Rakuten on April 16
Robin Li, founder, chairman and CEO of Baidu, delivered a keynote speech on "Everyone is a Developer" at the Create 2024 Baidu AI Developer Conference today, pointing out that "AI is setting off a creative revolution, and developing applications in the future is as simple as shooting a short video, everyone is a developer, and everyone is a creator." ”
Robin Li said that Baidu, as a technology company, is positioned to provide everyone with the development tools they need as much as possible to enhance social creativity, which includes a powerful basic model series and three AI development tools, which form a toolbox that allows developers to take away and use at any time.
The Wenxin model series includes: the flagship version of ERNIE3.5 and 4.0, and the lightweight version of ERNIE Speed, Lite, Tiny, etc.;
Robin Li revealed that since the release of Wenxin Yiyan more than a year ago, the number of users has exceeded 200 million, and at present, the Wenxin model has become the most advanced and widely used AI basic model in China.
Robin Li emphasized that the smaller size model cut out through the powerful Wenxin 4.0 dimensionality reduction is significantly better than the model directly tuned from open source, and the effect is significantly better under the same size, and the cost is significantly lower under the same effect, "so the open source model will be more and more backward." ”
At the same time, Baidu has also made three "out-of-the-box" AI development tools for everyone, including agent development tool AgentBuilder, AI native application development tool AppBuilder, and model customization tool ModelBuilder of various sizes. "These three tools all represent advanced productivity. ”
At the meeting, Robin Li shared Baidu's specific ideas for developing AI native applications based on large models, pointing out that MoE, small models and agents are three directions worth paying attention to. ”
The following is a transcript of Robin Li's keynote speech "Everyone is a Developer":
You can be a developer as long as you can speak
Hello everyone, welcome to the Create 2024 Baidu AI Developer Conference, which is the first time that the Create Conference will be held in the Guangdong-Hong Kong-Macao Greater Bay Area. More than 5,000 developers and tech enthusiasts are here today. In the past year, I have communicated with many entrepreneurs and developers, and I feel that everyone is in a state of "FOMO", that is, Fear of Missing Out, which is both excited and afraid of missing out. Indeed, large models and generative AI will revolutionize the developer community.
In the past, developers used code to change the world, but in the future, natural language will become the new general-purpose programming language, and as long as you can speak, you can become a developer and change the world with your creativity.
This day is not far away, and we have seen that because of the strong basic model, there are many low-threshold, or even zero-threshold development tools, and the productivity of developers has greatly improved.
For example, Comate, an intelligent code assistant based on the Wenxin model, not only supports more than 100 languages and all mainstream IDE platforms, but can also recommend code, generate code comments, find code defects, and give optimization solutions, but also deeply interpret the code base and associate private domain knowledge to generate new code. After more than a year on the job, Comate has entered tens of thousands of companies such as Himalaya, Mitsubishi Elevator, iSoftStone, etc., and the code adoption rate generated has reached 46%, and 27% of Baidu's new code every day has been generated by Comate.
Today, you can make an AI application without writing code, and you can make an agent without programming. AI is setting off a revolution in creativity, and developing applications in the future is as simple as shooting a short video, everyone is a developer, and everyone is a creator.
As a technology company, Baidu's role is to provide everyone with the development tools they need to continuously improve the creativity of society as a whole. Specifically, we provide a powerful basic model series, which is the Wenxin large model series, which includes the flagship version of ERNIE3.5, ERNIE4.0, and the lightweight version of ERNIE Speed, Lite, Tiny, etc.
We also provide tools to develop various applications based on large models, including agent development tool AgentBuilder, AI native application development tool AppBuilder, and model customization tool ModelBuilder in various sizes. These three tools all represent advanced productive forces.
Wenxin Yiyan has more than 200 million users, and the Wenxin model 4.0 tool version has been released
First of all, let's talk about the latest progress of Wenxin Yiyan and Wenxin model:
Wenxin Yiyan was released on March 16 last year, and it has been one year and one month to today. The number of our users has exceeded 200 million, the average daily API call volume has also exceeded 200 million, the number of customers served has reached 85,000, and the number of AI native applications developed using the Qianfan platform has exceeded 190,000.
