“Only the best large models can produce the best native applications of artificial intelligence.”
With the first batch of large model applications going online through registration at the end of August, China's artificial intelligence large model market has entered a new stage. Large model products developed through registered technology companies and institutions can provide services to all users, whereas before it could only be carried out with a limited number of users. Restricted testing.
Many large model practitioners believe that a large number of large model applications will appear in China. However, there are still many unknowns about how to use the power of large models to solve the common needs of a large number of users and create popular applications. In the past six months of exploration, some fast-moving large model companies and entrepreneurs are trying to provide reference.
AI native applications are the reason for the existence of large models
On September 5, the Baidu "Wenxin Cup" Entrepreneurship Competition, which lasted for more than three months, ended, and 15 teams were selected as winners. In addition to receiving tens of millions of yuan of investment from Baidu, they will also develop innovative technologies, products, and development strategies. , capital cooperation and other aspects to obtain long-term support from Baidu.
Baidu CEO Robin Li said in his award speech that a good artificial intelligence native application must have at least three conditions: support natural language interaction, support understanding, generation, reasoning, memory information, etc., and the interaction should not exceed a two-level menu. Artificial intelligence native Applications should be able to solve problems that could not or were not solved well in the past, rather than simply repeating mobile Internet applications or computer software.
"The model itself does not directly generate value. The applications developed based on the basic large model are the meaning of the model's existence." Robin Li said that in the mobile Internet era, the only operating systems are Android and iOS, but there are many particularly successful applications. In the era of large models, similar applications will also appear. situation - the basic large model is the operating system, and the new functions implemented based on the large model are "native applications" in the era of artificial intelligence.
In the new pattern of artificial intelligence, large model capabilities are the foundation, which directly determines the upper limit of large model applications. However, large models alone cannot form a prosperous ecosystem, and applications based on large models are also key.
Previously, China's large models were mainly based on model development, and the application layer was often ignored. But if everyone only focuses on model development and no one develops applications, it will be like a car without wheels. Overseas, the large model application layer has already begun to develop. In China, Baidu has made a head start.
"We hope that entrepreneurs can build popular applications in the AI era based on Wenxin's large model." Robin Li said. Baidu's entrepreneurial competition that started in the large-model testing phase has now come to an end, but there is no sign that Baidu's efforts to cultivate a large-model entrepreneurial ecosystem will stop.
15 winning apps cover multiple segmentation directions
Document-based Q&A assistant, design creative assistance, medical content generation, new material discovery, two-dimensional content creation, sales staff training... The projects of the winning team of Baidu Big Model Entrepreneurship Competition cover multiple subdivisions. Together, they demonstrate many characteristics of large-model application entrepreneurship.
First, entrepreneurs working on large model applications usually have a deep understanding of the scenarios they want to change. It is precisely because of an in-depth understanding of specific application scenarios that entrepreneurs can better understand the problems faced by existing solutions in these scenarios, and are therefore more likely to use large models to develop better solutions.
The ChatDOC company that won the first prize in this competition, Paoding Technology, was founded in 2017. In the six years since its establishment, it has been developing products related to processing financial documents, such as writing prospectuses based on a large amount of information provided by customers. For the ChatPPT project that also won the award, the person in charge, Zhou Ze'an, has 10 years of experience in PPT engine and function development. His previous entrepreneurial project, Pocket Animation, was acquired by WPS.
Similar to large model startups overseas that continue to gain customers and investments. For example, Jasper AI, a large model application company with a valuation of 1.5 billion, mainly uses large model capabilities to provide marketing support to customers. Its founding team has a strong marketing background. In November last year, Speak, a language learning application that received $27 million in investment from OpenAI, was created and sold by its founder in high school to assist learning and memory.
Secondly, the entrepreneur's strategy is not to build a new solution based on a large model, but to use the large model to optimize a specific link of the original solution. In this optimization process, they maximized the unique advantages of the large model while circumventing its remaining weaknesses.
