news

Text | Wang Zhiyuan

Recently, a friend gave me feedback:KimiThe advertising is really fierce.

Listening to NetEase Music and browsing bilibili, you can see that the product is for ordinary users, which is beyond doubt. However, only halfway through August, some actions of Kimi make people feel a little different.

What's the difference?

On the one hand, on August 2, Kimi's parent companyDark Side of the MoonMoonshot AI announced the official release of Kimi’s enterprise-level API. This enterprise-level model provides stronger data security protection and faster processing speed than the general user version, and can help enterprises handle complex tasks and large amounts of data.

Then, they announced that they would reduce the fee of Kimi Open Platform’s context cache technology from 10 yuan per 1M tokens per minute to 5 yuan. I understand that this price reduction has been in public beta on the open platform since July 1st.

After reading this, you may be curious about what cache is.

Simply put, its function is to store in advance texts or data that may be used repeatedly and frequently queried, thereby improving the reasoning efficiency of the model without increasing costs.

So, these actions made me wonder, has AI found an effective path in the B2B field? After a round of research, I think: Yes.

01

Let me first talk about an industry consensus:AI application is undoubtedly the key to the success of the model.

At the World Artificial Intelligence Conference this year, Baidu CEO Robin Li discussed his views on the application of large models. He said: Although C-end development is important, B-end application scenarios are where large models can achieve better results.

He believes that in the AI ​​era, applications that can profoundly impact industries and significantly improve efficiency are more valuable. He foresees that in the fields of healthcare, finance, education, manufacturing, transportation, and agriculture, customized intelligent entities will be developed based on their respective characteristics and data resources. In the future, the number of intelligent entities will reach millions, forming a huge ecosystem.

This year, Baidu won bids for 17 projects in various fields including medical care, finance, energy, environmental protection and transportation, covering large state-owned enterprises and industry-leading companies, and the amount was also very considerable.

Obviously, Robin Li believes that AI applications must be quickly implemented on intelligent entities.

Let’s not talk about whether it can be implemented on an intelligent entity. There are so many bids that need to be delivered one by one. If we don’t produce a finished product in a year or two, the customers will definitely not be happy.

Let’s see what Yang Zhilin, founder of Moonshot AI, thinks.

In June, during a conversation between Zhang Peng, founder of GeekPark, and Yang Zhilin, Yang mentioned that they did not completely exclude the B-end, but mainly focused on the C-end. His product Kimi has reached top traffic and usage in the field of AI.

Despite this, users often experience issues during peak hours due to insufficient computing power. To combat this, they have taken several steps to reduce operating costs and improve efficiency.

One of them is to optimize the model's inference performance through caching technology, which enables Kimi to respond faster when processing complex requests and reduces the need for repeated calculations.

This improvement not only improves the system's response speed and processing efficiency, but also ensures the consistency and accuracy of conversations or text generation; this will be particularly important in the future when it comes to processing a large number of data requests centrally when doing B-side business.

Therefore, Yang Zhilin believes that if you want to turn to the B-side, you must first solve the computing power problem, and ensuring the stability of computing power is the foundation.

On the other hand, I think that doing C-end business is not so sexy. It mainly exchanges conversions through advertising, which is not so eye-catching in the long run.

I learned from many advertising practitioners that starting from the first half of 2024, the per capita conversion cost has been rising, and the conversion cost of each Kimi PC user has basically reached 40 yuan per person. I have not verified this data with the official.

However, looking at the general trend, the increasing cost of competition in the AI ​​market in C-end scenarios has undoubtedly prompted many AI companies to re-evaluate their market strategies.

Therefore, although the C-end market has great potential, the application scenarios on the B-end are the key areas for realizing the far-reaching impact and efficient results of the big model. After all, only by achieving real cost reduction and efficiency improvement for enterprises can we promote the progress of the industry and even the entire industry.

02

Since the consensus is correct, how can intelligent agents, or AI, large (small) models effectively enter the B2B field?The first method is to be the upstream of the B-side.

What does it mean?

The so-called B-end upstream refers to the source of the supply chain. For example, suppose there is a pharmaceutical company that has the need and application scenarios to use AI, but it is difficult for a large model company to enter the market. What should it do?

At this time, you find that this pharmaceutical company is using some SaaS software, and the AI ​​big model company can consider cooperating with the software supplier; in this way, AI can be added to the existing software products, and the pharmaceutical company can smoothly transition to using AI while using the software.

This is the way many traditional B-side business people think.

In fact, B2B software can be deployed in many forms. First, local deployment, where the software is installed on the customer's own server or device, allows the customer to control data and security.

This method requires regular upgrades, which is troublesome and costly to maintain. For example, in the past, large manufacturers such as Mercedes-Benz and BMW would use traditional CRM systems as local deployments to manage their agents.

However, local deployments face many challenges in implementing AI integration, especially pre-trained models; once installed, such models can respond to customer queries even when not connected to the Internet.

Although many ToB companies are willing to try, the actual operation is very complicated. For example, pharmaceutical groups or hospitals have huge data management needs. Once they need to update the software and integrate new AI functions, the whole process will be extremely complicated.

In addition, project leaders have to explain to their bosses why they have to spend a lot of money to introduce AI capabilities.

for example:

Spending 20 million to build a model may sound impressive, as if we have AI capabilities, but in reality, it may seem like just an AI knowledge question-answering system. From this perspective, it doesn’t seem to be worth it.

Another point is that even if a company introduces a model, what if the employees don’t use it? Wouldn’t the money be wasted?In this case, no calculation is appropriate. Of course, some companies don’t even have a knowledge base, let alone the need to introduce AI capabilities.