Let's see, what are you all doing with Wenxin Yiyan?
The real people in the video are just the tip of the iceberg. We can see that Wenxin Yiyan is changing the work and life of more people.
The pedestal model that supports Wenxin Yiyan is the Wenxin model. Over the past year, it has undergone an evolution from version 3.0 to version 3.5 and then to version 4.0. Wenxin 4.0 has reached the industry-leading level in terms of understanding, generation, logic, and memory.
In recent months, Wenxin's large model has achieved further significant improvements in general capabilities such as code generation, code interpretation, and code optimization, reaching the international leading level.
Today, we officially released the tool version of Wenxin Model 4.0, and now, you can experience the code interpreter function on the tool version, through natural language interaction, you can realize the processing and analysis of 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.
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%.
In other words, customers can now make 1 million calls for the same cost instead of 10,000 calls a day. The media may not be excited about a 99% drop in costs. However, whether it is an enterprise or a developer, once it is used, the most important thing is the effect and cost.
We can reduce the cost of inference to 1% while improving performance, precisely because 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.
There is no doubt that the topic of large models will still be hot in 2024, and various technological breakthroughs will continue to emerge. The media will also continue to be keen to render with titles like "Shock Release" and "Epic Update". However, I would like to emphasize that large models do not directly create value on their own, and AI applications developed based on large models can meet real market needs.
Share specific ideas and tools for developing AI native applications: "Stepping on countless pitfalls and paying high tuition"
Today, I would like to share with you some specific ideas and tools for developing AI native applications based on large models. This is what our Baidu has done in exchange for stepping on countless pits and paying high tuition fees according to the practice of the past year.
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 are all technical and need to be matched for different scenarios of the application.
The second is the small model. The inference cost of the small model is low, the response speed is fast, and in some specific scenarios, the use effect of the small model after SFT fine-tuning can be comparable to that of the large model. That's why we're releasing three lightweight models: Speed, Lite, and Tiny. 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.
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, we can also allow different agents to interact and cooperate with each other to complete tasks with higher quality. We have developed these agent capabilities and are fully open to developers.
In the three directions of MoE, small model, and agent, Baidu has already done a good job of "out-of-the-box" tools for everyone. Below, I will introduce you to three different tools, namely: agent development tool AgentBuilder, AI native application development tool AppBuilder, and model customization tool ModelBuilder of various sizes.
AgentBuilder:
Agents are the most recent and mainstream way for everyone 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.
Baidu has just upgraded the Wenxin Twin platform. Up to now, more than 30,000 agents have been created, more than 50,000 developers and tens of thousands of enterprises have settled in. Our goal is to let everyone and every organization become an agent developer, and create the most complete agent ecosystem in China.
So how do you achieve this goal? It is to provide you with AgentBuilder, a zero-threshold agent development tool. Let's take the "Singapore Tourism Board" as an example to see how an agent is made.
First of all, we open the Wenxin intelligent twin platform, and the creation page has two modes: zero code and low code, and novices can directly choose the "zero code mode" and create an agent in a few sentences in natural language.
We first named the agent "Singapore Tourism Board", and then wrote in the settings that we needed to create a travel plan, answer questions, and provide hotel ticket booking services, all of which were used to guide the agent and tell it what it could do.
If you only need the basic agent, the platform will automatically complete the filling. But we want the Singapore Tourism Board to be a professional agent, so we need to make a high-level configuration. I can add all the Singapore Encyclopedia entries and links to the official website to the knowledge base and keep it updated every day. Then add some tools, such as hotel inquiry, attraction ticket purchase, etc., to enhance its service capabilities. At present, we have cooperated with Ctrip to provide travel service tools such as hotels, attractions, and ticketing. In this way, an agent of the Singapore Tourism Board is ready to be further previewed and tuned.
Now open the Baidu app and search for "when to go to Singapore with the least number of people", because everyone wants to avoid the crowds when they go out to travel. The agent synthesizes information from multiple sources to generate an answer, "January to March has the fewest people". We can also click on the agent to further interact with it, such as the precautions for traveling to Singapore, recommend the top three hotels in Singapore, and let it directly book tickets for Universal Studios Singapore, a one-stop solution to the needs, which greatly saves the user's time.
