图片来源:unsplash
Remember the earliest AI Agents?
In April 2023, less than a month after its release, Auto-GPT received 100,000 stars on GitHub, which OpenAI scientist Andrej Karpathy called "the next frontier of prompt engineering." Its latest Star count is 160,000, but the vast majority of it comes from before ChatGPT was released to the upgrade.
The reasons are varied, the first is the upgrade of the base model's capabilities, when OpenAI upgraded and updated GPT's Browsing, Code Interpreter and Plugins functions, Auto-GPT looked like a "fool".
At the same time, the value of Auto-GPT as an AI agent is too shallow, and the simple logic is to let the large model repeatedly decide what to do, and at the same time feed back the results of its thinking, which cannot meet the needs of individuals and enterprises at all.
Later, more AI Agent development platforms appeared, and the giants and emerging startups with heads and faces all stepped into the same track, and the industry generally recognized the importance of AI Agent, and AI Agent entered the early stage of contention.
Recently, Ng mentioned that "all people engaged in artificial intelligence should pay attention to AI Agent", in his opinion, through Agent, the types of tasks that artificial intelligence can do will be greatly expanded, even with lower parameters but faster response of large models, through more rounds of iteration, can also be better than larger parameter models.
The accuracy of the GPT-3.5 model is 48%, the accuracy of the GPT-4 model is 67%, the GPT-3.5 + Agent effect is higher than that of the GPT-4 model, and the performance of GPT-4 + Agent is much higher than that of the GPT-4 model.
In terms of market size, MarketsandMarkets believes that the revenue scale of the global autonomous artificial intelligence and autonomous agent market will exceed $4.8 billion in 2023 and is expected to reach about $28.5 billion by 2028, with a compound annual growth rate of 43.0% from 2023 to 2028.
All of these trends seem to indicate that an era of AI agents is coming. But the industry is more concerned about what preparations need to be made for AI Agent to transform from imagination to productivity, what engineering capabilities to combine, and most importantly, how to avoid the Auto-GPT-style dilemma.
"Scene, scene, scene"
If finding a good model is the goal, who defines what a good model is?
Especially for enterprises, a large model in a critical scenario can have a stable and outstanding effect, which is more valuable than the similar effect in multiple scenarios.
In the era of large models, AI agents are the carriers that solve requirements, and the starting point of this carrier is the scenario. Different scenarios will have different needs, and users are the source of needs, and in the same scenario, because of different users, the needs will be very different.
Scenarios, users, and requirements are the three elements of the product, and they are also the core problems to be solved by the productization of AI Agent. Breaking it down, there are many users who have a demand for AI Agent, and it is quite time-consuming and laborious for them to develop AI Agent from scratch, and it is difficult for most technology providers to meet the specific and micro scenario needs of a customer if they want to develop AI Agent.
This is currently a rift between customers and technical service providers, and customers who are willing to pay for large models and AI agents have very clear scenarios, but they do not want to "educate" large model manufacturers from scratch, and the two sides are getting closer to each other, provided that the customer wants the other party to take a big step.
图片来源:unsplash
In the process of communicating with many customers and manufacturers, Titanium Media App found that the pure large model effect is no longer the first consideration, and the comprehensive AI Agent creation process, the customer hopes that the manufacturer understands the industry know-how, has enough data accumulation, and preferably creates the AI Agent with a low threshold.
Zitui, the person in charge of DingTalk AI products, said that from the perspective of the platform building agent platform itself, the overall basic capabilities are not very different, and the future differentiation should first depend on whether the characteristics of the agent and the platform itself can be well combined, and the differentiated capabilities of the platform can be fitted to the development link of the agent itself.
For example, the knowledge base, most of the AI Agent development platforms, the knowledge base is mainly based on local files uploaded by users, which is a static knowledge base, and the knowledge base of DingTalk can be associated with DingTalk's online documents, which is equivalent to the AI assistant continuously updating knowledge. DingTalk's data capabilities, collaboration capabilities, time perception capabilities, and scenario integration capabilities are very rich in AI itself.
Just like the familiar saying "location, location, or location", whether it is the participants in the DingTalk AI Assistant Competition, or DingTalk's own intelligent transformation, as well as DingTalk's online AI assistant products and market, the enlightenment to the industry is that the key to AI Agent is "scene, scene, or scene".
A chance for the AI Agent to run out first
There is still some distance between the imaginary AI Agent and the real AI Agent, in the imagination, the AI Agent has the ability to perceive, remember, plan and act, as well as the ability to perform tasks across applications, but at present, most of the AI Agents have not been able to meet the actual needs, which is also the original intention of the DingTalk AI Assistant Competition, so that a group of AI Agents can run out first.
These AI assistants all have one thing in common, the incision found is small enough to achieve good results in specific scenarios. Taking the AI assistant of the gold medal cabinet - Zhixisheng 1.0 as an example, when Chen Zhiyong, the CIO of the gold medal cabinet, first came into contact with the large model, it was only used to check the information and quickly output some content, and he did not think that the large model would have any substantial intersection with his own business.
It wasn't until Chen Zhiyong began to hear about AI Agent that large models seemed to be able to meet the needs of some enterprise scenarios. "We also wanted to try it based on our demands, but we didn't expect to find that it could really be realized. ”
Gold Cabinet has more than 4,000 offline stores and dealers across the country, including store owners, shopping guides, investment managers, operation personnel, etc., and other roles, often to pass on the questions raised by consumers to the headquarters, these questions about delivery guarantee, product quotation, order flow, marketing policy, tripartite operation, online operation, investment meeting, learning and training and other scenarios, complex but just needed.
