DingTalk is becoming less and less like "DingTalk", and under the reshaping of large models, DingTalk is evolving to a new "AI Agent platform".
What is DingTalk?
Many people's understanding of it may still stay in the "office software", clocking in at work, work collaboration, etc. In fact, DingTalk is becoming less and less like "DingTalk".
In the past year, since the announcement of comprehensive intelligence, DingTalk is evolving to a new "AI Agent platform" under the reshaping of large models. In particular, on April 18, DingTalk launched the AI Assistant Market (AI Agent Store), where AI assistants developed by enterprises, individual users, and developers can be shared with more people, and the future direction of this route has become clearer.
This is also a way for the industry to be upgraded from "single-point application + AI" to "AI assistant + AI native application" after DingTalk is determined to "All in AI". Its significance, according to Ye Jun, president of DingTalk, is that DingTalk will become the birthplace of the next Midjourney and the next Pika.
He also said that based on the judgment that the AIGC wave has entered productivity and application scenarios, AI Agent is the best AI application entrance. DingTalk will build an AI intelligent assistant platform and application market, so that users, developers, and ISVs can develop personalized AI assistants with a low threshold on DingTalk, so that more excellent AI applications can emerge on DingTalk.
一、卷向AI Agent
This is no small goal. At present, the entire large-scale model track is still in the fierce knockout stage. OpenAI, Baidu and other leading domestic and foreign manufacturers are not confident to enter the next round of the competition. For example, Robin Li has expressed his "anxiety" in public many times, he said at an event, "In the past nearly a year, I have seen that the main excitement of the media, society, and the public is still on the basic model, and has not moved to AI native applications, which makes me more or less anxious."
From this point of view, DingTalk is squeezing the table in its own way in this round and winning the opportunity for the next round. So, why did DingTalk choose to "open the book" on the AI assistant, that is, the AI Agent?
In fact, it has only been more than a year since the concept of AI Agent was born. The industry-recognized origin is the release of Auto-GPT in March 2023, which OpenAI scientist Andrej Karpathy called "the next frontier of prompt engineering." However, the early products after the explosion of this large model are still very immature as AI Agents. When OpenAI upgraded some of its features, Auto-GPT looked more like a "fool" and could not effectively meet the needs of individuals or businesses.
But it is like opening a gate, more AI agent development platforms have emerged, and the industry has begun to generally recognize the importance of AI agents. A typical example is Bill Gates, who emphasized in a personal blog post at the end of last year that "Android, iOS, and Windows are all platforms, and AI Agent will be the next platform." He also predicts that with the increasing popularity of AI technology, various applications will be replaced in the next five years, and mobile phones or computers can handle requests based on instructions given by users in everyday language. In the not-too-distant future, anyone who goes online will be able to have an AI-powered personal assistant, known as an "AI Agent".
Meta founder Mark Zuckerberg has also said that he sees an opportunity to "introduce AI Agents to billions of people in a useful and meaningful way."
In addition, Andrew Ng, the former chief scientist of Baidu, also mentioned that "all people who are engaged in artificial intelligence should pay attention to AI Agent". He believes that through Agent, the types of tasks that AI can perform will be greatly expanded, and even large models with lower parameters but faster response can perform better than models with larger parameters through more rounds of iteration.
In this regard, Ng's team also asked the large model to write some code and run it, and finally compared the performance of different models and workflows: the accuracy rate of the GPT-3.5 model is 48%, the accuracy rate of the GPT-4 model is 67%, the GPT-3.5+Agent effect is higher than that of the GPT-4 model, and the effect of GPT-4+Agent is much higher than that of the GPT-4 model.
The rapid development of the AI agent market has also exceeded many people's expectations. According to MarketsandMarkets, the global autonomous AI and autonomous agent market revenue will exceed $4.8 billion in 2023 and is expected to reach around $28.5 billion by 2028, with a CAGR of 43.0% from 2023 to 2028.
It is not difficult to see that AI Agent is attracting the attention of the whole industry, and it is constantly iterating its capabilities to make it more accurate. It can even be said that an era of AI agents is coming. For DingTalk, embracing AI at the moment and constantly falling into it, rolling to AI Agent, is also hoping to "vigorously produce miracles" and follow the trend drumbeat.
Second, the short board becomes a long board
Not only that, DingTalk is also further exploring the topics that the industry is most concerned about: how does AI Agent transform from imagination to productivity? According to the information Tang Chen understands, DingTalk's exploration falls on specific actions, which can be divided into two stages:
The first stage is self-AI transformation: in April 2023, DingTalk announced that it will be fully intelligent, and all products will be reshaped with large models, and in the following more than 100 days, DingTalk's 17 product lines will complete intelligent reengineering.
The second stage is to move towards an open platform: after completing the self-AI transformation, DingTalk began to open the intelligent base (AI PaaS) to ecological partners and customers, and launched an innovative product based on AI PaaS, "Digital Employee", to continue the transformation to the ecological layer. Since then, the DingTalk personal version of AI has been updated, the 7.5 version of DingTalk has been released, and the AI assistant market has been launched, and the long-term goal of DingTalk AI has surfaced, and finally it has clearly fallen on the AI Agent platform.
According to DingTalk, the DingTalk AI assistant is built on a large language model and will have the ability to perceive, remember, plan and act. More importantly, the AI assistant can be seamlessly combined with DingTalk's rich applications, third-party applications, and enterprise-built applications, and can also be disassembled and orchestrated through workflows to disassemble and orchestrate the process of AI execution tasks at the time of creation, so that the AI assistant can take over the corresponding operations and perform more complex tasks.
