Mature "plaid shirts" and young "faces", these two elements are always indispensable in developer-related conferences, and Create 2024 Baidu AI Developer Conference is no exception.
In the past few decades, the emergence of new technologies such as programming languages and frameworks has shaped generations of young faces into mature plaid shirts, and the efficiency of software application development has indeed improved, but it has always been unable to keep up with the huge and complex demand, and it seems that developers will always be stuck in such a cycle.
Until the advent of large models.
"As long as you can talk, you can be a developer and change the world with your creativity. Robin Li, the founder, chairman and CEO of Baidu, was convincing.
Loud and sound. The mature plaid shirt is still thinking, the words are somewhat credible and the problems in between, and the young face is already convinced, eager to hear how to "speak" in order to become a developer in the new era.
The basic large model is indeed dazzling, and most people are still immersed in the basic model level, but when it is time to focus more on the application level, first new technologies such as large models provide new possibilities for application scenarios, and then the prosperity of the software application ecology to promote the promotion of large models again.
What Baidu is doing is to clear the barriers between large models and applications, which are also the three major tools that Baidu came up with at this developer conference - agent development tool AgentBuilder, AI native application development tool AppBuilder, and model customization tool ModelBuilder of various sizes.
Big model, application is king
"Large language models do not directly create value on their own, and AI applications developed based on large models can meet real market needs. Robin Li said.
As Robin Li said, in the past year, the large model is enough to attract people's attention, such as the evolution of Baidu's Wenxin model, from version 3.0 to version 3.5, and then to version 4.0, in terms of code generation, code interpretation, code optimization and other general capabilities, reaching the international leading level.
The Wenxin model has become the most advanced and widely used AI basic model in China. Not only that, compared with a year ago, the algorithm training efficiency of Wenxin's large model has increased to 5.1 times, the average weekly training efficiency has reached 98.8%, the inference performance has been improved by 105 times, and the cost of inference has been reduced to 1%.
However, large models alone are not enough to give birth to new business formats, just as the Internet was only a technical concept at first, and Internet applications have been integrated into the scenes of thousands of industries, creating the Internet giants that are familiar today. "They" did not invent Internet technology, nor did they make a fuss about the underlying infrastructure of the Internet, but they are the best representatives of the Internet.
The large model that Baidu provides for developers is like the Internet in the past, and Baidu has the motivation and necessity to provide the best possible large model, so as to lay a good foundation for large model application developers.
It is conceivable that Baidu is on the front line of the large model, and it is inevitable to step on the pit. Robin Li also bluntly said that Baidu "stepped on countless pitfalls and paid high tuition", and came up with some specific ideas and tools for developing AI native applications based on large models, which were directly shared with the industry.
The first is MoE. In the future, large-scale AI native applications will basically be MoE architecture, and the MoE mentioned here is not a general academic concept, but a mixture of large and small models, and does not rely on one model to solve all problems. However, when to call a small model, when to call a large model, and when not to call a model should be matched for different scenarios of the application.
The second is the small model. The inference cost of small models is low, the response speed is fast, and in some specific scenarios, the use effect of small models fine-tuned by SFT can be comparable to that of large models.
"That's why we released the Speed, Lite, and Tiny lightweight models. We use large models, compress and distill a basic model, and then use data to train it, which is much better than training a small model from scratch, and has a better effect, faster speed, and lower cost than the model trained based on open source models. Robin Li said.
The third is the agent. Agents are a hot topic at the moment, and with the improvement of their capabilities, a large number of new applications will continue to be born.
The agent mechanism, which includes understanding, planning, reflection, and evolution, allows machines to think and act like humans, to complete complex tasks autonomously, and to continuously learn, iterate, and evolve in the environment. In some complex systems, different agents can also interact with each other and cooperate with each other to complete tasks with higher quality. These agent capabilities, the ability to reflect, plan, and self-plan, have been developed and fully opened up to developers.
Three "artifacts" of AI application development
Robin Li said that large models and generative AI will revolutionize the developer community. In the past, developers used code to change the world, and in the future, natural language will become the new ubiquitous programming language.
Correspondingly, the developer's arsenal also needs to keep pace with the times to adapt to the software development system of the new era, and at the same time, Baidu has done a good job of "out-of-the-box" tools for the three major directions of MoE, small models, and agents.