For the headquarters, the gold cabinet first to arrange customer service docking, and then the internal correspondence to send a large number of manpower to solve different consultations, sometimes a delivery guarantee problem, may require ABC multiple roles, customer experience is not good, gold cabinet staff are very tired, the timeliness and level of solving the problem is also uncertain.
In September 2023, Gold Medal Cabinet officially established the AI assistant project. Previously, Gold Cabinet had several vertical systems, and the answers were scattered across different systems, but the AI assistant provided a new idea. "We only need to tell 'Xiaojin' (AI assistant) what needs now, and Xiaojin can output results, and DingTalk AI PaaS has already achieved system opening in the early stage, including ERP, WMS and other data, and the AI assistant's thinking ability and action are also stronger. Chen Zhiyong said.
2024 DingTalk AI Assistant Competition
In March this year, DingTalk held an AI assistant competition to solicit AI assistant works based on DingTalk for enterprises, ISVs and individual contestants in the whole society. By the end of March, a total of more than 700 entries had been received, and a total of 30 entries from the three tracks had entered the top 10, and a final roadshow and on-site awards were held in Shenzhen on April 23.
In the end, the "Public Security Government Assistant" of Hangzhou Municipal Public Security Bureau and the "Enterprise Site Selection AI Assistant" of Liye Cloud (Beijing) Smart Technology Co., Ltd. stood out and won the first prize in the enterprise track and office track respectively. The college life track was won by the "Cyber God of Wealth", and the results were voted by the audience after the trial in the exhibition area.
From the perspective of the DingTalk platform, he is also looking for a way to establish the AI assistant business model. Zitui said, "DingTalk pursues valuable scale, the core is not to be an AI assistant for the sake of scale, AI products are different from other Internet products, most Internet products are with the expansion of the user base, the platform effect is amplified, and the cost is getting lower and lower." But for now, as the number of AI assistant users increases, the cost will become higher and higher. ”
He added that the particularity of AI assistants makes DingTalk think about how to recycle business, gradually deepen the value of the platform, creators can get benefits, users are willing to pay for it, and the platform can continue to reduce costs or even realize benefits.
The best story requires the most elements
The large model is a good story, but only a few companies are qualified to tell it, just like the operating system in the mobile Internet era, AI Agent is a better story, similar to the Apple App Store, when everyone wants to tell the same story, the explicit and implicit requirements of AI Agent have also emerged.
"DingTalk itself is a platform-level application company, and the stronger the ability of the base model in the future, the better the blessing of DingTalk, the higher the possibility of creation, and the less investment in engineering. However, the unique scenarios, data, and actions capabilities of the DingTalk platform are also good, which are elements that model manufacturers do not have. Zitui said.
At present, many users have taken the lead in using DingTalk AI assistant, and have transitioned from the toy stage to the tool stage. Titanium Media learned that first, the team of CIO of large enterprises, based on the understanding of AI and relatively clear user needs, the AI assistant and business scenarios are deeply integrated.
The second is data analysis and business insight-related scenarios, a large number of enterprises are doing similar practices, based on permission design and business system integration, AI assistants can achieve very complex and time-consuming operations in the past.
There are also creators who are more college and may lack depth of scenes, but they often have unique insights and are willing to practice, such as the children's observation AI assistant won the second prize in this competition.
This is also a year of drastic changes in DingTalk driven by external forces, and DingTalk must first transform itself if it wants to do a good job as a platform. "The more industries we serve and the more customers we cover, the more we will find that we can't serve every scenario on our own. For example, manufacturing, process manufacturing, lean manufacturing, photovoltaic manufacturing, etc. involve many fields. Therefore, DingTalk is gradually becoming intelligent, making a differentiated AI platform, opening up platform capabilities and differentiated scenarios, and hoping that more creators will build assistants for their scenarios. Zitui said.
In the past year, DingTalk's AI products have gone through the process of changing from "+AI" to "AI+", connecting to the Tongyi large model, intelligently transforming DingTalk's own scenarios, fully accessing AI for documents, audio and video, schedules, etc., and using AI features to upgrade.
Subsequently, DingTalk realized that it is not enough to meet the original scenarios of DingTalk, DingTalk has a large number of customers and products in vertical fields, in addition to the blessing of AI and collaboration capabilities, there are also many business scenarios, and AI PaaS based on the open capability foundation extension of DingTalk base was born.
Immediately after DingTalk, it launched AI assistant products in January this year, and officially launched the AI assistant market (AI Agent Store) on April 18, where enterprises and individual users can find the AI Agent they need in the market, and they can also develop their own AI Agent, and DingTalk AI Assistant has entered the mass production stage.
"Only customers know what kind of intelligent blessing is best for their business, and comprehensively consider the cost, benefits, personnel, including the coordination of the organizational form. For example, for customers with strong capabilities and good digital awareness, cool applications and AI Agent are more creative, and in specific scenarios with low-cost transformation and many old systems, AI inside plus RPA is also a good solution. Zitui said.
In the future, different users can choose different solutions, but they can all be blessed by the capabilities of AI itself, which is a state that the platform is happy to see, and if you want to use a large model to redo all the systems, this may be a matter of the next ten or even twenty years.
In the current cycle of AI Agent, there are both optimists and pessimists, and DingTalk hopes to be a pragmatist of AI Agent, starting from specific scenarios, so that AI Agent can begin to enter thousands of industries.
(This article was first published on the Titanium Media App)