AI Large Models Continuously Iterate, Accelerating the Implementation of AI Applications
-
Recently, multiple enterprises, including telecom operators and internet companies, have accelerated the iteration and upgrade of general large models while launching industry-specific models for sectors such as industry, finance, and transportation. The accelerated implementation of industry-oriented large models further unleashes the industrial potential of artificial intelligence.
On November 10, China Telecom publicly released an upgraded semantic large model, Xingchen, with hundreds of billions of parameters. A relevant official from China Telecom explained that the Xingchen semantic large model, with its hundreds of billions of parameters, primarily addresses issues such as hallucinations and multi-round logical reasoning faced by semantic models with tens of billions of parameters during commercial implementation. It focuses on enhancing capabilities in text and image generation and comprehension, with a 30% improvement in Chinese contextual understanding and generation.
This is not an isolated case. Recently, domestic companies such as Baidu, Alibaba, and iFlytek have successively launched upgraded versions of their large AI models. For example, Alibaba Cloud released the 100-billion-parameter model Tongyi Qianwen 2.0, which shows significant improvements in complex instruction understanding, literary creation, general mathematics, and knowledge retention. iFlytek upgraded its Spark Cognitive Large Model V3.0 and initiated training for an even larger-scale Spark model.
"The development of artificial intelligence, represented by large models, is characterized by rapid technological innovation, strong application penetration, and intense international competition, demonstrating powerful enabling effects," said a relevant official from the Ministry of Industry and Information Technology. Currently, China's core AI industry continues to grow, with over 4,400 enterprises, and innovative achievements such as AI chips, development frameworks, and general large models are constantly emerging.
Meanwhile, industry-specific large models have attracted significant investment from numerous enterprises. Recently, China Mobile, in collaboration with China Information and Communication Technology Group, China TravelSky Holding Company, and China National Aviation Fuel Group, launched the "Jiutian·Zhongqing Base Large Model." This foundational model integrates specialized knowledge from eight industries, including communications, energy, steel, construction, and transportation, enabling tailored solutions for enterprises to build industry-specific large models and develop intelligent applications. China Telecom has introduced over a dozen industry-specific large models covering education, government services, transportation, finance, and tourism. Notably, its education model can reduce teachers' grading workload by 70%, while its grassroots governance model improves document drafting efficiency by sixfold.
"Compared to foundational large models trained on general datasets, industry-specific large models offer more specialized datasets, domain expertise, and workflow methods, making them crucial for accelerating industrial implementation," said Zou Debao, an analyst at CCID Consulting's Artificial Intelligence Industry Research Center.
From accelerating drug development and precise weather forecasting to shortening factory product delivery cycles and improving office efficiency, an increasing number of industry-specific large models are being rapidly implemented. For example, at the recently held 2023 World Internet Conference Wuzhen Summit, multiple companies showcased their latest achievements in large model applications.
In Huawei's exhibition area, the Pangu Weather Model caught people's attention. As the world's first AI model with accuracy surpassing traditional numerical forecasting methods, it can complete a 24-hour global weather forecast in just 1.4 seconds, achieving a speed improvement of over 10,000 times compared to traditional methods. Additionally, the intelligent cockpit model and development toolchain, developed based on the Wenxin large model, enable vehicles to possess higher-level autonomous driving capabilities, more powerful self-learning and memory functions, transforming human-vehicle interaction and enhancing user experience.
The "2023 China Generative AI Enterprise Application Research" report recently released by CCID Consulting predicts that China's generative AI enterprise adoption rate will reach approximately 85% by 2035. Manufacturing, retail, telecommunications, and healthcare sectors are leading the adoption, with rates of 82%, 90%, 65%, and 53% respectively.
"China's large model development is more industry-oriented, currently being applied across multiple sectors including manufacturing, energy, power, chemicals, and transportation. Generative AI is empowering all industries, creating unprecedented commercial value for enterprises," said Zou Debao. It's estimated that global generative AI deployment will generate $12.5 trillion in economic benefits by 2035, with China contributing $5.9 trillion, accounting for 47.2% of the global market share.
However, industry insiders also point out that constrained by factors such as applications, data, computing power, and algorithms, some large models still face difficulties like redundant construction, being numerous but not strong, and challenges in profitability. While further solidifying the foundational technologies of artificial intelligence, it is also necessary to unleash the potential of industry applications.
The relevant official from the Ministry of Industry and Information Technology recently stated that they will further strengthen the technological foundation of artificial intelligence. Through major scientific and technological innovation projects, efforts will be made to drive breakthroughs in foundational and original technologies such as large model algorithms and frameworks, improve the computing power of AI chips, and enhance the development of 'root' technologies. At the same time, they will promote the intelligent upgrading of key industries, deepen the integration of AI technologies throughout the entire manufacturing process, advance AI pilot demonstrations, expand specialized application scenarios, and accelerate 'intelligent transformation and digital transition' to form tangible productive forces.
"Applications are not only the purpose of technological development but also the engine driving its progress," said Zou Debao. He emphasized that breakthroughs in large-scale models still require massive industry-specific datasets for training to achieve more accurate reasoning capabilities and expand industry adoption and accessibility. Additionally, it is essential to promote deep integration of industry, academia, research, and application, strengthen collaborative innovation in high-performance computing and high-quality algorithms, and accelerate breakthroughs in key technologies and their industrial implementation.