01
Industry consensus: AI applications are key to model success.
Baidu CEO Robin Li believes that while C-end development is important, B-end application scenarios are more likely to yield good results for large models. He foresees the development of customized intelligent agents in fields such as healthcare, finance, and education, with the number of intelligent agents reaching millions in the future, forming a vast ecosystem.
This year, Baidu has won 17 projects in multiple fields, involving large state-owned enterprises and industry-leading companies, with considerable amounts. Li believes AI applications should be quickly implemented on intelligent agents.
Moonshot AI founder Yang Zhilin states that they don't completely exclude the B-end but mainly focus on the C-end. Kimi has achieved top traffic and usage in the AI field, but often faces problems due to insufficient computing power during peak hours. To address this, they have taken measures to reduce operating costs and improve efficiency, such as optimizing model inference performance through caching technology.
Yang Zhilin believes that shifting to the B-end requires solving computing power issues first, ensuring stable computing power as a foundation.
The increasing competition costs in the C-end AI market are driving many AI companies to re-evaluate their market strategies. B-end application scenarios are the key areas for realizing the profound impact and efficient results of large models. Only by truly reducing costs and increasing efficiency for enterprises can industry and even entire industrial progress be promoted.
02
How can intelligent agents or AI, large (small) models effectively enter the B2B field? The first approach is to focus on the B-end upstream.
B-end upstream refers to the source of the supply chain. For example, pharmaceutical companies have demands and scenarios for using AI, but it's difficult for large model companies to directly enter. In this case, they can consider collaborating with SaaS software suppliers used by pharmaceutical companies, integrating AI into existing software products, allowing pharmaceutical companies to smoothly transition to using AI while using the software.
B2B software has multiple deployment forms:
-
On-premises deployment: Installing software on the customer's own servers or devices, giving customers control over data and security. Requires regular upgrades, troublesome maintenance, and high costs. Faces challenges in implementing AI integration, especially for pre-trained models.
-
SaaS model: Users pay through subscription. SaaS companies can directly integrate AI functions, even bypassing small model companies to purchase services directly from large model companies for process transformation.
The SaaS model is the most convenient for integrating AI functions, as service providers can uniformly update and maintain AI functions on the backend, and customers don't need to worry about technical details.
Looking from top to bottom, companies providing AI solutions may win some orders, but the process of making customers successful can be quite challenging, often with disproportionate effort to reward.
03
Is there a new solution? Let's look from bottom to top.
New understanding: Using AI in a company is actually finding a breakthrough point, mainly to improve work efficiency. AI is generally used to enhance existing workflows rather than starting completely anew.
In the process of AI revolutionizing process reengineering, most of the time is spent dealing with work that humans are already familiar with, without the need to reinvent the wheel.
When companies start using AI models, these models need to be closely integrated with the company's own workflows. General large models may not be very suitable because each company's business and processes are unique, and the data AI needs is also specific.
At this point, small models or small assistants are more appropriate. For example, accountants in small and medium-sized enterprises use software like Kingdee or Yonyou, which already store a large amount of data. Operators only want to use AI to quickly find data or draw conclusions, without needing to greatly modify the existing software.
To apply AI in enterprise (ToB) scenarios, a good method is to break down complex business processes into multiple small tasks or specific small scenarios, and then use AI to help improve in each small scenario.
Companies like Microsoft and Salesforce haven't used AI to develop entirely new products, but rather use large models to assist and enhance existing business processes or product functions.
They refine small models into assistants or enhanced capabilities, better integrating and optimizing existing systems rather than completely replacing them. This is similar to the plugins many AI companies make for PCs, where a light mouse movement or pressing a shortcut key can summon AI for help, with the core function being to assist in making better decisions.
04
For ToB enterprises, what do they really need AI to do? Mainly use data to help make decisions in operations, management, decision-making, and marketing.
For AI companies, how to achieve both practicality and cost-effectiveness? The key is to achieve rapid replication, covering multiple scenarios at very low cost.
One approach is to package AI as intelligent agents that can call local data. This is why Kimi Chat reduces caching costs, as storing frequently used local text can improve the accuracy of small assistants.
But how to achieve accuracy, versatility, and low cost simultaneously?