"Volume" funds to grab talents, understand the competitive landscape of domestic AI large-scale models in one article

Original source: Titanium Media

Editor: Liu Yaning, Sun Cheng

Image source: Generated by Unbounded AI‌

On June 29, Meituan issued an announcement on the Hong Kong Stock Exchange, announcing the acquisition of the large-scale model company "Light Years Beyond" founded by Wang Huiwen, which also left speculation on the AI large-scale model market. Not only in China, the number of visits to ChatGPT has declined recently. According to data, ChatGPT’s global visits in June fell by 9.7% month-on-month, and the number of unique visitors fell by 5.7% month-on-month. In addition, the duration of users using ChatGPT is also declining. The duration of each visit decreased by 8.5% in May. AIGC seems to have ushered in a cooling-off period.

But on the other hand, on July 4th, the large-scale model company "Zero One Creation" organized by Kai-Fu Lee himself was officially launched, and the large-scale language model "Wen Xin Yi Yan" APP developed by Baidu was launched on the Apple App Market, and the enterprise check publisher check Large model - "Knowing the Enemy Alpha"... Various companies are making frequent moves in the field of large models, and they are in full swing.

How hot is the AI big model business? What is the investment layout of the capital market for this outlet? How is the performance of today's domestic large-scale models? This issue of Titanium Media·Tidu Tuwen tries to sort out the current development status of domestic large-scale models from the data level.

Domestic AI large-scale model "fairy fight" technology giants are competing to shoot

Since March this year, domestic AI models have sprung up like mushrooms after rain. The "China Artificial Intelligence Large Model Map Research Report" shows that by the end of May 2023, at least 79 domestic basic large models with a parameter scale of 1 billion or more have been released.

From a global perspective, the United States ranks first in the number of large-scale models in the world. As early as 2020, the United States has launched 15 large-scale models, including GPT-3. In China, starting from 2020, large-scale models have also entered a stage of rapid development. In 2021, 30 large-scale models will be launched, 28 large-scale models will be launched in 2022, and 19 large-scale models will be launched in the first five months of 2023, surpassing the 18 large-scale models in the United States. large model. Today, the number of large-scale models in China and the United States accounts for nearly 90% of the total number of large-scale models in the world, and the number of large-scale models in China has entered the first echelon.

The domestic large-scale model track is very lively. The "players" involved in the AI large-scale model track include Internet giants represented by Ali, Tencent, Baidu, etc. On March 16, Baidu led the way and launched the Chinese version of ChatGPT "文心一言", Subsequently, large models such as Ali, Huawei, Tencent, and 360 surfaced one after another.

In addition, there are AI technology companies represented by SenseTime and iFLYTEK, education companies like Kidswant, financial companies like Flush, and real estate companies such as I Love My Home. In addition, large models are also very popular in academia. As early as February this year, Fudan University released MOSS, the first ChatGPT-like model in China.

It is worth noting that with the popularity of ChatGPT, there has also been a boom in AI large-scale entrepreneurship in China. Many technology leaders and executives from major companies have devoted themselves to large-scale models and started re-entrepreneurship.

In February 2023, Wang Huiwen, the former co-founder of Meituan, released the "AI Hero List", a high-profile entry into the AI large model; also high-profile entry was Li Kaifu, CEO of Sinovation Ventures. On March 19, Li Kaifu announced in the circle of friends Established the Project AI 2.0 company and entered the AI large-scale model; subsequently, Sogou founder Wang Xiaochuan, former JD.com AI business pioneer Zhou Bowen, and Momenwen CEO Li Zhifei also joined the AI large-scale model entrepreneurial boom.

In addition to these technology giants, many executives from major Internet companies have also joined the competition. For example, Li Yan, a former AI core figure of Kuaishou, established the AI company Yuanshi Technology to conduct research and development of multi-modal large models; Li Kaifu's "closed disciple" and Yitu Technology's No. Reshape the model of human-computer interaction. Today, the domestic AI large-scale "fairy fight", the industry track is very lively.

Investment institutions are cautious, and the market recruits talents in the field of large models with high salaries

To enter the large model, you must first prepare money. After all, a large model startup company releases a relatively mature model, which needs to invest in the cost of large language model infrastructure, training and R&D talents. For start-ups, financing is the only way to survive.

According to the incomplete statistics of Titanium Media·Tidu Tuwen, as of July 4, 2023, the financing amount of AI large-scale startup companies that have obtained financing this year is generally between tens of millions and hundreds of millions. Among them, MiniMax, a large-scale model start-up company that received the most financing, completed a new round of financing exceeding US$250 million on June 1. The company's current valuation exceeds US$1.2 billion, and it is dubbed "Tencent's first investment." Entrepreneurship" title.

In addition, Light Years, which was just acquired by Meituan, completed its A-round financing of US$230 million on June 5, but this round of financing has also been returned in full after the completion of the acquisition by Meituan. It is worth mentioning that Xihu Xinchen, established in 2021, completed two consecutive financings in March and April this year, and the financing speed is very fast.

