Click to follow
Text｜Hao Xin Huang Xiaoyi, editor｜Liu Yuqi
Wudaokou, the "center of the universe," has regained its former glory.
In the Huaqing Garden at the corner of the crossroads, the entrepreneurial legends of various Internet tycoons, such as Wang Xing, Cheng Yixiao, Zhang Yiming, etc. seem to still linger, and the AI entrepreneurs who have moved in have continued a new chapter.
On an ordinary day in 2021, several young people from northern Qing Dynasty named Xu Zuobiao and He Yujian moved into a slightly cramped two-bedroom and one-bedroom apartment in Huaqing Jiayuan. They looked at the Tsinghua Science and Technology Park, a mecca for startup company incubation across the street, and began to Exploring AI and creative entrepreneurship. At that time, they did not expect that there would be a boom in large-scale models and AI applications, bringing madness beyond the Internet era.
However, unlike the chaos of the Internet era, today's entrepreneurs are the darlings of the new era, receiving the most attention and growing rapidly.
Take Nolibox as an example. In 2021, from just a vague idea to the product taking shape, Xu Zuobiao and others took the demo to the Baidu "Wenxin Cup" Entrepreneurship Competition, directly competed, and faced competitors and investors.
Entrepreneurs like them are coming.
On May 31, after the "Wenxin Cup" Entrepreneurship Competition was launched, in less than a month, it attracted 1,000 project registrations, of which more than 80% focused on the implementation of large model application layers, focusing on the direction of AI native applications.
The roadshow was crowded with people, and for a moment it was like traveling to the Silicon Valley AI project roadshow. On September 5, China’s first batch of AI native applications were unveiled in medical, second-dimensional communities, e-commerce, law, and finance.
"There is no point in rolling large models, and the application opportunities are greater." Robin Li, founder, chairman and CEO of Baidu, pointed out the new direction of current large model entrepreneurs.
Robin Li said, "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."
Today, large models have gone through a complete path of bottom model competition, middle layer capability enhancement, and model application deployment. The explosion of large model applications is just around the corner. Regardless of technology or market, the whole world is waiting for an AI-native “super APP” to be born.
01 The post-90s generation flooded into the AI entrepreneurship circle
The pool of AI entrepreneurship is already filled with people born in the 1990s.
Instant Noodle Park, Hello Tina, Mai Xiaoji, Nolibox, Genie AI subculture interest creation community... names that seem random but have their own identity are scattered in the Baidu Wenxin Cup venue, "Competition" "The scene" instantly transformed into a "young people's gathering".
According to Baidu, half of the founders of the 15 finalists were born in the 1990s, and the youngest was only 25 years old.
It is the post-90s generation who are at the forefront of the times. Living in the Internet era, they have no shortage of geek thinking and creative energy, and are closely aligned with the AI era.
Everyone is saying that AI will reconstruct everything, but this does not mean simply redoing past applications. Instead, we need to break through the old framework of the Internet, use more divergent thinking, explore new user needs, and create new information dissemination and interaction models.
However, being young does not mean that the threshold is lowered. AI entrepreneurship has a strong elitist trend.
At the scene, we can see a large number of top talents at home and abroad competing on the same stage, including star product managers from major domestic and overseas manufacturers, senior serial entrepreneurs, and students from Stanford University, Oxford University, Cornell University, Tsinghua University, Hong Kong Science and Technology Research teams with professors and Ph.D.s from prestigious universities and other universities, as well as senior figures from traditional industries such as medical care, finance, and law.
After further listening to various entrepreneurial demos, Guangcone Intelligence found that unlike domestic large-model entrepreneurial teams, which often have dozens of people, AI native entrepreneurial teams mostly adopt the "small but refined" entrepreneurial model of foreign Silicon Valley. Early entrepreneurial project teams The number of people is limited to 3-5 people.
This is because most of the current AI startups are in their early stages, and people are competing more on the implementation of ideas and scenarios rather than on expanding the scale of technology. In this regard, small teams not only have agile and flexible creativity, but also reduce labor costs.
According to Qichacha, among the participating teams in the "Wenxin Cup" Entrepreneurship Competition, many companies, including the Null Hypothesis Company, are small-scale. Therefore, when the number of people, capital investment, and resource investment are limited, participating in entrepreneurial competitions has become a "showcase" for young entrepreneurial companies.
The 15 teams competing in the finals all came well prepared.
According to Guangcone Intelligence, on the one hand, its products are relatively mature and have been tested by the market for users. For example, Buysmart.AI has gained thousands of users overseas since it was launched in early May. During the use of users, Continuously verify the product delivery model and adjust iterations. "The preliminary verification is not bad. Some people are willing to buy and are willing to come back and continue buying after buying." Buysmart.AI founder Chen Lifei told Guangcone Intelligence.
