Shulou(Shulou.com)11/24 Report--

"from the perspective of technological innovation, reproducing a GPT-4 does not make any sense, but to do technological algorithm innovation."

Artificial intelligence (AI) technology is becoming more and more valuable in industrial applications.

Global AI-related program spending is expected to grow to more than $500 billion by 2027, according to a report released on October 27 by IDC, a market research firm. The Fortune 2000 companies will then spend more than 40 per cent of their core IT spending on AI-related programs.

From the perspective of the domestic market, the "hundred model war" is becoming more and more intense. Internet technology giants such as Ali, Tencent and Byte Jump continue to pour in, one after another to GPT-4 upgrade AI general model. On the other hand, in the To B enterprise end field, the small and medium-sized model field is currently facing a complex situation such as accelerated financing of head enterprises, market "inner volume" competition and so on.

Now, a generative AI (AIGC) company, which has only been established for more than four months, has also entered this blue ocean track, which may bring new variables to the development of the AI large model industry.

During the 2023 Cloud Conference held from October 31 to November 2, Core RockAI, a subsidiary of 002195.SZ, a listed A-share company, announced for the first time the To B vertical domain industry and scenario plan based on AI model, including RockAI model brain, knowledge base question and answer, business assistant, intelligent customer service and other products.

Recently, Liu Fanping, CEO of Shiyama Mathematical Intelligence Company, held an exclusive conversation with titanium media App, detailing how the new "dark horse" broke through in the AIGC field.

Liu Fanping told titanium media App that from a technical point of view, the current core math model has been iterated to the fourth edition. Compared with other large model products, Core Mathematical Intelligence not only has a self-developed Transformer architecture model, but also has a non-transformer memory logic architecture model, which has been greatly improved in training efficiency, reasoning efficiency, model hallucination, application effect and other aspects, and has been used in government affairs, Internet and other fields.

"Core has the ability of self-development and model training and optimization, as well as abundant capital reserves and stable cash flow support of its listed companies, which makes the AIGC research and development of core intelligence strongly guaranteed." Liu Fanping once said.

It is reported that Core Digital Intelligence, founded in June 2023, is a new company established after Yanshan Technology Company announced this year to embrace the new wave of AIGC, transform to digital intelligence and change its name, focusing on building an one-stop AIGC digital intelligence service platform that empowers all industries. The RockAI team was formed internally as early as 2022.

Prior to serving as CEO of Core Mathematical Intelligence, Liu Fanping graduated from University of Science and Technology of China, where he worked and published technical works such as algorithms in the big data era, Neural Networks and Deep Learning applications, and Breakthrough algorithms. At the same time, as the inventor, he applied for more than 20 patents related to AI technology.

At present, core digital intelligence supports products with innovative technology, constructs the technical structure of "self-developed large model + multiple vertical models", realizes the strategy of "1 MaaS (large model as a service) platform, multiple application scenarios", develops reliable, high-performance and directly usable vertical models for industry customers, and supports customers in some fields to further customize their own data models according to their own needs. To achieve the rapid application of AIGC in many industries, it is expected to complete the deployment of its customer private domain exclusive model within 15-30 days.

Titanium media App has learned that the current core digital intelligence has its own model with tens of billions of parameters, thus providing a variety of intelligent products for the To B industry. For example, the RockAI model brain can enable enterprises to have a large model of private domain knowledge, achieve cross-document knowledge integration, real-time information recording, and can be used in banking, government office and other scenarios.

According to statistics, about 30% of the current domestic large models are general basic models with hundreds of billions of parameters, including Ali Tongyi, Wenxin, iFLYTEK and other products to provide services to customers. However, there are some practical problems in the basic model, such as high computing cost, single API commercialization model, and customers' needs can not be met, so it is urgent to solve users' pain point needs.

In this regard, Liu Fanping told titanium media App that from a market point of view, many basic models with hundreds of billions of parameters do have some strong general task capabilities, which is no doubt, but in fact these models can not solve the specific problems of many enterprises. The fundamental goal of the development of AI technology is to land on the industry and actual users and customers through the transformation of productivity. It is necessary to have industry scenario-type large model products that integrate enterprise data and target enterprise business in order to play a greater value.