"Our core lies not in the large language model, but in how to make an artificial intelligence competent to play the role of a human shopping guide." Buysmart.AI founder Chen Lifei said that the fundamental lies in how to make it correctly understand the user's problems and then recommend the products they strive for. . Buysmart.AI is another project that won first prize in the Baidu Big Model Entrepreneurship Competition. The approach they take is to combine large models with recommendation algorithms to leverage their ability to understand and process large amounts of information.
Lin Demiao, CEO of ChatDOC, said that similar smart document products on the market often fail to answer questions and generate answers by themselves. Therefore, they will limit the ability to generate large models in ChatDOC, requiring that every answer it gives must quote the original text. If the appropriate original text cannot be found, it will feedback "not found" instead of answering randomly.
In Magic Technology, which uses large models to assist in the discovery of new materials, the greatest value of large models now is to assist in the construction of a computing experiment platform that can use advanced algorithms and experiments without mastering the code. "The large language model reduces the cost of using software or operating each instrument to a certain extent, and can realize specific operations directly through voice." said Liu Yuyang, founder and CEO of Huanliang Technology.
Third, for large model applications, although it is not difficult to switch the underlying large model, dependence may occur if the same large model is used for a long time. Many of the teams participating in the Baidu Large Model Entrepreneurship Competition this time developed applications overseas based on ChatGPT in the early days. However, their domestic business has now switched to Wen Xinyiyan.
"It's not as difficult as expected to replace an overseas model with a domestic one. After we changed it, we didn't need to make too many changes, and the whole process was up and running." said Chen Lifei, founder of Buysmart.AI.
Peng Kangwei, CEO of Genie AI, which uses large models to assist in the creation of two-dimensional content, also has similar feelings. He switched from ChatGPT to Wen Xinyiyan, "The fine-tuning of the model and the expression of some keywords are not much different. From safety and compatibility From a Chinese perspective, Wen Xin Yiyan would be better."
However, in her opinion, if you use a large model, such as ChatGPT, for a long time, if you want to fully utilize its capabilities, you must set up the product architecture and build code around it. "Switching models over time is costly."
This is an often overlooked aspect of large model competition. Developers of large model applications may become dependent on a specific large model, which means that large model suppliers that enter the market earlier and attract entrepreneurs earlier will have greater advantages.
Large models gradually enter the era of AI native applications
Before the large model policy was implemented, most companies were quite cautious in promoting the application of large models. Their products for individual users are usually in the internal testing or invitation testing stage. Ordinary users cannot directly register or use them, and companies do not actively place advertisements to promote large-model products. These factors limit the spread of the products.
The new environment created by the implementation of the policy has transformed the competition of large models into a contest of comprehensive capabilities: the key to success is no longer just a company’s technical strength in training large models, but also its ability to understand market needs, develop matching applications, and excel. operational capabilities.
This is a test for every startup developing large-model applications. For companies that develop basic large models, the test is also directly related to their ability to build an ecosystem. This may be a direct reflection of their competitiveness.
Robin Li believes that “only the best large models can produce the best native applications of artificial intelligence.” He said that Baidu will soon launch version 4.0 of Wenxin Large Model, with the goal of "Baidu's goal is to build the basic capabilities of large models and support the development of native artificial intelligence applications."
Baidu’s investment in large model ecology is also continuing. It is understood that in addition to the "Wenxin Cup" Entrepreneurship Competition, Baidu has also launched the AI Studio Galaxy large model community, plug-in mechanism and Wenxin Large Model "Galaxy" co-creation plan for developers to attract more people to join Baidu's large model ecosystem .
According to data disclosed by Baidu, there are currently nearly 10,000 companies active on the Baidu Intelligent Cloud Qianfan large model platform every month, covering more than 400 business scenarios in finance, manufacturing, energy, government affairs, transportation and other industries.
At the end of August, Baidu’s Wenxinyiyan was officially opened to the public after being registered with relevant departments. Baidu said that Wenxinyiyan answered 33.42 million questions from netizens on the first day it was opened. "A large amount of real human feedback helps Baidu improve the basic model quickly and efficiently."
"I believe that the Wenxin Big Model will become the first choice for AI entrepreneurs and developers. More and more applications will be built on the model, and the entire ecosystem will be full of vitality." Robin Li said.