On September 1, the 2022 World Artificial Intelligence Conference (WAIC 2022) officially opened. On the afternoon of the 1st, the “AI Open Source and Industrial Intelligence Summit Forum” sponsored by the National Engineering Research Center for Deep Learning Technology and Application was successfully held. Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Research Center for Deep Learning Technology and Application, delivered an opening speech. Academicians He Jifeng and Wang Huaimin, academicians of the Chinese Academy of Sciences, discussed the value of AI open source and openness, and shared the practice of AI and different fields of integration and innovation. Suggestions and suggestions for industrial intelligence.

Wang Haifeng pointed out in his speech that the industrial model of the large model will be a “TSMC-like” model. The large model plus deep learning platform runs through the entire AI industry chain and is the foundation of industrial intelligence.

At present, as an important driving force for a new round of scientific and technological revolution and industrial transformation, artificial intelligence technology has shown strong versatility and has widely penetrated into the main links of economic production activities. The large model is an important direction for the development of artificial intelligence in recent years. It has the characteristics of good effect, strong generalization, and standardized R&D process, which further enhances the versatility of artificial intelligence and brings new opportunities for the further development of artificial intelligence.

At the same time, the research and development of large models also faces a series of challenges such as large data scale, uneven data quality, large model size, high training difficulty, and large computing power requirements.

In this context, how to accelerate the industrial landing of large models and make large models play a greater role? In Wang Haifeng’s view, enterprises with comprehensive advantages in algorithms, computing power, and data can encapsulate the complex process of model production, and provide large-scale model services for thousands of industries through a low-threshold, high-efficiency production platform. In this way, only a few companies need to worry about big data, big computing power, and big model capabilities, and thousands of industries can directly apply AI models.

Wang Haifeng took TSMC, Samsung and other companies as examples to further explain the industrial model of the large model, which can be compared to the foundry model of the chip industry. He said, “The chip manufacturing process has high technical barriers, and there are expensive production lines, which can standardize and automate mass production of chips according to customer needs, forming economies of scale. It has advantages in aspects such as massive data, large-scale computing power, etc. It can automatically and standardize the production of multi-scenario and multi-field models according to the needs of AI application parties. When it reaches a certain scale, it can form a healthy and sustainable large-scale model industry model.”

The development, training, reasoning deployment, and industrial implementation of large models cannot be separated from the support of deep learning platforms. The deep learning platform is connected to chips and applications, which is equivalent to the “operating system” of the intelligent era. The large model plus the deep learning platform can run through the entire AI industry chain from hardware adaptation, model training, reasoning deployment, and scene application.

The “TSMC-like” industrial model of the large model has been verified in Baidu’s large model practice.

Baidu has developed a series of Wenxin industry-level knowledge-enhanced large-scale models, including basic general-purpose large-scale models represented by Pengcheng-Baidu Wenxin, large-scale models in the field of biological computing, and large-scale models in the energy, finance, and aerospace industries. Tools and platforms for large-scale models to adapt to scene applications, explore Yanggu, a creative community for ecological co-construction, etc. Baidu Wenxin’s large model is also one of the “treasures of the eight town halls” of WAIC this year, making its hard-core debut in the exhibition area.

The successful development and application of Wenxin series of large models is based on the strong support of flying paddles. Flying Paddle, as my country’s first self-developed, open-source and open industrial-level deep learning platform, integrates core framework, industrial-level model library, development kit and tool components, as well as learning and training communities, and supports large-scale model production in a standardized and automated manner. and apply.

In terms of large model training, Paddle has developed an end-to-end adaptive distributed training technology, which automatically selects a parallel strategy according to the characteristics of the model and computing power platform to achieve the ultimate end-to-end performance optimization.

In terms of large model reasoning, Paddle has created a full-process deployment solution for large models of compression, reasoning, and service, which can widely support different types of model structures and achieve high-speed reasoning. At the same time, Paddle also provides automatic model compression tools to help save computing resources.

The Wenxin large model is open source and open based on the Flying Paddle platform, and the ecology is jointly built. There are more than 10,000 developers in total. Based on the Wenxin large model, more than 30,000 tasks have been created. At present, the Wenxin large model has been widely used in Baidu products, significantly improving product effects and R&D efficiency, and applied in finance, energy, medical care, manufacturing, corporate services and other fields. The ecology of the large model has begun to emerge.

“The industrial model of the large model will be a ‘TSMC-like’ model. The large model plus the deep learning platform, which runs through the entire AI industry chain, is the foundation of industrial intelligence.” Wang Haifeng concluded.

Reviewing editor: Peng Jing

Leave a Reply

Your email address will not be published.