The annual domestic "AI Spring Festival Gala"—— KLCII has once again kicked off.

On June 14, 2024, the AI industry event "2024 Beijing Zhiyuan Conference" opened at the Zhongguancun Exhibition Center. Kai-Fu Lee and Zhang Yaqin, Dean of the Institute of Intelligent Industry of Tsinghua University, had a fireside conversation, Wang Xiaochuan, CEO of Baichuan Intelligence, Zhang Peng, CEO of Zhipu AI, Yang Zhilin, CEO of the Dark Side of the Moon, and Li Dahai, CEO of Facing Wall Intelligence, and other key figures opened a large-scale discussion meeting on the most critical issues in the AI industry.

The iteration of domestic large-scale models is accelerated

In a recent interview and conference, KLCII President Wang Zhongyuan elaborated on the rapid development of large-scale model technology in China over the past year. He pointed out that the domestic large-scale model field has experienced a struggle to catch up with GPT 3.5 is now on average, surpassing GPT 3.5 and closely follow the significant progress of GPT4, especially in Chinese application scenarios, showing the ability to surpass the international advanced level.

However, in view of the continuous evolution of GPT4, such as the performance leap brought about by the release of GPT4o, the iteration and optimization of domestic large models still face urgent challenges.

KLCII has made a number of innovations in the exploration of large model technology, including the launch of Tele-FLM-1T, the first low-carbon trillion language model, and the launch of the BGE series model and the next-generation multimodal model Emu3 to solve the problem of large model illusion. Wang Zhongyuan also emphasized that although domestic large models have reached the usability standard, there is still a long way to go to achieve high optimization, especially in key technical problems such as computing power integration, core algorithms and system engineering.

About the accelerator of large model development, the law of scale (Scaling Law) has become a hot topic in the industry. Industry leaders such as Kai-Fu Lee believe that AI In the context of the 2.0 era, the law of scale is crucial to promote the intelligent growth of large models, and its potential is far from saturated.

However, some industry experts such as Yang Zhilin, Zhang Peng, Wang Xiaochuan and Li Dahai pointed out that although they agree on the importance of scale expansion, relying solely on the increase of data and computing power is not the only path to artificial general intelligence (AGI), and breakthroughs in learning paradigms, model training methods and systematic innovation are still needed.

Together, these perspectives paint a complex picture of both opportunities and challenges in the field of large models.

A comprehensive review from C-end potential to AI security

In the interview, Wang Zhongyuan is optimistic about the future application of large models, and expects that the next two or three years will usher in the widespread popularity of B-end applications, and admits that although the C-end market is looking forward to a blockbuster, it needs to wait patiently for the model to mature to a stage where it is both cost-effective and can directly meet the needs of users. He emphasized that the technical path towards the AGI era may transcend the limitations of a single language model, integrate multimodal understanding and decision-making capabilities, and eventually lead to in-depth applications in the field of embodied intelligence and science.

OpenAI Sora Team Lead, Aditya Ramesh expressed his recognition of the importance of language modality in his interaction with the academic community, and also foresaw the potential of fusing linguistic information with visual signals, indicating that the dependence of models on language may be reduced.

In response to the current boom in video generation models, Aditya highlighted the focus on security and social impact, revealing that the Sora team is focusing on enhancing the controllability of the model and reducing randomness in response to partner needs.

During the conference, AI security was a core topic, and Dr. Yeung reiterated its importance for the future, especially in the scaling of model capabilities In the context of the exponential growth of Law, preventing malicious exploitation by users and pre-built model behavior norms have become a dual-track strategy to maintain AI security.

This shows that while pursuing technological progress, the industry also needs to make forward-looking plans to ensure the healthy development of AI technology and the positive impact of society.

The power of open source and positive circulation

Li Dahai emphasized the multi-dimensional value of the open source ecosystem, pointing out that it not only brings together the driving force of original work, but also absorbs a wide range of demand input and feedback loops, forming a benign symbiotic environment.

He believes that actively participating in open source contributions can bring positive benefits to the company and promote a win-win situation for both technology and business.

Wang Xiaochuan analyzed the market demand, influence building and competitiveness maintenance of open source, and believed that open source action can not only enhance the industry influence of enterprises, but also share it in the rapid iteration of the model ecosystem.

On the contrary, this open attitude strengthens the company's market reputation and is in line with industry development trends. He praised the current many enterprises joining the ranks of open source to jointly promote China's AI ecosystem to quickly catch up with the international level, and called for continuous joint construction and sharing of a prosperous open source ecosystem.