In addition to Singapore, cultural and tourism agents such as Dalian and Shenyang are also online on the Wenxin Agent platform, as well as a variety of agents such as knowledge, creation, learning, and entertainment, which are made with AgentBuilder.
When Wenxin Yiyan was first released last year, I said that Wenxin Yiyan would affect every company. Because of its strong natural language understanding, expression ability, and reasoning ability, it can make any company closer to its customers.
Today, every merchant and every customer can have their own agent in Baidu. The whole process does not require programming at all, through the input of information similar to prompt words, and simple operation and optimization, you can quickly generate an agent and become a 7X24 hours online gold salesman.
Let's take a look at how a merchant agent is made.
Kai Tak Education is a well-known education company, with more than 60 branches across the country, as well as many overseas branches, covering a wide range of countries, and has high requirements for reception skills. How can we respond to customer inquiries 24 hours a day, improve the level of reception, and reduce operating costs?
Kai Tak Education used Baidu's AgentBuilder to create its own agent.
Let's take a look at how to build an agent with basic capabilities. It's very simple, fill in the avatar, name, business scope and welcome message of the agent on the platform, and then set some information that needs to be provided by the user, such as age and education. 5 minutes, zero threshold, an agent is ready.
Kai Tak Education also hopes that this agent is a study abroad consultant who understands business and students. It can make professional analysis and give accurate answers according to the different situations of students, such as whether they want to go to the United States or Australia, whether they want to go to the United States or Australia, whether they have a master's degree or a bachelor's degree, how many IELTS and TOEFL scores they have, etc. We can build a more advanced agent by adding the modules of knowledge, roles, and tools.
In the knowledge module, upload private domain knowledge, let the platform analyze it in real time, and automatically generate dialogue materials; in the role module, add some study abroad countries that are not within the scope of business to the filtering scheme, which can improve the effectiveness of user clues; in the tool module, add services such as appointment to store. Through such a few simple steps, a professional and capable Enlighten Education agent will be ready.
Now, let's search for "application conditions for studying in Australia", we can see that the agent quickly gives the required language ability, major selection and other seven necessary conditions, and also gives the corresponding study abroad consulting plan, which has all kinds of questions and answers.
In the first week of its launch, it was successfully distributed 1.55 million times and interacted with users 58,000 times, resulting in a linear increase in the number of lead conversions, a significant reduction in the conversion cost of effective leads, and a significant improvement in operational efficiency.
Next, I will introduce to you an intelligent twin in the home furnishing industry.
Sophia is a home furnishing brand that focuses on whole-house customization. As shown above, it can also create a basic merchant agent by filling in extremely simple information. But for the home furnishing industry, the offline experience of consumers is more important, so Sophia hopes to create a gold medal sales online and restore the offline reception experience.
Therefore, in the further setting, it selects the digital human as the display method in the role module, and then selects the appropriate background and voice for the digital human, and automatically summarizes a set of sales words in combination with the intelligent analysis ability of the platform. In the end, she has created a gentle, cordial, and professional gold medal salesperson, who can meet the various needs of users 24 hours a day and provide a high-level service experience.
When Baidu search users have decoration demands, Sophia Intelligent will use the ability of the Wenxin model to give priority to the answers to the questions. In addition, she will take the initiative to confirm specific needs with customers, such as decoration type, budget, etc., and recommend nearby offline stores.
Since the launch of Sophia's merchant agent, the cost of effective leads has dropped by 30%. That is, it gets a valid customer, if the cost used to be 100 yuan, now it only needs 70 yuan.
At present, more than 10,000 Baidu customers have merchant intelligent twins, covering more than 30 industries such as education and training, real estate and home furnishing, machinery and equipment, and business services.
Above, through three demos, I show how developers and merchants can use AgentBuilder to make agents in different industries.
Now, making an agent is really a matter of minutes. But here's the problem: If there is no traffic, no distribution, no finding, no use, then developers and merchants will have no revenue, and there will be no motivation without revenue. How to solve this pain point?