"Lingao Technology products were launched in March 2023. So far, it has been for five consecutive months, with user growth exceeding 100% every month. Users have also become diversified, covering startups, traditional companies and Internet companies, etc.", Lingao Li Qipeng, CEO of Ao Technology, talked to Guangcone Intelligence.
On the other hand, it has already received "Offers" in the capital market. For example, Nolibox has received investment from Chuxin Capital and Hillhouse Venture Capital; Null Hypothesis has received investment from Sequoia Seed, Yuanhe Origin, and Weikuai Venture Capital.
All in all, "sense of contradiction" is the key word for this group of AI-native entrepreneurs. The post-90s entrepreneurship, the company that is still in its infancy, the fearless drive and experience in the industry, the thoughtful commercial exploration, the down-to-earth product operation, etc. are all formed A sharp contrast. Looking up at the stars while walking on the road may be another footnote to the era of large-scale entrepreneurship.
02 The first batch of AI native applications are coming to light
How to define AI native applications? Robin Li believes that three basic conditions should be met: first, to be able to interact with natural language; second, to make full use of the understanding, generation, reasoning, memory and other capabilities of large language models; third, the interaction of each application should not exceed a two-level menu, making the interaction Become simpler.
By embedding large model capabilities into applications and changing interaction methods, a number of AI-native entrepreneurial projects emerged in the Baidu "Wenxin Cup" competition. Through these projects, we can get a glimpse of the current implementation of AI-native applications in China. latest progress.
Lightcone Intelligence selected the AI native projects of the 15 winning teams in this competition. After sorting out, it was found that the implementation of large-scale model applications has almost kept pace with foreign countries. This batch of AI native entrepreneurial projects largely overlaps with the popular entrepreneurial directions in Silicon Valley. The B-side , C-side applications are blooming everywhere, and marketing, medicine, law, AI assistants, etc. are even more popular at home and abroad.
Although the fields are common, when implemented in China, AI native applications have grown different fruits. Combined with the legal industry, an AI judicial mediation platform that helps people collect debts was born - Tianping Zhixing; combined with the two-dimensional dimension, a creative community with subcultural interests - Genie AI was born; combined with short videos, it was implemented using prompt generation Morph Studio for GIF animations and short videos.
Robin Li pointed out, "AI native applications must be able to solve problems that could not be solved or solved poorly in the past, rather than simply duplicating mobile Internet APPs and PC software."
Based on this, during the development process of AI native applications, Chinese entrepreneurs have further divided the use of models and technologies in different segmented scenarios.
Take Nolibox and Genie AI, which also use the SD Vincent graph model, as examples. Nolibox mainly targets e-commerce marketing scenarios, providing users with marketing material generation tools from creative design to typesetting combination; Genie AI mainly targets two-dimensional second-generation scenarios, generating comic works based on keywords based on large models, or one-click based on real-life videos Generate animation videos.
It is precisely based on the different scenarios and user demands that the two Wenshengtu capabilities are used in different directions. The focus of the application of Nolibox's graphic capabilities is on the generation process, that is, how to better achieve the effects desired by the user; while the focus of the application of Genie AI's graphic capabilities is on prompts, that is, how the model encoder converts and aligns the instructions after the user inputs the created story.
In this way, "generative AI capabilities" achieve different fission in different scenarios.
Go beyond the general scenarios and re-search for the subdivided implementation scenarios of some basic technologies such as "Wen Sheng Tu" and "Wen Sheng Video". Use large models and open source technologies as R&D capability libraries, and use them immediately. "Large models + science", "big models" "Model + Medical" and "Large Model + Finance", after being combined with the industry, various applications have grown on the same trunk.
In the final analysis, at this stage, China's native AI entrepreneurs have grasped the pulse of "segmented scenarios", which is why there are not a large number of homogeneous projects among the 1,000 projects.
In the roadshows we have participated in this year, Guangcone Intelligence has witnessed with our own eyes the transformation of AI native application entrepreneurial teams from PPT-style presentations to commercialization. Overall, China’s wave of AI native applications is not slow.
03 From large models to applications, the ecological war has ushered in a new starting point
"In the mobile Internet era, the only operating systems are Android and iOS, but there are many particularly successful applications. Only by generating enough AI native applications based on large models can we have a healthy ecological environment," said Robin Li.
As a latecomer, the application launch can catch up in a short period of time, which is also due to China's large model ecology.
"Wenxin Big Model has done a very good job in terms of overall compliance. It meets the current national requirements for generative AI, so you can feel more at ease when using it." "From a security perspective and Chinese compatibility perspective, it is Wenxin will be better." Nolibox CEO Xu Zuobiao, Genie AI CEO Peng Kangwei and other entrepreneurs expressed their reasons for choosing Baidu.
