360 Group Peng Hui: The development trend of large models is verticalization! Enterprise-level ChatGPT only needs these 4 steps...

Source: AI dark horse

Author: Vice President of Peng Hui 360 Group

01. The development trend of large models is "verticalization"

With its inclusive, ubiquitous and universal features, the large-scale model will enter thousands of households and empower thousands of industries in the future.

So, we have this view:

Every family, every government, and every enterprise will have one or more large models.

We also believe that the development of China and the United States in the To B market is very different. There will be no monopoly in China, and there will definitely not be only 3-5 large models.

In the future, big models must be ubiquitous, and future development opportunities must be in the enterprise market.

Everyone knows that digitalization has become a core strategy of our country, and industrial digitalization will be a huge decentralized incremental market in the future.

Therefore, in making large-scale models in China, we firmly believe that we must seize such a strategic opportunity to empower the industry, anchor the industrial-level market, and pull large-scale models from the so-called centralized market. To improve the productivity and production efficiency of the government and enterprises.

Of course, in this process, large models will encounter some huge challenges in the process of landing in the enterprise market.

I summarize it into four aspects:

1. Lack of professional knowledge.

Everyone knows that we will use a large amount of Internet corpus to feed the large model, it is like a high school student, at most an undergraduate graduate. However, it is extremely deficient in some professional domain knowledge, industry knowledge and enterprise internal knowledge. Not even updated in time. So, this is a big problem, a lack of expertise.

2. Occasional hallucinatory nonsense.

Everyone often says a word, and the big model will talk nonsense in a serious manner. It will have knowledge ambiguity and knowledge illusion. Because data and knowledge are like structured databases, they are encoded in the parameters and weights of our deep neural network in another matrix and vector form. However, I want to call it, use it, and actually need to do a better job of bootstrapping. In this process, its algorithm mechanism will produce an illusion of content, which cannot guarantee authenticity and credibility.

3. Security issues.

Enterprises are unwilling to contribute their unique skills to the public large model, or train them into a public large model.

4. Cost issues.

Now the supply of Nvidia H100 is out of stock, and ChatGPT claims to train tens of thousands of cards at a time. Therefore, this kind of investment is very difficult for an ordinary enterprise. We may have reduced manpower, but we haven't reduced costs.

So, how to solve these problems?

We believe that a development trend in the future must be towards verticalization and the creation of small but specialized vertical large models.

It is impossible to rely on a single, all-purpose general-purpose large model to solve all the problems of task decomposition, human-computer interaction and knowledge question answering. We must rely on enterprise-level data corpus and high-quality data to train small-scale, proprietary vertical large models.

Large models will become a standard configuration and component of all digital systems in the future.

**02 How can enterprise-level GPT be implemented quickly? **

We interviewed more than 100 corporate customers and partners, and everyone basically has a consensus:

Big models aren't everything at the moment.

So, the question becomes: how to better apply it to enterprise scenarios?

We need to professionalize the so-called generalists and become real government and enterprise experts.

At this time, we need to find a small incision and give full play to its strengths.

We believe that the current capabilities of the big model are mainly reflected in the two capabilities of text generation, or content creation and knowledge question answering. We can start with these two abilities.

More and more practitioners in the large-scale model industry believe that in relatively focused and narrow application scenarios, smaller and fine-tuned large models will meet the accuracy requirements of the To B end faster.

Therefore, we must proceed step by step, let the large model be a good assistant first, and let the large model be a good navigation first.

Focusing on such a scenario, we find corresponding application scenarios that adapt to these four products to quickly exert their productivity and effectiveness from the four dimensions of top, bottom, internal, and external.

1. In the internal scene, we think it is more about writing and summarizing the office.

2. In the external scene, a large number of digital people began to appear in the customer service scene.

3. In the above scenario, we emphasize the summary and analysis of information and intelligence.

4. In the next scenario, we can let the big model do a series of training on corporate knowledge and even job skills.

Therefore, during the entire implementation process, we realized a very important point. In the future, more than 80% of our business scenarios will be closely related to the enterprise's knowledge base.

In the past, when we worked on big data, we all stayed in the application of structured data. You must know that 80% of unstructured knowledge and data are abandoned or left alone. This part of big data will become the corpus for large model training.

Therefore, how to extract valuable knowledge and high-quality precision data from a big data base of the enterprise, transform it into a private domain knowledge base of the enterprise, and empower the large model through retrieval correction and enhancement , can truly produce credible content and timely content updates in the To B business scenario, as well as the security of content such as decentralization and division of domains.

