Mingshi, BAI und Ant Group, die Optimisten, Pessimisten und Mittelgesinnte in der AI - Investition, versammeln sich um einen runden Tisch.

Text | Zhou Xinyu

Editor | Su Jianxun

Today's AI investors can also be divided into "optimists" and "pessimists".

On September 12, 2025, at the Inclusion Bund Conference. At the round - table forum "The First Battle of AI Application Implementation: Has the Agent Era Arrived?", hosted by Feng Dagang, the CEO of 36Kr, several investors from leading institutions had a dialogue.

When Feng Dagang posed the first question: What has led to the explosion of agents? The biggest divide among the investors emerged.

△ Round - table forum "The First Battle of AI Application Implementation: Has the Agent Era Arrived?". Source: Official

Huang Mingming, the founding partner of Mingshi Venture Capital, who has invested in popular agents such as GenSpark, Lovart, and Sheet0, is a typical "optimist".

He mentioned that China has accumulated a large amount of product capabilities in the mobile Internet over the past 20 years. Two - thirds of the world's top agents in the future will come from China.

However, the "pessimist" investors see that today's agents still have many "inabilities".

In the view of Long Yu, the founding and managing partner of BAI Capital, the current explosion of agents is inseparable from users' unprecedented tolerance for AI.

But in scenarios with zero tolerance such as finance, she believes that agents still cannot be implemented on a large scale. And the "window period" of tolerance for agents is not long, "People will become more and more demanding."

Between these two types of investors, there is also a middle - ground group looking for opportunities.

Ji Gang, the vice - president of Ant Group and the president of the Strategic Investment and Corporate Development Department, gave advice to today's agent entrepreneurs: They can start with scenarios with high tolerance.

Representatives of such high - tolerance scenarios are tool - based scenarios that require interaction with real people, as well as companion scenarios for killing time and providing emotional value.

It is obvious that "Has the agent era arrived?" is one of the most concerned questions in the industry at present. The eagerness for an answer once filled the venue of this event with a limited number of seats, and there were also many audiences standing outside to listen.

However, as Feng Dagang summarized: The emergence of non - consensus precisely means that the development of agents is still in its early stage.

The following is a compilation of the conversations between Feng Dagang, Ji Gang, Huang Mingming, and Long Yu by "Intelligent Emergence", with the content slightly edited:

Two - thirds of the world's top agents in the future will come from China

Feng Dagang: The era of the explosion of AI agents and applications has arrived. I think this should be an undoubted conclusion.

What kind of inevitability do you think has led to the explosion of AI applications? Technology, cost, or revenue?

Ji Gang: Actually, it is the current level of technological development that has reached such an opportunity.

△ Ji Gang, vice - president of Ant Group and the president of the Strategic Investment and Corporate Development Department. Source: Official

Of course, on the other hand, our expectations for intelligence have far exceeded the previous generation's understanding of technology, such as some knowledge retrieval and the establishment of simple workflow - level work capabilities.

But today we see that in fact, AI potentially may be an intelligence that can surpass human wisdom. If there is a person around you who is smarter, more knowledgeable, and never sleeps 24 hours a day, what can he do? In fact, he can do everything.

Huang Mingming: I may be the only science - and - engineering guy in this forum today, so I'll talk a little bit about the technical details.

△ Huang Mingming, founding partner of Mingshi Venture Capital. Source: Official

We have been observing from the model layer. In the past two to three years, first, as everyone knows, the computational cost on the inference side has been reduced by about 280 times. Then, MoE (Mixture of Experts) has further reduced the computational cost by 80%. This is from the perspective of cost and efficiency.

From the model side, from DeepSeek's R1 at the beginning of the year to Claude 3.5 and 4.0, I think the most core ability is to enable the model to have the ability to call tools and plan.

For example, I remember when I was a child in biology class, the teacher always asked us what the most fundamental difference between humans and animals was. In fact, one thing is that when humans started using tools, it was a very important step in the evolution from primitive humans to Homo sapiens.

If we look at the development of AI today, when Claude 3.5 and Claude 4 enable the model to call a series of tools to complete a relatively complex task, it has a certain planning ability.

This suddenly increased the task completion rate of agents from most 0.5 last year or the year before to 30 - 40 points for the top agents this year, such as the Vibe Working or General Purpose Agent mentioned earlier, including GenSpark that we invested in.

From 0.5 to 30 or 40 points, it makes users' repeated usage rate and willingness to pay become very high.

When can it reach 60 points? We think it should be within the visible 8 - 12 months, so that more work can be unlocked.

I think you may need to add one more thing to your question: Why are many of the "firsts", the "world's first agents" in this wave from China?

Not only the first General Purpose Agent Manus we mentioned earlier, GenSpark of Vibe Working that we invested in, Lovart, the world's first Vibe Design agent, but also the first Data Agent sheet0, etc.

△ Lovart. Source: Lovart official website

I think this is because in the past 20 years, China has accumulated a large amount of product management capabilities in the mobile Internet, and at the same time, we have been keeping up with technology very quickly.

As mentioned earlier, agents use models to call tools, and there is a lot of engineering and high - speed iteration in the last mile. This is precisely the unparalleled advantage and ability of Chinese entrepreneurs in the world.

So we boldly predict that among the world's top - used agents in the future, although the headquarters of the final company may be in Silicon Valley, Singapore, or London, two - thirds of them must come from Chinese entrepreneurs.

Long Yu: Mingming gave a lot of popular science just now, but my answer is: The era of the explosion of agents has not arrived, and I'm not so optimistic about the prospects of Chinese entrepreneurs either.

