The Ambition Structure | Foundation | Model of the Big Model

What can ChatGPT do? Chat with someone? Answer a few tricky questions? Tell a joke? Of course, it can do all of these things. If you think it can only do these things, just simplify it. At the same time, you cannot understand why there have been dozens of big models similar to ChatGPT in China in just a few months, and more big models are still on the way.

At the 2023 World Artificial Intelligence Conference held on July 6th, the leaders of several major model manufacturers coincidentally mentioned MaaS in their speeches. Their goal is to turn the big model into a service platform. This service platform can be used in various unrelated industries such as finance, office, urban management, and healthcare. In the past, there was talk about Internet platforms, but today, the ambition of all major models is increasingly clear - to be an AI platform.

The big model behind ChatGPT, GPT, belongs to the general big model. The general model has mastered a lot of general knowledge, but lacks in-depth understanding of various professional knowledge. In the past few months, the development of such universal large models and generative artificial intelligence has achieved strong language understanding and reasoning abilities, and can generate complete paragraphs, exquisite images, videos, and even code according to prompts, making artificial intelligence a more powerful personal assistant.

However, there are still many problems in applying the general large model to a variety of different industries. Tang Daosheng, senior executive vice president of Tencent Group and CEO of Cloud and Intelligence Industry Group, said at the World Conference on Artificial Intelligence on the 6th: "General large models are generally trained based on extensive public literature and network information. Online information may be wrong, rumored and biased, and many professional knowledge and industry data are not accumulated enough, resulting in insufficient industry pertinence and accuracy of answers and relatively broad information output." He believes that the general large model can solve 70%-80% of the problems in 100 scenarios, but it may not be able to meet 100 of the needs of a certain scenario of the enterprise.

Different industries and enterprises need large models that serve their respective fields. This type of industry model and exclusive model is based on a universal large model, but requires further fine-tuning with professional data. The finely tuned model can become a platform, serving thousands of industries. And this is precisely the "caution machine" of the big model.

Tencent Cloud recently released a panoramic view of Tencent Cloud's MaaS service, which is based on the Tencent Cloud TI platform to create industry models and select stores, providing more than 50 solutions in 10 major industries including finance, culture and tourism, government affairs, healthcare, media, and education; Launch an industry wide model fine-tuning solution to help model developers and algorithm engineers solve tasks such as model invocation, data and label management, model fine-tuning, evaluation testing, and deployment in a one-stop manner, reducing the pressure of creating large models. Based on these models and tool platforms, enterprises can quickly generate exclusive models by adding their own scene data.

Huawei's Pangu large model has a similar idea. Huawei has introduced a new three-tier large model structure, namely, the basic large model, the industry model, and the scenario model. The model has penetrated into more than 10 industries such as finance, manufacturing, government affairs, electric power, coal mining, medical care, and railways, supporting the landing of artificial intelligence applications in more than 400 business scenarios. Hu Houkun, the rotating chairman of Huawei, said, "The first layer of the basic big model is vividly called Reading Ten Thousand Books, which is to learn a lot of basic knowledge. The second layer of the industry model and the third layer of the scene model are called Wanli Road." There are still many difficulties to overcome from reading thousands of books to traveling thousands of miles. The key point is to fully match and integrate the knowledge of all walks of life with the big model.

The application of large models in various industries can be compared to chip foundries. Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Research Center for Deep Learning Technology and Applications, said that chip foundries have expensive equipment, very long supply chains and complex processes. But for chip design companies, they should not consider these, as long as the chip design scheme is handed over to the OEM factory, and all subsequent production processes are completed by the OEM factory. In the future, the industrial model of large models may be very similar to chips. The large model platform requires expensive computing power and data. After enterprises that produce large models have built the large model platform, those that need to apply large models only need to raise their demands, and then use a small amount of data to fine tune and adapt the large model to meet the industry's needs.

Wang Haifeng introduced that Baidu's ERNIE Bot has been actively used in many scenes, such as energy, finance, education, office, media, etc. In the process of implementing large model industries such as ERNIE Bot, the mode of "intensive production and platform application" can be adopted, that is, enterprises with comprehensive advantages in algorithm, computing power and data can encapsulate the complex process of model production and provide large model services for thousands of industries through a low threshold and efficient production platform. In addition, Baidu's Wenxin Big Model 3.5 has added a plugin mechanism, and the plugin ecosystem will gradually open up in the future. Developers can build their own applications based on the Wenxin Big Model.

Some people compare the future of big models to car rental platforms, where anyone can obtain applications that meet their needs. In 2023, the platform battle for domestically produced large models has already begun.