The global new wave of artificial intelligence (AI) technology, led by large models and generative AI, continues to heat up. The advancements in large model technology both domestically and internationally are surging, with increasing industrial competition and deepening application implementations. As we enter 2024, China’s AI development exhibits a series of new phenomena and trends that deserve our attention and ongoing focus.
Recent New Phenomena in Domestic AI Development
In the early stages of large model development, China’s focus was primarily on catching up with foundational model capabilities. Since 2024, the industrialization of AI has begun to attract attention as a new hotspot, with continuous advancements and achievements emerging in vertical large models, digital humans, embodied intelligence, and AI for Science, among other directions. At the same time, balancing development with governance and promoting the implementation of AI governance has also become a focal point in this new era.
(1) The Application of Vertical Large Models is Accelerating
Currently, the challenge of large-scale commercialization of large models has not been fully resolved domestically or internationally. The path to industrialization of large models in China differs significantly from that in the United States, with faster development of industry-specific models. Since 2024, numerous large model research teams have begun to invest efforts in providing high-quality specialized solutions tailored to specific business scenarios by further training or fine-tuning foundational models using expertise related to those scenarios. For instance, the Pangu Steel Model, developed by Huawei and Xiangtan Iron and Steel Group, won the International Telecommunication Union (ITU) Global AI Excellent Innovation Award, empowering Baowu Steel’s digital transformation and achieving cost savings of 1 billion yuan in just one year in the pig iron production segment. The visual large model developed by the State Grid has already been applied in dozens of scenarios such as drone inspections and channel visualization, completing inspections over 300,000 kilometers and improving average recognition efficiency by 10%. In addition, a large number of industrial large models, mining large models, logistics large models, port large models, and medical large models have started to deepen their implementations in their respective fields, leveraging their vertical advantages to bring significant impacts on China’s industries, especially in manufacturing.
(2) Digital Humans Rapidly Emerging with Broad Applications
Large models are accelerating the maturity of digital content generation capabilities and multi-modal processing of images, language, and text, providing stunning experiences in areas such as character representation, voice generation, animation generation, audio-video synthesis, and human-computer interaction. Digital humans are active in commercial scenarios including education, healthcare, financial services, e-commerce, online streaming, and customer service sales, as well as in large-scale memorial events. For example, during the 2024 “6·18” event, JD launched over 18 brand president digital avatars in its live broadcast, and Ant Group’s Lingjing digital human platform assisted CCTV Finance in creating AI anchors for the Two Sessions. The “2024 White Paper on the Development of China’s Virtual Digital Human Industry” predicts that by 2025, the market size driven by “digital humans” in China will exceed 640 billion yuan and 48 billion yuan in core market size, representing growth of over 90% and 130% compared to 2023.
(3) The Status of Embodied Intelligence is Rising in AI Development
Since 2024, embodied intelligence has become a hot topic in various important forums in China, a priority for research institutions, and a focus for AI-related companies and venture capital investments. Multi-modal large models are key drivers for the development of embodied intelligence, enabling interaction with the environment and evolutionary learning capabilities. Following language dialogue and digital content generation, further exploring the potential of Transformer large models in multi-task learning and multi-modal capabilities, combining multi-modal large models, reinforcement learning, and the perception, action, and environmental interaction capabilities of robots is becoming a new form of AI in the post-large model era, and is regarded by many as a necessary path to general artificial intelligence. A batch of embodied large models have rapidly emerged in China, such as the multi-modal embodied large model RobotGPT launched by Datar Technology, the second-generation embodied intelligence technology LPLM large model developed by Yulu Robotics, and the Ruoyu·Jiutian robot brain introduced by Ruoyu Technology, all utilizing multi-modal large models to achieve functions like understanding human intentions, high-frequency human-computer interaction, and complex task planning. Advanced autonomous driving and humanoid robots are expected to benefit from breakthroughs in embodied intelligence, leading to new capabilities.
(4) Humanoid Robots Are Gaining Momentum
Humanoid robots are a typical product concept of embodied intelligence and a carrier of embodied intelligence technology development. The generalization capability of large models in various scenarios and the explosion of embodied intelligence technology provide significant driving force for the development of more versatile humanoid robots. Although humanoid robots are still in a low-version state of embodied intelligence both domestically and internationally, the technological iterations are rapid. Humanoid robots are advancing in parallel in hot areas such as brain, cerebellum, spatial intelligence, limbs, and upstream core components, with domestic manufacturers emerging in robot bodies, core components, and software algorithms, continuously forming innovative results. For example, Yushu Technology’s Unitree G1 humanoid robot adopts an embodied intelligence approach to break free from pre-programmed algorithms, enhancing its ability to interact with the physical world. Stanford University has developed the HumanPlus humanoid robot based on Yushu Technology’s domestic hardware platform; Zhiyuan Robotics’