liang zhihui: 360 pioneered the coe technology architecture, and model automatic calling was launched two months before openai
2024-09-27
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
on september 25, at the ai large model ecology and computing power conference, liang zhihui, vice president of 360 (hereinafter referred to as "360") group, announced that 360's original coe (collaboration-of-experts) expert collaboration model technical architecture has many the model collaboration capability can reach the openai o1 level in terms of effect, and it is currently fully connected to 360ai search and 360ai browser. in addition, openai's recently launched model automatically calls the "auto" mode, and 360 has already implemented the coe architecture two months ahead of the launch as early as the end of july.
it is understood that 360’s original coe expert collaboration architecture has been connected to 16 of the strongest large-scale models in china, including 360 zhi nao, wen xin yi yan, tong yi qian wen, doubao, kimi, etc. based on the search data and user behavior accumulated by 360 in the past, 360 intelligent brain has established a powerful intent recognition model to accurately guess the intent and purpose of user questions. after intent recognition, it uses the task decomposition routing model to automatically perform automatic analysis on many expert models. calling and intelligent scheduling realize "let the strongest model solve the most difficult problems". its intelligence, flexibility and high efficiency far exceed the capabilities of a single large model.
analysts pointed out that the coe technical architecture coincides with the technical concept of openai's latest o1 model. the o1 model is based on reinforcement learning and "thinking chain" technology, which improves the ability of large models to solve complex problems. this is consistent with the o1 model released by 360 at the end of july. the coe technical architecture concepts of the two companies are similar, the methods are similar, and the practical effects are relatively close.
at the isc.ai2024 conference at the end of july, zhou hongyi, founder of 360 group, announced that "using an agent-based framework to create a slow thinking system will enhance the slow thinking capabilities of large models." 360's original coe expert collaboration technology architecture allows major models to collaborate and realize "slow thinking" of large models. through multi-model collaboration, it can obtain capabilities that the single large model it accesses does not have. this is similar to how the opena o1 model simulates the working principle of the human brain based on reinforcement learning and "thinking chain" technology, which can be said to be different approaches to the same goal.
at present, the coe architecture has been fully integrated into 360ai search and 360ai browser products. the ai assistant of 360ai browser also supports user-defined calling models, providing users with functions such as arena mode, team pk, and multi-model collaboration, which are convenient large model capabilities for users to experience different scenarios.
the independent ai product list "ai product list" shows that 360ai search user visits connected to the coe architecture ranked second in the domestic list and third in the global search engine list in august. combined with similarweb public data, in august 2024, 360ai search user visits exceeded 200 million, more than three times that of perplexity ai, and continued to be the world's largest ai native search engine; monthly unique visits (uv) exceeded 80 million, far exceeding domestic a number of ai products; and with a monthly growth rate of 113%, it has become the fastest growing ai search engine above scale in the world.
correspondent xiaohan
proofreading by sheng yuanyuan
