Brockman, and young researcher Chelsea Wirth, knocked on the door of Bill Gates' $130 million mansion with trepidation.
Although he has long since stepped down from Microsoft's board of directors, Bill Gates, as the founder, still maintains close ties with the company's executives, and his views on artificial intelligence will also influence Microsoft's investment decisions to a certain extent.
Five months ago, Gates told Altman and Brockman at a dinner in his mansion that he had expressed doubts about the effectiveness of the large language model developed by OpenAI, which had struck a bit of a blow to their hearts.
Five months later, when the three young people were about to show Gates the latest technology they had developed, Gates called several Microsoft executives as well. They sat in the spacious living room, while on a large digital display, Researcher Voss demonstrated a technology called GPT-4 to those present.
GPT-4 demonstrated the results of a multiple-choice test in advanced biology with a total of 60 questions, and GPT-4 got only one wrong.
According to the New York Times report, this shocked Gates. His reaction was not excessive, and just three months later, OpenAI launched ChatGPT, a real-time conversational and interactive app that has wowed the world.
The birth of ChatGPT is destined to be mentioned repeatedly as a milestone event in the long history of artificial intelligence. It not only allowed artificial intelligence to successfully cross the inflection point, revealing the huge potential and broad market of artificial intelligence, but also sounded the clarion call for global technology companies to carry out large-scale model "arms race", heralding the arrival of a new era of artificial intelligence.
02
The "100 model war" in full swing
In 2023, in the field of Chinese artificial intelligence, the "100 model war" is definitely an unavoidable topic.
Readers who don't usually pay much attention to large models may have a question, is it the explosion of ChatGPT around the world that has made Chinese technology companies quickly join the battle?
That's only half the answer to the question.
Some start-up AI companies have indeed come to the frenzy caused by ChatGPT, and have begun to invest in the research and development of large models after rapid financing. However, for some large domestic Internet companies, they have been deeply engaged in the field of artificial intelligence for a long time, and have already carried out detailed layout and preparation for many years.
◎ Alibaba Cloud
Since 2009, Alibaba Cloud has developed its own big data computing service (ODPS), which has built leading technical capabilities in large-scale data processing and large-scale computing applications, which promoted the birth of the artificial intelligence platform PAI in 2014.
In terms of the research and development of large models, Alibaba Cloud has been following up since 2018, when it was still mainly based on pre-trained models. In 2022, Alibaba Cloud recognized that generative AI may be the mainstream of the future, and officially began to lay out the entire "generative task-centric" research line in March of that year.
Half a year later, on September 2, 2022, Alibaba Cloud released the "Tongyi" large model series for the first time, including a variety of general models such as Tongyi M6 and Tongyi AliceMind.
Two months later, Alibaba Cloud took the lead in proposing the concept of "Model as a Service" in the industry at the 2022 Apsara Conference, that is, Model as a Service (MaaS), firing the "first shot" to lower the threshold for AI development.
In April 2023, the Tongyi Qianwen model was unveiled at the Alibaba Cloud Summit. According to Zhou Jingren, Chief Technology Officer of Alibaba Cloud, Tongyi Qianwen can realize services such as multi-round dialogue, copywriting, logical reasoning, multi-modal understanding, and multi-language support, and realize the generalization of AI capabilities through API plug-ins.
Since then, the Tongyi Qianwen model has been continuously optimized. On October 31, Alibaba Cloud officially released the Tongyi Qianwen 2.0 version, with parameters reaching 100 billion yuan. On the 10 mainstream evaluation sets such as MMLU, AGIEval, and C-Eval, the comprehensive performance of Tongyi Qianwen 2.0 exceeds GPT-3.5, accelerating to catch up with GPT-4.
◎ Baidu
In July 2017, after Baidu launched its "All in AI" strategy, its research and development efforts in the field of artificial intelligence have also increased significantly.
On March 16, 2023, Baidu officially announced the launch of a new generation of large language model Wenxin Yiyan.
Although Wenxin Yiyan was initially criticized by users for its slow response and answering questions that were not asked, after the release of version 4.0 in October 2023, with the comprehensive upgrade of its basic model, it has significantly improved its comprehension, generation, logic, and memory capabilities.
◎ Tencent
On September 7, 2023, at the annual Global Digital Ecology Conference, Tencent's hybrid model was officially unveiled. The product is self-developed by Tencent Fulllink, with a parameter scale of over 100 billion and a pre-trained corpus of more than 2 trillion tokens. Tencent's hybrid model includes computer vision, natural language processing, multimodal content understanding, copywriting generation, and Wensheng video.
