Baidu Completes AI Transformation of Its 'Full Suite' with Over 10 Native Applications Launched
-
On October 17, 2023, seven months after the release of Ernie Bot, Baidu conducted a comprehensive review of its AI achievements at the Baidu World Conference.
Compared to the initial launch of Ernie Bot on March 16, 2023, Robin Li appeared much more confident and at ease during the World Conference.
Robin Li (Image source: Baidu)In his opening speech, he announced the upgrade of the Ernie large model to version 4.0 and shared Baidu's assessment of its comprehensive capabilities:
"Its overall capabilities are on par with GPT-4."
The enhanced foundational capabilities of the model have led to updates and restructuring in the application ecosystem. Robin Li emphasized that "understanding, generation, logic, and memory" are the four core capabilities that form the foundation of AI-native applications.
"Out with the old, in with the new" has been the theme of Baidu's initiatives following the release of the Wenxin Large Model 4.0.
On one hand, the Wenxin Large Model has fully upgraded its mobile ecosystem, including search, office, and marketing, with all products now AI-enhanced. On the other hand, as one of the first batch of large model manufacturers to obtain service licenses in August, Baidu has introduced its first self-developed AI-native applications based on the Wenxin Large Model, deploying them in scenarios such as Baidu Netdisk, Ruiliu, Qianfan Large Model Platform, GBI, and Smart Cockpit.
The primary methods for iterating the Wenxin Large Model to version 4.0 include leveraging greater computing power and data, as well as expanding parameter scale. By early September, Wenxin Large Model 4.0 had met the launch standards and was made available for limited testing.
Baidu CTO Wang Haifeng stated that the improvements in the understanding and generation capabilities of Wenxin Large Model 4.0 are comparable, with particularly notable enhancements in logic and memory. The improvement in logic is three times that of understanding, while the improvement in memory is twice that of understanding.
Compared to version 3.5, version 4.0 has enhanced knowledge processing in both input and output stages: during input, the search engine improves its understanding and retrieval capabilities of knowledge graphs; during output, database feedback mechanisms enhance the accuracy of results. It is reported that since March, the efficiency of training algorithms has increased by 3.6 times.
Now, the competition in large models has entered its second phase—focusing on implementation and applications, and Baidu clearly does not want to fall behind. At the World Conference, Qianfan also launched its own 'App Store'—an AI-native application marketplace, providing a trading platform for applications developed based on the Wenxin large model.
Currently, the app store has launched Baidu's self-developed and customer-developed AI applications for five scenarios: smart office, marketing services, industry functions, productivity improvement, and analytical decision-making.
AI-native App Store. Source: Baidu'China has rich application scenarios, and Chinese users are naturally willing to embrace new technologies. With advanced foundational large models, we can build a thriving AI ecosystem and jointly create a new round of economic growth,' summarized Robin Li.
To build applications, tools are essential. In this regard, the Qianfan platform has launched an AI-native application development platform. This development platform is divided into two parts: the application framework and application components, providing "showrooms" for application development and tools, respectively.
As of now, the Qianfan platform has onboarded 42 mainstream large models, including Llama2, from both domestic and international sources. Over 17,000 enterprises have already utilized APIs from various models to develop applications.
Meanwhile, starting today, enterprise clients can apply for the beta test of the Wenxin Large Model 4.0 API on the Qianfan platform.
Under the application framework, Baidu provides Retrieval-Augmented Generation (RAG) and Agent capabilities. This marks Baidu's first introduction of an RAG framework, which works by segmenting documents into slices and unifying them into trainable data via vector databases. Based on RAG, enterprises can develop applications like document Q&A tailored to business needs using large models.
Regarding RAG pricing, Baidu Group Executive Vice President and President of Baidu Intelligent Cloud Business Group Shen Dou revealed that document Q&A development costs only a few hundred yuan, with single-article analysis priced at just a few cents.
Robin Li, Baidu's CEO, considers plugins a special type of AI-native application—low-barrier, user-friendly, and capable of enhancing large model functionalities. Previously, Baidu Cloud launched the Lingjing Matrix plugin platform, which has already attracted 27,000 developer applications across over 20 fields, including legal services and language practice.
Industry consensus since ChatGPT's release identifies search engine revamps as one of the most immediately viable implementation scenarios for this technology.
Baidu Search can now provide integrated answers at the top of search results through its large language model. For numerical or trend-related queries, it can also generate corresponding dynamic charts using multimodal capabilities. After answering, Baidu Search recommends more relevant follow-up questions and supports multi-turn interactions where users can continuously ask further questions.
Similar to ChatPDF, Baidu Wenku's AI transformation primarily targets users' document comprehension and organization needs. During a recent communication session, Wang Ying, Baidu's Vice President and Head of Interactive Entertainment & Vertical Categories, revealed that Baidu Wenku now implements 2-3 new feature iterations weekly. In the past month alone, it has rolled out functions including document-to-PPT generation, text-to-chart creation with analytical insights, PPT chart generation, and intelligent document summarization with Q&A capabilities.
