Ignoring small models and knowledge bases will lead to a dead end for enterprise AI applications

In 2022 , the emergence of ChatGPT made many people exclaim that "the era of general artificial intelligence has arrived"! Now, three years have passed in a blink of an eye, and the craze for artificial intelligence is still there, but how has the application of artificial intelligence developed in actual production and life? Have we really entered the era of general artificial intelligence, or is it still a mess under the glitz? Personally, I think that the popularity of large models and general artificial intelligence has indeed begun to change some of people's behaviors, but it is far from reaching the threshold of true "intelligence", and even whether it represents the correct path for the future development of artificial intelligence is more questionable than three years ago.

1. The emergence of large models has indeed opened a new door for the application of general artificial intelligence

The essence of a big model is a neural network model that contains ultra-large-scale parameters (usually more than one billion) and complex computing structures. It is trained using massive amounts of data and exhibits human-like intelligence based on its huge data and parameter scale. The training of a general big model is based on massive amounts of data available on the Internet or public channels, so it has a relatively comprehensive understanding of the common sense knowledge currently mastered by humans. In addition, it uses new reasoning methods such as thought chains in text organization, making it appear "both knowledgeable and natural" in conversations with people.

These advantages of the general big model have undoubtedly brought some convenience to people's work. For example, if there is any common sense that you don't understand, you can directly ask the big model, and the big model will find the answer for you. Now even my son who is only in the fourth grade knows that he can ask DeepSeek if he doesn't understand anything . For example, when I write bidding documents, I will ask the big model to give me a draft for some chapters that have no reference materials, and then I will modify and supplement them, which really improves efficiency. Therefore, I think the emergence of the big model can indeed be said to have opened a new door for the application of general artificial intelligence to everyone, and let everyone enjoy the benefits brought by artificial intelligence technology.

However, we should also note that the current areas where large models and general artificial intelligence are better applied are mainly concentrated in knowledge questions and answers, data search, text, image and video generation, etc. These scenarios have a characteristic that large models are not responsible for the results, and right or wrong will not affect actual production and life. Therefore, large models are still mainly used for auxiliary purposes, and are far from being responsible for their own behavior, and still cannot do without people. In addition, large models often talk nonsense, and they still look serious. It will graft what Kangxi did onto Qin Shihuang, and will randomly quote formulas in papers to explain why it is so hot today. These behaviors will mislead people, especially if they teach the wrong children, so you still have to be very careful.

2. General AI and large models are difficult to apply to core production and management of enterprises

For an enterprise, the core production and management work is very rigorous, and many of them are process-based and standardized. This means that in the actual production process of an enterprise, the amount of redundancy that can go wrong is extremely small, and some cannot even have the slightest error. Under such requirements, the shortcomings of large models and general artificial intelligence will be infinitely magnified, and even become a disturbing "time bomb".

On the one hand, the illusion problem of the big model can never be effectively solved. This means that when the same thing is handed over to the big model to execute, the results each time may be very different. The existence of such differences is fatal to corporate production. Imagine that you are now reporting the operating indicators of the first half of the year to the company's top leader, but the big model gives you different results every time. How do you report to your leader?

On the other hand, many results given by large models are unexplainable. This means that even if you know that there is something wrong with the results, you don’t know where the problem is and how to correct it. Let’s take the example of business indicators. The results are different every time, and you don’t know how they are calculated each time. You don’t know which one is right and which one is wrong, and how to correct it. In the end, it’s better to do the statistics yourself.

In addition, there is another most fatal problem, which is that they do not understand the profession. Big models and general artificial intelligence are "generalists" but not "specialists". If you ask them about common sense, they know more than you do, but if you ask them about the system norms, professional knowledge, and process parameters in the enterprise, they have no knowledge at all because no one has taught them. In this case, they will "do things randomly" and become an uncontrolled "unstable factor."

Therefore, under current conditions, it is impossible to directly apply big models and general artificial intelligence to enterprise production management activities and play a core execution role. Therefore, it may be far from what many enterprise leaders think: I have introduced big model services, localized big models, and connected to DeepSeek , and our enterprise's production management can be upgraded to intelligent ones, which can solve a lot of labor costs.

3. Small models that are both precise and specialized are the key to improving enterprise productivity in the future

So, if companies really want to use artificial intelligence technology to improve productivity and efficiency, what should they do? I believe that compared with large models, small models with stronger "professional capabilities" and certainty are the key to future enterprise intelligent upgrades.

