“The real value in the era of large models lies in AI native applications”
Following the launch of Tencent Cloud's high-performance application service HAI and the creation of AIGC applications in 10 minutes, Baidu Cloud also released the AI cloud application platform AppBuilder at the Intelligent Computing Conference yesterday, which is known as zero-threshold and minute-level creation of intelligent agents. The AI track is so lively. The editor will take you to test it today.
In the second half of the "First Year of Big Models", the topic focus has shown a tendency to shift from technology-oriented basic models to application keywords such as "agent".
It can also be seen from the latest patterns rolled out by major model manufacturers——
Just now, Baidu Intelligent Cloud has taken the lead in showing its hand: creating an "interviewer" agent in minutes.
Such capabilities come from the " Baidu Intelligent Cloud Qianfan AppBuilder " that was showcased at the Baidu Intelligent Computing Conference today, a workbench that can efficiently develop AI native applications.
Sun Ke, chairman of the Baidu Intelligent Cloud Technical Committee, clicked a button and an Agent with a personality, built-in code interpreter, Q&A and Vincent diagram capabilities was automatically created.
You only need to enter the "role command" and select a plug-in tool such as a code interpreter, and the "Interviewer Assistant" will be online to work on the spot.
What exactly can be done? Just like this, just throw it a resume, and the agent will help HR focus on the job description, score and evaluate the candidate's resume from five dimensions.
You can also further put forward requirements, draw a candidate's ability radar chart, and recommend interview questions.
It will even automatically and thoughtfully help you write the code to save the picture.
That's right, it doesn't take 3 months or 3 weeks to build such an intelligent agent, it can even be achieved in minutes .
If you feel that the simple low-code graphical interface is not customized enough, Qianfan AppBuilder also provides a more "black technology" code form.
Through the free replacement and combination of various components, programmers can create large-model applications according to their needs with the same efficiency.
At today's Intelligent Computing Conference, Sun Ke used a few lines of code to quickly screen out the candidates who were most suitable for JD from 10 resumes.
Isn’t this demonstration interesting? We also heard that not only agents, but also the popular RAG (Retrieval Enhanced Generation) and GBI (Intelligent Data Analysis) can be quickly implemented based on such a workbench.
Not much to say, of course we entered the Qianfan AppBuilder official website for the first time to test and see for ourselves~
Within minutes, AI colleagues were successfully put on the job
In the first round of actual testing, let’s try to see if it is difficult to use the “RAG Question and Answer Application” to create a “qubit assistant”.
The so-called RAG (Retrieval Augmented Generation) refers to a technology that combines generative models and retrieval modules to allow generative AI to obtain information from external knowledge sources to improve accuracy and reliability.
Taking the creation of a "qubit assistant" as an example, using the RAG framework, we do not need to fine-tune the model. We only need to prepare some internal qubit reference materials for the large model, so that it can "open the book" to answer questions when faced with the large model. Can answer some domain-specific questions more accurately.
In actual operation, click the "Upload Knowledge" button in the interface and import any knowledge document (supports four formats: .doc, .txt, .docx, and .pdf), and you can create your own knowledge collection.
After simply feeding some Qubit Think Tank reports and internal knowledge documents, a RAG application was created in less than 5 minutes. We have officially entered the testing phase.
It should be noted that in order to verify how "fool-like" it is to create such a RAG application, the documents uploaded this time have not been processed. What they will look like to others is what they will look like to the large model.
Q1: What are the top ten technology trends of the year nominated by qubits?
The answer is very accurate and more detailed than the answer given by Wenxin Yiyan 4.0 based on Internet search.
Let’s ask some more technical details?
Faced with the question "What are the mainstream technologies for 3D generation?", the answer given by the "Qubit Assistant" that can be produced in minutes is Aunt Jiang's:
Compared with the general version of the large model, after referring to the internal knowledge documents of the qubit, the Qubit Assistant can answer more targeted questions.
Let's increase the difficulty again and ask this little assistant some questions that "Baidu" cannot.
For example, how should Qubit editors review topic selections every morning?
Looking at the posture, the boss no longer has to worry about me not learning the document well.
Similarly, some techniques to improve efficiency can also be mastered in the quick version of the "Qubit Assistant".
After this wave of testing, I just want to say to this AI colleague:
In addition to agents and RAG, at Baidu, BI (Business Intelligence) has also been reconstructed by large models. Now, the new GBI (Intelligent Data Analysis) capability has also been put on AppBuilder, enabling small-scale testing.
To put it simply, GBI focuses on "chatting to get everything done": ask any one or more data tables with any questions and get conclusions quickly.
For example, what should you do if your boss requires you to process massive data within 10 minutes, provide accurate cost estimates, and even make decisions based on the analysis results? Of course, let the big model do the calculation~
The specific creation process is the same as RAG, which can be completed through simple graphical clicks. It is mainly divided into three steps: data source access, data range selection, and intelligent analysis.
Next, ask questions and test and you're done. Even insightful analysis of complex problems can be completed quickly.
In addition, this GBI framework also supports professional knowledge injection and complex calculation logic, which can meet more advanced analysis needs.
