Wenxin Big Model 4.0 Released! Claims to benchmark GPT 4.0

Article source: Geek Park

Author | Li Yuan

Edit | Jing Yu

From search, library to business analysis, AI big models have become Baidu's "vertical line".

"Welcome to the era of generative AI! **」

On October 17, 2023, wearing a white shirt and small white shoes, Robin Li, founder, chairman and CEO of Baidu, announced the arrival of a new era on stage.

At this press conference named "Generating the Future", Robin Li officially released the latest 4.0 version of the company's big language model, Wenxin Big Model, and taught people how to use prompt words to make the upgraded Wenxin big model apply, helping people use Beijing provident fund to buy houses in Hebei, make advertisements and videos, and create online novels.

Robin Li directly benchmarked the Wenxin Model 4.0 with GPT-4 at the meeting|Baidu

At the meeting, **Robin Li confidently stated that the capabilities of version 4.0 of the Wen Xin model are "no less than GPT-4".

At the same time, the statement that "all applications are worth reconstructing with large models" announced half a year ago also had results at the meeting on the same day. From the main business search, to Baidu library, network disk, map and other Baidu family bucket applications, they have been connected to the Wenxin big model, showing better interaction and logic capabilities.

On the B side, Robin Li also exposed GBI (generative business analysis) tools, as well as office assistants with large model capabilities "Ruliu".

If generative AI activates the entire tech world, then Baidu may be the giant that benefits the most.

01 Wen Xin 4.0 Direct Benchmark GPT-4

As soon as the press conference began, Robin Li first announced the release of version 4.0 of the Wenxin model.

Baidu divides four defining criteria for the capabilities of large models: understanding, generation, logic and memory. This time the Wen Xin big model 4.0 upgrade, the infrastructure is still the same as the 3.0 and 3.5 versions, but claims to be greatly improved in terms of logic and memory.

Robin Li announces Wenxin Grand Model 4.0|Baidu

According to Baidu CTO Wang Haifeng, the improvement of comprehension and generation ability of Wenxin Big Model 4.0 is similar, while the improvement of logic is 3 times that of comprehension, and the improvement of memory is 2 times that of comprehension**.

Four different capabilities can improve the efficiency of different application scenarios. For these abilities, Li Yanhong showed it on the spot.

Understanding is the basis for conversational AI to help users, and it is very important in government affairs, marketing, customer service and other fields.

In terms of comprehension ability, the scene used a word order reversal and vague expression (prompt) to test the model's ability: "I want to return to Chengde to buy a house, can I use the provident fund loan?" What about the procedures? I work in Beijing."

To understand this sentence, AI must understand that "working in Beijing" and "returning to Chengde to buy a house" actually have "paying provident fund in Beijing, and the hukou is in Chengde." This kind of subtext with Chinese characteristics can make accurate answers that users need. And sure enough, Wen Xin quickly understood the key to the question and made a correct answer.

Robin Li explains the four core capabilities of AI|Baidu

The generation ability can mainly improve the efficiency of brand marketing, copywriting and creative work.

At the scene, Li Yanhong showed that according to a picture, with natural language prompts, you can perform background transformation, subject blurring, and generate posters and copywriting according to official website information.

In addition to these traditional image processing, Baidu also demonstrated its ability to generate video. Through natural language, in the live demonstration, Wen Xin generated a digital human oral video ad with almost no delay. The video incorporates product pictures, adds a lot of transitional background, and a digital person in a suit appears from time to time to introduce the product features collected from the official website.

Live display of Wenxin Model 4.0 advertising generation capabilities|Baidu

The effect that could only be achieved by the cooperation of multiple AIGC products was seamlessly integrated in this display. One advertising film, 5 ad copy, and a poster took less than 3 minutes to generate.

Logical reasoning, usually manifested as a test of mathematical logic. In this showcase, Baidu highlighted its potential in education.

Li Yanhong gave an application problem involving the transformation of conical volume into cube volume, and Wen Xin not only gave the solution, but also solved the problem step by step, and analyzed the knowledge points involved in each step.

Using generative artificial intelligence to tutor children with homework is simpler|Baidu

For the display of memory ability, Baidu's choice is more special.

Baidu chose to let Wen Xin write the outline setting of a martial arts novel. After the writing is completed, on the original outline, let the big model increase the relationship between the characters and increase the drama conflict, to show that the big model can remember the original outline setting and character ability after adding complex information, without using imagination aimlessly.

Baidu also shared the technical support for the improvement of Wenxin's large model capabilities.

Baidu has previously announced that the Wenxin model is the first large model trained using the Wanka cluster in China, and many people speculate that the parameter scale of the Wenxin model 4.0 is expected to exceed the trillion level. However, at this conference, Baidu did not emphasize the parameter level of the large model.

In addition to Wanka training, Baidu CTO also mentioned that the weekly average of Baidu's algorithm training stability has exceeded 98%, and knowledge point enhancement technology has been carried out in terms of input and output.

02 "Refactoring" Baidu Family Bucket

Although they are shown separately, in fact, more often than not, the four basic capabilities of large models are applied in combination.

In May, Baidu announced the use of large models to reconstruct Baidu's applications. At this conference, Baidu also demonstrated the latest achievements of Baidu's application of Wenxin large model reconstruction.

Among them, the most amazing is the refactoring of search.

In February, Microsoft launched New Bing based on GPT's technology to refactor its search. In his latest testimony, Microsoft's Nadella said Microsoft's share of the search market has barely changed since adding AI capabilities to Bing.

