Shulou(Shulou.com)11/24 Report--

On October 17, Baidu World 2023 with the theme of "Generating the Future" was held in Shougang Park, Beijing. Wang Haifeng, chief technology officer of Baidu, interpreted the key technologies and latest developments behind Wenxin Model 4.0. Wang Haifeng said that the four abilities of understanding, generation, logic and memory of Wenxin Big Model 4.0 have been significantly improved. In September, small flow has been launched, and the effect has increased by nearly 30% in the past month. Since August 31, Japanese Heart Word has been open to the whole society, with 45 million users, 54,000 developers, 4300 scenarios, 825 applications and more than 500 plug-ins.

Wenxin Big Model 4.0 was launched, and the ability of understanding, generation, logic and memory was significantly improved.

It is understood that Baidu released a knowledge-enhanced large language model on March 16. Wenxin Yiyan integrates learning from trillions of data and hundreds of billions of knowledge to obtain a pre-trained large model. On this basis, it adopts supervised fine tuning, human feedback reinforcement learning, prompt and other technologies, and has the technical advantages of knowledge enhancement, retrieval enhancement and dialogue enhancement.

The basic model of Wenxin Yiyan was upgraded to Wenxin Grand Model 3.5 in May, and innovative breakthroughs were made in basic model upgrading, fine adjustment technology innovation, knowledge point enhancement, logical reasoning enhancement, plug-in mechanism, etc., and the effect and efficiency were improved.

The Wenxin Grand Model 4.0 released today has significantly improved the four abilities of understanding, generation, logic and memory compared to version 3.5. Among them, the improvement of understanding and production ability is similar, but the improvement of logic and memory ability is greater, the improvement of logic is nearly 3 times that of understanding, and the improvement of memory is more than 2 times that of understanding. Baidu has developed an intelligent code assistant Comate based on Wenxin Big Model. From the internal application effect, the overall code adoption rate reaches 40%, and the code adoption rate of high-frequency users reaches 60%.

It is reported that Wenxin large model 4.0 in September has started small flow online, the effect in the past more than a month has increased nearly 30%. The efficiency of the training algorithm has increased by 3.6 times since March, and the weekly average training efficiency exceeds 98%. The basic technical architecture of Wenxin Grand Model 4.0 is in line with versions 3.0 and 3.5, and further innovation breakthroughs in several key technical directions.

Running the flying propeller platform on the 10,000-calorie computing power supports the stable and efficient training of large models through the soft and hard collaborative optimization of cluster infrastructure, scheduling system and flying propeller framework. A multi-dimensional data system has been built, forming a closed loop of data mining, analysis, synthesis, labeling and evaluation, fully releasing the value of data and greatly improving the model effect. Multi-stage alignment based on supervised fine tuning, preference learning, reinforcement learning and other techniques ensures that the model is better aligned with human judgments and choices. Regenerative training technology saves training resources and time effectively and speeds up model iteration speed through incremental parameter optimization.

In addition, Wenxin Model 4.0 performs knowledge point enhancement in both the input and output phases. On the one hand, it understands the questions input by users, disassembles the knowledge points needed to answer the questions, then searches for accurate knowledge in search engines, knowledge graphs and databases, and finally assembles these found knowledge into Prompt and sends them to large models with good accuracy and high efficiency. On the other hand, the output of the large model is reflected, knowledge points are disassembled from the generated results, and then the search engine, knowledge graph, database, and the large model itself are used to confirm, and then the erroneous points are corrected.

On the basis of powerful basic large model, Baidu has further developed intelligent body mechanism, including understanding, planning, reflection and evolution, which can achieve reliable execution and self-evolution, and white-box the thinking process to a certain extent, so that the machine can think and act like a human being, complete complex tasks independently, and continuously learn to realize independent evolution in the environment.

Wen Xin's words help activate the cultural resources of the national library collection and upgrade the AI auxiliary training system of the national diving team.

Baidu has carried out strategic cooperation with the National Library of China to give full play to their respective advantages in resources, technology and services to jointly promote the intellectualization of cultural resources and knowledge services and benefit the public. The National Library has the world's largest collection of ancient chronicles. It uses the Wenxin model to learn ancient chronicles and genealogy data, helping Chinese around the world to obtain more clues to their roots. Wang Haifeng shared the story of overseas Chinese Zhao Peijuan continuing his grandfather's homesickness and successfully "seeking roots."

China national diving team coach, Olympic hardware Wang Chen Ruolin and Olympic champions Quan Hongchan, Chen Yuxi, Wang Zongyuan also came to the world congress site, and Wang Haifeng eager interaction, decryption dream team training "unique skills." Based on Baidu big model technology such as Wenxin Yiyan, the AI auxiliary training system of China National Diving Team has been comprehensively upgraded. Through learning from mass data and knowledge, it has mastered rich diving knowledge, understood and executed the complex instructions of coaches and athletes, provided accurate information in time, scored movements in real time, accurately quantified analysis, and provided the most valuable guidance information to assist the efficient training of China National Diving Team.

On the spot, Chen Ruolin was entrusted by Zhou Jihong, chairman of China Swimming Association, to award Baidu as "artificial intelligence partner of China National Diving Team." Chen Ruolin said that China's national diving team and Baidu have carried out in-depth technical cooperation since 2019, and have gone through the Tokyo Olympic Games, Fukuoka World Championships and Hangzhou Asian Games hand in hand, and are currently preparing for the Paris Olympic Games next year. "Thanks to Baidu's technical support and service, both parties will continue to use artificial intelligence technology to help diving teams train intelligently and improve their competitive level through scientific training. "

Baidu Talent Training Galaxy Plan Released, Retraining 5 Million Model Talents

Wang Haifeng released the latest data of Baidu artificial intelligence talent cultivation on the spot. Baidu proposed to cultivate 5 million AI talents for the whole society in 5 years in 2020. Up to now, Baidu has cultivated 4.2 million AI talents.

At present, large models have become the hot direction of artificial intelligence development, further accelerating industrial transformation, AI talent training also needs to keep pace with the times. Wang Haifeng released Baidu's talent training Xinghe plan. He said,"We will cooperate closely with all walks of life in industry, university and research, deepen the integration of production and education, cultivate 5 million model talents for the society, let the 'flower of innovation' of artificial intelligence technology bear more 'fruit of industry', serve the national strategy, serve the social development and serve the people's well-being! "

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.