In a race for Chinese GPT models, a "state-owned team" has emerged.
BEIJING, June 5 (TiPost) --Zhongke Wenge, an artificial intelligence (AI) technology company owned by the Chinese Academy of Sciences" Institute of Automation, released on Saturday an enterprise-level exclusive large-scale Yayi (雅意) model, with five core capabilities, including real-time online Q&A, domain knowledge Q&A, multilingual content comprehension, complex information extraction, and multi-modal content generation.
(相关资料图)
It can quickly connect to government and enterprise data and generate large-scale model-specific application services. This model can be applied in media, finance, publicity, governance, security, and other fields.
Wang Lei, the chairman of Zhongke Wenge, said that almost all large-scale models are basic models that are geared towards general tasks and do not consider deep application in specific industries. There are many problems, such as high computing costs, difficult localization deployment, high data security risks, and weak domain-specific knowledge. Yayi is designed to understand complex data, not just chat; it is designed to focus on specific industries, not just general task capabilities; and it supports proprietary data and private domain deployment. Yayi has four core advantages: high integration, exclusive ownership, security and reliability, and deep domain cognition.
Wang Zhangsheng, the head of Zhongke Wenge"s delivery center, told TiPost App that the Yayi model focuses more on vertical domain knowledge capabilities, supports offline, online, and provides local training and recommendation integration for private domain deployment. Overall, the Yayi model supports high-quality instruction fine-tuning capabilities, and supports high-quality data training in the fields of finance, media, and governance, thereby better solving practical application problems in these areas.
Zhongke Wenge is a company incubated by the Chinese Academy of Sciences" Institute of Automation in 2017, positioning itself as a cognitive and decision-making intelligent technology enterprise, based on multi-modal (text, image, video, voice) fusion semantic analysis, focusing on complex data analysis and AI-assisted decision-making, and has created a data and decision-making intelligent basic platform DIOS with independent intellectual property rights, providing multi-language, cross-modal, and deep cognitive intelligent big data and AI basic platform and solutions for media, publicity, public opinion, finance, governance, security, and other fields.
Zhongke Wenge has recently advanced AI technology iteration and product implementation, including DIOS Central Data Operation System, Yayi Domain Big Model, and Zhiji Digital Person, the three major digital requirements. The company has more than a thousand paid customers, mainly including state-owned enterprises and universities such as State Grid Jiangsu Electric Power, State Grid, Tsinghua University, China Media University, Bank of Beijing, and China Merchants Bank.
As of now, Zhongke Wenge has completed six rounds of financing, raising over 1 billion yuan. Investment institutions include China Development Bank Financial, CEC Capital, Yingfutaike, DeepVC, and Chinese Academy of Sciences Capital. According to the official website of the China Securities Regulatory Commission in January of this year, Zhongke Wenge has entered the listing counseling stage, with Haitong Securities and others as counseling agencies.
Wang Zhangsheng explained to TiPost App that the Yayi big model has universal capabilities and supports three service forms: MaaS (Model as a Service), knowledge connection mode, and knowledge enhancement mode. It supports cloud use, local integrated machine deployment, and autonomous private training deployment.
Not only that, around the Yayi industry vertical ecology, Zhongke Wenge has also launched the "5+N" plan, covering scenario business in the fields of media, publicity, finance, governance, and security, and generalized to industries such as home, medical care, and education, providing intelligent topic selection, intelligent manuscript creation, event summary, macro policy interpretation, government data governance, and event element analysis services. In addition, Zhongke Wenge has also released the high-fidelity interactive digital person "Zhiji" with a brain.
Wang Zhangsheng told TiPost App that Zhongke Wenge has a clear positioning, a multimodal, lightweight, and controllable big model, which truly enables customers to use it. Moreover, the Yayi big model has real market demand, technical needs, and a clear business model.
Regarding the relationship between the development of generative large models and decision-making AI technology,Wang Zhangsheng believes that the core technologies of GPT currently include intelligent learning, CoT prompting in thinking chains, reinforcement learning, and Markov decision processes (MDP). At present, large models enhance decision-making intelligence, but they cannot replace it, or the two will develop together for a long time. "You need to use large models to enhance some links in the decision-making chain, or to generate some toolboxes through large models and use these toolboxes to enhance the various links of decision-making. But if you ask whether there will be a replacement in the future, the direction of technological development may change."
At the press conference last Saturday, Zhongke Wenge, Yuxin Electronics, Contemporary China and the World Research Institute, Huichen Stock, Juzhang Data, Ruishi Fangda, and Xiongdong Zhijia signed contracts to jointly build the Yayi large model ecosystem.
Wang Zhangsheng emphasized that today"s breakthroughs in general AI technology and new developments in cross-modal and large language model technology have brought the company closer to the vision of empowering hundreds of industries, thousands of cities, and millions of enterprises with AI. With the help of the Yayi large model, the company is committed to creating a safe and reliable enterprise-level exclusive large model for every customer. (This article was first published on the TiPost App, author | Lin Zhijia)
