news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Lei Gangbai sent from Aofei Temple
Quantum Bit | Public Account QbitAI

In the wave of generative AI, scene players are a group that is easily overlooked, but they are often the first to reap the benefits in a low-key manner.

Overseas, this is the case for both Microsoft and Adobe; domestically, Meitu (HK.1357) is also showing this trend.

As a company that has produced many popular products and whose name "Meitu" is used as a verb, Meitu is showing a new look in the AI ​​wave.

There are models. We were the first to hand over large video models in China. After Sora came out, we were the first to upgrade and update the DiT architecture.

There are applications. At this year's Photo Festival alone, as many as 6 products were released;

There are achievements. The number of VIP users worldwide has exceeded 10 million. The revenue of a single AI application will exceed 100 million in 2023, and large-scale profits will be achieved through AI.

However, these changes are only phenomena after the final result. How will AI refresh Meitu? Where does Meitu hope to go in the AI ​​era? All these questions need to be answered.

andWu Xinhong——The founder, current chairman and CEO of Meitu is the undisputed leader.

In the latest conversation with Quantum位, Wu Xinhong confirmed the changes in Meitu in the AI ​​opportunities and shared the insights gained from the practical implementation of AI.

He repeatedly talked about Meitu's role and positioning in AI, as well as its goal, which is both its starting point and ultimate destination: to create products through AI to help users make money.

In the AI ​​era, "beautify it" may have new connotations, representing productivity, cost reduction and efficiency improvement, and users making money through beautifying it.

Dialogue with Wu Xinhong

“Meitu has always been an application company, even in the AI ​​era”

Quantum bits: Meitu has already conveyed the changes in the AI ​​era, but the outside world is also constantly asking: What are Meitu’s advantages in AI?

Wu Xinhong: First of all, Meitu has always been an application company. When we released the first version of Meitu XiuXiu in 2008, we were still in the era of PC applications. We quickly entered the era of mobile applications, and now AI applications.

Essentially, Meitu is a company that excels in AI applications. You can implement these vertical scenarios and quickly realize them. So many people compare Meitu with this big model company, saying, "Why can you compare with OpenAI?"We didn't intend to compete with it., including many new large model companies in China, because they are essentially not on the same dimension, because everyone has their own strengths.

Quantum bits: But you also developed your own large models?

Wu Xinhong: Meitu's self-developed big model is more to make our AI applications more competitive, such as pursuing the ultimate effect, including deep integration with products, etc. So this is actually a misunderstanding. Many people may wonder why Meitu is so successful, but I was originally an application developer. However, the whole era is changing, from PC Internet, mobile Internet to the current AI era. In short, we just want to do a good job in application and monetization in every major technological change.

Many people may think of AGI when they mention big models, and they want to make general big models, but in fact we are making vertical big models in the imaging and design tracks. We think that general big models are definitely running towards the goal of AGI, and in the future they may become an AI assistant with a super brain.

We do vertical work, such as image and video generation and applications, and in the future we will have a collaborative relationship with the general large model, because we know that it is impossible for a general large model to train all capabilities in one model, which means that it has extremely high training and reasoning costs, and it will also become very slow when actually providing generation services. Therefore, it is likely that the general large model will still be used to do things like the brain and central control, and then call on social resources and use various tools to complete tasks.

Quantum bits: Many companies that focus on positioning applications and products will choose not to develop large models themselves and invest less.

Wu Xinhong: First of all, if there is a ready-made API available, we will also access it. We are a company that is very open to cooperation. Sora does not have an open API, and other domestic manufacturers have not opened it before self-research, and these products need to be optimized in the early stages in terms of controllability, cost and other issues. We may not be able to wait, because as a company with many application scenarios in AI applications, especially in imaging and video as one of the important branches, we can only do it ourselves first, but in the future if there is a suitable API, such as a large video model, we will also consider accessing it. It can even be said that in one product, we will provide our own and multiple model options, giving users the right to compare.

Including Adobe's PR video editing software, it has previously announced that it may be connected to models like Runway, Pika, and Sora. In fact, as an application or tool, we are extremely open to this, but now it doesn't exist, so we have to do it ourselves.

Why do we need to develop our own big model? I just mentioned the effect. Confidence in the product and a lot of underlying knowledge can support the competitiveness of your product. Because in the era of generative AI,Depth of CognitionIt’s quite crucial.

Quantum bits: Only by doing it yourself can you understand the know-how?

Wu Xinhong: If you don’t work on the model side, and you just use some APIs in a general way, your understanding may not be very deep. You have to suffer and be beaten before you can gradually build up your understanding and thus improve your competitiveness.

“We don’t need to use a hammer to find nails. The window period for AI application is only two years.”

Quantum bits: Being able to be the first to come up with a large video model is itself a reflection of the accumulation of AI capabilities and competitiveness.