In AI large-scale model investment, leading VCs such as IDG Capital, Matrix Partners, and ZhenFund have made fewer moves.

According to the incomplete statistics of Titanium Media and Titanium News, Tencent Investment, Innovation Works, Qiji Chuangtan, and Sequoia China are the most active investment institutions, with three investment companies, followed by BV Baidu Ventures and Zhipu AI , investment companies are 2 respectively, and most institutions invest in only 1 company, and Hillhouse Venture Capital and Shenzhen Venture Capital, which have always been active, have no public data showing that they have invested in AI large-scale model tracks. It can be seen that investors are interested in The track of large-scale model entrepreneurship is more cautious.

In addition to financing, technical capabilities are also at the heart of large-scale model competition. Large models have high requirements for algorithms, computing power, and data. AI large models require high training costs and R&D investment. Some major Internet companies and technology companies have deeper capital and technical reserves.

According to the incomplete statistics of Titanium Media·Tidu Tuwen, in 2022, Huawei will invest 161.5 billion yuan in R&D expenses, becoming the company with the most R&D investment funds; secondly, Tencent ranks second with 61.4 billion yuan, and Ali ranks with 55.5 billion yuan third. Internet giants invest tens of billions in R&D every year. In addition, they also have a strong R&D team. They are well-deserved "first echelon" players on the large-scale model track.

In addition to capital and technology, talent is also an indispensable part of large-scale model competition.

In the first quarter of this year, various companies launched a competition for top AI talent. Wang Huiwen said that he would use 75% of the shares to invite top R&D talents. Li Kaifu called for recruiting world-class talents around the world. Baidu recruited AI large-scale model algorithm engineers with a salary of 25-40k/month, and recruited model architecture with a salary of 20-40k/month Engineer; Ali recruits large-scale model training and algorithm engineers with a salary of 40-70k/month, and major companies are eager for talents.

Relevant data show that in the first quarter of this year, the number of postings related to AI large models increased by 10.16% year-on-year, which is the field with the largest growth related to ChatGPT. Among them, the median annual salary of two positions of machine learning and algorithm engineers has reached more than 400,000 yuan .

China-made large-scale model development challenges still have some areas that outperform international models

In the past six months, large models have been released one after another. What is their current strength? According to the evaluation of 10 large-scale AI models at home and abroad by relevant institutions, ChatGPT has a relatively obvious leading edge overall, ranking first in terms of comprehensive data. In terms of sub-sectors, the domestic large-scale model surpasses the foreign model in terms of word and sentence understanding and knowledge questions, that is, the domestic AI large-scale model has a stronger basic cognition and learning ability for text.

Among the six domestic large-scale models, the comprehensive data performance of "Wen Xin Yi Yan" is relatively excellent, especially leading the domestic large language models in terms of word and sentence comprehension, programming, knowledge questions and translation; in terms of business writing, domestic products perform the best The best is "Xunfei Xinghuo"; in terms of literature questions, the domestic product "Tiangong 3.5" has the highest score rate of 88.33%, which is on par with ChatGPT; in terms of logical reasoning, "Wenxinyiyan" and "Xunfei "Spark" performed fairly well.

Domestic large-scale models perform very well in some subdivided fields, but there is still a certain gap with similar foreign products, and there is still a lot of room for improvement in core technologies such as data and landing scenarios.

At the data level, the development of large models requires high-quality training data sets. At present, the total amount of Chinese data required for large-scale model training based on Chinese semantics in China is insufficient, and the quality is not high. The scarcity of high-quality Chinese large-scale model training corpus also widens the gap.

In terms of computing power, based on data center capacity, the leading companies in this field in 2022 are Google, Microsoft, Amazon, and Meta. These four data center platforms account for more than 70% of the total capacity of self-built data centers in the world; and At present, although some leading technology companies such as Alibaba, Baidu, and Tencent have completed the construction of data centers in China, there is still a certain gap in terms of capacity compared with foreign countries.

In terms of landing scenarios, the cost of landing in different subdivision scenarios is also a problem. In order to obtain a sufficiently effective and easy-to-use large model, sufficient and targeted corpus needs to be invested, which means that the implementation of different subdivision scenarios requires a lot of cost investment.

It is estimated that the cost of training a large model is between US$2 million and US$12 million. Due to high training costs, the artificial intelligence development company OpenAI will lose about 540 million U.S. dollars in 2022. Therefore, it will take time for the commercial realization of AI large models.

On this highly competitive track, even though there are many difficulties, domestic large models are moving towards a more perfect technical direction. Moreover, the competition in the track will also accelerate the pace of development of domestic large-scale model technology and promote the overall technological upgrading of China's AI industry.

*Data source: Structure Research, InfoQ Research Center, Liepin Big Data, "China Artificial Intelligence Large Model Map Research Report" and other public information are not completely compiled.