Fang Yi of MobTech: From DI to AI, there are at least four levels of entrepreneurial opportunities above large models.

People are stumbling towards AI. Are you worried about where the development space for ordinary people lies when surrounded by large models? What kind of productivity transformation will be brought about when AI is more integrated with scenarios? The emergence of intelligent agents has led to a fierce collision between the original App ecosystem and open protocols. Will the next stop of the Chinese Internet be "closed" or "symbiotic"? This article is compiled based on the keynote speech of Fang Yi, the CEO of Daily Interaction, at the "High - quality Development Forum of Artificial Intelligence Terminals" of the 2025 China Internet Conference. It will discuss with you the most hardcore survival rules and ecological codes in the era of large models. The following is the transcript of Fang Yi's on - site speech:

I. From "Data Intelligence" to "Artificial Intelligence": Observations from an Industry Practitioner

Today, I've seen a lot of technical inferences and in - depth speculations about large models on - site. Everyone has shared many systematic summaries. As an industry practitioner, I'd like to share my own feelings.

Looking back at the history of AI development, from the Dartmouth Conference to the present, it has experienced three waves of upsurges and two waves of downturns. Regarding the translation of "Artificial Intelligence", I think "Artificial Intelligence" easily makes people think of "replacing human labor" or "human + intelligence". Perhaps "Artificial Intelligence" is a more appropriate translation, and its corresponding term is "Divine Intelligence". For a long time, the development of AI has mainly focused on anthropomorphism and human - likeness, such as chess games. This leads to a key question: Scientists have extracted highly abstract features from humans (such as the Transformer model) to shape today's AI. Then, how can humans relearn these abstracted essences from AI?

II. Insights from the Three Stages of AI Development: Master the Rules, Think a Few Steps Ahead, and the Early Bird Catches the Worm

We can draw wisdom from the development process of AI to understand its evolution:

The first is to master the rules (like in Gomoku): Just as freshmen can write simple human - computer battle programs, the core is to master the limited rules (such as blocking and forbidden moves). Remembering the rules is enough. In reality, it's the same when people take exams for certificates. They study the rules of various industries carefully, obtain the "certificate", and figure out things within the scope of the rules.

The second is to think a few steps ahead (like Deep Blue): The chess records of international chess can be exhaustively memorized, but why can't humans beat "Deep Blue"? Because Deep Blue can calculate 12 steps, far exceeding the human limit of 10 steps. In reality, it's very rare to think more than others. In the workplace, those who can "think two steps ahead" (such as predicting equipment failures and coordinating resources in advance) are more likely to stand out and become experts. Their value lies in whether they can foresee your next problem and prepare a solution - just like Deep Blue, winning by calculating two more steps.

The third is the early bird catches the worm (like AlphaGo): Facing Go, which cannot be exhaustively calculated, AlphaGo first learned from human chess records and then surpassed humans through reinforcement learning (self - play + reward function) until AlphaGo Zero reached an invincible state. This reflects the power of "starting early and making continuous progress". So when learning AI, we should learn from the "early bird". Every step is a step forward, just as the old Chinese saying goes, "Make a little progress every day, and all efforts will pay off."

III. Core Challenges in the Era of Large Models: Attention, Data, and Ecosystem

Today, many experts have mentioned that a very important aspect of AI is "All About Attention" (attention mechanism). An important part of AI also achieves efficient information processing through multi - head attention (Multi - Head Attention) and mixture of experts model (MoE). Currently, Kimi also performs very well. It uses fewer multi - head attention models but more MoE models. Usually, the power consumption of the human brain is only twenty or thirty watts (about forty or fifty watts during the college entrance examination), while the energy consumption of a single NVIDIA graphics card is as high as 4000 watts. So, there is still a hundred - fold improvement space for machines to imitate the human brain. But on the other hand, Moore's Law or Super - Moore's Law will drive its rapid progress. So in the future, it's meaningless to compare computing speeds with large models. What we need to do is to master the understanding of data and become human experts.

