Huaxia Times reporter Lu Xiao reports from Beijing
Not only the roll parameters, but also the number of rolls has become an increasingly important competition direction for large model manufacturers.
"ERNIE Bot has reached 300 million cumulative users and 500 million daily calls." On June 28, Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Research Center for Deep Learning Technology and Applications, disclosed this group of figures when he released Wenxin 4.0 Turbo at the WAVE SUMMIT Deep Learning Developers Conference. In April of this year, when Baidu disclosed these two data to the public, they only exceeded 200 million each.
Big model manufacturers are looking forward to rapidly expanding the vast AIGC ecosystem, which is intertwined with upstream developers and downstream users. One important reason is that AI big models are bringing visible commercial growth opportunities to manufacturers, and the ecological competition among big model manufacturers is becoming increasingly fierce.
Continuous evolution of large models
The big model is the core of the AIGC ecosystem, and its evolution is ongoing.
As the earliest manufacturer in China to release the universal big model, on June 28th, Baidu released the 4.0 Turbo version of the Wenxin Big Model to the public. Our reporter learned from Baidu that through the continuous innovation of core technologies such as data, basic model, knowledge enhancement, retrieval enhancement and dialogue enhancement, as well as the joint optimization of PaddlePaddle Wenxin, Wenxin 4.0 Turbo is faster and has better effect. It is reported that the web version, APP and API of Wenxin 4.0 Turbo have been opened to users in succession, and enterprise users can use API services immediately by logging into Baidu AI Cloud Qianfan model platform.
The continuous evolution of big models aims to make themselves more useful and attract more developers and users.
Public information shows that Baidu has been fully deploying artificial intelligence since 2010. In March 2019, it launched the Wenxin Big Model 1.0, and in June 2023, it released the Wenxin Big Model 3.5 version. In October of that year, it upgraded to version 4.0, and in April of this year, it released the Wenxin Big Model 4.0 tool version. In addition, the Wenxin Big Model also includes lightweight versions such as ERNIE Speed, Lite, and Tiny.
This series of models has their own positioning. Wu Tian, vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and Application, told reporters such as the Huaxia Times on June 28 that in terms of specific capabilities, Wenxin's lightweight model is suitable for solving the problem of determining scenarios, Wenxin 3.5 is suitable for daily information processing and text generation tasks, while Wenxin 4.0, a larger and more capable model, is suitable for providing professional and in-depth help. Wenxin 4.0 tool version is based on agent technology and is good at comprehensively using a variety of tools and data to complete very complex tasks as required.
How to make the big model evolve better? Industry insiders believe that effectiveness and cost are the two key points of concern for the upstream and downstream of the AIGC ecosystem when communicating with reporters from Huaxia Times.
From the perspective of Baidu, Robin Lee, the founder, chairman and CEO of Baidu, said at the Baidu AI Developers Conference Create 2024 in April this year that "the big model and generative AI will completely change the group of developers. In the past, developers used code to change the world; in the future, natural language will become the new universal programming language." He also disclosed at the time that compared with the Wenxin big model a year ago, the algorithm training efficiency has increased to 5.1 times, the weekly average training efficiency has reached 98.8%, the reasoning performance has improved 105 times, and the reasoning cost has decreased to 1%.
On June 28th, Wu Tian also revealed that in the past six months, the daily average number of questions asked by users of the Wenxin model has increased by 78%, and the average length of questions has increased by 89%. As of now, the Wenxin Model has generated a total of 7 billion lines of code, created 590 million articles, compiled one million professional research reports, answered 170 million learning questions, and assisted 130 million people in their work.
In addition, our reporter also learned from Baidu that currently 80% of Baidu's engineers are deeply using their own intelligent code assistant, Wen Xin Kuai Code, with a code adoption rate of 46% and a new code generation rate of 29%.
Intense ecological competition
Behind the rapid expansion of the ecosystem, the commercial growth brought about by the big model is being seen.
As the earliest technology manufacturer in China to release AI big models, Baidu Wenxin's performance in financial reports has always been a window for external observation of the commercialization progress of big models. The financial report shows that Baidu's core revenue of 23.8 billion yuan in Q1 this year increased by 4% year-on-year, with Baidu AI Cloud's revenue of 4.7 billion yuan increasing by 12% year-on-year, and continuing to achieve operating profits on a non GAAP basis. Robin Lee also proposed in the Q1 financial report teleconference that model reasoning is one of Baidu's long-term opportunities and will become a key growth driver of AI cloud revenue in the future. The set of numerical comparisons he proposed at the time was that in April, the Wenxin Big Model processed about 200 million API calls per day, compared to only about 50 million in December.
Robin Lee also said at the above financial report meeting that 2024 will be the second year for Baidu to embark on the road of general artificial intelligence. Baidu is transforming its Internet centric business into AI oriented business. Baidu is also actively promoting its 2C business and 2B business to adopt the cultural heart model. "In the long run, AI will bring us sustained revenue and profit growth."
The visible commercialization potential has prompted big model manufacturers to engage in an ecological war.
In terms of costs that developers are concerned about, in May this year, ByteDance's Doubao Big Model, Alibaba's Tongyi Qianwen Big Model, Baidu's Wenxin Big Model, and iFlytek's Starfire Big Model all launched a B-end price war. Among them, Baidu announced that its two lightweight main models, ERNIE Speed and ERNIE Lite, are fully free of input and output.
Ding Daoshi, an Internet analyst, believes that it is reasonable for big model manufacturers to wage price wars. He told a reporter from Huaxia Times that the entire cloud service industry is engaged in a price war, competing for users and markets. If the technical strength of major model manufacturers is not inferior, price will become an important factor in ecological selection. "According to Moore's Law, as the technology industry develops, all high-precision technologies will become cheaper and the application scale will become more and more widespread."
In terms of market cultivation, big model manufacturers have also put in a lot of effort. According to reporters, during the release of Pangu Model 5.0, Huawei Cloud CEO Zhang Ping'an announced that Huawei plans to develop over 3000 Huawei developer evangelists in the next three years. In addition, Huawei Cloud will provide each new ecosystem developer with a free cloud host, a set of development tools, and 5GB of cloud storage space. On June 28th, Baidu also announced that it will continue to promote the development of the Wenxin Big Model Xinghe Community Data Co creation Plan, and will continue to hire a team of thousands of experts to join Wenxin mentors to guide Wenxin Big Model in learning more diverse professional knowledge.
However, based on the current penetration rate of big models, the ecological competition among big model manufacturers not only requires competition in technology and funding, but more importantly, patience.