In the latest narrative of leading technology manufacturers, ToB's empowerment of thousands of industries is the focus of the large-scale model landing.

After Tencent and Huawei, another large-scale model "arrived although it was late" was long overdue.

On July 13, at the 2023 JD Global Technology Explorers Conference and JD Cloud Summit, JD.com officially launched the large model "Yanxi". The platform is currently open for appointment registration and is expected to be officially launched in August.

So far, the large-scale model products of leading Internet companies have all appeared: Baidu has Wenxin, Ali has Tongyi, Tencent has to make industry large-scale models, Huawei has Pangu, JD.com has Yanxi, and Byte has Volcano Ark.

Looking around the world, hundreds of large-scale models have been launched; more than 80 models have been released in China alone.

Image source: "Development Trends of Artificial Intelligence Large Models in my country" released by China Mobile Research Institute

In the 2023 WAIC (World Artificial Intelligence Conference) exhibition hall held a few days ago, more than 30 large models are scattered throughout it. In addition to the first batch of general-purpose large-scale models released by Baidu Wenxin, Ali Tongyi, Xunfei Xinghuo, and Shangtang Rixin, there are also a bunch of industry-wide large-scale models. Everyone seems to be talking about big models.

But compared to the frenzy when ChatGPT became popular, people have become relatively calm now, thinking and discussing more pragmatic questions such as "what can the big model do?" "how does the big model land?".

Major technology companies such as BAT, Huawei, and JD.com have made their own choices on key issues such as "ToC or ToB" and how to implement them. In their latest narrative, facing the B-side and empowering thousands of industries is the focus of the big model.

It is not difficult to understand that technology companies have been chasing AI for many years, and solving problems is more suitable for the present than poetry and distant places.

Large model path of a large technology factory

JD.com made it clear at the press conference on July 13 that "JD.com wants to be an industrial model rather than a general model." Xu Ran, CEO of Jingdong Group, said, "The real value of large models must be in industrial applications."

This means that JD.com's "Yanxi" large-scale model does not focus on ToC, but on ToB, focusing on solving industrial problems.

Not long ago, Tencent and Huawei also clearly chose the ToB direction. On July 7, Huawei released version 3.0 of the Pangu model. Zhang Pingan, CEO of HUAWEI CLOUD, clarified for the first time the positioning of the Pangu model - "born for the industry", focusing on B-end industry customers.

Although Tencent has more than 1 billion social users and has a large Hunyuan model in reserve, it is unexpected that Tencent did not launch a general-purpose large model on the ToC side, but directly faces the B-side, focusing on large-scale industry models.

Tencent's logic is that the general-purpose large model can solve 70%-80% of the problems in 100 scenarios, but it may not be able to 100% meet the needs of a certain scenario of the enterprise. Therefore, it is the enterprise's responsibility to build its own exclusive model based on the industry large model. More preferred.

Bytedance is a player with extraordinary AI technology strength among major technology companies. How do bytes make a large model?

On June 28, ByteDance unveiled the mystery: its Volcano Engine released a large-scale model service platform "Volcano Ark".

What Volcano Ark needs to do is to build a large-scale model "store" that aggregates third-party manufacturers, and provides services such as model fine-tuning, evaluation, and reasoning for enterprises. Different from BAT's self-developed large model, Byte has chosen a lower cost and faster entry path.

In a nutshell, Tencent, Huawei, JD.com, and Byte have already said no to the "Chinese version of ChatGPT".

However, Baidu and Ali, which launched large language models earlier, currently insist on "not being biased", requiring both ToC and ToB, both general and exclusive large models.

In March of this year, Baidu launched Wenxin Yiyan, an important member of the Wenxin large model, a knowledge-enhanced large language model, becoming the first company in the world to benchmark against ChatGPT.

Baidu started the research and development of artificial intelligence in 2010, and began to develop the pre-training large model in 2019, launching the Wenxin large model version 1.0. Baidu has invested more than 100 billion yuan in AI so far. For more than ten years, Baidu has been exploring the implementation and commercialization of AI.

The popularity of large models gave Baidu the excitement of "finally waiting for you", and thus showed the momentum of who else to give up. From technical strength to scene implementation and ecological creation, Baidu seems to be an "all-around player" and strives to be the first in everything.

On July 7th, at the World Artificial Intelligence Conference, Baidu CTO Wang Haifeng disclosed in his speech the latest progress in the effects, functions, and performance of the Wenxin large model version 3.5—including: the model effect has been increased by 50%, and the training speed has been increased. 2 times, the reasoning speed increased by 30 times.

It is not difficult to see that one of the messages that Baidu wants to emphasize to the outside world is: the key to making a large model depends on technology. In addition, in Baidu's latest speech, the key words also include "empowerment" and "ecology".

Baidu launched a one-stop enterprise-level large-scale model platform "Wenxin Qianfan" in March this year. Wenxin Qianfan mainly provides two major services: with Wenxin Yiyan as the core, it provides large-scale model services and helps customers transform products and production processes. Enterprises can develop their own exclusive large models based on any open source or closed source large models on Wenxin Qianfan.

Up to now, more than 300 Baidu Smart Cloud ecological partners have participated in the internal testing of the "Wenxin Qianfan Large Model Platform".

This means that Baidu's large-scale model not only requires ToC, but also through the underlying technology iteration and platform ecological opening, to realize the landing of the scene, and ToB empowers thousands of industries.

In addition, in May, Baidu also established a 1 billion yuan Wenxin investment fund, the purpose of which is to promote the ecological prosperity of large models.

