Preview Chengdu Big Model｜Dialogue with Mingtu Technology: Metaverse is still an important development trend, and the catalyst should be "digital human"
2024-08-07
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Preview Chengdu Large ModelThe development of artificial intelligence is in full swing, integrating with thousands of industries and generating new technological power. As the key to the competition of artificial intelligence, big models will enter the first year of monetization this year.Red Star Capital noted that in this industry competition, the compound growth rate of Chengdu's artificial intelligence industry has exceeded 40% in the past three years, and its "competitiveness of artificial intelligence technology industry" has ranked sixth in the country. Sichuan Province has even listed artificial intelligence as the "No. 1 Innovation Project" this year.Recently, Red Star Capital interviewed several AI large model manufacturers in Chengdu and discussed with them the "Chengdu power" in the artificial intelligence trend.
In May this year, the "Mingtu WorkGPT" large model of Chengdu Mingtu Technology Co., Ltd. (hereinafter referred to as "Mingtu Technology") passed the filing and approval of the "Interim Measures for the Management of Generative Artificial Intelligence Services". Earlier in February this year, Mingtu Technology's work content generation algorithm successfully passed the filing of the deep synthesis service algorithm of the State Internet Information Office. Since then, Mingtu Technology has become a large model manufacturer with "double filings".
Recently, Red Star Capital Bureau had a dialogue with Xiao Xuesong, Chairman of Mingtu Technology. He believes that "reinforcement learning" should be regarded as a core capability in vertical models. Digital humans are an important catalyst for the re-emergence of the metaverse and its development into an emerging industry.
Xiao Xuesong photo provided by the interviewee
The "school" scene may explode sooner
Red Star Capital Bureau: What is the company's original intention to create a digital human avatar? What is the company's vertical model better at? What aspects does the training focus on?
Xiao Xuesong:The original intention of making digital human avatars is to use artificial intelligence to quickly assist and replace everyone's repetitive and inefficient work. Currently, there are three forms of display.
The first type is APP, which is equivalent to everyone creating an avatar. You can feed your digital avatar with knowledge and give it basic and business capabilities, such as helping you write and make videos.
The second form is to combine the digital human with a box or a desktop office robot to form an intelligent terminal. For example, you can embed the digital human avatar into a large screen, and then you can talk to the other party and arrange tasks through a box.
The third form is implanted in a robot, which can perform certain actions according to your instructions, such as helping you get things, receiving customers, etc.
We have a comprehensive model structure called "Mingtu WorkBrain", and the cognitive intelligence part belongs to the vertical model. Its parameters are between 15 billion and 20 billion. We use office data for training, including knowledge such as personnel responsibilities, job responsibilities, work processes, management systems, and project management specifications.
In the early days, we focused on the office field and aggregated and analyzed the data from the office process. Recently, we have also expanded to hospitals, schools and other fields. For example, the digital avatars of doctors, students and teachers can quickly enhance the relevant intentions.
Red Star Capital Bureau: Where are the application scenarios for digital human avatars, and what is the fastest way to achieve implementation?
Xiao Xuesong:For example, in office scenarios, the most common applications are around tasks such as employee weekly reports and work summaries. The digital human avatar can complete the work quickly according to instructions, and can also use the digital human avatar to achieve group collaboration capabilities.
If you are a student, you can directly communicate with your avatar about your learning problems by combining your learning ability with your digital avatar. We have developed an exam system with the Cognitive Intelligence Laboratory of the University of Electronic Science and Technology of China, which combines the ability to grade exam papers and homework with digital analysis. This digital analysis can also help students automatically check which knowledge points they have not mastered during the test. The digital avatar will remind students to focus on the knowledge points they have not mastered.
In a psychological counseling scenario, communicating with one's digital avatar does not involve privacy issues and may be easier than involving psychological institutions.
In the health field, especially for the elderly, the genetic data of this group of people can be combined with the digital human avatar, including physical examination content. On the one hand, it ensures the security of their own health information, and on the other hand, based on health information, it can help users maintain daily healthy living habits. We have also been integrating with their original smart information systems in some hospitals to assist doctors and hospitals in providing patient services.
In addition, we think that the "school" scenario may explode faster, because students are more likely to accept new things. Some domestic universities have already tried to make digital avatars. The "avatar" is equivalent to your own "assistant". You can "feed" it with your own data about study, work, and health, and it will help you complete your plan or make preliminary preparations.
Red Star Capital Bureau: How is this digital human different from the ones we have seen before? What level of interaction can it achieve? What state do you hope to achieve in the future?