Our Wenxin intelligent twins platform provides developers with a channel to monetize their 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.
AppBuilder: You can develop an app in three natural language steps
Next, I would like to introduce you to the second development tool, AppBuilder. It is currently the best AI native application development tool. On AppBuilder, we have encapsulated and preset various components and frameworks required for developing AI-native applications in advance, greatly reducing the development threshold.
In just three steps, developers can develop an AI-native application in natural language, which can be easily published and integrated into a variety of business environments. Let's look at a few examples:
At the beginning of this year, we held an AI native application development challenge, which used AppBuilder to create a "playground queue planning assistant" to help visitors better understand the playground queue situation, design personalized play routes, and get the best play experience in a limited time.
The champion of this competition, without writing a line of code, developed an application and won the 100,000 yuan prize provided by Baidu. If you can write code for this topic, it is not difficult for you to write one, but you can still make it without writing a single line of code, which still depends very much on the ability of the basic model and the AppBuilder tool.
Let's take a look at how to use AppBuilder to make this AI application.
Let's start with a review of the questions. It assumes the queuing time and excitement index of each program of "Universal Studios", so getting the experience with the highest thrill index in a limited time is the problem to be solved in this issue.
First, open the development interface of AppBuilder and name the application "Playground Queuing Assistant", the second step is to describe the specific requirements in the character instructions, including calling the code interpreter, calculating the best combination within a fixed time, and the output result, etc., and the third step is to add the code interpreter to the tool component to help with the operation.
Now, let's test the effect, type the question "I have 3 and a half hours, how can I play the most exciting?" on the right, and you can see that the code interpreter translates the problem into code, and then calls the data understanding tool to analyze the known conditions. After a series of calculations, it was concluded that the combination of "Harry Potter Forbidden Journey", "Jurassic Adventure", "Decepticon Roller Coaster" and "Bumblebee Rondo" has the best effect. The test is fine, click publish, and with zero code, an app is generated.
Now, AppBuilder has been further upgraded, and in the process of creation, all of them can be automatically helped to optimize role commands, component configurations and other links through the "AI Optimal Configuration" function, further improving development efficiency.
Let's look at another example.
At the beginning of this year, North China Electric Power University proposed to provide intelligent and exclusive services for teachers and students of the university. Based on Baidu's AppBuilder, we jointly created a Huadian AI assistant, below, let me show you how the Huadian AI assistant is made:
Step 1: Open AppBuilder, go to the creation page, and click AI to automatically generate configurations. First, set basic information such as the name, introduction, and avatar for the app.
Step 2: Describe specific requirements in the role instructions through natural language, including tasks, component capabilities, requirements, and constraints.
Step 3: Insert customized components such as book borrowing query, class schedule query, and student score query, so that the campus assistant can have the ability to provide intelligent services. Then add an opening line for the Campus Assistant so that the app is configured.
Next, in the preview interface, we debug the assistant based on the user's questions. For example, ask about the registration time of the fourth and sixth level exams, and test the effect of automatic invocation of each component.
As you can see, with these simple operations, the app is built. At present, it has been launched on a small-scale test, and has been connected to high-frequency scenarios such as checking systems, checking courses, charging meal cards, and borrowing books, serving the teachers and students of the school. Subsequently, we will work with North China Electric Power University to further deepen application cooperation and provide more abundant and convenient services.
Baidu also has many years of technical accumulation in terms of cross-modality. On AppBuilder, we also provide certain cross-modal capabilities, you only need to give a piece of text, or a few sentences, you can quickly create drawing applications, such as comics, children's picture books, etc.
The process is also very simple: open AppBuilder, click to enter "App Creation", enter the role command, select the "Add Wensheng Diagram" component, and then enter three recommended questions, and then click Publish. After the application is ready, we only need to enter a character or plot that we roughly want, and AppBuilder can automatically generate a story and output pictures.
Baidu Wenku's latest smart comic and artbook generation feature takes advantage of the kind of components available on AppBuilder. Let's take a look at how Baidu Library's comic generation function allows everyone with creative ideas to make good works.
Let's take the classic "Removing the Three Evils Around the World" as an example.
