Looking at the big model to accelerate the release of AI potential
Since Baidu released the Flying Paddle Xinghe Community in August 2023, it has launched over 4,000 innovative AI applications based on Wenxin Big Model.
◇The Water Affairs Bureau of Lintao County and Baidu AI Cloud jointly built the first "Artificial Intelligence Emergency Rescue System for Drowning Prevention" in China. In the month when it was launched, the first life rescue of the drowning prevention system was successful, and many high-risk behaviors have been successfully prevented so far.
Text | "Outlook" Newsweek reporter Hu Yongshun
Hearing impaired people are using "Acoustic Bridge AI Language Training" to practice speaking.
"After installing a cochlear implant, can I use AI to help me learn to speak?" Due to genetic reasons, Li Pengcheng from Inner Mongolia lost his hearing when he was born. After the cochlear implant was installed, he had to pay a "learning to speak" fee, because after the cochlear implant was installed, he could not immediately understand what others were saying. To learn to speak through a rehabilitation teacher, "it costs at least 5,000 yuan a month, which many people can’t afford."
Li Pengcheng’s needs have been solved in the "Sound Bridge AI Language Training" team. Tang Xuan, the head of the team, said that they designed a product that uses AI technology to help hearing-impaired people correct their voices. "By recognizing the voices of hearing-impaired people, I give specific suggestions for unclear or wrong parts, guiding them to modify and make progress, which reduces the cost of learning to speak."
"Sound Bridge AI Language Training" helps the hearing impaired learn to speak, which comes from the functional extension of Baidu’s ERNIE Bot language model. At present, the large model based on strong algorithm, large computing power and big data has become the mainstream direction of artificial intelligence development, providing a new base for artificial intelligence technology and application.
Recently, at the WAVE SUMMIT+ Deep Learning Developers Conference, Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Research Center for Deep Learning Technology and Application, announced the latest progress in promoting AI value creation: the scale of users in ERNIE Bot exceeded 100 million; By the end of December 2023, Flying Paddle had gathered 10.7 million developers, served 235,000 enterprises and institutions, and created 860,000 models based on Flying Paddle. Since Baidu released the Flying Paddle Xinghe Community in August 2023, it has launched over 4,000 innovative AI applications based on Wenxin Big Model, and "Sound Bridge AI Language Training" is one of them.
The foundation of large model is gradually compacted.
Since 2023, the wave of AI big model technology has continued to be hot. Some universities and innovative enterprises have increased their research efforts, and the large-scale model technology has been iteratively upgraded. Internet companies such as Baidu and Alibaba, as well as scientific research institutions such as Fudan University, have launched their own large models.
According to public information, as of October, 2023, 238 large models have been published in China, which can be divided into two categories: general model and industry vertical model. By learning common knowledge from massive data, the general large model has become a model base with universality and generalization ability.
"Artificial intelligence has many typical abilities, among which understanding, generation, logic and memory are the core basic abilities. The stronger these four abilities are, the closer they are to general artificial intelligence, and the big language model has these four abilities, which brings dawn to general artificial intelligence." Wang Haifeng said.
Taking ERNIE Bot as an example, in March, 2023, Baidu released the ERNIE Bot Grand Language Model, which was widely used by users, ranging from welcome speech, speeches to plans, manuals, flowcharts, mind maps and so on, covering many aspects of work and life.
ERNIE Bot’s grand language model is a part of Wenxin grand model series. Since 2019, Baidu has been deeply involved in the research and development of pre-training models and released version 1.0 of Wenxin Big Model. ERNIE Bot’s basic model is Wen Xin Da Model 3.0. Since then, Wenxin model has been rapidly upgraded to versions 3.5 and 4.0, and the four basic AI abilities of understanding, generation, logic and memory have been comprehensively improved.
Wang Haifeng introduced that the upgrade of Wenxin Big Model is based on further innovation breakthroughs in several key technical directions, and knowledge enhancement, logic enhancement, plug-ins and agent mechanisms are added on the basis of knowledge enhancement, retrieval enhancement and dialogue enhancement.
