CloudWalk's big model: what is the relationship between the big model and the AI platform? Why build an industry model?

Original: Dear Data

On May 18 , 2023 , at the coordinates of Nansha, Guangzhou, the " Yuncong Rongrong Large Model " from Yuncong Technology was officially unveiled.

Since then, Yuncong Technology has "advanced" from CV Four Tigers to an AI large-scale model company. At the same time, don't forget that Yuncong has another identity that is often mentioned: the first AI platform company to land on the Science and Technology Innovation Board.

In the first speech, Zhou Xi, chairman of CloudWalk Technology, led the technical management department Ye Mao to bring a lot of interesting large-scale model tests to the scene.

At the same time, Zhou Xi also publicly stated that in the previous " multi-point technology closed-loop " stage, the degree of project customization was heavy, which led to an imbalance in the input-output ratio and eventually a loss-making situation.

The good news is that the era of large models that is currently entering can further promote the standardization of artificial intelligence technology, and the marginal effect has risen, enabling the intelligentization of massive real-world scenarios more quickly and efficiently.

At the press conference site, the expression "Comongrong large model plans to launch a self-developed general multi-modal large model and a series of industry large models such as language, vision, voice, code generation, and image generation large models" was printed on the introduction back panel; Print the English words " Make AI Great Again" (Make AI Great Again) on the gift bag.

The horn blew, and the morale ready to fight rushed to the face.

In the first speech, Zhou Xi introduced that there are three types of interaction methods in the era of large models: question and answer, accompanying and hosting.

He said: "Question and answering is the passive participation of AI ; accompanied by AI interacting with you in real time, providing its suggestions and reasons, and humans making decisions; hosting means that we humans mainly hand over this matter to AI ."

Questions and answers, accompanying and hosting in Yuncong Technology's view are not only their consistent technical concept and technical direction, but also the realization of earlier forward-looking judgments.

Zhou Xi believes that the nature of these three forms of interaction is progressive, which corresponds to a large number of technological breakthroughs. "What does it take to actually be able to do these forms of interaction?" he asked.

The answer: "Large models across modalities."

Zhou Xi said: "In the future, the large model must be able to understand, understand, and understand everything in the physical world, so that it can interact with us. The future needs action systems, digital humans, physical robots, etc., at any time Interacting with our human entities. So, a lot of technology needs to evolve."

What does self-study mean?

There is a pre-trained large model, which is significantly different from the development based on open source large models. AI companies like Yuncong have the ability to fine-tune models with pre-trained layers as a "compulsory course".

At the same time, model-level services (private deployment) are different from those provided in the form of APIs in the past. They are more sticky and have greater business expectations.

AI companies starting from algorithms will have to develop their own AI platforms sooner or later, and the sooner the better. The early AI platform was the integration of data algorithm computing power.

The AI platform completely separates non-self-developed algorithm teams from companies with strong R&D capabilities. For easy understanding, many companies refer to AI platforms as "operating systems", which is a metaphor.

Now it seems that the emergence of many AI platforms is the product developed after the AI company's own internal needs and the greatest commonality are found.

In this regard, the national level also pays more attention to it, because the product is very influential and demonstrative, and the hope is that it can be promoted and applied on a large scale and contribute value to the economy.

My understanding is that the government will pay more attention to the open development of the platform, such as: how many ecosystems are supported, how many active users are there, and how many industries are supported.

We often talk about "open platform", so what is the open platform ?

The answer is: the capabilities of the platform.

In the speech, Zhou Xi also introduced: "In 2019 , we started to build the entire CWOS operating system. In 2020 , we launched the human-computer collaboration open platform in Guangzhou. In 2021 , we launched the AI wizard. In 2022 , we will undertake the Ministry of Science and Technology's " National Artificial Intelligence Open Platform for Audio-Visual Interaction".

Here is a background knowledge: the platform is an interactive platform, and the bottom layer is the operating system of CWOS .

Taking the development of CloudWalk as an example, after entering the era of large models, what is the relationship between artificial intelligence platforms and large models? Some people might say that the platform is eliminated. The understanding of Yuncong is that the platform is still the foundation.

Zhang Li, vice president of CloudWalk Technology, told me: "For the AI platform, it is nothing more than the ability to unlock the platform's large model, one of the capabilities of the large model. Technology unlocking brings new explosive points."