Open Baidu Library,Enter the theme "Eliminate the Three Evils of the Week",This is written in the "Book of Jin" and "The New Language of the World",After the library search,The story will be generated according to the original text,We can also modify the content of the story;Then click on the AI toolbar on the right,Start creating this comic。
Enter the comic production interface, the library will automatically generate comic storyboards for us according to the storyline, and then choose the comic style that best fits the story from a variety of styles such as light and shadow, realism, and cartoons, and finally select different characters according to the characters, so as to complete the generation of comics.
After the comics are generated, we can browse the complete comics in the intelligent comic editor of Baidu Library. At the same time, the library supports editing, modifying and fine-tuning each picture, for example, select the screen "Zhou and the White-fronted Tiger" on the left, click Edit, and add the picture description "The character has a clear face and the picture is bright", and you can fine-tune the comic that better meets your needs. As you can see, Baidu Library's comic function has done a very good job in terms of stylistic consistency of characters and scenes.
Baidu Library's intelligent comic capabilities have greatly improved the efficiency of creating comics, lowered the cost and threshold of comic creation, and allowed more people with ideas and creativity to realize their dreams of comic creation.
In addition to generating comics, Baidu Wenku can now also help users create drawing books with zero threshold. You may not know that the average number of picture books read by Chinese children per capita is only 10 a year, while that in Europe and the United States is about 50. Now, AI is able to allow parents who have no drawing ability at all to create their own children's picture books for their children. Let's take a look at this illustrated audio picture book!
Since last year, we have reconstructed Baidu Wenku with AI, making it a "starting point for content production" for users. Now, with the support of AppBuilder, Baidu Wenku's newly launched smart comics and smart picture book functions have extended the scene to a more interesting cross-modal creation field.
Just now, I used three cases to show how to use Baidu's AppBuilder to make AI native applications. You should be able to feel two distinct advantages of AppBuilder:
First, it is powerful. Relying on Wenxin 4.0's ability to understand and follow instructions, our AppBuilder can ensure that cold start-up can reach a good level, and will not take a long time to optimize due to poor results, which greatly reduces the development threshold. Relying on the retrieval enhancement technology RAG, in typical scenarios such as knowledge Q&A, our Q&A accuracy and friendly response degree have reached more than 95%, which greatly exceeds 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. and third-party APIs in some mainstream scenarios, such as flight query, paper query, etc. We've also just added support for custom components, so customers can connect directly to any of their 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. We also support open-source SDKs, which is convenient for you to carry out secondary development.
ModelBuilder: Efficient and low-cost model production
Next, I will introduce a tool that is more suitable for professional developers, that is, ModelBuilder, which can customize models of any size according to the needs of developers, and further fine-tune the SFT model according to the subdivision scenario, so as to achieve better results. One of the most important things for developers to use large models well is to master the method of model fine-tuning.
Let's take the essay correction scenario in the education industry as an example. Because there are clear scoring criteria for essay grading, and the requirements and scoring criteria for essays are different for different grades. This requires fine tuning of the model, so that the output of the large model can better meet the specific requirements.
Let's take a look at how ModelBuilder implements the fine-tuning of the model for essay grading.
Step 1: Create a dataset. The effectiveness of model fine-tuning is highly dependent on the quality of our data. In this case, there were only 180 pieces of raw data, and the quality was not high enough. There are three features that we need to use: data cleansing, data annotation, and data augmentation. Data cleaning can quickly remove problems such as data gaps and garbled characters. In terms of data annotation, we have added more dimensions to the composition, such as the depth of the content, the writing technique, etc. Data enhancement capabilities can generate identical but non-duplicate data for data enrichment. After the extension, ModelBuilder generated 920 pieces of high-quality data for us.
The second step is to enter the model fine-tuning stage. First of all, we need to choose a base model for fine tuning, and here we have chosen ERNIE Speed as the base model. Then, we configure the parameters according to the recommended values of the platform, this time the data is close to 1000, according to the recommended iteration rounds, configure it to 10 rounds, and then we can let the model start training.
Step 3: Deploy the model on the platform, so that the whole fine-tuning process is complete.