AgentBuilder: The most popular way to use large models
The first is AgentBuilder, an agent development tool. Based on a powerful basic model, agents can be generated in batches and applied in a variety of scenarios.
Taking the Singapore Tourism Board as an example, open the Wenxin intelligent twin platform, create a page with zero code, low code two modes, novices can directly choose the "zero code mode", with natural language, a few sentences can create an intelligent body, to achieve hotel inquiry, scenic spot ticket purchase and other service capabilities, Baidu has cooperated with Ctrip, providing hotels, attractions, ticketing and other tourism service tools.
In addition to Singapore, cultural and tourism agents such as Dalian and Shenyang are also online on the Wenxin Agent platform, as well as various agents such as knowledge, creation, learning, and entertainment.
At present, more than 30,000 agents have been created, more than 50,000 developers and tens of thousands of enterprises have settled in. The Wenxin Intelligent Twin platform also provides developers with a channel to monetize traffic. In addition to Baidu search, other products in the Baidu ecosystem, such as Xiaodu, map, post bar, car machine, etc., can access the relevant capabilities of the agent, solve the worries of traffic distribution for developers, and obtain real benefits.
With distribution, there will be data feedback, and with data feedback, the flywheel will spin, and the agent will be able to iterate autonomously, and the more it is used, the smarter it becomes. Wenxin Intelligent Twins platform has also launched the data analysis and Q&A tuning module of the agent, and more new capabilities will be launched in the near future. The Wenxin Intelligent Twins platform will drive the agents to form a positive cycle of better quality, better traffic, and greater benefits through the data flywheel of distribution, diagnosis, and revenue.
In addition to Singapore, cultural and tourism agents such as Dalian and Shenyang are also online on the Wenxin Agent platform, as well as various agents such as knowledge, creation, learning, and entertainment, which can be made in AgentBuilder.
AppBuilder: You can develop an app in three natural language steps
AppBuilder encapsulates and presets various components and frameworks required for the development of AI-native applications in advance, which can greatly reduce the development threshold, without writing a single line of code, and developers can develop an AI-native application in natural language in as little as three steps, and can easily publish and integrate into a variety of business environments.
Take the playground queuing assistant as an example, open the development interface of AppBuilder, the first step is to name the application "playground queuing assistant", the second step is to fill in the role instructions, including calling the code interpreter, calculating the best combination within a fixed time, and the output result, etc., and then insert the required tool components in the third step. In this way, with zero code, an application is generated.
Since last year, Baidu has reconstructed Baidu Wenku with AI, making it the "starting point of content production" for users. Now, with the support of AppBuilder, Baidu Wenku's newly launched smart comics and smart picture books have extended the scene to a more interesting cross-modal creation field.
On AppBuilder, Baidu also provides certain cross-modal capabilities, developers only need to give a piece of text, or a few sentences, you can quickly create painting applications, such as comics, children's picture books, etc. Baidu Library's newly launched manga generation and picture book generation features take advantage of these components available on AppBuilder.
From an industry perspective, Baidu AppBuilder has two obvious advantages:
First, it is powerful. Relying on Wenxin 4.0's ability to understand and follow instructions, Baidu AppBuilder can ensure that it can reach a good level in the cold start state, and will not take a long time to optimize because of poor results, which greatly reduces the development threshold.
Relying on the retrieval enhancement technology RAG, the accuracy rate of Q&A and the degree of friendly reply have reached more than 95% in typical scenarios such as knowledge Q&A, which greatly surpasses other similar products. AppBuilder also provides a wealth of complete component tools, including 55 components such as Baidu search, which are based on Baidu's years of technology accumulation, large model capabilities, and Baidu's exclusive business components.
In addition, Baidu also provides third-party APIs for some mainstream scenarios, such as flight queries, paper queries, etc., and the latest support for custom components allows customers to directly connect with any of their own proprietary tools and data. These rich components together support the efficient development of AI-native applications.
Second, it is simple and easy to use. With AppBuilder, you can quickly create and distribute apps in just three steps. Baidu also supports open-source SDKs, which is convenient for everyone to carry out secondary development.
ModelBuilder: Produces models efficiently and at low cost
ModelBuilder is a tool that is more suitable for professional developers, which can customize models of any size according to the needs of developers, and further fine-tune the SFT of the model according to the subdivision scenario, so as to achieve better results.