Wu Xinhong: Meitu is an application company, but many of these applications are driven by AI, so we actually have a very strong and large AI vision team in the backend. We believe that our talent has always been first-class in China, but we have not had much exposure or publicity about our AI capabilities in the past. So in terms of video generation, we think we are capable of doing this. You can also see that a team in China has started to do it.We are now on the right track.It's just a matter of time, and we believe we have the ability to do it well.

Quantum bits: The presence of applications and scenarios is also a guarantee for AI talents to be able to use their skills?

Wu Xinhong: Meitu has many very good applications, which is not necessarily the same as the need for AI companies. Because many companies may have hammers and I look for nails. But we are already there,We actually have a clear idea of ​​what users want and what effects we need to achieve.。As long as we work towards the direction of AI technology, we can achieve the results that users want.. This is a matter of time, and I think this is the most important first point.

second,Our technical capabilities are also first-class in ChinaWe have won many awards and participated in many international competitions, and it is not difficult to get some honors. So whether it is a horizontal or vertical comparison, the team is actually quite confident. We feel that we are capable of bringing some of the most advanced AI technologies to the forefront, and we can do it if others can do it.

Quantum bits: I am actually most interested in Kaipai and MOKI, which seem to have opened a new window for Meitu. MOKI is an AI short film workflow specially created for video creators; while Kaipai is a live broadcast scene for selling goods, which can help farmers make selling videos in less than half an hour.

Wu Xinhong: Yes, we are actually exploring several paths for the implementation of AI video large models. Whether it is Kaisha or MOKI, they are both solutions we have given.

pictureStart shootingAfter its launch last year, it has become a benchmark for similar products within a year. This is because it was indeed early in exploring AI workflows, that is, how to build and how to transform the original need to use several products into a single product that can solve all needs.

MOKIWhen we were training large video models and proofing various AI short films, we discovered a big pain point: the materials generated by large video models could not be turned into films or stories in one click. However, we actually have all the technical capabilities required to make AI short films, so why not connect them into one product?

Therefore, this kind of AI workflow is used to solve some of the current pain points in different vertical scenarios.

Quantum bits: When we started shooting this product, we targeted a granularity that even allowed for oral broadcasting, which reflects your ability to discover pain points and define them when making products and applications.

Wu Xinhong: Yes, in fact, the essence of developing applications is to provide services. If you want to serve your customers well, you need to constantly obtain their feedback.We especially like negative feedback. It’s the user’s complaints. The harsher they are, the more valuable they are to us, because they can help us optimize our products quickly.

we think,The window period for AI applications should be only two years, and a year has passed. So there is not much bonus left for developers. The window period is gradually ending, and the track is gradually saturated. After saturation, everyone will go for it. You just mentioned various experiences, and then continue to iterate and optimize based on data. Now it is generally still in a relatively extensive and barbaric growth stage. Because the AI ​​applications in various tracks are actually far from saturated, everyone is now roughly and quickly positioning themselves.

Quantum bits: For a while, it was difficult for me to download new apps. Recently, because of AI, I downloaded a lot of them again to experience them. It seems like a new cycle?

Wu Xinhong: In the future, the survival of the fittest will gradually take place. Because now many people will start from a single point when applying AI, but it is easy to be covered by stronger products and larger companies. These products or companies may not intend to attack them, but they are indeed naturally covered by others. Therefore, we may need to deeply root in an industry, a vertical scenario, and do it very deeply when we make this AI application. This depth includes that your products and technologies in the vertical scenario must be good enough, and then your cognition, because you spend a huge amount of time on it every day, your cognition must be deeper than others. I think this is likely to continue to gain a foothold in the increasingly fierce competition in the future, because there are too many bonuses coming in now, but they may not necessarily combine some of their own advantages. There may be many people chasing the trend, but they are not necessarily real advantages.

Quantum bits: It’s more about chasing the trend and the excitement.

Wu Xinhong: This is still quite easy to be impacted, so we still have a strong sense of boundaries. Our imaging and design track is already big enough, and we even need to continue to go more vertically and deeper in this track to defend our position. Otherwise, once you become impatient, it is easy to expand your territory and be impacted.

“Whether or not a new product can help users make money is the primary criterion for making new products.”

Quantum bits: Is there anything that is clearly not being done now or a sense of boundaries?

Wu XinhongSo now we emphasizeReuse of capabilitiesFor example, the most basic model generation capabilities, including the middle-office capabilities, must be reused in different products to avoid the need for a lot of customized development for each product.

We are refining the strong commonality of products, that is, what capabilities they have, such as the middle platform module. We also pay attention to the scale effect. We must make these investments to obtain better income by serving more users, so that we can have a certain advantage in the competition. Because any industry is a competition of scale effect in the end.

So to sum up, we are refining strong commonalities, reusing underlying technologies and middle-office capabilities, and forming economies of scale in terms of user revenue and scale. Combined with what I just said, we need to go deep into the industry and vertical scenarios.