Let's use a middle school essay to test the effect of the finely tuned model. The title of the essay is "Looking forward to growing up", the subject is not limited, and it is not less than 800 words.
We input a student's essay and grading format requirements into the two models before and after fine-tuning. Let's take a look at the comparison between the two models and the teacher's comments. It can be seen that the fine-tuned large model has a more professional teacher's comment thinking and better format compliance. Essay correction tasks, which used to take a long time for teachers to complete, can now be completed quickly in a few seconds with the help of large models, which greatly improves the speed and efficiency of giving feedback to students. At present, this kind of fine-tuned model has been applied in many leading enterprises in the education industry.
As I mentioned earlier, large-scale AI-native applications are generally MoE. So, how to combine these models to get the best cost performance? Next, I will use Xiaodu to show a case of MoE model routing. Let's take a look first,With the blessing of the large model,Is Xiaodu getting smarter:
Robin: Small
Xiaodu: Yes
Robin: What's the news today?
Xiaodu: Xiaodu broadcasts 3-5 news flashes of the day
Robin: Small. I heard that you've been getting a lot of skills lately?
Xiaodu: I can not only chat with you, but also be your AI schedule assistant
Robin: Then I'm going to test you. Help me arrange my schedule, after today's event, I will have dinner with Mr. Zhang, and then there will be a management meeting in the hotel, which will take about 1 and a half hours, and I hope it can be finished before 10 o'clock. I'm going for a morning run tomorrow morning to see what the weather is like, and then there's a morning meeting at 8 a.m., please help me arrange my schedule.
Xiaodu: According to your needs, Xiaodu will make the best schedule for you as follows
o You and Mr. Zhang have dinner from 18:00 to 19:30, after which you will attend a management meeting at 20:00 and end at around 21:30
o The alarm clock will remind you to wake up and go for a morning run at 6 o'clock tomorrow morning, and the running time is controlled to less than 1 hour, so as to ensure that you are not late at 8 o'clock tomorrow morning
o Tomorrow morning the weather will be sunny and the temperature will be 25 degrees, so it is recommended that you go out in short sleeves and do a good job of sun protection
o See if I can be your secretary, I have more new skills
Robin: That's good, great (thumbs up to Xiaodu), let's sync these schedules to my phone.
Just now, I asked Xiaodu several questions, including simple news inquiries and complex scheduling needs. It seems to be communicating with the same small degree.,In fact, different models are called behind it.。 When it received my question, the small model ERNIE Tiny first performed the work of "model routing"; for weather problems, it called the fine-tuned model based on ERNIE Lite based on the results of the weather query, and also through the ERNIE Lite of SFT, and gave dressing suggestions; for more complex schedules, it called the best large model, Wenxin 4.0, to calculate the schedule from tonight to tomorrow morning.
Today,Every question we ask Xiaodu,It will be assigned to different models to execute in this way,When calling the application's API interface,ERNIE Functions model will be used,When teaching children,Use Wenxin model3.5or4.0,The creation of intelligent assistants uses ERNIE Character model,Improve the consistency of personality、Stimulate users' desire to chat。
Through this combination of large and small model scheduling,Xiaodu not only successfully completed"Brain change action",Installed a new AINnative operating systemDuerOS X,And formed the optimal combination of effect、Speed and cost,Compared with the flagship version that uses all Wenxin large models,The response speed is increased2 times,The cost has been reduced99。 The small Tiantian AI tablet robot that I just talked to,Yesterday it was fully sold on major platforms,Interested friends can place an order to experience it immediately。
These examples of ModelBuilder demonstrate Baidu's ability to produce models efficiently and at low prices. Because Wenxin 4.0, the most powerful basic model, we can tailor a smaller model suitable for various scenarios according to our needs, taking into account various considerations such as effect, response speed, and inference cost, and support fine tuning and post pretrain. In this way, the model cut out through dimensionality reduction is significantly better than the model directly retrieved from open sources, and the effect is significantly better under the same size, and the cost is significantly lower under the same effect. In the past, people used to think that open source was cheap, but in fact, in the large model scenario, open source is the most expensive. So the open source model will fall further and further behind.
