In the post-DeepSeek era, how can AI applications navigate the "Year of Surpassing"? | 2025 AI Partner Conference

On April 18th, the 2025 AI Partner Conference hosted by 36Kr grandly kicked off at the Shanghai Modu Space. The theme of this conference was "The Arrival of Super APP", focusing on the disruptive transformation of AI applications in all industries. The conference was divided into two major chapters: "The Arrival of Super App" and "Who Will Be the Next Super Application", covering seven major topics such as "Growing Up in the AI World" and "In 2025, Compete in AI by Competing in Super Applications". It included more than 10 keynote speeches, 3 round - table discussions, and the release of two lists of excellent AI case companies. It deeply analyzed how AI technology reconstructs business logic and reshapes the industrial pattern, and explored the infinite possibilities brought by AI super applications.

At the beginning of 2025, the popularity of DeepSeek and Manus opened the prelude to the AI market. When there is a generational improvement in model capabilities, there is no doubt that AI applications will explode again this year, and the concept and implementation path of AI native applications will also be refreshed.

At the site of the AI Partner Conference, Long Ling, the co - founder of Quwan Technology, Liu Yan, the chief product designer of Microsoft, Chen Liping, the co - founder and senior vice - president of Silicon Intelligence, and Li Sen, the senior vice - president and chief expert of Huice Solutions, jointly participated in a round - table discussion titled "2025: The Year of Transcendence for AI Applications".

Image source: 36Kr

The following is the transcript of the round - table discussion, sorted and edited by 36Kr:

Deng Yongyi: Hello, everyone. Welcome to today's AI Partner Conference. I'm Deng Yongyi, the host of 36Kr. I'm very glad to invite guests from different roles and industries to this round - table discussion today.

In 2025, the most pleasant surprises in the Chinese AI market must be the two good starts at the beginning of the year, one is DeepSeek and the other is Manus. There have also been some breakthroughs in embodied intelligence.

These innovations all indicate that AI applications will undergo very profound changes this year. Whether it is the breakthrough of DeepSeek in the model or the popular phenomenon - level application like Manus, they are actually closer to people's lives and more deeply implemented in applications, with breakthrough progress. That's why we want to discuss "The Year of Transcendence for AI Applications" with you guests today.

For the first question, I hope each guest can briefly introduce themselves and the companies they work for. We have also set up a small session where we hope everyone can share an AI moment related to themselves. You can talk about any notable innovations in AI applications in the past year or the moments when you were amazed by AI.

Long Ling, co - founder of Quwan Technology. Image source: 36Kr

Long Ling: Hello, everyone. Thank you for the invitation from 36Kr. I'm Long Ling from Quwan Technology. Our company was established in 2014 and has been deeply involved in the vertical track of mobile Internet interest social networking. We operate a product called TT Voice.

Following the wave of the AI era, we have also made some technological breakthroughs and product explorations around AI voice. Some time ago, our first AI voice creation platform, "Quwan Qianyin", which is equipped with our self - developed model "MaskGCT", was officially opened for external invitation testing. This product is positioned to provide a one - stop intelligent voice solution, integrating functions such as text - to - speech, video translation, and multilingual synthesis, and is committed to realizing the globalization of global content.

The development speed of AI in the past year has been too fast. What touched me deeply is not the breakthrough of certain technical parameters, but the fact that I feel AI has really reconstructed the production chains of some traditional industries. For example, in the scenario of short - drama going global, after a leading platform used our product, the production cycle of dubbed dramas was shortened from 30 days to 3 days, the efficiency was increased by 10 times, and the cost was reduced by 90%. As a result, our product has become the first in the industry to achieve industrial mass production. This kind of feedback from users really surprised me.

Liu Yan: Hello, my name is Liu Yan. I'm usually based in Silicon Valley, USA. I just came back yesterday. I'm very happy to be back home. Thank you for the invitation from 36Kr.

I'm currently in Silicon Valley, USA, working as the chief product designer at Microsoft, mainly responsible for Copilot and a series of other AI intelligent application products.

I think most of you have heard of Copilot, so I won't say much about it. For me, an application I really like this year is Notion AI. I'm a person who likes to take notes and do knowledge management. I think AI has given great help in knowledge collaboration and management. Notion has many functions that can help you organize knowledge bases with one click, provide prompts, and help you complete tasks. It's a product that has really surprised me and I use it quite frequently this year.