The large model has attracted leaders from all walks of life. In addition to the above three major manufacturers, Huawei, JD.com, ByteDance, Meituan, iFLYTEK and other technology and Internet giants have also stepped up their entry into the game, in addition, including the three major operators of Mobile, Unicom, and Telecom, Tsinghua University, Fudan, Harbin Institute of Technology and other scientific research institutes and universities, as well as Daguan Data, Baichuan Intelligence, Fourth Paradigm, Mobvoi and other artificial intelligence innovation companies, they will all release their own large models in 2023.
From January to July 2023, a total of 64 large models will be released, and domestic large models are growing explosively. According to the "2023 Survey Report on the Status of Large Models" released by CCID Consulting, as of the end of July 2023, a total of 130 large models have been launched in China, and the situation of "100 model wars" has emerged. By October, the total number of large-scale models published in China had reached 238. The competition of large models began to enter the stage of "vigorously producing miracles".
03
Chinese companies are vying to be the first to get a position in the market
When the domestic "100-model war" is in full swing, are technology companies in other countries just watching the fire from the other side?
Of course not. If we look at the world, we will find that there is a high degree of uniformity in the application and development of large models, but from the perspective of achievements, it can be said that China and the United States are the undisputed leaders in large model technology.
According to the "2023 Comparative Study of the Application of AI Large Models in China and the United States" released by Titanium Media International Think Tank, the advantage of the United States is that the technology giants represented by Google have long been committed to the research of the basic theory of artificial intelligence, making the United States lead the development trend of artificial intelligence.
In the primary market, technology giants represented by Nvidia, Microsoft, and SAFTSE have become the most important "unicorn hunters" in the field of AI in the United States, and their large investments have accumulated strength for the development of artificial intelligence in the United States.
According to the White Paper on the Innovative Application of Large Models in the Artificial Intelligence Industry in Beijing (2023) released on November 29, 2023, the total number of general large models released in the United States and China accounts for 80% of the global releases. However, in addition to China and the United States, other countries and regions such as EU member states, the United Kingdom, Canada, Singapore and other countries and regions are still in the early stage of trying large models, and only a few leading enterprises have begun to apply them.
Compared with most foreign technology companies that are still in the wait-and-see stage of large models, why are Chinese companies so enthusiastic about joining the large model competition, and why are they advancing so fast?
The answer is simple - to be the first to "get stuck" in the market.
Large model R&D companies have a vision: the faster their large models are released, the more user data they can get first, so as to feed themselves better large models, so as to achieve the "data flywheel effect".
Nvidia founder and CEO Jensen Huang once said, "We are in the 'iPhone moment' of AI", indicating that an immeasurable AI market is in front of all technology companies.
According to IDC, the global AI computing market will grow from $19.50 billion in 2022 to $34.66 billion in 2026. Among them, the generative AI computing market will grow from $820 million in 2022 to $10.99 billion in 2026.
The fundamental role of large models in the development of artificial intelligence has become a recognized fact, and any technology company with lofty ideals and R&D capabilities wants to take the initiative of large model research and development in their own hands.
However, a corresponding argument soon emerged: the large model itself does not directly generate value, and the application developed based on the basic large model is the meaning of the model's existence.
This argument is not unreasonable. No matter how the research and development of the large model is "rolled", it will eventually return to the essence of business, and the company cannot blindly burn money, but must move towards the goal of profitability.
In 2023, among the application products launched based on 130 large models in China, only the Miao Ya camera launched by Alibaba's entertainment team has successfully broken the circle and become a recognized phenomenal AIGC product. As for applications with worldwide influence such as ChatGPT and Midjourney, it is believed that they are also the common goal of the Chinese AI industry.
In order to standardize and create an environment conducive to the better development of large models, before the implementation of the Measures for the Administration of Generative AI Services (Draft for Comments) on August 15, 2023, relevant national policies require that generative AI products that have not been declared to the CAC for security assessment and have been filed with algorithms shall not be open to the public.
Since August 31, products including Baidu Wenxin Yiyan, Alibaba Cloud Tongyi Qianwen, Baichuan Intelligence, SenseTime Consultation, Tencent Mixed Yuan Model and other products have been opened to the public one after another.
The "100 model war" instantly entered the fierce "hand-to-hand combat", and the large models of various enterprises have a good use experience and strong commercial landing ability, and have reached the stage of "whether they are mules or horses, pull them out for a walk".