Xiao Xuesong:Digital humans have gone through four eras. The first stage is to bind a two-dimensional image, which is equivalent to IP; the second stage is the virtual digital human stage, combining human expressions, voices, and forms with digital humans; the third stage is the initial integration with large models, such as the digital humans we often see playing programs in cultural tourism and TV stations; the fourth stage is "industrial digital humans", which are integrated into the service industry, production industry, agriculture, etc., so that the digital humans are deeply bound in the subdivided fields, and the industrial digital humans can also work together with upstream and downstream to carry out group collaboration. "Industrialization of data knowledge" and "industrialization of group intelligence" are two typical signs of "industrial digital humans", and we are now in the fourth stage.
We are also making a "box" to achieve virtual-real interaction, which is somewhat similar to AR and VR. Through your expression, face recognition, gesture recognition, etc., it can interact between a virtual digital person and a real person in reality. For example, if a colleague cannot attend a virtual meeting, his digital avatar is bound to the "box". When the meeting begins, he can plug his "box" into the TV and his digital avatar will appear.
The digital avatar image was provided by the interviewee
During a meeting, you can make speeches or give presentations based on the process, which is very similar to the scenes we see in movies. Now the scene can be initially realized, but there are still some technical difficulties to be overcome, such as how to identify mixed voices in multiple scenes and how to interrupt.
After these difficulties are overcome, the state of "being able to be interrupted at any time" can be achieved. After being interrupted, the digital human can answer the question and then return to the previous logic to continue. After the technology matures, teachers, live sales, unattended operations and other aspects can reach a commercial level.
Digital humans are the catalyst for the resurgence of the metaverse
Red Star Capital Bureau: This year is also known as the first year of big model application. What do you think of the current trend of the entire industry?
Xiao Xuesong:At present, large models still have problems such as illusion and lack of pertinence, which makes it unlikely that large models can promote the implementation of application scenarios in the short term by relying on their own intelligence level. Judging from the voice of the entire industry, the application of vertical models is a direction.
In the past, large models emphasized "deep learning" to improve the general intelligence level, but now we find that "reinforcement learning" may be more important in practice. We should regard "reinforcement learning" as a core capability in vertical models.
When solving a single application scenario in the early stage, the large model may seem stupid, but after using it for a long time, you will find that it becomes smarter and closer to what you want. This is where "reinforcement learning" plays a very important role.
If a mechanism can emerge on top of the integration of reinforcement learning, it may also be able to speed up the implementation of large-model artificial intelligence.
Judging from the current explosive applications on the C-end, I feel a little pessimistic. There is also a new saying abroad that it depends on whether the "world model" is built. Its construction will lead to new explosive applications and change the early development methods of artificial intelligence. Even some concepts will be overturned, but this is very difficult to achieve.
On the B-side and G-side, for more concrete applications in certain scenarios, I expect that some general AI applications will emerge this year.
Red Star Capital Bureau: Mingtu is also involved in the Metaverse. Compared with the current AI craze, the Metaverse is relativelycold” What do you think of the two markets of artificial intelligence and metaverse?
Xiao Xuesong:The reason why the Metaverse was not popular in the early days was that, when there was no artificial intelligence, it was more about integrating virtual reality and building a new experience in the form of digital twins. There were no organisms or intelligent entities in it to discover various resources.
If there is no "human" element in any virtual world, it is a static virtual world. Digital human avatars actually combine human wisdom with digital humans. If digital human elements are added to the static metaverse, they can communicate with each other, socialize virtually, and engage in commercial activities, and the world will become active.
Digital humans are a very important catalyst for the metaverse to rise again and develop into an emerging industry. There are also future virtual industries in the metaverse, such as those related to cultural creativity and digital products, including literature, art, games and other virtual industries. The metaverse is a very important carrier and development trend, but its catalyst should be digital humans.
Capital needs to continue to invest in technology
Red Star Capital Bureau: What are Chengdu's advantages in developing artificial intelligence? As an enterprise, what policies and support are necessary?
Xiao Xuesong:"Artificial intelligence" is the "No. 1 Innovation Project" of Sichuan Province this year. Chengdu is rich in talents in the software industry and digital industry, and has advantages in talent and application needs. At the same time, the government is now opening up some scenarios, and government affairs are also encouraging everyone to use artificial intelligence to upgrade existing information technology.
It is necessary to drive information technology upgrades. After so many years of information technology construction, enterprises actually have a large amount of data and a relatively mature process support management system. The current goal is to use artificial intelligence to further tap the value of data elements, upgrade the original more complicated content, and expand the demand market.
In the capital market, I think we still need to create a more relaxed environment and make more sustained investments. In the fields of reinforcement learning and verticals, we still have many technical details that we have not mastered, and we need to make certain investments to support intervention. This is a long-term investment, and it cannot be solved by scenarios or subsidies.
Red Star News reporter Wang Tian
Editor: Deng Lingyao
(Download Red Star News and get a reward for reporting!)