Alibaba Cloud released the "Tongyi" large-scale model series in September last year. In April this year, its large language model "Tongyi Qianwen" was officially launched. Ali is the second major company in China to release ChatGPT-like products.

On June 1 and July 7, Alibaba Cloud launched the large-scale model application product "Tongyi Hearing" focusing on audio and video content and the large-scale AI painting creation model "Tongyi Wanxiang".

In three months, Alibaba Cloud has launched three different large-scale model products, extending the modality of large-scale models from text and voice to images, and further approaching multi-modal models.

Alibaba Cloud believes that it is difficult to solve all problems with a general model. Therefore, in its three-tier service architecture (IaaS+PaaS+MaaS), the MaaS layer includes both the basic model/general model and the enterprise-specific model.

Alibaba Cloud CTO Zhou Jingren emphasized in his speech on July 7 that the logic of Alibaba Cloud's large-scale model "is not to compete with other large models." He said that the large-scale model of Alibaba Cloud is to serve the enterprises and ISVs on the cloud well. At the same time, Alibaba Cloud also serves the innovators and model developers of self-developed models, providing them with tools and ecology.

Zhou Jingren emphasized that promoting the ecological prosperity of China's large-scale models is the primary goal of Alibaba Cloud.

The Mota community is an important carrier for Alibaba Cloud to create a large-scale model ecology. The Mota community has gathered 1.8 million AI developers and more than 900 high-quality AI models contributed by more than 20 top artificial intelligence institutions.

Alibaba Cloud's goal is to cultivate the Mota community into China's largest "big model free market".

ToB is the key to landing a large model

In May of this year, the China Mobile Research Institute mentioned in the "Development Trends of my country's Artificial Intelligence Large-scale Models" that from the perspective of large-scale model applications, most domestic enterprises mainly focus on internal applications in the early stage, and then mainly expand services to B-end enterprises. , It is expected that a small number of enterprises will form a scale in the C-end market.

Judging from the current path choices of major domestic technology companies, ToB has become the focus of large-scale model implementation.

This is not the same as the market expectations when ChatGPT first became popular.

ChatGPT attracted 120 million users in less than three months after its launch, making it the app with the fastest user growth in history. Backed by generative AI, it amazes everyone with its powerful ability to answer questions.

Microsoft CEO Nadella believes that ChatGPT will reshape all software categories, and the search engine may be the first to be subverted. Therefore, Microsoft fired the first shot of the ChatGPT product - releasing a new version of the Bing search engine embedded in ChatGPT. Microsoft even sees this as a new era opportunity to reshape the search engine landscape.

The popularity of ChatGPT also made Baidu and Google, the world's two largest search giants, nervous and anxious at first. Robin Li, founder of Baidu, set the key task of OKR in this year's Q1 as "leading the intergenerational change of search experience". Google also panicked, and the two founders returned urgently to lead the company in a counterattack.

After half a year of soaring, ChatGPT not only failed to subvert the search engine, but experienced a decline in traffic.

According to data from SimilarWeb, in June this year, ChatGPT’s global visits fell by 9.7% month-on-month. In June, ChatGPT downloads on iOS in the US dropped 38% month-on-month, and Bing downloads dropped 38%.

At the same time, the market share of Google's search engine has been rising year-on-year, and has now exceeded 92%; Bing's market share has dropped by 0.4% year-on-year to about 2.8%.

In terms of profit model, ChatGPT launched ChatGPTPlus for the C-end in February this year, with a monthly subscription fee of $20. In April this year, OpenAI disclosed on its official website that it will launch an enterprise-oriented ChatGPT Business in the next few months, and paid subscriptions are still the revenue-generating model on the ToB side.

However, compared with ChatGPT's high R&D and operation and maintenance costs, the current subscription revenue is almost negligible. OpenAI pays up to $700,000 a day to keep its robust infrastructure running, according to research firm SemiAnalysis.

However, using a large model to obtain the incremental market of cloud business, Microsoft Cloud has already tasted the first soup.

In January of this year, Microsoft management disclosed in a conference call after the financial report that Microsoft is using AI models to innovate computing platforms, and Azure ML (machine learning platform) revenue has grown by more than 100% for four consecutive quarters, which is higher than that of AI. Cloud and enterprise software business.

In China, the large model also brings new growth momentum to the cloud business. Taking Baidu as an example, after the launch of Wenxin Yiyan, more than 300 ecological partners have achieved test results in more than 400 specific scenarios. Baidu Smart Cloud sales leads increased by more than 400% year-on-year in March.

Looking around the world, cloud vendors hope to be the first beneficiaries of the big model era. The large-scale models of technology giants are basically dominated by the cloud business department, or deeply bound with the cloud business department.

In China, from Baidu, Ali to Tencent, Huawei, JD.com, and Byte, the big models are all dominated by cloud businesses.

In 2022, the global cloud service market will be sluggish, and cloud service giants will "fall" one after another, and their performance growth will slow down. In Q1 of 2023, the cloud service market will grow by 19% year-on-year, and the weak state continues.

Image source: Gartner

When the era of large models comes, the mainstream business model of cloud computing will change from IaaS (Infrastructure as a Service) to MaaS (Model as a Service). AI technology redone it again. This means that the large model will bring a new round of growth momentum to the cloud business.

Today, when growth is scarce and uncertainty is normalized, for major technology companies, writing poems is far more urgent than selling cloud products as soon as possible, and realizing commercial realization.