In order to make it easier for professional developers to get started quickly, ModelBuilder presets the most comprehensive and abundant large models. It includes ERNIE3.5 and ERNIE4.0 flagship large models, which are suitable for general complex scenarios and have powerful capabilities;
In addition, there are three large lightweight models, Speed, Lite, and Tiny, as well as two vertical models – ERNIE Character for role-playing, and ERNIE Functions for external tool use and business function calls in dialogue or Q&A scenarios. Of course, it also includes third-party mainstream models at home and abroad, with a total of 77 models, and ModelBuilder is the development platform with the largest number of large models in China.
Take Xiaodu Tiantian's AI robot scheduling as an example,Different models are called behind it。 First of all, the small model ERNIE Tiny performed the work of "model routing" - the weather problems in the morning run were assigned to the fine-tuned model based on ERNIE Lite, and the information of 25 degrees of temperature and sunny weather was quickly queried, and at the same time, the more complex schedule was assigned to the best large model - Wenxin 4.0 to calculate the schedule of various events of the day.
The key to ModelBuilder is the ability to produce models efficiently and at a low cost. Enterprise customers can tailor smaller models suitable for various scenarios according to their needs, taking into account various considerations such as effect, response speed, and inference cost, and support fine tuning and post-pretrain.
Compared with the model directly retrieved from open sources, the effect is significantly better under the same size, and the cost is significantly lower under the same effect.
When AI applications prosper, large models can evolve again
The prosperity of AI applications can not only prove the value of large models, but also promote the re-evolution of large models.
When a large number of AI applications are created, more and more data is generated and collected, and models have more opportunities to learn application knowledge in different scenarios, the popularization of dedicated hardware and performance improvement brought about by scale, making it more feasible and efficient to train and run large models, and diversified application scenarios bring positive business benefits, and business success will be transformed into resources for AI applications and large models.
Baidu Wenxin model is embarking on this road, using the mutual promotion of large models and AI applications to achieve a positive business cycle of large model ecology.
Robin Li said that the Wenxin model has become the most advanced and widely used AI basic model in China. Baidu has a full-stack layout in the four-layer architecture of chips, frameworks, models, and applications, and continuously reduces costs through end-to-end optimization, so that more people can use large models to do AI applications efficiently and cheaply, and continuously improves the efficiency of Wenxin large models and Wenxin Yiyan through end-to-end optimization capabilities.
"Actual combat is the best training ground for large models, and the huge call volume will give more feedback to the Wenxin large model, which in turn promotes the ability improvement of the Wenxin large model, forms a flywheel effect, and further expands the gap between the Wenxin large model and its domestic counterparts. He said.
Wenxin Yiyan was first launched on March 16 last year, and in one year and one month, the number of users has exceeded 200 million, the number of API calls per day has also exceeded 200 million, the number of customers or enterprises served has reached 85,000, and the number of AI native applications developed using the Qianfan platform has exceeded 190,000.
Baidu has also released the tool version of Wenxin Model 4.0, where users can experience the code interpreter function on the tool version, and through natural language interaction, they can process and analyze complex data and files, and can also generate charts or files, which can quickly gain insight into the characteristics of the data, analyze the change trend, and provide efficient and accurate support for subsequent decision-making.
In addition, the entrepreneur ecosystem is also a vital force of the large model. In May last year, Baidu launched the "Wenxin Cup" Entrepreneurship Competition, the first "Wenxin Cup" Entrepreneurship Competition, received nearly 1,000 entrepreneurial team registrations, and Baidu provided nearly 100 million yuan of investment support for 15 of the winning teams, and continued to provide all-round support in technology, team and resources.
Robin Li announced at the scene that the second "Wenxin Cup" Entrepreneurship Competition was officially launched, and this time it will expand the scope of project screening, set up sub-venues, and recruit entrepreneurial and innovative teams for the global market and college students.
At the same time, Baidu has also increased its support for entrepreneurs, providing more adequate investment funds and richer business resources, and has also set up a "special award" for the first time, where particularly outstanding projects will have the opportunity to receive up to 50 million yuan in cash and resource support.
Whether it is a mature "plaid shirt" or a young face, they are all part of China's large model industry, and Baidu has developed three "artifacts" through AI applications, showing a road to the future of the large model industry. On this road, there is Baidu, and there are thousands of developer groups.