Baidu's large model plug-in, the savior of workers?

Text | Zhou Xinyu

Editor | Anita Tang

Image source: Generated by Unbounded AI

On August 16, 2023, Baidu 2023WAVE SUMMIT, a methodology for building a large-scale model development ecology, replaced the model itself as the protagonist.

Generally speaking, Baidu’s attempt to implement applications based on Wenxin’s large model is to follow Microsoft’s path again: releasing supporting plug-ins and Copilot for productivity scenarios, while building a cloud-based developer ecosystem.

For example, the newly released 5 plug-ins correspond to nearly 70 plug-ins connected to Bing and ChatGPT; the launched plug-in development tool set and AI application development kit Comate Stack correspond to Copilot Stack; The office platform “Ruliu” corresponds to Teams connected to Copilot.

The strong market demand and the large-scale model technology in the high-speed iteration period constitute the main contradiction in the application of large-scale models. The application of the model must first face the high cost of iteration.

The solution found by many manufacturers is: plug-in.

Compared with the traditional application development process, plug-ins do not need to develop a complete set of systems, but can provide services through interfaces, and are easy to modify and update. In a sense, plug-ins are the best carrier for AI application innovation.

Based on this idea, Baidu made a series of attempts at 2023WAVE SUMMIT around “plug-ins”, from self-research, to developer community construction, to landing work scenarios.

Ecological construction is another focus that Baidu emphasizes.

As the progress of large-scale model development rapidly moves from research and development to the implementation stage, developers are faced with an increasingly complex industrial chain, and implementation applications also put forward higher requirements for the coordination of upstream and downstream, software and hardware.

As an infrastructure provider, Baidu has iteratively upgraded the Fei Paddle open source framework in terms of large model training, reasoning, hardware adaptation, etc., increasing the training speed of Wenxin large model 3.5 to 3 times, and the reasoning speed increased by more than 30 times . Its purpose is to gather developers from all links of the industrial chain, and build a development ecology based on the large model base and cloud platform.

Wu Tian, vice president of Baidu Group and deputy director of the National Engineering Research Center for Deep Learning Technology and Application, believes, “After half a year, I think the role division of the entire industry will gradually become clear.”

** “Wen Xin Yi Yan” superimposed drawing buff, access to 5 Baidu self-developed plug-ins **

In addition to the “Baidu Search” and “Lanjuan Documents” (formerly ChatFile) released at the Wenxin Qianfan Large-scale Model Platform Media Roundtable Conference on August 2, Baidu also launched the “E Yanyi” on Wenxin Yiyan. Three AIGC plug-ins, “Figure”, “Speaking Graphics”, and “Yijing Liuying”, are respectively used in chart generation, image generation and understanding, and video generation scenarios.

Wenxin Yiyan uses the “Browse Document” plug-in to interpret the document. Source: Baidu

Wenxin Yiyan uses the “Eyan Yitu” plug-in to generate pie charts from the data in the document. Source: Baidu

However, due to high energy consumption, the internal testing of the video generation plug-in “Yi Jing Liu Ying” has not yet been fully carried out. In the future, Wenxin Yiyan will be open to more third-party plug-ins.

Flying paddles “sell water”, launch a large-scale model plug-in development platform

Starting with five self-developed plug-ins, what Baidu wants to do more is to establish a complete developer ecosystem linking upstream and downstream.

Wu Tian believes that the time has come to build a developer ecosystem: "On the one hand, I think the core technology system itself must be mature enough to support the construction of the ecosystem. Wenxin Yiyan has reached this stage; ** on the other hand On the one hand, the biggest role of ecology is to make applications more prosperous.**”

Relying on the model capability of “Wen Xin Yi Yan”, Baidu has launched a large model plug-in development platform for developers. The platform provides a plug-in development tool set (ERINE Bot SDK), which can support the development of plug-ins for information services, tools, and LLM-based innovations.

ERINE Bot SDK. Source: Baidu

At the same time, Flying Paddle AI Studio launched the Xinghe large-scale model community this time. This is like the domestic version of Microsoft Azure AI Studio. Baidu, as the “water seller” of large-scale application, provides developers with development, deployment, marketing and other process services.

Xinghe big model community. Source: Baidu

On “Ruliu”, do articles to improve productivity

What is the “killer” application in the AI era? At present, the industry may not be conclusive. However, the development of AI native (AI Native) applications is the basis of “killer” applications, and it has increasingly become a consensus.

The so-called AI-native applications simply mean that without AI, applications with core capabilities cannot be provided—this undoubtedly puts forward higher requirements for developers: simply iterating on original applications is not enough.

In the 2023 WAVE SUMMIT, Baidu handed over the original AI answer sheet, which was about reshaping productivity. On Baidu’s intelligent work platform “Ruliu”, Baidu launched “Super Assistant” and “Comate X Intelligent Programming Assistant”.

The function of the “Super Assistant” is similar to the DingTalk magic wand released by Ali and the AI assistant “My AI” released by Feishu. It can complete document processing, chat content search, and message sending based on the needs of natural language input.

The “super assistant” completes the travel arrangements. Source: Baidu

“Super Assistant” can analyze documents. Source: Baidu

Similarly, Comate X, which assists in programming, is also based on natural language descriptions and generates code fragments that can realize corresponding functions. Currently, Comate X supports more than 30 languages and more than 10 IDEs (Integrated Development Environment).

For the development of AI native applications, Baidu has also established a developer ecosystem. The AI native application development kit Comate Stack launched this time provides developers with an evaluation platform iue, an AI application development platform IPlayground and a dataset hosting platform iDateSet.

It is worth mentioning that when developing AI applications on the Comate Stack, natural language is still the interface for human-computer interaction. For example, to develop a plug-in that can implement vacation policies, only two steps are required to set up and arrange the process.

The AI application development platform IPlayground supports 2-step generation of AI plug-ins. Source: Baidu

The rapid launch of AI applications means that only Baidu employees and a few big B customers can use the two productivity tools. But at the same time, this is also a signal of Ruliu’s active commercialization.

Li Ying, Vice President of Baidu Group and Chief Information Officer of Baidu Group, introduced that Ruliu will increase its commercial promotion efforts in the future with the help of super assistants and programming tool chains.