Solving "just need" with big model
The research and development threshold of large models is high and difficult. Only by truly integrating into thousands of industries, solving the "just need" of industrial development, and allowing industries to reap value from AI, large models can achieve sustainable development.
The White Paper on Innovative Application of Big Model in Beijing Artificial Intelligence Industry (2023) proposes that from the perspective of model evolution, the general big model tends to converge, and vertical industry application has become the key track for the big model industry to land. At present, the development of large-scale model presents a development path from technology to product, and then to commercial application, and continues to penetrate into vertical industries.
In the process of going deep into the vertical industry, the large model relies on the comprehensive support of algorithms, computing power and data, and industrialization is facing challenges. Wang Haifeng said that enterprises with comprehensive advantages of algorithm, 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 and high-efficiency production platform.
At present, Wenxin Big Model has been widely used in Internet products such as search, information flow, smart speakers, etc., and has been used in all walks of life such as manufacturing, energy, finance, communication, media, cities, education, etc. through the open source platform of paddle flying. With the further expansion of application scenarios, Wenxin Big Model has built more than 10 industry big models with head enterprises and institutions in various industries to help enterprises reduce costs and increase efficiency, and accelerate the transformation and upgrading of digital intelligence in the industry.
In Lintao County, Gansu Province, the Taohe River runs through the whole city, and there are drowning incidents every year. In June, 2023, the first domestic "Artificial Intelligence Emergency Rescue System against Drowning" jointly built by Lintao County Water Affairs Bureau and Baidu AI Cloud was launched. Researchers analyzed the test data through AI video, built a drowning prevention model, and trained the model with a large number of high-quality scene data, which can identify and warn the dangerous behaviors such as crossing the railing, approaching the water flow, wandering in dangerous areas and so on at the first time, thus gaining valuable time for subsequent emergency treatment and rescue. On-line month, the first life rescue of the drowning prevention system was successful, and many high-risk behaviors have been successfully prevented so far.
Baidu also upgraded the AI-aided training system of the national diving team based on the Wenxin model. The system can not only understand and execute the complex instructions of coaches and athletes, but also provide accurate information in time, and can also score and accurately quantify the movements in real time. In 2023, China Swimming Association awarded Baidu the title of "Artificial Intelligence Partner of China National Diving Team".
Wen Xin Da model is also applied to Chinese root-seeking. Through the cooperation with the National Library, Wenxin Big Model has learned a lot of ancient local records and genealogy data, and recognized and understood the characters. At the same time, it has comprehensively applied the knowledge map of location, occupation, diet, important deeds and other information, and launched the service of "ancient prose asking today" in ERNIE Bot. Users only need to input the root-seeking information, and they can get the corresponding clue feedback.
Provide all-factor support for AI native applications
In order to promote the big model to produce more native applications, Baidu has recently carried out a series of new upgrades to the Xinghe community around the community ecology.
The reporter learned from Wu Tianchu, vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and Application, that the newly released Xinghe Community Large Model Tool Center, including the flying paddle industrial model library, Baidu Brain AI capabilities, ERNIE Bot tools, etc., also supports eco-tool access, provides a visual interactive interface, flexible parameter configuration, and real-time preview effect, providing developers with all elements of AI native applications, including integrated services of development, experience, promotion and communication.
In terms of ecological co-creation, Baidu released the Wenxin model Xinghe co-creation plan, which will jointly activate the value of data resources with developers and ecological partners, jointly build large model plug-ins and extensively innovate AI applications.
In addition, in order to accelerate the training of AI talents, in 2020, Baidu proposed the goal of "training 5 million artificial intelligence talents for the whole society in five years", and the number of talents currently trained has reached 84% of the target. The rapid development of large-scale model technology has also put forward higher requirements for AI talents. In 2023, Baidu released a new initiative for AI talent training-Galaxy Project.
"We will work closely with all walks of life in Industry-University-Research to deepen the integration of production and education and train another 5 million model talents for the society." Wang Haifeng said that the reason for doing this is to make the "flower of innovation" of AI technology bear more "fruits of industry" and serve the national strategy, social development and people’s well-being.
(Outlook, No.03, 2024)
Reporting/feedback