Overseas large models are chosen for their performance, while domestic ones pay more attention to safety and compliance. This has also become the primary consideration for Chinese companies when choosing large models.
Many AI application entrepreneurs report that "using ChatGPT for English and Baidu Wenxin Big Model for Chinese" is becoming the norm.
ChatDOC CEO Lin Demiao believes that Baidu has given enterprises more opportunities to understand and use the Wenxin large model. “Before the big model came, we could only do document intelligence in specific industries, because once the circle is expanded, the cost of model training will increase. Very high. But now Baidu helps us do the work of training large models, so as long as there are application scenarios and needs, combined with our technology, we can quickly rush to the front."
It is reported that the shortlisted players of the Demo development stage of the participating teams can carry out the demo development on the Baidu Intelligent Cloud Qianfan large model platform, and can call the Wenxin large model API and Baidu Intelligent Cloud computing power resources. Many entrepreneurial teams said, "It was the first time I came into contact with a large domestic model. I originally had a trying mentality, but unexpectedly found that the gap between domestic and foreign countries was not as big as imagined."
"In the future, I think China will have its own ecosystem, with its own large language model and basic model. There will be at least one, possibly two or three basic large models in China that can support people in developing various AI native applications. ." Robin Li hopes that the Wenxin model can become one of the default basic models for AI developers like ChatGPT.
In addition, Robin Li also revealed that version 4.0 of Wenxin Large Model will be launched soon. "Baidu's goal is to build the basic capabilities of large models and support the development of native AI applications."
The second is to improve its own diversity, which is reflected in the diversity of tools and the diversity of ecological partners.
The newly upgraded Baidu Intelligent Cloud Qianfan large model platform is equivalent to a one-stop large model toolbox that integrates multiple models and services. Enterprises can independently select models and data sets on the platform and combine them with their own business data, and ultimately quickly applied to business flow.
"In the past, it took customers at least a week to verify the effect of large models. After the upgrade, based on the one-stop tool chain, it only takes one day to run through a model and see the effect." Baidu Group Executive Vice President and President of Baidu Intelligent Cloud Business Group Shen Dou expressed.
In terms of ecological partners, there are currently more than 500 ecological partners using various large model services through the Baidu Smart Cloud Qianfan large model platform. The number of monthly active enterprises on the Qianfan large model platform exceeds 10,000. Like the contestants this time, it covers both 2B companies and 2C companies. The number is not only large, but also penetrates into all walks of life.
Finally, there is the division of ecological roles with clear responsibilities. A healthy large model ecosystem should have both upstream and downstream members of the industry chain performing their respective duties. Baidu has positioned itself as a "water seller" early on, opening up large model capabilities and providing development tools, while also maintaining a win-win cooperative relationship with enterprises.
"We used a relatively low price in the industry to use Baidu Wenxin Wenshengtu, Wenshengwen and other capabilities. Baidu not only gave the company a big discount on the price, but also set aside some time for fine-tuning and testing. At present, We have some commercial cooperation, and we hope to add some new implementation scenarios and cooperation directions in the future," Nolibox CEO Xu Zuobiao said.
Li Qipeng also revealed to Light Cone Intelligence that he has reached cooperation with Baidu in terms of technology, scene implementation, customer development and other aspects.
"Technically, when our engineers encounter difficulties in model parameter adjustment and API use, Baidu will respond immediately; when the scene is implemented, we and Baidu's architects will complement each other and create together to jointly serve user scenarios; customers Expansion is a long-term strategy that we hope to develop with Baidu Wenxin, accumulate customer cases in various industries, and form a solution template."
"Baidu's emphasis on this entrepreneurial competition exceeded my expectations. Robin Li personally came to the finals. We originally participated with the mentality of giving it a try, but we didn't expect to actually get the bonus. It seems that Baidu is really serious this time. I hope to develop a large-scale model ecology," Li Qipeng lamented.
Robin Li said that Baidu will invest tens of millions of yuan in the above-mentioned team and provide long-term support in terms of technology, products, development strategies, capital cooperation, etc. "I believe that 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."
The pace of Baidu ecosystem construction is still accelerating. After Baidu Wenxinyiyan APP was announced to be open to the whole society, it quickly topped the Apple App Store free rankings within 12 hours, becoming the first Chinese to top the App Store top list. AI native applications.
The enthusiasm of ordinary users for Wen Xinyiyan APP has made more and more entrepreneurs and developers deeply aware of the broad development prospects of generative AI in this wave of natural language processing technology.
At this point, China's large model ecological measurement standards have gradually become clear: safety and compliance, proactive openness, clear rights and responsibilities, and diversity. From large model technology competition to the implementation of AI native applications, the battle for large model ecology has ushered in a new starting point.
#baidu##ai#
Welcome to follow Light Cone Intelligence and get more cutting-edge knowledge of science and technology!