The Heart of the Machine Original
Heart of the Machine Editorial Department
The vigorous wave of large models seems to be entering a period of torment.
In mid-June, Microsoft made the surprise news that they would be discontinuing their Copilot GPTs service, which had only been available for three months, in July.
Copilot GPTs can be understood as intelligent assistants created for various needs, such as groceries Copilot GPT can make a list of purchases based on the meal plan you send it. This service is very similar to OpenAI's GPT Store. And, likewise, the latter failed to become a market darling: digital solutions company CTOL. A report from Digital estimates that GPT Store has less than 2% of the traffic to the normal ChatGPT UI. You know, when this product was released, many people shouted that the "iPhone moment" of AI has arrived, and this kind of GPTs is considered to be an effective way for large models to benefit the masses and improve social productivity.
While announcing the shutdown, Microsoft also revealed that they have decided to shift GPT's application center to commercial and enterprise-level scenarios. Whether this shift will succeed remains to be seen. But across the ocean, the same route is unlocking the huge potential of the AI application layer.
It is undeniable that there is still a gap between Chinese AI and the United States in terms of models, algorithms, and computing power, which has become a consensus. Just like in the early days of the mobile Internet, the iPhone and some popular apps on it also first appeared in the United States.
But then things changed: China's mobile Internet has achieved remarkable success in many aspects, including application ecology, business model, profitability, etc. Even in some areas, Chinese Internet applications have performed better than their American counterparts. Super apps such as Didi, WeChat, Alipay, DingTalk, and Meituan cover almost every aspect of Chinese life.
Today, one of the major directions of AI models is to build killerAPP in a thriving application ecosystem based on the achievements of the mobile Internet era. In China, this soil is very fertile. Taking this as an opportunity, developers can more quickly combine the latest capabilities of AI large models with ever-changing real-world scenarios, and launch, distribute, and even monetize AI applications as soon as possible. From this point of view, the wrestling between the AI application layer in China and the United States may show a similar trajectory to the development of the mobile Internet.
Can China's AGI repeat the story of the mobile Internet?
What is the difficulty in the AI application layer that Microsoft and OpenAI can't play?
The so-called iPhone moment has been teased many times in the AI space over the past year.
The reason why this moment is expected is that in the last revolution, it marked the arrival of the mobile Internet era. Until then, surfing the internet was often seen as an activity, but the advent of the iPhone has made it possible for us to be online almost at any moment, like breathing. The way people live and how they work and live has changed dramatically: With an intuitive touch interface, a diverse ecosystem of applications, and seamless hardware and software integration, iPhone has redefined smartphones and made it easy for everyone to take advantage of the mobile internet. At the same time, various mobile applications such as Didi, WeChat, DingTalk, Taobao, etc., have greatly improved the efficiency of users and enterprises, and changed the traditional business model and lifestyle.
Therefore, more than a decade later, it is expected that the "iPhone moment of AI" will bring about similar changes, greatly improving the efficiency of life and work.
Microsoft, OpenAI's GPTs are also working in this direction to a certain extent, but the results are not ideal. Many OpenAI users have reported that the quality of GPT in the GPT Store is different, and many GPT functions are simple and cannot really save time and improve efficiency. For enterprise users, this situation is only getting worse, as OpenAI is not a company in an enterprise use case. The needs of enterprises are more diverse, and they also have a large number of data assets that need to be connected, which cannot be solved by simply setting up a prompt and then answering questions with an AI assistant.
OpenAI has also proposed solutions to these shortcomings, such as the recent acquisition of a search analytics startup called Rockset. By integrating Rockset's real-time search and analytics technology, OpenAI wants to provide enterprise customers with more powerful data processing and analysis capabilities to further enhance the application value of its AI models. In other words, they want to be closer to the enterprise and the scene through this move, and make their large model grow on the company's data.
And this goal is precisely the longboard of many Chinese Internet technology companies. Among these companies, if you want to find a target to observe Chinese AI applications, DingTalk has long followed a similar AI agent route with OpenAI and has grown in thousands of industry scenarios, which is a sample worth observing.
DingTalk – a noteworthy sample of China's AI application layer
Zhou Zichao is a digital engineer at Modern Agriculture, Animal Husbandry and Food Group Tieqi Lux. In order to help cooperative farmers breed scientifically and detect risk factors such as "swine fever" in a timely manner, he developed an AI assistant called "Smart Farming Expert".
On the surface, this AI assistant is no different from what we've seen in the OpenAI GPT Store. But the difference is that he developed it on the DingTalk platform, and DingTalk is a tool that the company has been using since 2018, and a lot of the company's data is on it. With the help of this AI assistant, farmers can not only consult farming knowledge anytime, anywhere, but also see the data differences between their own pig farms and their brother pig farms and seek improvements. Moreover, because there is a need before development, this AI assistant has been a user since its birth, and there is no need to wait for their application scenarios like many GPTs in the GPT Store. In addition, there are many mature applications and tools that can be used on the DingTalk platform, which makes its interaction not limited to the simple form of one question and one answer. This is the advantage accumulated by domestic technology companies like DingTalk in the mobile Internet era, and it is also a large part of the reason why they can go through the road of AI assistants.
It is reported that in just two weeks after the launch of the "Smart Breeding Expert", it has attracted more than 100 people to use it, and the number of calls has exceeded 5,000 times. In fact, similar things are also happening in enterprises that use DingTalk to carry business, such as Gold Medal Cabinet, Western Airport, Ligong Industry, Yousao Food, Belle, and Jiawo.