Chen Liping: Hello, I'm Chen Liping from Silicon Intelligence. Thank you for the invitation from 36Kr. Silicon Intelligence is a company focusing on AI technology R & D and AI application innovation. It has been established for 8 years.

This year, the biggest impact AI has brought to me is the release of DeepSeek R1. I believe everyone here has also felt the shock, which is no less than the moment when ChatGPT was released three years ago. Everyone has a new understanding of artificial intelligence.

In my opinion, there are three aspects of AI progress that have greatly impacted me. First, a production model for low - cost and high - efficiency AI models has emerged.

Second, when we look at R1, we can find that the answers given by previous large text models seemed correct, but in fact, DeepSeek shows the reasoning process of "Why", which also allows us to see the originality of future AI reasoning and the emergence of self - thinking ability.

Third, why have we been discussing DeepSeek? In fact, it's a process where open - source defeats closed - source. On the day of its release, DeepSeek became a star - rated project in the open - source community. After DeepSeek went open - source, Silicon Intelligence also open - sourced its underlying digital human large model and real - time interaction system. It exceeded 10,000 stars on GitHub within a month.

Now, the HeyGem digital human model we released has become the "DeepSeek" in the digital human field. We also hope that through the open - sourcing of these underlying large models, the ecosystem will become more prosperous, and future applications can be developed based on these underlying large models and technologies.

Chen Liping: Hello, I'm Chen Liping from Silicon Intelligence. Thank you for the invitation from 36Kr. Silicon Intelligence is a company focusing on AI technology R & D and AI application innovation. It has been established for 8 years.

This year, the biggest impact AI has brought to me is the release of DeepSeek R1. I believe everyone here has also felt the shock, which is no less than the moment when ChatGPT was released three years ago. Everyone has a new understanding of artificial intelligence.

In my opinion, there are three aspects of AI progress that have greatly impacted me. First, a production model for low - cost and high - efficiency AI models has emerged.

Second, when we look at R1, we can find that the answers given by previous large text models seemed correct, but in fact, DeepSeek shows the reasoning process of "Why", which also allows us to see the originality of future AI reasoning and the emergence of self - thinking ability.

Third, why have we been discussing DeepSeek? In fact, it's a process where open - source defeats closed - source. On the day of its release, DeepSeek became a star - rated project in the open - source community. After DeepSeek went open - source, Silicon Intelligence also open - sourced its underlying digital human large model and real - time interaction system. It exceeded 10,000 stars on GitHub within a month.

Now, the HeyGem digital human model we released has become the "DeepSeek" in the digital human field. We also hope that through the open - sourcing of these underlying large models, the ecosystem will become more prosperous, and future applications can be developed based on these underlying large models and technologies.

Li Sen: It's a great honor to discuss with you all here. I'm Li Sen from Huice Wangdiantong. Huice Wangdiantong is a SaaS - type e - commerce ERP service provider. It processes more than 100 million e - commerce orders on the platform every day. That is to say, about one - third of domestic e - commerce orders are fulfilled through our platform.

I'll talk about how our customers in the e - commerce scenario use large models in various fancy ways to do some really amazing operations in my opinion.

First, I'll share the story of a cross - border e - commerce customer, some post - 90s guys. Their thinking in cross - border e - commerce is different from others. It often snows in the United States and Canada. He uses AI tools to capture weather forecast data from various places as a reference. For example, when it snowed heavily in Utah last time, he immediately increased the prices of snow shovels, snow gear, and gloves by 30%, and the sales volume increased by 300%. He made money using AI tools, mainly by taking advantage of the information gap.

Second, many of our domestic live - streaming customers draw on the capabilities of multi - modal large models. By analyzing the pictures of the live - stream hosts, the intonation of the hosts, and the emotions and order volumes of consumers in the bullet comments, they optimize the live - stream scripts. These are all things I've seen with my own eyes.

There are even some supply - chain customers whose application of AI has amazed me to what extent? He said he can use a camera to capture the wrinkle degree of product packaging in the warehouse to judge the turnover rate of the goods in the warehouse. There are now all kinds of ways we can't even imagine to process business in various industries through large models, which really broadens my horizons and surprises me.

Deng Yongyi: Are these scenarios using DeepSeek, or are other models also being used?