The changes in the era of large models are obvious, from process ( IT )-centered development to knowledge ( Knowle )-centered development, the former is essentially customization; the latter is essentially knowledge capability.

According to insiders of Yuncong, Zhou Xi has repeatedly emphasized within the company that a general large model is the foundation, while specialized knowledge of different categories requires an industry large model.

In the first speech, Zhou Xi also gave a special explanation on this. His analogy is: "The basic quality of a person is very important. With the foundation, we can do professional things. We study professional knowledge in university, and we study law to become lawyers. , we study medicine and become doctors.”

In response to this, Yuncong revealed at the event that it will cooperate with China Inspection and Quarantine to launch a large-scale quality model, cooperate with China Information Technology to launch a large-scale financial model, cooperate with Shenzhen Newspaper to launch a large-scale entertainment model, and cooperate with Jiadu Technology to launch a large-scale traffic model . Models, jointly launched manufacturing large models with Jin Shiyuan, cooperated with Youzu Network to launch game large models, and cooperated with Aiden Technology to launch medical large models.

The basic large-scale model is the foundation of capabilities, used to catch up with the world's leading large-scale model basic technology, focusing on R&D investment, while the industry large-scale model can obtain application practices and orders.

Yao Zhiqiang, co-founder of CloudWalk Technology, believes: "The cost of iteration of a general-purpose large model is very high. Although many traditional enterprises have the urge to try it out, they will calm down gradually. Independent development, operation and maintenance iteration is really not cost-effective. That’s the opportunity for industry megamodels.”

Large-scale model technology has a wealth of application scenarios in intelligent manufacturing, and Cloudwalk pays more attention to it. At the same time, the "large-scale scenarios" that Cloudwalk has always attached importance to and accumulated years of experience include finance and government affairs.

From the perspective of Yuncong Technology, there are many highly complex things in smart manufacturing.

Yao Zhiqiang introduced: "There are more and more equipment, involving different manufacturers, and corresponding to a large number of complicated work manuals. Communication and positioning need to solve problems and cannot be as inefficient as looking for a needle in a haystack. For example, common equipment inspection problems are located as an incision , large models will have better applications.”

To some extent, the large model absorbs a large amount of knowledge, and the large model provides high-quality knowledge services for intelligent manufacturing.

The large model analyzes and models the data in the production process, which can help enterprises realize refined control and prediction; the large model conducts multi-modal analysis and modeling of equipment operation data, which can realize intelligent fault detection and equipment maintenance .

The visual large model can analyze the equipment to be tested. After the abnormal analysis results are generated into text, the language large model can be used to obtain the problem maintenance plan and historical maintenance records. Maintenance personnel can carry out targeted maintenance based on this information to improve efficiency and quality.

Taking the intelligent operation and maintenance of equipment as an example, the work includes the name of the equipment to be tested, testing time, testing content, testing standards and many other details.

Utilize the industry knowledge that the essence of the large model is compressed mode, and the knowledge comes from various equipment documents, maintenance manuals, and operation and maintenance records. Automatic generation of equipment inspection plans can be realized in the future.

Yao Zhiqiang, co-founder of Yuncong Technology, believes: "The problem has always been within sight, and it was not solved before because it was limited by technological development."

In addition, Yuncong has incubated several large-scale model application entrepreneurial projects internally. After accessing the Congrong large-scale model, the products have a new look, such as the Damai digital human live broadcast platform.

Li Jiayu is an "old colleague" of Yuncong, and joined Yuncong's internal entrepreneurial digital person project.

She introduced at the scene: "The field of live broadcast is not only about the anchor alone, but also the test of the cooperation of the entire live broadcast team. The Damai digital human platform can monitor the data of the live broadcast room in real time, so as to adjust the market strategy and ultimately improve the GMV . "

After accessing the calm large model, the Damai digital human platform can generate a one-stop complete solution for people, goods, and venues in addition to generating smart words for the anchor. Match products, pick live styles and more.

Zhou Xi said that Yuncong does not mean that the large-scale model is only available today. We have been thinking about this matter. We proposed the three-wave AI technology route a few years ago, but the technology has not developed to that point. "

Zhou Xi emphasized: "The path is like this, and the rest is just a matter of time."

-Finish-