With the emergence of ChatGPT, the World's Artificial Intelligence Conference (WAIC), which resumed its normal pace after the pandemic, has seen a surge in popularity.
On July 8th, EqualOcean reporters followed the surging crowd through the security checkpoint and entered the Shanghai Pudong World Expo Exhibition and Convention Center. They saw a scene of crowds of people and visitors gathering around multiple well-known internet and AI company booths. According to official data released by WAIC, as of 3 pm that day, the number of offline visitors to the conference exceeded 177,000, breaking the historical record, while in 2022, during the pandemic, the whole exhibition had only 30,000 visitors.
At the same time, the most striking feature of this year's WAIC is its "large" and "precise" scale. It is reported that this year's exhibition area is 50,000 square meters, twice that of last year. More than 400 companies participated in the exhibition, twice as many as last year. In terms of "precision," the hotly discussed concepts of the metaverse and AR in the past two years have taken a back seat, replaced by Large-scale Language Models (LLM) that have been driven and dominated the internet headlines since the beginning of this year by OpenAI and ChatGPT. As a result, the derivative products of LLM, virtual humans, have become the main push products of various companies.
AI Companies Pursuing LLMs are Now More Rational
Walking in the over 50,000 square meter exhibition venue, the most direct feeling for EqualOcean reporters is that there are really too many companies focusing on LLM. The three words almost occupy every inch of the exhibition hall, from the product brochures of participating companies, the theme introductions of booth press conferences, the billboards inside and outside the exhibition hall, to the chatting voices of passers-by. We can clearly smell the strong taste of LLM. From a data perspective, the popularity of LLM is global. According to the "China Large-scale Language Model Map Research Report" released by the New Generation Artificial Intelligence Development Research Center of the Ministry of Science and Technology, more and more research and development teams in Europe, Russia, Israel, South Korea and other places are investing in the development of LLM, in addition to China and the United States. In terms of the global distribution of released LLM, China and the United States are far ahead, accounting for more than 80% of the total.
According to EqualOcean's inventory, among the participating companies, there are 29 companies that have released products with over 1 billion parameters of LLM, mainly divided into the following categories:
Internet and cloud computing companies: Baidu (百度), Huawei (华为), Alibaba (阿里巴巴), Tencent (腾讯), JD.com (京东), NetEase (网易), 360, CE Cloud (中国电子云);
Computer vision and speech companies: SenseTime (商汤科技), iFlytek (科大讯飞), Mobvoi (出门问问), 4Paradigm (第四范式), Unisound (云知声);
Telecom operators: China Mobile (中国移动), China Telecom (中国电信);
New generation of Large-scale Language model startups: Frontis (衔远科技), Silicon Intelligence (硅基智能), Langboat (澜舟科技), Zhipu AI (智谱AI);
Big data and scenario companies: Emotibot (竹间智能), MIDU (蜜度), Data Grand (达观数据), LuchenTech (潞晨科技), SoundAI (声智科技), Transwarp (星环科技), WAYZ (维智科技), Yoo (必优科技) , Wenge (中科闻歌), ZKJ (中关村科金).
Huawei
Although ChatGPT has set off a technological trend of generative AI, and during the hot period at the beginning of the year, various internet giants competed to come up with a Chinese version of GPT, as companies fought over NVIDIA GPU resources and LLM training costs surged, this trend has gradually returned to rationality.
Currently, models with over tens of billion training parameters are mainly developed by large companies such as Huawei, Alibaba, and Baidu, as their products have monthly active users counted in the tens of millions. Other mid-sized companies and AI startups focus more on small but precise models with billions of parameters, which are dedicated to providing industry solutions.
B2B Solutions: AI Development Platforms, Offices, Virtual Customer Service, Databases
Looking closely at the industry customer solutions presented by companies showcasing LLM, we can see that most of the deployment of LLM has been focused on upgrading existing customer groups. The solution types include the following categories: AI development platforms, offices, virtual customer service, and databases.
Huawei, occupying the largest booth at the event, brought their PanGu LLM, undoubtedly the most eye-catching star. EqualOcean saw at the event that PanGu's practical cases in various scenarios, such as mining, railways, meteorology, and finance, were all showcased on the booth. According to the introduction, PanGu has already been deployed in hundreds of scenarios in more than ten industries, including government affairs, manufacturing, coal mines, railways, and pharmaceuticals. In addition, the Ascend AI platform, as the technical foundation for various software and hardware development tools, provides computing power support for more than 30 LLMs.
Microsoft recently launched the groundbreaking product Office Copilot. To seize market opportunities, Kingsoft (金山软件) also brought WPS AI to its WPS office users, which is the first ChatGPT-style application for domestic collaborative office. According to the staff, the new WPS AI includes two modules: AIGC and Codex. The former includes content generation, semantic understanding (products: text, intelligent documents); content creation, layout beautification, and note generation (products: PPT presentation); content recognition, semantic understanding, and knowledge Q&A (products: PDF, snapshots). The latter includes advanced functions such as data Q&A and analysis, table operation, and formula generation, all of which will be integrated into tables and intelligent tables.
Kingsoft
Virtual assistants also appeared on the booths of many companies with both AI and intelligent hardware businesses. Tencent Cloud Intelligence presented the small-sample digital human from its YouTu Lab; Mobvoi showcased its latest virtual assistant LLM "Sequence Monkey," and developed AI dubbing assistant "Magic Sound Workshop," digital avatar "Wonderful Origin," writing assistant "Wonderful Text," and painting assistant "Words and pictures" on this basis. As a leading intelligent hardware manufacturer for overseas markets, Mobvoi told EqualOcean that the company will integrate these LLM products into its smart watches and bluetooth earphones in the future.
Baidu
In addition, the models for general solutions in the B2B industry include Alibaba's AI painting creation model Tongyi Wanxiang, Baidu Wenxin, SenseNova from SenseTime, and the enterprise software LLM "Shi Shuo" from 4Paradigm.
In the database and model underlying related fields, tool company Emotibot created Model Store and brought the LLM training fine-tuning platform EmotiBrain; Transwarp provides Transwarp Scope, an enterprise-level interactive data retrieval and statistical analysis platform for industry customers; and Colossal-AI, an open-source, low-cost AI LLM development system, was brought by LuchenTech, incubated by Sequoia Capital and Innovation Works. The system is based on PyTorch and can reduce the cost of developing and applying AI LLM training/fine-tuning/inference.
Summary
Based on observations and discussions at the WAIC event, EqualOcean has summarized the following trends and perspectives:
Although the LLM war has only been going on for less than half a year, the differentiation of the tracks has already emerged. Internet giants are showcasing their muscles in chasing after C-end general products, while small and medium-sized companies have shifted their focus to B-end solutions, where being small and precise is enough to support their business.
LLM cannot do without the support of underlying computing power and terminals. Therefore, in this wave of hype, general-purpose computing chips, GPUs, and enterprise software are expected to have a share of the pie.
Compared with the metaverse boom of the past two years and the current lack of attention, the industry should think more about how to truly leverage LLM to improve business efficiency and avoid being left with a mess.