The ability to spread technology is very important in a country's technological strength, especially in the context of AI. Research has shown that compared to the dispersed and more conducive to diffusion ecosystems in the United States, China's diffusion ability lags far behind its innovation ability - ranking nearly 30 places lower than the latter. At present, China is planning to enhance its diffusion capacity through large-scale investment in AI education to cultivate a mid-range, industrial focused AI workforce.
The United States has recently made continuous breakthroughs in Large Language Modeling (LLM) and Diffusion Modeling. These successes have driven a surge in US venture capital for generative AI startups focused on the technology and service industries, such as HarveyAI, Jasper, and Runway. This optimistic sentiment in the United States is in stark contrast to the relatively slow AI startups in China. For example, some analysts suggest that the language model of the highly anticipated Alibaba supported startup, Zero One All Things (01. AI), is largely based on Meta's open-source LLaMA Foundation model.
However, China's delay in the field of generative AI can be explained by the government's strategic priority of prioritizing industrial applications over service industries and traditional knowledge work applications. Especially, the government hopes to combat the decline in industrial productivity growth through AI investment in the industrial sector, in order to break free from the middle-income trap.
For this reason, the government has requested the Ministry of Education to promote cutting-edge machine learning technologies in the domestic industrial sector. Therefore, Chinese universities have established over 2300 undergraduate AI programs, most of which focus on applied AI projects for industrial applications. The AI higher education of the Ministry of Education includes two objectives:
The unprecedented investment by the Ministry of Education in applied AI higher education projects is due to the government's recognition of a fundamental problem faced by most AI companies focused on industry: building AI solutions for specific industries requires a significant amount of time and funding, often unable to be achieved through horizontal business models. This issue is common because, apart from niche industries such as recycling (successful Western startups such as AMP Robotics have already emerged), most industrial sectors have company specific data infrastructure. Manufacturing companies typically run on software systems called Manufacturing Execution Systems (MES) and Monitoring and Data Acquisition (SCADA). These systems often combine a chaotic combination of internal tools and software services built by traditional suppliers such as SAP or Siemens.
Therefore, in order to build an AI model that can be practically deployed, these different systems must first be unified through cross system data channels. Most of this work is customized and involves the tedious process of building powerful data channels between traditional, disconnected systems, which typically come from different providers with different data patterns. Only after completing this arduous task can computer vision models begin training and deployment. Even after the completion of these complex data channels, industrial process training for AI systems will bring its own challenges:
Given the multitude of products and processes used by industrial enterprises, each model must be trained on a specific system within the company or product line. In other words, the large-scale and industry wide diversity limits the scalability of industrial AI solutions.
In addition, every time a company changes its processes or switches to new products, it not only needs to establish new manufacturing processes, but also needs to retrain its computer vision models - if this is a new product or process without a large amount of existing training data, this is usually a daunting task. Therefore, without a robust dataset to adjust machine learning algorithms for new processes, manufacturers are often forced to manually test and generate fitness training data until the model becomes sufficiently reliable.
In the field of industrial AI, due to the need for continuous updates, limited customer base, and high upfront costs, the lucrative economic benefits of Software as a Service (SaaS) companies are no longer applicable. Li Kaifu's AI industry startup, AInovation, developed AI systems for industrial enterprises such as China Steel International and construction giant China Railway No. 4. As Innovation Intelligence expands its customer base, it is forced to increase software spending and deploy more proprietary low profit hardware, resulting in a sharp drop in gross profit margin from 62.9% in 2018 to 29.1% in 2020.
Due to the fact that AI in the industrial sector cannot be adopted through a one size fits all horizontal solution, the industry requires a large number of new intermediate AI talents - more skilled than regular data analysts, but with lower costs than ML engineers and researchers. China's AI education strategy aims to fill this niche market for applied AI talent, so that every industrial company can hire internal teams to build its own machine learning infrastructure.