Industry trend｜Slow thinking AI era, CoE architecture to accelerate the big model change?
With the rapid development of science and technology, new technologies are still emerging, refreshing people’s infinite imagination of science and technology. For example, in the field of artificial intelligence, the rapid iteration of AI technology and the emergence of new technologies are constantly subverting people’s cognition and imagination space.
Not long ago, the United States OpenAI company officially launched the all-new model o1, which has attracted wide attention in the industry. It is reported that the o1 model uses a new optimization algorithm and a new training data set specially customized for its training, the use of “reinforcement learning” approach, and through the “chain of ideas” to deal with user queries, which makes it better at dealing with complex problems.
Different from previous models, the o1 model adopts a “slow thinking” approach, that is, it will repeatedly think, disassemble, understand, reason and other complex in-depth thinking before answering questions, and then provide high-quality answers.
It is worth mentioning that the concept of “slow thinking” was not invented by OpenAI. As early as the ISC.AI 2024 conference, Zhou Hongyi, founder of 360 Group, proposed, “Build a slow thinking system with an agent-based framework, so as to enhance the slow thinking ability of large models.” At the same time, 360 launched its first Expert Collaboration (CoE) technology architecture and hybrid large model, which has been applied to a number of products, such as 360 AI search and 360 AI browser.
What is the CoE architecture?
CoE (Collaboration of Experts) architecture, as an innovative technology architecture, the core idea is to decompose a large model into multiple “expert models”, each expert model has a deep accumulation and precise capabilities in a specific domain. These expert models can be dynamically combined and coordinated according to actual needs to provide optimal solutions to improve the overall performance and intelligence level of artificial intelligence systems.
Unlike the MoE architecture, the CoE architecture emphasizes the collaborative work between multiple expert models, and there is no binding relationship, so the architecture is more flexible and scalable. In the CoE, a single entry accesses multiple models at the same time, and an intent identification step is added before model analysis to determine which models or models are responsible for the task.
As an innovative technical architecture, CoE architecture intelligently optimizes resource allocation through the collaborative work of mixed large models and expert models, so that the most suitable models can handle the most suitable tasks, significantly improve the accuracy and integrity of answers, and also improve the overall performance and efficiency of the system, injecting new vitality and momentum into the development of artificial intelligence field.
360 teamed up with domestic large model manufacturers to defeat GPT-4o
At present, with the rapid development of artificial intelligence technology, AI large models have become the new highland of the current global scientific and technological competition. With its large number of parameters and strong generalization ability, large model has shown great application potential and value in many fields.
But the current situation is that there are too many big models and not enough applications. The big model must not only run, but also use. Therefore, how to make the large model better applied to the specific scene and landing, to create actual value for various industries, has become particularly important. However, the large model of a single enterprise often has limitations, and it is difficult to meet the complex and changing market needs.
In response to this, 360 proposed CoE (Collaboration-of-Experts) technical architecture, bringing a new mode of thinking for the application of large models.
It is reported that in addition to the access to 360 self-developed large model 360 brain, CoE architecture also includes intelligent spectrum AI, Sensetime Technology, Baichuan intelligence, Bytebeat, Baidu, Tencent, IFLY, Huawei, MiniMax, zero all things, wall intelligence, Ali Cloud, magic square quantization, good future, the moon’s dark side and other major domestic manufacturers of large model products.
At the same time, 360 announced a strong cooperation with 15 domestic large model developers to launch an AI search application product — 360 AI Assistant, which will be fully built in 360 family bucket products and 360 AI search, without installing additional plug-ins can directly call each large model capabilities. Achieve one-stop experience of up to 15 of the country’s most powerful models.
Of course, this “super model” data architecture that integrates a number of large model capabilities also lives up to expectations, and it has excellent performance and strong strength in many aspects. From the actual effect, the AI assistant that sets the strengths of various companies has a total score of 80.4 in the test, significantly surpassing the GPT-4o’s 69.22 points, and is comprehensively ahead in 11 ability dimensions.
As one of the leaders of AI technology, 360’s innovation and exploration in the field of large models not only brings competitive advantages for itself, but also broadens ideas for the development of the entire industry and brings more leading and standardized references. In the future, with the continuous progress of technology and the continuous expansion of application scenarios, we have reason to believe that large model technology will play a more important role in the future, promoting the sustainable development and popularization of AI technology.
This paper is from Ulink Media, Shenzhen, China, the organizer of IOTE EXPO (IoT Expo in China)
