In early 2023, it emerged as the brightest star in the field of science and technology, demonstrating the huge potential of Kai AI technology to transform the economy and society. After more than a year of technological breakthroughs and market verification, entering 2024, the huge value of large models has been recognized by mainstream countries, and mainstream technology companies have made plans. Countries such as the United States and China are in the future of leading the development of the large model industry.

Among domestic technology companies, Baidu, which was the first to release Wenxin’s large model, has achieved extraordinary results. At the Create 2024 Baidu AI Developer Conference held today, Robin Li, founder, chairman and CEO of Baidu, revealed in his keynote speech that the number of Wenxin Yiyan users exceeded 200 million, and the Wenxin Big Model has become China's leading application. The most extensive AI basic model.

At the same time, Robin Li also plans to turn all big model users into big model developers. "In the future, natural language will become a new universal programming language. As long as you can speak, you can become a developer and use your creativity to change the world." Robin Li said passionately at the meeting.

One billion users means 1 billion AI developers

According to our understanding, Wenxin Yiyu exceeded 100 million users at the end of last year, ranking first in the country. It exceeded 200 million in less than half a year, achieving doubling growth, showing that the penetration rate of large models is accelerating. Robin Li also revealed a set of data: the average daily API calls of Wenxinyiyu exceeded 200 million, the number of service customers reached 85,000, and the number of AI native applications developed using the Qianbi platform exceeded 19 million.

"Compared with a year ago, the algorithm training efficiency of the Wenxin large model has increased to 5.1 times, the average weekly training efficiency has reached 98.8%, the inference performance has increased by more than 105 times, and the cost of inference has dropped to The original 1%. In other words, customers used to adjust parameters 10,000 times a day, but now they can adjust parameters 1 million times a day at the same cost.” It is a leader in the development of the domestic large model industry. At the same time, Robin Li has been calling on the industry to create a healthy large-scale industrial ecosystem: vigorously develop AI native applications, create AI super applications that exceed imagination, and achieve industry leadership in the next era.

Robin Li believes that large models and generative AI will completely change the developer group. In the past, developers used code to change the world; in the future, natural language will become the new universal programming language. As long as you can speak, you can become a developer and use your creativity to change the world.

"Today's China has more than 1 billion Internet users, strong infrastructure, enough AI application scenarios, and the most complete industrial system in the world. The country is also vigorously encouraging and supporting 'artificial intelligence' 'Action. Every individual and every company can unleash unlimited creativity and productivity by making full use of these tools." Robin Li pointed out, "Everyone can become a developer!"

In the past ten years. For years, video production and distribution has been dominated by agencies. Nowadays, with the rise of short video super APPs such as Douyin and Kuaishou, everyone can shoot short videos and become a video creator, which has generated significant commercial value. In Robin Li's view, AI is setting off a revolution in creativity, and developing applications in the future will be as easy as shooting a short video.

Three major AI development artifacts unveiled

In order to help startups and individuals become developers, Baidu has launched three AI development artifacts that can be used "out of the box": the agent development tool AgentBuilder , AI native application development tool AppBuilder, and model customization tool ModelBuilder of various sizes.

Li Yanhong pointed out that there are three main points in developing AI native applications based on large models: First, MoE. In the future, large-scale AI native applications will basically be MoE, which means a mix of large and small models, without relying on one model to solve all problems; the second is small models. Small models have low reasoning costs and fast response times. In some specific scenarios, the fine-tuned small models can be used as effectively as large models; the third is the intelligent agent. The intelligent agent mechanism includes understanding, planning, reflection, and evolution. It allows machines to think and act like humans. They can complete complex tasks autonomously, continue to learn in the environment, and realize self-iteration and evolution. As the capabilities of intelligent agents improve, a large number of new applications will continue to emerge.

"This is the result of Baidu stepping on countless pitfalls and paying high tuition fees based on the past year's practice." Robin Li said that Baidu has released three lightweight models: Speed, Lite, and Tiny. Wenxin’s intelligent agent platform has been upgraded. So far, more than 30,000 agents have been created, more than 50,000 developers and tens of thousands of companies have settled in.

Based on the above understanding, the three major AI development tools launched by Baidu will help developers improve the development efficiency of AI native applications and accelerate the application popularization of large models.

Among them, based on AgentBuilder, merchants do not need to program. By inputting information similar to prompt words and adjusting a few simple steps, they can quickly generate an exclusive agent and become a "gold medal salesman" online 24/7. At the conference, Robin Li demonstrated three smart agent cases from the Singapore Tourism Board, Kai Tak Education and Sophia, teaching developers step-by-step how to use natural language to create an intelligent agent in 5 minutes with zero threshold.

AppBuilder encapsulates and presets various components and frameworks required to develop AI native applications in advance, greatly reducing the development threshold. Robin Li introduced that in just three steps at the fastest, developers can develop an AI native application using natural language, and it can be easily released and integrated into various business environments. In addition, in addition to integrating Baidu's own rich AI components, it also supports open source SDK to facilitate secondary development.

At the conference site, Robin Li demonstrated the creation process of an AI native application through three cases of "Amusement Park Queuing Assistant", North China Electric Power University's "Huadian AI Assistant" and Baidu Wenku Intelligent Comic Generation. You can create an AI native application in three simple steps: setting the name, filling in role instructions, and inserting components.

ModelBuilder can customize models of any size according to developer needs, and further fine-tune SFT on the model according to subdivided scenarios. Robin Li showed a case of essay correction in the education industry. After data processing and fine-tuning of the model, the "Essay Correction Assistant" can not only have a more professional teacher's comment thinking and follow the format, but also be better than the fine-tuned model. The final model score is closer to the real teacher score. ModelBuilder also supports domestic and foreign third-party mainstream models, and is the development platform with the largest number of large models in China.

By releasing three AI development tools, Baidu once again showed the industry its determination to continue to build a strong model industry ecosystem. It can be believed that the creativity inspired by "everyone is a developer" will accelerate the arrival of the AI ??era, help developers win at the starting line, and seize the opportunities in the AI ??industry that transcend the mobile Internet era.