As for the challenges of the industry, Liu Fanping believes that algorithms are more challenging than innovation in computing. Innovative AI algorithms are decreasing, and many existing algorithms not only can not meet the practical application needs of customers, but also cause some pressure on the cost of large models. He said that at present, computing power accounts for nearly half of customer delivery costs, while core math uses a self-developed non-transformer memory logic model architecture to reduce delivery costs to about 30% and 40%, and the final model effect does not change much.

"the fine-tuning based on the open source model can not meet the actual needs of users, as is the case with many enterprise feedbacks. therefore, Core Mathematical Intelligence hopes to set up its own team to build an industry model. at the same time, we hope to produce 'intelligence emergence' not only in the general model, but also in professional and industry fields to improve the model effect." Liu Fanping said.

As for market competitiveness, Liu Fanping said that the application of the large model needs to go deep into the business of the vertical industry, which is a very important process. Customers' information construction and digital construction are insufficient, and their needs are not API, an interface of MaaS platform, or selling a large model software, but products and services. Core Digital Intelligence, with a long-term cooperative mentality, is willing to help enterprises provide their own solutions, and the big model is only one part of the solution.

Liu Fanping stressed that in the next step, Core Mathematical Intelligence hopes to build a multimodal real-time human-computer interaction system for future intelligent devices based on a large model that can be controlled and trusted. The system not only runs on the terminal equipment, but also has independent and intelligent cognition and control ability through natural language, image and video. This will enable AI technology to help enterprises reduce costs, improve efficiency, and achieve better results.

The following is part of the communication between Core Mathematical Intelligence CEO Liu Fanping and titanium media App:

Titanium Media App: what is the market advantage of Core Digital Intelligence over other large models?

Liu Fanping: I think there are four main advantages in the market for large model industry solutions based on core math intelligence and RockAI model brain products.

First, independent intellectual property rights. At present, most of the domestic use of open model fine-tuning, and we have our own intellectual property rights of AI algorithm technology, such as non-transformer memory logic model architecture, so we can do this thing, and can carry out in-depth model transformation and business adaptation.

Second, the model is controllable and reliable. Core digital intelligence realizes the controllability and credibility of the model in the aspects of content decision-making, technological innovation and high-quality data, and can accurately judge the security of the generated content. In addition, in specific domain tasks and diversified output, our large model can ensure that the output results are relatively fair and inclusive, as well as follow social ethics.

Third, the deployment of privatization. We not only have privatized solutions, but also can achieve secure rights management, including access to different users' applications in the model brain.

Fourth, rich enterprise application plug-ins. For enterprises, customized plug-ins for large model aids are still needed, including weather, search engine, stock, reimbursement process, leave, and other plug-ins, so as to meet the extensive needs of customers, thus enabling them to be included in the internal process tools of the enterprise. improve production efficiency.

Therefore, if compared with the industry, our differences and technical advantages probably reflect these aspects.

Titanium Media App: based on private data deployment, the model data needs to be updated, so how do customers upgrade?

Liu Fanping: the upgrade of privatization is indeed a problem in the early days, and now we are also slowly solving this problem. We can let customers train in their private cloud service providers, and we can access the training directly on their models without flipping through their data. It is done directly inside the privatization. Of course, this process requires the client's server computing support.

Titanium media App: recently, many big models are competing with the strategy of "low price". Will core digital intelligence follow?

Liu Fanping: this is a problem of commercial nature, and we do differential management in two aspects to deal with it. First, it is technologically leading, not only the new algorithm, but also our next step is to layout a multimodal real-time human-computer interaction system, which has great market advantages in terms of technological innovation and price.

The second is to better serve customers. We hope to integrate into the customer's business chain, and the product solutions under the business chain can get greater benefits one step ahead.

From the perspective of technological innovation, reproducing a GPT-4 does not make any sense, but should do technological algorithm innovation. I predict that in the next 1-2 years, some enterprises will make new innovative technologies at a relatively low cost. In the longer term, everyone may have their own big model in the future, and each big model is different. Therefore, in such a development trend, we embrace technological innovation, only by going to the forefront can we get more opportunities.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.