Source: Quantum Bit

Image Source: Generated by Wujie AI

This year, the last shot of domestic large models belongs to Baidu:

Released only 2 months ago, WENXIN large model 4.0, the capability has increased by 32%. The user scale of WENXIN YIYAN has exceeded 1 billion!

This is the latest data about the WENXIN large model disclosed by Baidu CTO Wang Haifeng at the "WAVE SUMMIT+ Deep Learning Developer Conference 2023", which is like the "Spring Festival Gala" of deep learning.

It is worth mentioning that this "Spring Festival Gala" is different from the past. Starting from 2019, it has already reached the milestone of the 10th session at a pace of "twice a year".

In the five years of progress, in addition to the WENXIN large model gradually iterating from 1.0 to 4.0, another protagonist of each WAVE SUMMIT, PaddlePaddle, has also reached a milestone.

From the data of the PaddlePaddle ecosystem announced by Wang Haifeng, we can sense the following:

• Developer scale: Increased by an order of magnitude from 1.5 million in 2019 to 10.7 million;
• Number of enterprises: Served 235,000 corporate units;
• Number of models: Created 860,000 models based on PaddlePaddle.

So, while Baidu presents this "big homework", how do WENXIN YIYAN and PaddlePaddle perform specifically?

Let's take a look together.

WENXIN YIYAN is stronger

First is WENXIN YIYAN.

Baidu Group Vice President Wu Tian also summarized a set of related data at the conference:

In 2023, a total of 3.7 billion characters of text were created.

What does this concept mean?

The scale of characters is equivalent to 10 "Yongle Encyclopedia", 500 sets of "Complete Works of Lu Xun", and 10,000 copies of "The Three-Body Problem".

In terms of code writing, WENXIN YIYAN has also outputted 300 million lines of code, covering all mainstream programming languages.

In addition, focusing on more specific tasks, WENXIN YIYAN has also completed a cumulative total of 400 million characters of professional contracts, formulated 5 million travel plans, and provided advice and support 2.4 million times.

Not only that, in order to make WENXIN YIYAN "smarter", Baidu has also integrated AI Agent, that is, intelligent agent technology.

Baidu's approach to intelligent agents is to develop two systems in WENXIN YIYAN:

System one: Based on models and memory, it provides users with direct responsive reply generation; System two: Enhances a series of abilities such as understanding, planning, reflection, and evolution.

With the "double buff" of the two, WENXIN YIYAN can use knowledge and tools more flexibly, and analyze problems more deeply.

Even better news is that now, the intelligent agent mode of WENXIN YIYAN has been invited for testing by professional users!

As for the "new release", this time Baidu is aiming at the ecosystem of the large model—Star River Community has been completely upgraded.

From the bottom up, in terms of computing power, Star River Community has provided support for heterogeneous computing power for developers, including NVIDIA, Intel, and Sugon.

In the aspect of general components, it can achieve multi-language programming environment and service deployment.

At the model development level, Star River Community provides the PaddlePaddle industrial model library and the full-process development tools of PaddlePaddle, allowing developers to develop models in a more flexible way.

Finally, in the aspect of application development, Baidu also provides various modes, including zero-threshold development methods for AI painting and AI dialogue, as well as a large model tool center and multi-tool intelligent orchestration.

The large model tool center mentioned earlier is one of the key new contents in the Star River Community this time.

For example, with the support of the large model tool center, developing a "travel assistant" now becomes as simple as "clicking".

First, go to the application section of the PaddlePaddle AI Studio Star River Community page, click "Create Application→Zero-code Development" in the upper right corner, then click "Multi-tool Intelligent Orchestration", and enter the project name:

To make the "travel assistant" more professional, you can also upload more professional documents to the knowledge base, allowing the large model to become an expert and improve its question-answering ability in specific fields.

You can also click "Tool Mounting" to add many mounting tools provided by Baidu, including multi-modal input and output.

Finally, just click "Publish", and an application is built.

It is easy to see that now, even novices who do not know how to program can operate with ease to develop an application.

Behind this, Baidu has actually integrated hundreds of models from the industrial model library accumulated by PaddlePaddle, covering finance, manufacturing, industry, and more; it also includes Baidu Brain's AI capabilities, covering mainstream AI technology directions such as speech, vision, and natural language processing.

Not only that, in order to make WENXIN YIYAN stronger, Baidu has deliberately let it "apprentice" under 10 experts in 10 industries.

The fields covered by these mentors are also extremely wide-ranging, including art design, science and technology communication, transportation, literary creation, medical health, and more.

Wu Tian stated:

Under their guidance, I believe "WENXIN YIYAN" will continue to improve its capabilities in the corresponding professional fields, and eventually provide more thoughtful and extensive support for everyone.

Of course, making WENXIN YIYAN stronger also relies on the support of PaddlePaddle.

PaddlePaddle Upgraded Again: Lowering the Development Threshold

At this WAVE SUMMIT, the PaddlePaddle open-source framework also appeared in a new form—V2.6.

Let's first take a look at the overall upgrade.

First, at the bottom layer, PaddlePaddle has solidified a highly scalable intermediate representation system, which can better support more extreme performance optimization.

Secondly, in terms of dynamic-to-static training, PaddlePaddle has made further upgrades, achieving a 100% success rate through adaptive graph construction technology.

To address the difficulty of distributed development, PaddlePaddle has developed unified automatic parallel programming for dynamic and static, allowing developers to easily develop related mixed parallel training code by understanding tensor segmentation.

At the same time, the suite for large models has been fully optimized, with improvements in all aspects from pre-training to fine-tuning, compression, inference, and deployment.

Finally, in the most challenging aspect of computing power, PaddlePaddle has also made corresponding upgrades in adaptation.

