At the 2024 Zhiyuan Conference, four leading figures in the field of large models in China—BaiChuan Intelligence CEO Wang Xiaochuan, ZhiPu AI CEO Zhang Peng, CEO of Moon's Dark Side Yang Zhilin, and CEO of Wall-Facing Intelligence Li Dahai—engaged in an in-depth discussion on the topic of "whether large models are the cornerstone on the path to Artificial General Intelligence (AGI)." They shared their views on the role of large models in the development of AGI, and here are the highlights of the discussion.

Large Models: Stepping Stones to AGI?

Wang Xiaochuan believes that large models are the cornerstone to AGI, but relying solely on the Scaling Law cannot achieve AGI. A paradigm shift is needed, such as innovations in data algorithms and computing power, as well as exploring new learning paradigms and compression modes. He proposed that the definition of AGI could be whether it can create doctors, as doctors are one of the professions with the highest intellectual density.

Zhang Peng stated that although it is currently uncertain whether large models can help humans reach the pinnacle of AGI, at this stage, large models are effective and the Scaling Law will continue to play a role for a considerable period in the future. He believes that AGI is a dynamic concept, and its connotation and extension will continue to evolve.

Yang Zhilin emphasized that large models are the first principles, and intelligence can be generated by increasing the model scale and better compression. However, he also pointed out that in areas where data is scarce or non-existent, the data issue needs to be addressed.

Li Dahai believes that large models are the furthest that current technology can go in the direction of AGI, but whether they can directly achieve AGI still has unknown factors. He proposed that large models currently mainly deal with the work of the human brain's System 1, and in the future, the capabilities of System 2 need to be externalized or internalized through agent technology.

Definition and Development of AGI

During the discussion, the CEOs offered different views on the definition of AGI:

Yang Zhilin believes that the definition of AGI is important, but it is difficult to quantify precisely in the short term. It is necessary to break down the evaluation dimensions to better measure the progress of AGI development.

Wang Xiaochuan uses the ability to create doctors as an indicator to evaluate AGI, as the medical profession requires multimodal processing, low hallucination, memory, reasoning, literature review, and other abilities.

Li Dahai defines AGI from an economic perspective as having a marginal cost of zero for performing any task and believes that the intelligence density and miniaturization of large models are also important directions for development.

Large Models and AI Safety

With the development of large models, AI safety has also become a focal point of discussion:

Yang Zhilin proposed that AI safety is very important and needs to be prepared in advance, focusing on the issue that the model itself may make inappropriate actions due to malicious intentions of users.

Wang Xiaochuan emphasized the three levels of ideological security, human civilization security, and practical security, believing that the current security issues should focus on ideological security.

Zhang Peng and Li Dahai both stated that AI safety is an important topic that requires the joint efforts of the industry to ensure that technology is used for the right purposes.

Conclusion

Through this roundtable discussion, we can see that although large models play an important role on the path to AGI, whether they can become the decisive factor in achieving AGI, and the exact definition of AGI, are still open questions. At the same time, the issue of AI safety cannot be ignored and requires joint efforts from both inside and outside the industry to ensure the healthy and sustainable development of technology.

The Zhiyuan Conference, as a grand event in the field of AI, not only provides a platform for the industry to exchange ideas and share insights but also points out the direction for the future development of large models and AGI. With the continuous advancement of technology and the increasing maturity of the market, we have reason to believe that large models and AGI will bring more possibilities and opportunities to human society.