Microsoft's New Bing mainly launched a system of conversational bots that can chat with New Bing to ask questions to get integrated information with links. Google's Bard is similar.

Baidu uses AI to reconstruct its main business search|Baidu

However, Baidu's search reconstruction this time goes deeper into the entire search system. Baidu describes it as "ultimate satisfaction, recommendation stimulation, and multiple rounds of interaction".

The ultimate satisfaction is reflected in entering a question in the search box, Search can no longer give a link, but directly generate the best answer.

In the presentation, Robin Li raised the question of what is the ranking of industrial added value of various countries in the past 20 years.

Unlike New Bing and Bard, which may give a linked data answer, the new Baidu can directly give a dynamic table graph, in the form of a bar chart, showing the industrial growth values of different countries. This graph is even dynamic, growing and changing over time.

The recommendation excitation function is somewhat equivalent to the relevant questions of the current search engine, which can prompt the user to continue to understand some related questions according to the prompt, such as "What is the relationship between industrial added value and GDP?" "What is the impact of industry on the development of the national economy?".

Robin Li shares AI reconstructs Baidu family bucket application experience|Baidu

And the multiple rounds of interaction are also very interesting.

In the current wave of big language model entrepreneurship, one of the many entrepreneurs is working hard to use, that is, to use large language models with recommendation engines to conduct multiple rounds of dialogue to provide users with the best choice.

In September, Baidu held the Wenxin Cup entrepreneurship competition project, and the first prize winner Buysmart.AI was the leader in this direction. Users use natural language and clicks to constantly clarify their needs, and Buysmart.AI uses the recommendation engine to ultimately recommend the products that users need most.

The reconstructed Baidu search directly adds a function of similar direction to the search.

In the demo, Baidu's search prompt is asking "Where to go hiking around Beijing?" After giving multiple answers such as Baihuashan, Haituo Mountain, etc., the search engine allows users to further click to supplement and choose their own situation. For example, if you choose to add parent-child hiking novices, the search engine will change to recommend places such as Xishan and Baiwangshan, which are relatively easy to climb and more friendly to parent-child activities.

In addition to the reconstruction of search, Baidu also showed the reconstruction of Baidu network disk, Baidu map, Baidu library and other applications.

Baidu Network Disk's cloud a personal cloud assistant has been launched before. As the world's first personal cloud assistant, it currently has 20 million users. You can use natural language to communicate with the assistant, find a video in the personal cloud in one sentence, understand the video content, find a certain content in the video, summarize the golden sentence of the video, and so on.

Baidu Map, according to Baidu's promotion, is the world's first AI native map product. Talking to the map's assistant makes it possible to access thousands of services in a multi-level menu in one step. You can also recommend restaurants with suitable locations, choose from the environment of the restaurant, and finally book a taxi directly.

Relying on billions of past manuscript resources, Baidu Wenku can directly select the type of article needed, serious academic literature or general public materials after users search for information on specific topics, and generate one-click articles.

The reconstructed Baidu library also adds the function of PPT generation, which can understand whether the views are juxtaposed or progressive, and switch PPT style style with one click, Baidu claims that "far beyond other PPT generation tools on the market."

03 Power B-side

In this demonstration, Baidu also showed some new B-side applications.

Among them, Baidu focused on launching a business intelligence product. Baidu GBI, Generative Business intelligence.

This is a new product launched by Baidu, which is the first generative business intelligence product in China, with the ability to support natural language interaction, cross-database analysis and professional knowledge learning, shortening the data analysis work that business analysts can complete in a few days to minutes.

Baidu GBI products targeting the B-side|Baidu

In the commercial, the question "What is the estimated cost?" What is the price floor without losing money? The customer asked us to complete the delivery within 3 months, can we do it? How long is the fastest? If the competition is right, such as our low price, what can be done?" For this series of related financial analysis, project interaction, and user analysis questions, Baidu GBI can directly give answers through natural language dialogue, and generate illustrated answers.

No expert is required, and no additional operations are required to access data across databases and tables. In addition, companies can also train them to learn professional knowledge and become industry experts.

Another B-side product is Ruliu. After using generative AI for refactoring, such as Flow can generate meeting minutes with one click, summarizing the content of thousands of work groups. Combined with the company's CRM system, propose project background and project discussion for managers. According to personal itinerary, plan work plans, send out meeting invitations, etc.

In addition to enabling the office, Baidu also demonstrated the empowerment of large models for autonomous driving, intelligent cockpit and government intelligent monitoring projects.

Since its release for more than half a year, Wenxin has rapidly iterated to reconstruct Baidu applications, and at the same time is gradually establishing the Wenxin ecosystem.

Baidu also introduced the recently launched Lingjing platform at the press conference. Whether it is personal or enterprise data or applications, it can be quickly turned into a plug-in on the Lingjing platform, and the API can be used to access the ability of the Wenxin large model.

Robin Li predicts the coming AI ecological era|Baidu

Baidu introduced that in the current month since the launch of the Lingjing platform, 27,000 developers have applied to settle in, covering more than 20 fields, including legal consultation, resume generation, brain map production, speaking practice and other native applications in various scenarios. Enterprise private data can be easily and quickly accessed to the capabilities of this state-of-the-art large model without the risk of leakage.

"China has a wealth of application scenarios, and Chinese users are naturally willing to embrace new technologies, and with advanced basic big models, we can build a thriving AI ecosystem and jointly create a new round of economic growth." Li Yanhong said.