This is our current sense of boundaries or our understanding of future competitiveness.

Quantum bits: Boundaries convey stability, but they also represent a visible ceiling, and many people prefer to convey that there is no ceiling.

Wu Xinhong: We are a company that does not like to make empty promises. The Meitu Design Studio and Kaipai application cases mentioned at the press conference were all promoted by Yiwu e-commerce sellers and villagers from Leshui Village, and they are very real. We are not afraid of people thinking that we are LOW, because this is what the public needs. We are a company that creates products and provides good services for the public, because the company's values ​​are to seek truth and be pragmatic, and to work hard to win, so what we present is what we think in our hearts.

Now when we think about whether to make a product, the first thing we consider is whether it can help users make money.This is the premise. Only when we help users make money will they be willing to pay and we can make money. It is very realistic. I think many users are in urgent need of making money, so this is the most basic consideration when we decide to make a product.

Quantum bits: This is different from your earliest well-known products Meitu XiuXiu and Meipai. Previously, they were more about meeting user needs, even interactive entertainment needs, but now they are all productivity tools?

Wu Xinhong:becauseGenerative AI is naturally closer to industryFor example, Meitu Cloud Repair helps photo studios make money, Kaosei helps talk show hosts, and Meitu Design Studio helps small and micro e-commerce sellers. These are ways to help them make money and improve efficiency, and can indeed help the industry to reduce costs and improve efficiency.

Second, it has service costs.Of course, many of them are still in the cloud. Of course, using this kind of terminal computing power may effectively reduce costs in the future, but at least for now many AI applications need to cover their generation costs through subscriptions and single purchases.

Therefore, it is naturally suitable to be used as a productivity tool. It can help the industry reduce costs and improve efficiency, and the cost can also be covered through subscriptions.

The combination of generative AI and productivity tools is also an inevitable choiceAt least you have to make your entire business model work, the so-called growth flywheel, and you can make money to feed back the investment in R&D. When it comes to R&D investment, it will be even greater, just like the big models mentioned earlier, or our investment in a strong middle platform.

"Say goodbye to the mentality of hot products, the competition for productivity value is endless"

Quantum bits: Do you think the outside world has a fair sense of the changes that AI is bringing to you? Meitu is being refreshed by generative AI.

Wu Xinhong: To be honest, fairness is not something we can talk about all day. You need to understand me objectively, but if you really make strong products and gain good user and revenue growth, people will naturally change their views of you after seeing the results. In fact, many people think that Meitu is the company that has been hit the hardest in the AI ​​era.

Quantum bits: Really? Did someone tell you that?

Wu Xinhong: For example, after Apple's AI debuted, I wondered why Apple users would need your Meitu products in the future? But this is a question of depth, because the depth of AI assistants is actually limited, and the links are relatively shallow. So if you want to go deep, you can have a complementary relationship.

we sayAI AgentIt uses various tools and capabilities to complete specific tasks, rather than being able to do everything on its own. So I think this requires the whole industry to grow together, and users will gradually establish a relatively objective expectation of AI in actual use.

Quantum bits: Do you think there will be a time point or period when people will realize that Meitu’s AI products are starting to explode?

Wu Xinhong:usChoosing to create productivity tools in the era of generative AI is a long-term choice.Because productivity tools aren’t thatFashionIt doesn’t mean that we can create a hit product every day and become famous overnight. But it is indeed a long-term and valuable thing, so we continue to accumulate capabilities and iterate products. Maybe as users gradually accumulate and product strength improves, it will gradually reach a stage where it gets better and better. So it doesn’t necessarily mean that one day it will really explode and completely change. In fact, it is a so-calledWater does not compete for the lead, but for the continuous flow.We don’t necessarily rush to grab it at this time and say that we are the first.

For example, the capabilities of video big models will gradually catch up with each other in the next year. For image big models, Midjourney is far ahead at the beginning, but there will not be much difference in capabilities later. The key is how to implement the application and how to monetize it. So we will keep a low profile and do things well. We may not deliberately talk about how to lead the industry. In fact, it is not important.

Quantum bits: My mentality is very peaceful now. Meitu has had its peak. Do you feel a sense of disappointment now?

Wu Xinhong: No, I’m in a good mood now.We really want to help users create valueWe think this is the right thing to do, but it is a relatively long process. So again, since we have chosen to make productivity tools in the era of generative AI, we must be able to endure some of the long-term efforts in the early stages, and we must break away from the inertial thinking of making C-end products and blockbuster products.

Because in essence, they are two different things. We are now identifying the right industry, the right use scenario, and the right target user group, and constantly strengthening the product and serving them well. We won't go for anything else, and we don't necessarily have to take all the opportunities. We just need to do our best.

Therefore, whether it is Kaipai or Meitu Design Studio, under this concept, they are developing very well, and the user income is growing very fast. This is value, real value.