When the "100 model war" entered the second half, another advantage of the big factory began to be mentioned repeatedly. We all know that the three key factors of a large model are: computing power, algorithms, and big data, but in fact, for a large model, it is the scenario application that needs to be paid more attention to.
At the 2023 Alibaba Cloud Summit, Daniel Zhang, then Chairman and CEO of Alibaba Group and CEO of Alibaba Cloud Intelligence Group, proposed another perspective to intervene in the application of large models: for the AI era, all products are worth upgrading with large models.
Alibaba also took the lead in demonstrating, and "Tongyi Qianwen" has been connected to the test of DingTalk, Tmall Genie and other products, becoming the first batch of products or applications of "early adopters" of large models in China.
This fully reflects the advantages of Internet manufacturers, whose own business scenarios are rich and diverse, and even if there are no external customers, large-scale model products have enough scenarios to verify and optimize in their own business.
In September 2023, Wu Yongming, the new CEO of Alibaba and CEO of Alibaba Cloud Intelligence Group, showed Alibaba's determination to fully embrace AI through a letter to all employees: "Alibaba has established two strategic priorities: user first and AI to drive the next decade...... There are a large number of user scenarios in each of our businesses, and we must turn these scenarios into the best application scenarios for AI technology, and bring breakthrough user experiences and business models through technological innovation. ”
In the era of AI, the computing model is facing changes, so for cloud computing vendors, AI has also brought about an explosive increase in the demand for computing power.
Based on AI, new software architectures, software methods, and computing architectures will be in the process of subversion and reshuffling for a long time to come. In the future, the application construction of all industries supported by cloud computing, as well as the computing power storage scheduling behind it, will most likely be driven by the AI models of various vertical industries as the underlying system.
The biggest variable in the next decade will undoubtedly be the profound industry-wide changes brought about by AI.
04
Large factories entered
Compete in the "Large Model Ecosystem" stage
We understand the logic of the Internet giants vigorously promoting large models, in fact, there is another dimension, large manufacturers are generally linked to the industry, there is a huge downstream application market, and the development of large model products can be quickly applied to the industry.
Therefore, in the second half of the "100 model war", the Internet giants who have been in the field of artificial intelligence for many years have begun to build a large model ecosystem: on the C-side, the large manufacturers provide services to creators and developers; On the B side, it mainly provides solutions, especially focusing on finance, cultural tourism, media, medical care, government affairs and other industries, and also provides infrastructure services such as computing power and data management for the rest of the large model research and development enterprises.
Taking Alibaba Cloud as an example, its large-scale model ecosystem provides products and services in four directions:
The first layer is the lowest layer, which is the PAI-Lingjun intelligent computing platform built on Alibaba Cloud's global infrastructure, as well as the data management service system.
The second layer is the two basic models of Tongyi Qianwen and Tongyi Wanxiang, as well as the eight industry models and applications developed on their basis.
The third layer is Alibaba Cloud Bailian one-stop large model service platform (hereinafter referred to as Bailian).
The fourth layer is the artificial intelligence open community magic pairing.
Zhou Jingren, chief technology officer of Alibaba Cloud, made a summary: "Alibaba Cloud is striving to become the infrastructure of the AI model era by serving entrepreneurs, developers, and enterprise customers in the era of AI large models." ”
More small and medium-sized enterprises realize that there is a systematic gap with large factories in terms of resources, talents and capital, and begin to find vitality from market segments. For example, companies such as SenseTime, Mobvoi, and 4Paradigm have begun to focus on migrating large model capabilities to their previous advantageous business areas.
Nowadays, the "100-model war" has entered the deep water area of competition for ecosystem construction, and it can be said that the fierce battle is in full swing. This also objectively makes the field of Chinese artificial intelligence show a prosperous trend, but the consensus in the industry is that the development of large models is only the first step of the long march, to let the large model achieve commercial landing, truly let the large model into our daily life, is the proper meaning of research and development of large models.
In 1994, Kevin Kelly, the founding editor-in-chief of Wired magazine, in his book "Out of Control", had extremely accurate predictions about artificial intelligence technology:
"We hardly notice the ubiquitous text on the labels on the fruit, movie subtitles, etc. When the motor first appeared, it resembled a gigantic and proud beast; But since then, they have shrunk into microscopic things, integrated into (and forgotten in) most mechanisms.
The most profound techniques are those that are unseen and weave themselves into the minutiae of everyday life until they become a part of life. ”
Now, AI technology is surging in the direction of "becoming a part of our lives". Under the wave of AI, there are a large number of entrepreneurs in almost every subdivision in China who are thinking about how to make innovative applications of AIGC.