For example, in terms of computing execution, it can support parallel operator scheduling for multiple streams; in terms of hardware vendor development, it can flexibly access "custom acceleration operators" to customize different granularity large operators based on its own hardware characteristics.

In the process of combining with the adaptation of the WENXIN large model, PaddlePaddle has also carried out software and hardware collaboration, supporting hardware vendors in building a Transformer large operator acceleration library at the hardware level, and assisting hardware vendors in accelerating the improvement of the software stack.

Based on the above upgrades, Baidu AI Technology Ecology General Manager Ma Yanjun also brought three "new releases"—the reconstruction of the large model development toolchain, bringing three new development paradigms.

The first new release is Comate AutoWork—develop a program to receive Comate trial rights in 2 minutes.

In simple terms, developers only need to provide requirements, and the rest of the work can be handled by Comate AutoWork.

For example, if we provide a requirement like this:

Add a button called "Limited-time Feedback" on the Comate official website homepage, and when this button is clicked, a pop-up window will display a wheel to receive rights.

And then attach a PRD document to help Comate AutoWork better understand the requirements.

As can be seen, Comate AutoWork will automatically analyze the PRD document, summarize, and refine the work.

For unclear areas, it will also ask questions, and all we need to do is to further explain and clarify the requirements. For example:

There are a total of 6 rights, ranging from 1 to 6 months of trial vouchers.

After that, AutoWork will provide an execution plan and make dynamic adjustments based on the developer's feedback.

Subsequently, AutoWork will mark the code files that need "modification", making it convenient for developers to make further adjustments.

Developers can also make adjustments to the files that need modification using natural language; ultimately, developers can check the code, adopt it, and test it locally.

The second new release from PaddlePaddle is the upgrade of the low-code development tool PaddleX to version 2.2.

Taking character recognition as an example, in the past, even when using PaddleOCR, the recognition result for more complex documents might look like this:

In version 2.2, PaddlePaddle specifically developed PP-ChatOCR. When performing the same task again, the result is completely correct:

Ma Yanjun stated at the event:

Through this version, everyone can truly experience how to fully integrate classic AI application development with the capabilities of large models, and improve their own research and development efficiency. In this version, PaddleX currently has more than 40 industrial-grade selected models, covering ten major AI tasks, and we also support multiple mainstream hardware for the corresponding software, making it convenient for everyone to run in the cloud and locally.

As for the last release from PaddlePaddle—the WENXIN YIYAN development mechanism for the ecological community of developers.

Users of the Star River Community can use the basic capabilities of the WENXIN large model through API and SDK, and support plugin, multi-tool intelligent orchestration development, to achieve AI native application development with low-code and zero-code development interfaces.

The WENXIN YIYAN development mechanism has also been upgraded with supporting development tools, optimizing the registration and access automation process, supporting developers in deep effect tuning, and assisting application promotion, comprehensively reducing the threshold for application development. The typical application development time is shortened by more than 40%, and the end-to-end effect is improved by more than 30%.

For example, if we want to create a "dynamic ranking chart of the top 10 provinces in China by resident population over time", can Prompt ask WENXIN YIYAN to do it? Yes! This also involves using the WENXIN YIYAN plugin "Code Interpreter".

After uploading some data, we can express our requirements in natural language:

Look up the province with the highest resident population in 2003.

As can be seen, the model will automatically generate the corresponding Python code based on the above requirements. Then we can continue to express our requirements:

Generate a dynamic ranking chart of the top ten provinces by resident population from 2003 to 2022.

Finally, after receiving this command, the dynamic chart we wanted to see is presented:

The above is the most significant upgrade at this WAVE SUMMIT.

So the final question is:

In Five Years and Ten Sessions, What Has Baidu Achieved?

First and foremost, a very intuitive feeling is that Baidu has significantly lowered the threshold for "development" through the power of AI technology.

Looking at all the development-related content at the tenth WAVE SUMMIT, almost everything is done through natural language prompts or "clicks".

However, it is obvious that lowering the development threshold to this extent is not something that can be achieved overnight.

If we unfold the core of the five-year journey of the WAVE SUMMIT, Baidu's path in AI development will become clearer:

• 2019: Proposed that deep learning is the operating system of the intelligent era; deep learning drives artificial intelligence into the industrial mass production stage.
• 2020: Building a new type of AI infrastructure, integrating cloud and intelligence to accelerate industrial intelligence.
• 2021: Integration and innovation, lowering the threshold.
• 2022: Accelerating large models on the deep learning platform, solidifying the foundation of industrial intelligence.
• 2023: Large language models bring dawn to general artificial intelligence.

Although the "theme" of Baidu's WAVE SUMMIT may change every year, it is not difficult to see the "constant" amid the changes—building a solid foundation platform for the era of artificial intelligence.

Whether it is in the stage of industrial mass production or the current era of large models that everyone can develop, the combination of "WENXIN large model + PaddlePaddle" seems to play the role of a strong engine at the bottom, providing a continuous source of energy for the blossoming of upper-level applications.

The emphasis on ecological construction in previous WAVE SUMMITs is also the same.

Its importance is as significant as the ecological battle between Apple and Android in the mobile era; in the era of artificial intelligence, the parallel importance of technology and ecology is particularly prominent.

Perhaps this is also the reason why Wang Haifeng mentioned "WENXIN plus PaddlePaddle, elegantly heading to the Star River" twice at this year's WAVE SUMMIT, and this "Star River" is the vast universe of general artificial intelligence.

In conclusion, in the "five years and ten sessions" of the WAVE SUMMIT, Baidu has indeed established a solid foundation platform for the era of artificial intelligence.

On this foundation, in the ever-changing future of technology, it is worth looking forward to what kind of technological revolution Baidu will bring.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。