Li Sen: There were previous OCR (Optical Character Recognition) models. After DeepSeek went open - source, some basic e - commerce companies also connected to it for analysis. There are also some models based on MCP used to handle relatively complex tasks similar to Agent context. We've seen all of these.

How to understand "AI + Application" and "AI Native"?

Deng Yongyi: Today's theme is "The Year of Transcendence for AI Applications". When there is a generational update of models, we also need to look at AI native applications with a brand - new perspective. So, for the first question, I hope each guest can share their understanding of this concept. In the past, we talked more about "AI + Application", but after the emergence of large models, in fact, everyone is talking about the concept of "AI Native". Since you guests come from different industries, you must have different understandings. I'd like you to talk from your own perspectives.

Li Sen: I'll briefly share my experience. I think last year it was more about "AI + Application", like adding a single - point plug - in. For example, wearing a vest when it's cold. For native applications, starting from our e - commerce ERP business, we need to penetrate deeply from the underlying AI core, such as vector databases and streaming pipeline data, and integrate some algorithms, strategies, and rules in vertical fields at the native level. Compared with wearing a vest, it's like exercising to improve resistance and cold - resistance ability, which is an essential difference.

Chen Liping: Since we are also involved in the technology underlying layer, our understanding of "AI native applications" is that, on the one hand, AI itself cannot generate applications. It must be combined with industries. All "AI + Applications" are actually developed on the basis of industries.

What will grow on the industrial basis? As mentioned just now, "AI +" is one way. We also think that the "AI +" approach is more like an AI assistant. Without changing the original underlying system architecture, it strengthens the application logic, and through the way of AI assistance, it adds some functions at the application level to make it more convenient to use, improve efficiency, and reduce costs.

AI native, on the other hand, starts from the underlying architecture and the industrial demand side. It reconstructs the underlying architecture. The demand may be refined by AI. At the data architecture level, demand - side capture and data - side analysis all use AI technology to reconstruct the underlying layer. So, it starts from AI in the design stage and is more future - oriented.

People's usage habits are cultivated. Since there was no AI before, most people have accepted the application forms they can access. Sometimes, the interaction is inconvenient. You need to input 100 words to get the answer you want. Now, with the emergence of AI, you may get the answer with just a few simple prompts. These are all improvements in the experience level brought by AI.

Therefore, AI native is a future direction. It's easier to provide us with a better user experience, and people will get used to such an experience more and more.

Just now we talked about AI assistance. On the basis of AI native, more different interaction methods between users and applications will emerge in the future. I think this is also the subversive direction that AI native brings in terms of interaction interfaces and business models.

Deng Yongyi: What General Manager Chen means is that it will have a very big impact from product positioning to design and specific application links.

Liu Yan, chief product designer of Microsoft. Image source: 36Kr

Liu Yan: I'd like to share a different view first. In fact, I feel that the United States is still stuck in the tool stage and hasn't reached the so - called "AI Native" concept.

In terms of products, our country is already much stronger than the United States. There are more diverse application scenarios and modalities. Abroad, they still regard AI as more of a tool and assistant. Even Copilot, which is a relatively advanced product, can be seen from its name "Copilot (co - pilot)" that it doesn't want to be a real substitute for humans.

Abroad, a big difference is that they don't expect AI to eventually replace humans or have products completely driven by AI. Most people still regard it as a tool.

However, there are still some new directions, which may be more scenario - based. We call it context - driven. They expect to input some prompts, and it can automatically recommend some things according to your usual work habits, scenarios, and workflows. This is already a relatively advanced practice abroad.

It will still take some time to truly achieve AI native.

Currently, what's very popular and often discussed is AI Agent. Agent may meet such expectations in the next few years and really become your assistant or help you coordinate more things. But it's still in a relatively early stage to play a dominant role.

Deng Yongyi: You mentioned that the focus of the AI native concept in China and the United States is different. Is it because in the previous Internet era, China mainly focused on the to - C consumer Internet, while in Silicon Valley, people talked more about SaaS or software - driven models? Or is there some other reason?

Liu Yan: I can share my recent observations on the obvious differences between China and the United States. In the United States, the concept of Super APP doesn't work. They can't create a Super APP, which is related to people's usage habits. They prefer the professionalism and simplicity of each tool.

For example, the workflow is Teams or Slack plus Notion, Zoom.