Loading...
On the one hand, the formation of this situation is due to the inherent advantages of DingTalk's "application scenarios and business data", and applications similar to GPTs grow on platforms like DingTalk without a sense of suspension; On the other hand, it also benefited from the speed and strength of DingTalk's attack in this direction.
As early as April 2023, when the "100 model war" had just started, DingTalk quickly entered the market in an environment full of uncertainties, announcing that it was fully intelligent, and all products would be reshaped with large models. In the following more than 100 days, DingTalk's 17 product lines completed intelligent reengineering. On August 22, DingTalk opened the Intelligent Base (AIPaaS) to ecological partners and customers, marking that DingTalk Intelligence has fully entered the ecological layer. In November, more than 500,000 enterprises joined the DingTalk AI invitation test, and DingTalk AI was launched, with 17 product lines, 60+ scenarios, and nearly 100 AI skills fully open to users for testing, which is the first national-level work application in China that fully opens up AI.
This speed of decision-making and execution has enabled DingTalk to incorporate cutting-edge concepts such as AI Agent into its product thinking early on. Therefore, in January 2024, just two months after OpenAI announced the GPT Store, DingTalk launched its own AI assistant, and combined its own advantages in applications and data to enhance the perception, memory, thinking, and action of AI agents. The results of this enhancement can be seen in DingTalk's most recent major version update (version 7.6).
The new version gives people two main feelings: one is that the AI assistant is now more anthropomorphic, it can know who you are and what you have done, and it is closer to "customized".
For example, we always get some long documents in the work group, and regular AI assistants can usually organize them into minutes, but DingTalk's AI assistant goes a step further and can organize them according to your role in the company. For example, when a marketer uses an AI assistant's speed reading, the AI assistant will focus on the marketing part of the financial report. When used by production personnel, the AI assistant focuses on analyzing data on production costs.
This feature actually activates all the information you have on DingTalk, including your name, position and other personal information, the company's organizational structure, your work content in the past week, the chat summary of the work group, etc. That way, it can really give you what you need, like a weekly report, rather than a weekly template.
The second is that the things it can do and analyze are more complex, and multiple agents can be used together to complete a multi-step workflow.
For example, if your company wants to plan a large-scale marketing campaign, it needs a comprehensive solution that integrates public opinion monitoring, market analysis, and creative ideation. On DingTalk, this task will be planned by the AI assistant into a workflow composed of several sub-tasks, and then relayed by different AI assistants (market analysis assistant, creative idea assistant, event planning assistant), during which the AI assistant can also call the RPA capabilities on the DingTalk platform.
In this simple example, we can actually see several agent design concepts mentioned by Ng in his previous talk, including tool use, planning, and multi-agent collaboration. Andrew Ng is particularly optimistic about the prospects of multi-agent collaboration, pointing out that the advantage of multi-agent collaboration is the ability to decompose complex tasks into subtasks and complete them through collaboration between agents. This approach has been proven to be effective in practice and is able to provide better performance than a single agent. DingTalk's AI assistant is the first to prove this in the enterprise scenario.
This kind of AI assistant with memory, interaction, and task execution capabilities is actually closer to the concept of Large Action Models (LAM), which is hotly discussed in Silicon Valley. Compared to traditional LLMs, LAM not only has advanced language understanding and generation capabilities, but also can translate these capabilities into actual actions and task execution, providing a more comprehensive and automated intelligent experience in a wider range of application scenarios. This will make DingTalk's AI assistant more useful in enterprise scenarios, making it favored by more enterprise users. This kind of rapid experimentation and implementation of cutting-edge technologies is also a major reason why DingTalk AI is worth paying attention to.
The opportunity for AI in China is at the application layer
At yesterday's 2024 DingTalk Ecosystem Conference, DingTalk released a set of important figures: the number of DingTalk ecosystem partners has exceeded 5,600; The total number of AI assistants created has reached 500,000, and the number of AI assistants on the AI assistant market has exceeded 700. DingTalk AI has been invoked more than 10 million times a day. At a time when the large model has entered the stage of difficulty in landing, this is undoubtedly very encouraging.
In an interview with Tencent News "Periscope" in March this year, Zhu Xiaohu, managing partner of GSR Ventures, predicted that there are many opportunities for innovation in China's AI at the application layer, because China far surpasses the United States in terms of data and application scenarios. Although the underlying model of Chinese AI is not powerful enough, you can "add more things on it", you can do value-added services on it, and you can monetize it for customers immediately. This is the unique advantage of China's AI in the application layer, and it also makes companies like DingTalk have high hopes.
In the face of such expectations, in addition to its own efforts, DingTalk is also trying to rely on the strength of the community.
Due to the large number of industries faced by DingTalk and the wide variety of user needs, it is difficult for a general large model to meet all the needs. Therefore, at the conference, DingTalk announced the opening of the base model, and will join forces with Orion Star, Zhipu AI, MiniMax, Tongyi, Dark Side of the Moon, Baichuan Intelligence and Zero One Everything Seven large model partners to collect the large model "Dragon Ball" and jointly create the most open AI ecosystem in China. For DingTalk, this move is conducive to improving the capabilities of its underlying model and improving the user experience. For the above-mentioned large-scale model enterprises, DingTalk is not only a connector, but also an amplifier, which is conducive to them to quickly find the application scenarios suitable for their own large-scale models and maximize their own value.
We also hope to see these companies work together to repeat the story of the mobile Internet in China.