In the first half of 2023, why did nearly half of the world's venture capital invest in artificial intelligence?

Image source: Picture Worm Creative

Economic Observer reporter Li Xiaodan Global AI enterprise venture capital is slowing down as a whole, with the amount of venture capital disclosed in the first half of 2023 reaching US$24.6 billion, down 14.6% from the same period last year. In the first half of 2023, the proportion of venture capital received by global AI companies in the total global venture capital reached 18.9%, a new high in recent years.

According to the report "Artificial Intelligence Global Transformation Prospects: The Transition Point is Coming (2023)" released by KPMG and Zhongguancun Industry Research Institute, from the perspective of investment and financing, the proportion of artificial intelligence in global venture capital increased rapidly in the first half of 2023.

On 14 December, Jiang Liqin, Head of Client & Business Development, KPMG China, and Zhang Qingjie, Head of Digital Enablement, KPMG China, were interviewed by The Economic Observer.

Jiang Liqin said that in recent years, the world economic pattern has been deeply adjusted, the global capital market has become more cautious, and the overall investment action has slowed down. However, with the technical characteristics of the integration and application of artificial intelligence with various cutting-edge technologies such as the Internet of Things, blockchain, and Web3.0, the potential market space is broad, and the growth certainty is relatively obvious, which is generally optimistic about investors, especially the generative AI applications represented by ChatGPT will usher in a round of explosion in 2023, which will raise the popularity of artificial intelligence in the capital market to a certain extent.

There are currently 36,000 AI companies in the world, with China and the United States leading the way in the number of companies, 12,925 and 5,734 respectively. As of the end of June 2023, the total number of unicorns in the global AI field reached 291, with 131 in the United States and 108 in China, accounting for 45% and 37% respectively.

"Fundamentally speaking, the technology of artificial intelligence subdivision is becoming more and more mature, and the scale effect of scenario-based application is gradually emerging as the underlying driving force for capital flow to related fields, taking the equity investment in the field of artificial intelligence in Chinese in the past ten years as an example, in terms of technology, the growth rate of venture capital in the four subdivisions of computing power, data platform, natural language processing, computer vision and image has accelerated significantly; There are many investment events in logistics and warehousing, accounting for more than 75% of all investment events in artificial intelligence. Jiang Liqin said.

According to the report "Artificial Intelligence Global Transformation Prospect: The Transition Point is Coming (2023)", Chinese enterprises in the field of artificial intelligence were intensively born between 2015 and 2018, about two-thirds of the core enterprises in the field of artificial intelligence were established in 5-10 years, and with the gradual weakening of effective investment growth, the number of newly registered enterprises in the field of artificial intelligence reached a peak of 528 in 2017, and decreased year by year, and the number of newly registered enterprises decreased to 63 by 2022.

Combined with the regional layout, Chinese artificial intelligence enterprises are mainly concentrated in Beijing, Guangdong, Shanghai, Zhejiang and other places, forming a three-legged pattern of Beijing-Tianjin-Hebei, Yangtze River Delta, Guangdong, Hong Kong and Macao, of which there are more than 1,600 artificial intelligence enterprises in Beijing. In terms of AI unicorn companies, there are 41 AI unicorn companies in Beijing, ranking first in the country. Shanghai and Guangdong Province ranked second and third, with 24 and 23 AI unicorns, respectively.

Zhang Qingjie, Head of Digital Enablement, KPMG China, said that as the pace of large model launch in the market slows out, the large models that have been launched will enter a round of competition, and the focus of the competition is no longer on the improvement of the number of parameters in a single modality, but on multimodal information integration and in-depth mining, so that the model can more accurately capture the correlation between different modal information through the exquisite design of pre-training tasks.

Zhang Qingjie introduced that there are three main development ideas: one is to use single-modal models such as LLMs (large language models) to mobilize functional modules of other data types to complete multi-modal tasks. HuggingGPT, etc.; the second is to directly use image and text information to train a multi-modal large model, typical representatives are KOSMOS-1, etc.; the third is to organically combine LLMs with cross-modal encoders, etc., to integrate the inference retrieval ability of LLMs and the multi-modal information integration ability of encoders, typical representatives are Flamingo, BLIP2, etc.

"In the domestic market, although the domestic large model is slightly inferior to the GPT series model and PaLM-E in terms of market influence, it has local advantages in Chinese corpus training and Chinese cultural understanding. In addition, real industries such as domestic manufacturing provide rich training data and application scenarios for large models. In the future, in terms of large-scale model empowerment for the industry, China's large-scale model is very likely to be the last to come, and it will also be one of the key factors in the competition of domestic large-scale models. Zhang Qingjie said.

The shortage of computing power has aroused widespread concern from all walks of life, and intelligent computing power is regarded as the main technology to solve this problem. With the rapid explosion of demand for intelligent computing power, the classical computing theory represented by Moore's Law and Feng's structure system has entered a bottleneck period, and the transformation of traditional computing paradigms has become an inevitable trend, and the industry is accelerating the innovation of chips and computing architectures. In this context, this report proposes that intelligent computing power will be ubiquitous in the future, specifically presenting four characteristics: "multiple heterogeneity, software and hardware collaboration, green intensification, and cloud-edge-end integration".

In the MaaS model, users on the demand side can focus on their own business logic and user experience, without having to pay attention to the underlying technical details, which is conducive to solving the key bottleneck of AI "usable" but "not easy to use", while on the supply side, it is expected to form a "general large model + field large model + industry large model + enterprise/individual small model" This basic business format promotes the implementation of AI in thousands of industries, and finally realizes AGI (artificial general intelligence).

Jiang Liqin pointed out that the risks faced by the Technical Ethics and Social Forum in the development of artificial intelligence show that there is a long way to go in the development of safe and credible artificial intelligence, and in the process of solving AI risks, technological innovation opportunities such as explainable AI and federated academic Xi have emerged.