In other words, this capability is not only within DingTalk, it also has the ability to execute across applications, and users can create an AI Agent that "shuttles freely" between DingTalk, third-party and enterprise-built applications according to their own ideas and needs.
At present, DingTalk presets official AI capabilities such as intelligent Q&A, image generation, content creation, and data statistics for the creation of AI assistants, which users can use with simple configuration. For developers and IT teams, it supports the rapid development of customized AI capabilities through DingTalk AI PaaS, and connects with original systems such as SaaS applications and local systems through DingTalk's open APIs and connectors.
Some media described these features as "this is a brand new, and even looks very different from the previous DingTalk", and its AI capabilities are built into DingTalk, which can be switched at any time to avoid complicated downloads. At the same time, it provides a number of functions such as AI dialogue, AI drawing, etc., which are also the capabilities of AI Agent but full of to C flavor.
In DingTalk's view, Al assistant will become the mainstream form of future applications, and a rich value business exchange model must be formed. DingTalk's official Al assistant, built by enterprises, ecosystem partners and developers, will become the three main components of the DingTalk AI assistant market. The reason why DingTalk dares to take such a big step lies in the attributes of DingTalk itself. Ye Jun introduced, "Generally, a new technology is easier to land on the production side, on the tool side, and on the B-side. ”
Zhou Hongyi, chairman of 360, has said many times that with the development of open source large models, large models have begun to "step down from the altar", from the business of selling "atomic bombs" to the business of selling "tea eggs", and the real barriers have changed from technology to scenes and data. He also pointed out that enterprises should not use large models to advance rashly, but to use AI to gradually transform their business, step by step, and in practice, they should split the specific analysis of scenarios, find the right entry point in the business process, and choose the business link that matches the mature ability of the large model.
According to his line of thought, Agent is a powerful technical solution to solve complex scenes, and even must rely on complex scenes to "survive". The power of an agent is maximized when a complex but well-defined problem is presented in front of you.
It is worth mentioning that in the past, the scenes of DingTalk were too scattered and too complex, but in the AI era, they have become the advantage of the scene for its landing, which is tantamount to a reversal of the short board to the long board.
3. Nail the table
According to the latest data, DingTalk has nearly 200 AI assistants on the shelves in the first batch, including C-end and B-end assistants, covering creative design, learning and education, operation and promotion, sales and customer service, personnel administration, finance and taxation legal affairs, e-commerce and foreign trade, manufacturing, enterprise services and other fields, of which more than 30 are from industrialized professional scenarios.
This also shows that DingTalk has a natural AI application scenario, and it reversely docks with large models to make AI assistant products, rather than holding a hammer to find nails. To put it simply, with application scenarios and business data at the top, DingTalk has developed its own differentiated play.
Back to a key question: What are the advantages of building an AI agent on DingTalk compared with building it directly on the base model? According to the capabilities of DingTalk's AI assistant, there are four main performances:
First of all, the AI assistant can be deeply bound and combined with DingTalk. For example, the AI assistant and the DingTalk field are fully integrated, and they can be added to the address book, pulled into the group chat, @ in the document, added to the processor list in the OA approval, and pulled into the audio and video conference just like the organization members. On this basis, the AI assistant can perceive the identity/position/responsibility of the user and related people, as well as the context of each field of DingTalk: for example, when being pulled into a group chat, the group members and group identities of the current group. With more accurate environment awareness, the AI assistant will be significantly improved in terms of intent recognition, skill routing, and inference planning.
Second, solve the problem of traffic or rationality. DingTalk itself has the needs and scenarios of all walks of life, users naturally exist in the scene, and there are needs in the scene, and the current problem such as GPTS and large models is the lack of clear user needs, and users only go to AI when they have needs. For developers, it also means that potential users already exist.
Third, the DingTalk AI assistant market is a further upgrade of DingTalk's AI-oriented open capabilities, and integrates with the original open connectors, APIs, low-code and other systems, rather than an independent open system. DingTalk's original open capabilities, such as openAPI, connectors, data asset platform, 1000w+ low-code applications, and 5000+ ISVs, have verified the rationality of the ecosystem's business path.
Fourth, it is difficult for ToB to exist in a single phenomenal application, but tens of thousands of assistants in different colors and industries to meet specific user groups. This determines the DingTalk AI assistant market, does not make full recommendations, only recommends selected AI assistants, and has more industry attributes, action capabilities and professional capabilities.
These have also become the core starting point of DingTalk's technical adjustment, that is, the cost of allowing people to build these agents on DingTalk, the cost of release, and the cost of use are all minimized. An agent can be developed through natural language dialogue, and after development, it can be pulled to the DingTalk group for use.
Obviously, DingTalk is building a basic platform for AI Agents to "network" agents similar to stand-alone applications to achieve resource interoperability, provide users with rich resources, and lower the threshold for AI applications.
Now it seems that in the year of AI, DingTalk is indeed becoming less and less like "DingTalk", it has become an AI application platform. This is also a new role that DingTalk is trying, and using it as a focus to send itself to the table of the large-scale model application competition.
References:
Silicon Star Pro, "An "Office Software" Is All in AI? No, DingTalk's ambition is even bigger than that"
Columnist
Tang Chen, WeChat public account: Tang Chen, everyone is a product manager columnist. Content links, insights and interpretations, focusing on Internet technology and business stories.
This article was originally published on Everyone is a Product Manager. Reproduction without permission is prohibited
The title image is from Unsplash and is licensed under CC0
The views in this article only represent the author's own, everyone is a product manager, and the platform only provides information storage space services.