△ Long Yu, founding and managing partner of BAI Capital. Source: Official

We do have great advantages, but the concept of agents is still in the process of being defined and implemented. In fact, we are still in a process of perseverance and continuous progress.

In fact, on the user side, whether it is consumers or enterprises, they have given the so - called "agent" concept extremely high and unprecedented tolerance.

But in previous enterprise - level services, people would never accept any product that says "I may be able to complete your task". How can I take such a risk?

So this time, the whole world has embraced AI exploration with a very tolerant attitude. There are countless inputs on the development side, countless product definition experiments, and countless feedbacks from the user side, constantly making fine - adjustments, all promoting development with a very tolerant attitude.

But I think the window of opportunity left for us is not long. In fact, after the inference cost has been greatly reduced this year, the possibilities of redesign, improvement, optimization, and workflow architecture have emerged.

But in the end, people will become more and more demanding about the completion rate of agents. Because after all, there is a considerable amount of data flow and data that must be completed accurately in a fully controllable and strongly executable coding environment. For example, in finance, there is no room for any tolerance. It's not just about doing a good job.

Therefore, I think in fact the industry is starting to differentiate, models are starting to be industrialized, and everyone is starting to think deeply.

Entrepreneurs on both sides of China and the United States have made many surprising progresses. Chinese entrepreneurs in this field have shown super - strong coping abilities and very fast execution speeds. I think their contributions are huge, and they will definitely have their place.

But in fact, compared with agents, the intelligence concentration of the model itself, as well as the cliff - like and runaway - style lead, and what they launch - whether it is called an agent, an app, or a function - is still far ahead.

Feng Dagang: First of all, I don't think it's a pessimistic situation today, but I very much agree that the industry is still in a relatively early stage. Having different views is precisely a characteristic of the early stage.

Let's get back to the issue of applications. Ant has developed many applications and agents, including those for pension, medical care, and family. Our theme is "The First Battle of Implementation". Where do you think "the first battle" is?

Ji Gang: In response to what Mr. Huang and Ms. Long said, I'll give my answer to this question.

I'm a liberal arts student. Please don't discriminate against liberal arts students. Liberal arts students have some visionary ideas in this world and may have more opportunities than science students.

Long Yu mentioned that the whole world today has a high tolerance for agents, including in work and in life.

Today, there are a large number of offline opportunities for interaction with people. When you travel, you look for a travel consultant; when you look for a house, you look for a so - called agent to help you. When you interact with them, you definitely hope they can provide you with relatively accurate information.

If there are deviations in this information, whether due to knowledge limitations or for some other reasons to give you wrong information, you can judge and tolerate it.

Actually, I think in vertical fields, this is the relationship between us and agents today.

There is also another scenario with high tolerance. Now there are many agents replacing people closer to you, such as friends, relatives, and partners. When you spend time with a real person, you have a high tolerance. So I think in this scenario, people may also have a high tolerance for agents.

Except for work scenarios that require extreme precision, in many other scenarios, agents can gradually evolve in a low - tolerance way.

Feng Dagang: Will the explosion of applications start from the periphery first?

Ji Gang: There is a relationship of distance. For example, an offline real - estate agent is far from you in terms of relationship, but your interaction frequency with them is very low, maybe only twice a year.

There may be new opportunities for agents here, a kind of tool - like service that combines real people, tools, and data.

Another type is closer to you, helping you kill time or providing emotional value. There are opportunities for high - tolerance scenarios in both of these areas.

On the contrary, for those scenarios that require accurate data to complete work and form a closed - loop, I think it still takes some time. The capabilities of models and agents have not fully developed.

Speaking of your question, we have indeed developed agents for finance and medical care. These actually have relatively high requirements for accuracy.

When you use them today, you will find that most of the interactions still suggest that you finally go to the hospital to register and interact with a real doctor. They may only provide some basic information consultations.

Since finance and medical care are our own businesses, we will definitely develop agents in this area. Whether it can meet everyone's expectations, achieve a closed - loop, and be extremely accurate definitely takes a process. But we also hope to explore more opportunities with low tolerance.

It is unacceptable for an AI efficiency tool not to be profitable within one year

Feng Dagang: Mr. Huang, I know you have invested in a lot of hard - tech projects.

You mentioned just now that you have invested in many good projects. What are the fundamental differences in this wave? Including investment logic, investment opportunities, and our own judgments.

Huang Mingming: Actually, there are quite a lot of commonalities.

For example, the people we like are still super product managers with non - consensus views, like Li Xiang and many others we have invested in who have a deep understanding of user needs.

If we have to say what's different, in the past, Internet applications were more about connecting more information with people and building a production relationship.

Agents are productivity tools. As both of you mentioned, the two most core indicators for measuring productivity tools are the complexity of the tasks completed and the task completion rate. So you need to deliver results.

For those who do well, if we evaluate them, they can achieve 30 - 40 points in their fields. But the tolerance of global users is indeed very high.

What I mean by "tolerance" is not just trying it once and then giving up. It means that the rate of repeated use and repeated payment should exceed 50%. That is, 50% of the people who have paid once should come back to use it the next month. You don't need to regard it as a full - time employee, but at least you can use it. This is a very important measurement indicator.

In fact, the best agents today can only achieve 30 - 40 points in many fields. In other words, even in some very niche fields, if you have the opportunity to fully exploit the model's capabilities and do the "last - mile" dirty work, reach 50 or 60 points, then a large number of users will be willing to pay for it.