The SaaS model, as another form, allows users to pay fees through subscription, such as collaborative tools such as Feishu. In this model, SaaS companies can directly integrate AI functions and even have the ability to bypass small model companies and directly purchase services from large model companies to transform processes.

Among all these forms, the SaaS model is the easiest to integrate AI functions, because service providers can uniformly update and maintain AI functions on the back end, and customers do not need to worry about technical details.

Therefore, from a top-down perspective, companies that provide AI solutions may win some orders, but the process of truly making customers successful may be quite difficult, and the effort and reward are often not proportional.

03

So, is there a new solution? Yes. Let’s look at it from the bottom up.

We need to establish a new understanding:

The use of AI in the company is actually a breakthrough. It mainly helps us improve the efficiency of doing things. AI is generally used to strengthen existing work processes rather than starting from scratch.

What is a clean start?

It means starting from scratch and completely redefining the process of something without relying on any existing structure or plan; in the context of workflow or project management, it means abandoning old methods and systems and adopting a completely new approach to solving problems or performing tasks.

For example, a company originally used a very traditional customer relationship management (CRM) system. As time went by and the business grew, this system could no longer meet the company's needs.

The company decided not to modify or upgrade the old system and decided to develop a brand new CRM system from scratch. This new system will use the latest technology stack, design functions that are more in line with modern business processes, and can better integrate the latest data analysis tools.

This is an example of a clean slate. But you know what?AI subverts the process of process reengineering. Most of the time, AI handles tasks that humans are already familiar with, without having to reinvent the wheel.

What does it mean?

When a company starts using AI models, these models must be closely integrated with the company's own workflow. For example, some companies want to use some advanced large models in the medical or education fields, but this process may encounter many difficulties.

Because each company's business and processes are unique, and the data required by AI is also unique, a general large model may not be appropriate, and it is usually difficult to truly apply these broad concepts in a company.

what to do?At this time, it would be easier if you have a small model or assistant.

For example:

Small and medium-sized enterprises usually use accounting software such as Kingdee, UFIDA, and Chanjet. These software already store a lot of data; for operators, they just want to use the power of AI to quickly find a piece of data or draw a conclusion, without having to make major changes to the original software.

Therefore, to apply AI in enterprise (ToB) scenarios, a good way is to break down complex business processes into many small tasks, or specific small scenarios, and then use AI to help make improvements in each small scenario.

We also see that companies like MicrosoftOpenAI’s investors, as well as Salesforce, the leader in the CRM industry, they did not use AI to develop new products.

Although they continue to label their products with new features, in reality, they are using big models to assist and enhance existing business processes or product features.

That is, they refine the small model into an assistant or an enhancement that can better integrate into and optimize existing systems rather than completely replace them.

This is a bit like the plug-ins that many AI companies have made on PCs recently. You just need to slide the mouse or press a shortcut key, and AI will pop up to help you. Their core function is to help us make better decisions.

04

If you don’t believe it, you can also think about this: For ToB companies, what do they really need AI to do?

After Zhiyuan’s research, it was found that they mainly use data to help make decisions in operations, management, decision-making and marketing.So, how can AI companies achieve both practicality and cost-saving? The key is to achieve rapid replication and cover a variety of scenarios at a very low cost.

How to copy?

One way is to encapsulate AI into an intelligent agent that can call local data.This is why Kimi Chat reduces the cost of cache; because saving frequently used local texts can help improve the accuracy of the assistant.

But there is a problem here: how can we make it accurate, universal and low-cost at the same time?

For this, we can define “high accuracy” as follows:It’s fine as long as the business can be used and the accuracy is high enough. Even if errors occur occasionally, users can recognize them and explain why they happened.

"Low cost" means that the initial cost of the project must be low, it can be deployed on a single machine, and it must be easy to use and maintain.Nowadays, the hardware costs of many large models are very high. If a company is not sure how much money it can make, it will definitely not be willing to invest too much.

The advantage of large models is that they solve the universality problem of early AI algorithms.

In theory, one model can handle all scenarios; in reality, we need to fine-tune it. Therefore, we define "high versatility" as being able to be tooled and automated, so that it can be replicated on a large scale and implemented at low cost.

A large model with billions of data points has achieved an accuracy rate of over 95%. Then, we derived a small model from this large model and encapsulated it into an assistant, which solved the problem of using the entry point.

Imagine that the Big Model company has a super powerful Big Model solution, which is like a parent model that supports the broad needs of the entire enterprise.

This large model is very powerful and can handle complex data analysis and give very accurate insights. It is particularly suitable for B-side companies that require a lot of computing resources and large-scale data processing. But in fact, not all B-side companies need such a large system.

For more small and medium-sized enterprises, a smaller, more focused model is sufficient. We can think of it as a "sub-model" of the big model.

These small models inherit the core capabilities from the large models, but have been optimized and adjusted so that they can better adapt to specific business needs and small-scale operating environments.

Small models do not require as much computing power as large models, but instead focus on specific tasks and can be adjusted according to different needs, such as generating intelligent analysis reports or performing intelligent analysis.

This allows small businesses with limited resources to use AI to improve work efficiency and decision-making quality.

Therefore, the small model is like the essence extracted from the large model. It only needs basic knowledge base capabilities. The key is to make it into a small assistant and embed it into the tool for use.

Don’t think that this means the logic is correct. According to Zhiyuan’s research, AI products have already been implemented in B-side enterprises; therefore, for TOB, the application scenario of AI is the intelligent body and the plug-in.

Another benefit of doing this is that it is easy to deliver without incurring high costs.

Summarize

TOB AI, not following the old path.

If the small model is the key point of TOB, then making it into an intelligent agent or assistant is the key link to connect the last mile of TOB.After all, bottom-up is people- and task-centric, not software-centric.