As the name implies, a small model is a model with smaller parameters than a large model. The biggest difference between large and small models in terms of application is that large models tend to be all-round and general, while small models generally tend to solve a specific problem in a vertical field. For example, a small image recognition model is specially trained to recognize license plate numbers, and can have a good recognition accuracy for license plate numbers. However, a large image recognition model can not only recognize license plate numbers, but also most of the pictures we see in our lives. From our human perspective, it seems to have its own understanding of the content in the picture and seems to have a higher level of intelligence. Compared with large models, small models usually have fewer parameters, the learned features and patterns are relatively simple, and the training does not require high computing resources.

For enterprises, a feasible approach is to divide the internal production management work of the enterprise into professional divisions, build a small model for the core work content of each profession, and use the existing business data for training, so that this small model can execute standardized workflows and accurate business judgments. As an auxiliary, the big model plays its own advantages and provides support in reasoning and text generation outside the norm. Under this architecture, the key to enterprise productivity lies in whether the professional small models are professional enough, rather than relying on the computing power, algorithms, and knowledge of the big model.

4. Knowledge base and high-quality data sets are the core competitiveness in the AI ​​era

Whether it is a large model or a small model, they mainly solve the problem of thinking and execution. In terms of people, it is like two people, one is more knowledgeable and the other is more specialized, but if no one teaches them the relevant knowledge, they can't do anything. This is also a problem that many companies are gradually discovering now. They have large models and artificial intelligence, but they can't do anything except question answering. This is because compared with the model, what really determines the effectiveness of artificial intelligence application is the enterprise-level knowledge base and high-quality data set.

In my opinion, knowledge base and high-quality data sets are broadly unified, and both are knowledge that can teach large or small models to do things. In a narrow sense, high-quality data sets are the basis of knowledge bases. The knowledge in knowledge bases should be further refined on the basis of high-quality data sets, including business objects, attributes, relationships, rules, and other contents.

For example, when inspecting poles in telecommunications companies, it is necessary to promptly identify existing risks or faults that have already occurred. In this scenario, I picked out 2,000 high-quality pictures from previous inspection photos, covering various major hidden dangers and fault manifestations. Then I need to mark these hidden danger points and fault points, either manually or with automated annotation tools. These 2,000 labeled high-quality pictures form a high-quality data set for this scenario. Based on this data set, we further refine relevant knowledge, such as grading hidden dangers and faults, corresponding different processing procedures for different hidden dangers and faults, different notification levels, etc., which form a knowledge base for this scenario. Then, these high-quality data sets and knowledge bases are used to train small models so that they can accurately identify hidden dangers and faults in a new photo, and can even initiate corresponding work order processes, give recommended measures, or simply operate online.

From the above examples, we can see that in a specific business scenario, the core value lies in high-quality data sets and knowledge bases. Regardless of whether it is a high-quality data set or a knowledge base, most of them need to be built by the enterprise itself, and no one else can help.

5. AI applications may require greater resource costs in the short term

Since others cannot help, companies need to invest their own resources and costs to do it.

This is a very contradictory thing. Most business owners want to introduce artificial intelligence, thinking about how much manpower can be replaced, how much cost can be saved, and how much profit can be obtained. I can only say that if the ultimate ideal state can be achieved, it is indeed possible. However, the road to the ideal state is not easy. Building a company's own high-quality data set, knowledge base, and training professional small models all require a large amount of initial investment, repeated verification in the middle term, and subsequent continuous maintenance. This means that before artificial intelligence can really play a role in actual work, there will be a large amount of silent costs to be invested, and this investment may not be lower than the cost of using traditional manpower in a certain period of time (such as three or five years). In other words, the return on investment of business owners in artificial intelligence may not be seen until three to five years later. Including central state-owned enterprises, how many companies can accept such a return cycle?

Therefore, the so-called artificial intelligence applications of many companies are actually mainly concentrated in some general fields such as knowledge question answering, data statistics, auxiliary reporting, etc. This is also the reason why I said at the beginning that the current artificial intelligence applications are far from reaching the "threshold".

6. Small models that cannot be “continuously cultivated” will also become “artificially mentally retarded”

Finally, I would like to say that no matter whether it is a large model, a small model, or a high-quality data set or knowledge base, it all needs to be "continuously cultivated". For high-quality data sets and knowledge bases, it is necessary to update or supplement the content regularly, especially after business changes, data and knowledge with new features should also be included to ensure the comprehensiveness of data and knowledge; for large and small models, new data and knowledge are needed for training, and rapid adjustments and optimizations are also needed in combination with problems encountered in practical use. If you cannot keep up with the times, even a small model will become "artificial intelligence", which will not be able to help you solve problems but will continue to create problems.