Wu Tian, vice president of Baidu: The Wenxin big model has been applied to search and other products, and more than 60,000 developers have been developed

Wu Tian, Vice President of Baidu Group and Deputy Director of the National Engineering Laboratory for Deep Learning Technology and Application (Source: Baidu Official)

Since OpenAI released GPT-3, big models have become the target of almost all of the world's leading artificial intelligence (AI) companies.

Titanium Media App April 19 news, Baidu Group Vice President Wu Tian recently said in an interview with Titanium Media App and other interviews that Baidu's industry-level knowledge enhancement model "Wenxin" has been applied to Baidu's internal products on a large scale, including search, information flow, Xiaodu smart screen, Baidu map, etc. The number of individual and enterprise developers of the "Wenxin Big Model" has exceeded 60,000.

Wu Tian stressed that "this year is a key year for the landing of the Wenxin big model industry."

It is reported that "big model" is currently one of the hottest topics in the field of AI research in the world. AI technology has developed to this day, and models with huge parameters such as GPT and BERT have been developed, and they have made unprecedented achievements in fields such as computer vision and natural language processing.

With data blowouts, algorithm progress and computing power breakthroughs, pre-trained large models with strong generalization capabilities and versatility are becoming the key direction of AI technology development, and have become an important driving force for AI industry applications, which is expected to deeply integrate AI technology with differentiated scenarios in various industries, so that most enterprises have less labeling data, higher development efficiency, and lower application costs, thereby greatly reducing the application threshold of AI.

In March 2019, Baidu released China's first officially opened pre-trained model ERNIE1.0; in December 2021, ERNIE 3.0 was upgraded to a knowledge-enhancing 100 billion model "Pengcheng- Baidu Wenxin", with a model parameter scale of 260 billion, which is currently the world's largest Chinese monomer model.

At the same time, in December 2021, a new panorama of Wenxin big model was released, which includes NLP (natural language understanding) big model, CV (computer vision) big model, cross-modal big model, and tools and platforms.

In Wu Tian's view, as an industry-level big model, the core value of the "Wenxin Big Model" is to drive the large-scale application of AI technology.

Therefore, in order to further reduce the difficulty of application, Baidu has also developed an easy-to-use and lightweight deployment tool platform for the "Wenxin Big Model", including providing various development kits, zero-threshold AI development platform EasyDL, full-featured AI development platform BML, etc., so that different groups can achieve AI technology applications at a low threshold.

Wu Tian said that through the Baidu AI open platform, there are now nearly 1400 capabilities open to enterprise developers.

Wu Tian told the Titanium Media App that based on the Baidu Flying Propeller platform and the Baige cluster, the "Wenxin Big Model" has achieved independent innovation at the level of algorithm, framework and computing power. Through the integration of large models and domestic deep learning frameworks. Supporting the training of the "Wenxin Big Model" is the end-to-end, adaptive, distributed training framework and 4D hybrid parallel technology independently developed by Baidu Feipao Platform. Baidu has built an AI foundation for independent innovation, which can drive large-scale application of AI.

At present, the "Wenxin Big Model" has been applied through the Flying Propeller Platform and Baidu Intelligent Cloud to empower industry, energy, finance, communications, media, education and other industries. Among them, in the intelligent manufacturing scenario, the large model can be applied to the natural language processing scenarios such as quality inspection and inspection computer vision scenarios and the operation and maintenance of data equipment.

In terms of specific cases, Baidu cooperated with Chinese Life to extract key fields for the text of terms in insurance contracts. Based on the Wenxin model, the intelligent parsing of insurance contract terms is realized, and the key fields of nearly 40 dimensions are automatically extracted, and the efficiency of business processing is greatly improved.

"Based on the annotation information accumulated by the enterprise itself, and then using the 'Wenxin Big Model' to conduct secondary training together with the previous data, it will help customers do some data enhancement work." Then through multiple customer feedback to construct data, model iteration, to achieve a practical state. Wu Tian told the Titanium Media App that when it is really used, the ability of the big model is embedded in the private cloud, a function is embedded in the workflow of the insurance personnel, telling him some suggestions after analysis, and then basically you can quickly use the "Wenxin big model" capability.

In terms of delivery form, the "Wenxin Big Model" mainly has three delivery modes: by calling APIs for developers to use, nested tools on platforms such as Baidu EasyDL and BML Full-featured AI Learning, and delivered in some scenario-oriented products, such as Intelligent Document Analysis TextMind, Intelligent Authoring Platform, etc.

In terms of revenue sources, as a basic work, the revenue source of the "Wenxin Big Model" enters the revenue of Baidu's intelligent cloud on the one hand through the combination of vertical scenarios of industry customers on the one hand, and on the other hand, through the combination of vertical scenarios of industry customers.

However, Wu Tian told the Titanium Media App that the "Wenxin Big Model" consumes computing power in the early training, and in the Baidu search scenario, the Kunlun core is used to calculate the reasoning of Wenxin ERNIE, and every day is more than 100 million traffic calculations. But not all enterprises have such a large-scale computing power platform, to the enterprise scene to really use the big model, many as long as the secondary training can be, the secondary training will not be as large as the training of hundreds of billions of basic models.

For the repeated construction of large models and the problem of healthy competition, Wu Tian said that the value and role that each enterprise and institution will eventually produce are actually their own focus. There is still a lot of room for innovation.

(This article was first published on titanium media App, author | Lin Zhijia)