Baidu opens up AI native application development secrets
After some testing, do you feel that Baidu Smart Cloud’s AppBuilder does have something to offer? In fact, our actual testing only experienced some of the more basic functions.
It is not difficult to see that AppBuilder is actually Baidu's experience and tools in developing large-scale model applications, and has been transformed into a workbench with complete tools.
Specifically, this workbench mainly consists of two parts: application components + application framework .
Application components are mainly componentized encapsulation of cloud and AI capabilities. These include:
- Multi-modal AI capability components cover various AI capabilities such as text, images, and speech. Speech recognition and AI painting can all be realized through this type of component.
- AI capability components based on large language models include question and answer pair extraction, query rewriting, nl2sql and other large model capabilities.
- Basic cloud capability components include traditional cloud service capabilities such as vector database and object storage.
The application framework is oriented to specific scenario tasks and is a series combination of application components. Currently, Baidu Smart Cloud provides the following commonly used AI native application frameworks in AppBuilder:
- Retrieval enhancement generation RAG
- Agent
- Intelligent data analysis GBI
- Basic generation applications, such as multi-style marketing copy generator, question and answer generator, etc.
That is to say, in the application framework, Baidu Intelligent Cloud has packaged all the AI and cloud capabilities required to build AI native applications such as agent and RAG in advance. As we have actually experienced, developers can use it out of the box without even writing a line of code .
In addition to being "fast", Shen Dou, executive vice president of Baidu Group and president of Baidu Intelligent Cloud Business Group, previously revealed that the cost of building such a RAG "assistant" is only a few dollars, even if it needs to process thousands of documents with a length of 10,000 words. One hundred yuan; after that, every time the user consults, the cost is only a few cents.
Zhongtian Steel has also built an intelligent "enterprise dispatch center" based on the Agent framework to realize the automatic perception, decomposition and execution of task instructions.
For example, when it is discovered that steel production does not meet the standard, you only need to ask once, and the large model can automatically call various resources and APIs managed by the platform to find out the reasons for not meeting the standard, adjust the production schedule in a timely manner, and send an email to notify the dispatcher.
It is worth mentioning that in addition to "low-code state" services such as application frameworks, AppBuilder also provides "code state" services for developers: providing a full set of programmable development kits, including SDKs for calling components, development environments, and related debugging tool.
Dr. Sun Ke also introduced a noteworthy detail in this "low code state + code state" development combination:
In addition to allowing developers to quickly test application effects, the low-code kit also includes a code interpreter module that can write code based on task requirements, and these codes are directly available.
Referring to these code examples, programmers can develop and debug by themselves in code mode, or replace them with any application components they want.
This means that programmers can carry out deeper AI native application development and more fully display their creativity.
“The real value in the era of large models lies in native applications”
Baidu's roll AI native application has actually been traced for a long time.
Baidu CEO Robin Li has expressed his opinion more than once: On top of the basic model, there must be thousands or even millions of AI native applications before the value of the large model can be reflected.
Just like the relationship between iOS/Android in the mobile Internet era and the killer apps WeChat and Douyin, there may only be three or four final competitors for the basic model, but the application market above this actually contains more potential. Huge space and potential for innovation.
This view is now arousing wider resonance in the industry.
The trend has already emerged, but Baidu, as the earliest entrant into the game, has also discovered new problems in practice: not all application requirements can be realized through simple prompts. There is an engineering gap between large models and AI native application construction. Complex processes such as design, strategy design, interface solution definition, integration, etc. also require the use of speech, translation, graphics and other capabilities and a complete development environment in conjunction with large models.
Dr. Sun Ke explained to us:
How to build an agent? We asked many developers this question, and they really didn’t know.
Take an application framework as an example. To ensure the stability and controllability of large model output, you need to write some complex prompts and combine them with external functions.
If Big Model wants to make the first "Angry Birds" (a hit app), it needs a series of AI native application development kits to help all developers more quickly understand what Big Model can achieve at its deepest level. things. This is a very missing piece for developers right now.
For such problems, it is far from enough to provide MaaS interfaces for large models.
Baidu Smart Cloud Qianfan AppBuilder covers common modes, tools, and processes for large model development of various applications, precisely to bridge the gap between such large models and AI native application construction.
As Sun Ke said, application components are like loose Lego bricks for senior developers. Everyone can build the entire scene with their own imagination; while the application framework is like a scene version of Lego, which is aimed at more ordinary people. Developers and non-programmers provide examples of application building, allowing more people to use the framework to quickly build AI native application prototypes.
Based on such a set of tools, there are new solutions to the problems of low efficiency and high threshold faced by the current development of AI native applications.
The explosion of the App market in the mobile Internet era is inseparable from the basic support of various technical frameworks. Now, it can be said that workbenches like Qianfan AppBuilder are also laying the foundation for "hot apps" in the big model era.
Big model talk has been running throughout the year, but the real storm of change has just begun.
Now, the stage for the first opportunity point has gradually taken shape, and we can look forward to a wave of killer apps in the new era.
ps Now you only need to enter the official website of Baidu Smart Cloud Qianfan AppBuilder to experience all the components and frameworks.