The data is divided into three gates:

The first gate may be open Internet data, the second gate is semi-public industry data, or enterprise data, and part of it is confidential data within the enterprise.

For this kind of enterprise's confidential data and authorized data, we must put it in an enterprise's knowledge base, or put it in a vector database, so that it can generate a kind of management with authority and audit, through A kind of authority management of classified and hierarchical enterprise knowledge, enhanced by the retrieval of large models, so as to provide more accurate knowledge and empowerment.

Another aspect is the application. At the application level, everyone has been exposed to ChatGPT. Do you think it is easy to use?

Why some time ago all the brains were mentioning that in the future there will be a lot of posts and roles that suggest engineers, in fact it is very complicated.

We want it to write a good article, and we have to give it a lot of hints, central ideas, abstracts, and outlines before it can write a good article. We let it make a picture, use Midjourney, I You even have to tell it how many millimeters of lens, focal length, aperture, and depth of field you need to use, what kind of environment can make a really beautiful generated picture like this, but such a hint project is actually only usable, But it is very awkward and difficult to use.

Therefore, in the future development process, don't be superstitious about the so-called language UI, and more traditional interfaces will not be eliminated.

Moreover, it will be generated in large numbers in future scenarios such as office writing, picture creation, marketing creativity, etc., and even government knowledge question-and-answer scenarios, because it provides more intuition and applicability.

For example, Lao Zhou often said that he wanted to eat a plate of shredded potatoes. I wanted fried ones instead of vinegared ones. The click may be at the level of a second, and I have finished this matter.

360 will release the entire product system framework of its own enterprise-level GPT. At the bottom layer, we still believe that data and knowledge will become a base for large enterprise-level models in the future, which cannot be separated from the accumulation of all data in the past. It is just that we need to adapt to the needs of large models and unstructure the data accumulated by all enterprises in the past. Content and documents, including multimedia audio and video graphics, image data, through multiple data connectors and knowledge tracking robots, a processing engine that promotes multi-source data incorporates it into our enterprise knowledge base, through the vector Indexes, abstract indexes in the traditional sense, text indexes and multi-modal indexes, build a knowledge base of an enterprise-level large model, and then through our search and knowledge enhancement, empower our professional vertical enterprise large-scale The model provides services upwards.

03, 3 practices and the best process for landing large models

1. Office writing.

We will hide the complex projects behind the different 15-category large templates and nearly 80-category subdivided document templates. Use such a tool to efficiently complete official document writing, and effectively solve the problems of time-consuming and low-quality official document writing.

2. Government services.

Through the large model and the knowledge base of government affairs, we can make the large model understand semantics like a human through multiple rounds of dialogue, supplement relevant information through follow-up and follow-up questions, and finally form a question and answer. It can objectively and accurately answer all the questions that ordinary people have in the process of doing business.

3. Cultural tourism digital people.

Lao Zhou also mentioned Wenlv Digital People on many occasions. Everyone has done travel planning. Can a so-called travel itinerary plan solve your travel problems?

What we care about is that after landing at a destination, we have a local friend and a local tour guide. The sights, food, anecdotes, jokes I care about, how do these things become a destination-centric digital companion? We want to create such a digital companion. In the future, under the leadership of the government, we will gradually open up the capabilities of China and Taiwan, and after connecting our OTA, local hotels, and restaurants.

Combined with 360 Group itself, and more than 100 corporate customers and partners, we have formed the implementation process of the best practice so far.

The first step is business analysis and scenario selection.

The second step is data collection and cleaning preparation.

The third step is to train the enterprise proprietary large model.

The fourth step is to develop enterprise scenario applications.

We believe that when all large-scale models are implemented in enterprise-level scenarios, one of the top priorities is still business analysis. It is no different from digitization.

We still need to find the pain points of the business in the process of business analysis, so as to find a suitable scenario, and after selecting this scenario, define our solution.

Then, collect and clean data and knowledge around this scene to form our high-quality, labeled data. After entering the database, part of it is fed to our vertical large model as corpus for training, and part of it enters our enterprise knowledge base to do a search enhancement of knowledge. Next is the development of smart assistants, digital employees, and digital humans, through the application arrangement of models and the opening of APIs to integrate with existing business systems.

During the implementation process of the enterprise-level large model, we emphasize that it is inseparable from the close collaboration of the business and technical experts of both parties.