In order to make it easier for you to get started quickly, ModelBuilder presets the most comprehensive and abundant large models. It includes ERNIE3.5 and ERNIE4.0, which are the flagship large models, which are suitable for general complex scenes and have powerful capabilities, as well as three lightweight large models, ERNIE Speed, Lite, and Tiny, and two vertical scene models, ERNIE Character is suitable for role-playing, and ERNIE Functions is suitable for external tool use and business function calls in dialogue or Q&A scenarios. Of course, ModelBuilder also supports third-party mainstream models at home and abroad, with a total of 77 models, making it the development platform with the largest number of large models in China.
The second "Wenxin Cup" entrepreneurship competition was officially launched: a special prize of 50 million yuan was set up
For developers, we provide financial and resource support in addition to these development tools.
In May last year, Baidu launched the "Wenxin Cup" entrepreneurship competition, hoping to promote the prosperity of the large model ecosystem, create a more dynamic ecosystem, and help entrepreneurs and developers develop various AI native applications. In the first "Wenxin Cup" Entrepreneurship Competition, we received nearly 1,000 entrepreneurial teams to sign up, 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.
Today, I announce the official launch of the second "Wenxin Cup" Entrepreneurship Competition. This time, we will expand the scope of project screening, set up sub-venues, and recruit entrepreneurial and innovative teams for the global market and college students, as long as your entrepreneurial direction is AI native applications, you can register for the competition on the official website of the competition. At the same time, we have also stepped up our support for entrepreneurs, provided more adequate investment funds and richer business resources, and 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.
Chinese entrepreneurs and developers are very good at using new technologies to develop applications. I believe that the Wenxin model will become the first choice for Chinese AI entrepreneurs and developers, and more and more applications will be built on top of the Wenxin model. I also look forward to more entrepreneurs and developers joining us to build a thriving AI ecosystem.
The biggest application scenario of the large visual model is autonomous driving
Most of the tools and cases we talked about earlier are based on large language models. Looking to the future, I believe that multimodal large models, or the integration of text, images, voice, video, etc., are a very important long-term development direction for basic models, and they are the only way to AGI. Baidu has a long-term investment in these areas and will keep up to date with technological progress.
I have a very different judgment: the biggest application scenario of the large visual model is autonomous driving. Baidu is the best in this direction, a global leader in autonomous driving, and we train AI not only on how to generate video, but also on AI to understand what's happening in the real world and predict the future.
Based on more than 100 million kilometers of road test mileage data in China's complex cities, Baidu trained the Apollo visual perception model. It has four basic capabilities: detection, tracking, understanding, and mapping. This gives Baidu a smarter, more adaptable, and safer autonomous driving solution.
Baidu Maps is also the first to apply the visual perception model to the field of mapping. Now, the world's largest lane-level map data has been launched in 360 cities across the country. As long as Baidu map navigation can go anywhere, intelligent driving can be opened.
After the Spring Festival this year, Baidu's radish run has achieved the "first cross" of the Yangtze River, we have extended the service from the north bank of the Yangtze River to the south bank, in some areas of Wuhan, we have achieved 7X24 hours of all-weather operation, and also plan to deploy 1,000 unmanned vehicles in Wuhan within the year.
This is a landmark event for autonomous driving to be truly commercialized, which is no longer just a regional demonstration, but has entered a new stage of city-level application demonstration. With a population of more than 3,000 square kilometers and 7.7 million people in Wuhan, Radish Express is the world's largest autonomous driving operation area.
"Take it away, take it away"
Just now, I showed you the Baidu Wenxin large model series, as well as three development tools - AgentBuilder, AppBuilder, and ModelBuilder, which form a toolbox, which you can pack and take away immediately, and use it at any time.
At this moment, I am standing here, in fact, I am also a developer, an entrepreneur, and I am as excited as everyone else. Today's China has 1 billion Internet users, a strong basic model, enough AI application scenarios, the world's most complete industrial system, and the state is also vigorously encouraging and supporting the "artificial intelligence +" action.
Today, anyone can be a developer. And the future will be a future created by developers together!
———————————————
Lei Di was founded by Lei Jianping, a media person, if it is reprinted, please indicate the source.