Notably, the new Baidu Wenku can not only process document content but also automatically generate PPTs in different styles based on article content. Robin Li noted that the revamped Baidu Wenku has already attracted over 2 million users.
For content creators, Baidu has launched an AI creation engine within its app, generating copy and images tailored to various needs such as marketing, guides, and content sharing.
In its renewed push into live-streaming e-commerce, Baidu has introduced the AI Native marketing platform 'Qingge,' built on its ERNIE model. 'Qingge' integrates AI tools like the intelligent marketing chatbot 'Brand BOT' and offers businesses the 'AIA Marketing Science' methodology for the generative AI era. It also includes 'Yangji,' a commercial engine providing user insights, analysis, and content creation tools.
Previously, Baidu Netdisk rolled out 'Cloud One File Assistant,' an in-app AI assistant for managing cloud-stored content. Users can input natural language commands to locate documents, videos, and other files. Unlike AI assistants focused on language and image processing, 'Cloud One' excels in video handling, offering features like summarizing video content, adding subtitles, exporting subtitles as documents, and generating titles for transcripts.
The most significant upgrade to Baidu's 'Cloud One' capability lies in its integration with the WeChat ecosystem, enabling intelligent summarization of any public account articles shared by users. According to Baidu, Baidu Netdisk's Cloud One has already reached 20 million users.
Following DingTalk and Feishu's launch of AI assistants for workplace scenarios based on large models, Rulu has also been enhancing its AI functionalities. Naturally, the features of these three platforms are quite similar, primarily targeting major scenarios such as meetings, leave management, business trips, and knowledge bases.
With WeCom, Feishu, and DingTalk already holding over 90% market share, Rulu faces significant challenges in making a comeback. Baidu's strategy is to first enable Rulu's super assistant to handle more complex tasks and possess autonomous decision-making capabilities.
He Wei, Director of Baidu's Intelligent Office Platform Department, believes that while the previously launched AI tools on Rulu were highly functional, they didn't constitute truly intelligent work. Baidu's data shows that the 'AI Meeting Insights' feature on Rulu has increased meeting content readership by 3.5 times.
Business Intelligence (BI) is essentially the interaction between analytical systems and databases, serving as an indispensable part of commercial data analysis. Traditional BI systems have high barriers to entry, requiring users to adapt to the system and input professional queries to obtain accurate data.
Previously, Baidu's Sugar BI had already integrated AIGC capabilities, launching "Wenxin Wenshu Sugar Bot," which allows users to perform data retrieval and report generation through conversational interfaces.
With the support of large language models, Baidu has now introduced GBI (Generative BI). The "generative" aspect of GBI is reflected in its support for natural language conversational interactions, lowering the barrier to entry.
GBI. Image source: BaiduAt the same time, GBI offers strong scalability, enabling connections to data sources across various industries. To address industry-specific jargon, enterprises can train GBI based on professional knowledge databases to meet specialized analytical needs.
GBI's advantages also include its diverse access methods. Companies can upload data to the GBI platform for querying and analysis or opt for a full private deployment of GBI capabilities. For businesses with highly sensitive data but limited budgets, GBI can provide only its intelligent query interface, allowing direct integration with existing systems like OA and ERP.
By integrating hundreds of millions of POI (Point of Interest) data points and trillions of traffic insights, Wenxin Yiyan has enhanced the functionality of Baidu Maps.
At the World Conference, Baidu Maps launched intelligent recommendation and planning features for travel destinations, meeting points, transportation modes, and routes.
Baidu Maps AI Guide. Source: BaiduThis time, Baidu also brought the Jiyue 01, equipped with the Apollo AI Cabin model, to the World Conference.
Jiyue 01. Source: AuthorThe most notable feature of experiencing the Jiyue 01 is that voice has become the primary mode of interaction inside the vehicle. The AI can recognize the needs of different users in noisy or multi-person conversation environments, applying to scenarios such as navigation, vehicle usage, entertainment, and information services.
According to Su Tan, General Manager of Baidu's Intelligent Vehicle Business Unit under the Intelligent Driving Group (IDG), the human-machine relationship in car cabins during the era of large models is essentially an interaction between humans and virtual entities.
Leveraging the Wenxin large model as its foundation, Baidu has introduced the Apollo Intelligent Cabin Large Model and its development toolchain, building upon Apollo's vast dataset from millions of smart vehicles across various scenarios. Currently, besides the JiYue 01, Apollo's in-car voice products powered by this model have been mass-produced in vehicles like the Cadillac LYRIQ, Buick E5, Geely Galaxy L7, and Geely Galaxy L6.
The integration of large models into Xiaodu has spawned three more specialized hardware categories.
For home companionship scenarios, Xiaodu launched the Tiantian Home Robot, which intelligently transforms between robot, tablet, and speaker modes to meet needs like caregiving, entertainment, and scheduling. The Tiantian Home Robot is slated for release in spring 2024.
Meanwhile, for educational settings, Xiaodu introduced the Qinghe All-in-One Learning Machine, while for users with higher audio quality demands, it unveiled the premium 'Tiantian Casa' smart speaker line.