For example, in the mixture of experts model, taking medical image recognition as an example, currently, the recognition accuracy of AI for CT scans can reach 95% or even 97% or higher. In contrast, the accuracy of a human expert is usually around 92% - 93%. This leads to an interesting phenomenon: Human experts have now learned to "strike after the opponent has shown their hand" - they won't rush to give their opinions before the conclusion is out. Instead, they will talk about something else first. After the AI's conclusion comes out, the experts will carefully review it and find the crucial 3% errors. This process precisely proves the indispensable value of human experts.

On the contrary, if human experts give their opinions first, AI is likely to find 7% errors. So the future trend is that humans will increasingly understand the wisdom of "speaking later" - let AI show its abilities first, and then humans can give full play to their unique judgment and error - correction advantages to form more efficient cooperation. This may also be the most important key point when humans coexist with AI in the future.

In the process of practicing big - data applications, we have deeply thought about its core components and summarized them into three key elements: Data, Machine, and People.

Data is the foundation, and machines are the carriers on which various algorithm models are deployed. The most crucial thing is people's ability to understand scenarios. This point has been emphasized by experts in the industry and will never go out of date. As recent industry discussions have shown: In the past, it might take one product manager and seven programmers to implement an idea, while now one product manager may only need half a programmer. This shows that the pure technical implementation part is becoming relatively less important.

This reminds me of the framework proposed by the American company Palantir: Data, Logical, Action. This model is well - known in the industry and echoes the five - layer architecture in the AI era summarized by the China Academy of Information and Communications Technology and other experts.

Regarding the evolution of the knowledge hierarchy, there is a consensus in the industry: In the era of data intelligence, data turns into information, information turns into knowledge, and knowledge turns into wisdom. In the AI era, the mainstream view, such as the classification by "Queen of the Internet" Mary Meeker, describes the progression of abilities: from chat interaction, to logical reasoning, then to intelligent decision - making, and finally to becoming an innovator and an organizer. Of course, different experts may have different expressions.

This leads to a practical question: How to choose a career in the future? Recently, when my niece was filling out the college entrance examination application form, she asked me, "Uncle, which major should I choose?" My advice is: Either root in the underlying disciplines that are difficult for AI to replace, such as physics and mathematics, or become an 'organizer' in the AI era. She finally chose the film and television directing major. In the future, the value of jobs similar to film and television directing may become more prominent - because the core of film and television directing lies in 'what to express' (What), 'why to express' (Why), and 'how to express' (How), which is the embodiment of creativity. As for the specific 'how to achieve' (How to do) and 'what to do' (What to do) at the execution level, it is very likely to be solved by AI in the future.

IV. Future Path: Edge - Cloud Collaboration and an Open Ecosystem

We believe that the key to future AI applications lies in the collaboration between the edge (device side) and the cloud (center side). During the Spring Festival, we invited the major AI companies in Hangzhou (jokingly called "gathering the seven Dragon Balls") for a four - hour in - depth discussion. Two major technical schools emerged from the discussion - the "cloud school" advocates that models should be centrally deployed in the cloud; the "local school" (like us) believes that models should be closer to users and deployed locally.

In the practice of promoting local deployment, we observed that enterprises' willingness to invest is related to their market value - they are generally willing to invest about one - ten - thousandth of their market value (for example, a company with a market value of 100 billion is willing to invest 10 million, and a company with a market value of 1 billion is willing to invest 100,000). However, the cost of achieving "full - blooded" (high - performance) local deployment with domestic computing power is still as high as five or six million, which is too high a threshold. So, to lower the threshold, we also self - developed a low - cost device at an early stage and successfully stuffed a complete 671B parameter model into it. But this brought some performance compromises - slow output, limited processing ability, and sometimes it would "slack off". This experience is far from ideal.

In the process of implementing large - model applications, we have also accumulated a lot of our own experience and lessons. We found that "blindly trusting AI is worse than having no AI", and there is also a saying "Countable numbers are not the real numbers, and the uncountable ones are". This means that when using AI, the quality of the data itself is very important. Because in addition to hallucinations, models also have the "good - student syndrome" - they will "innocently" regard any information input by humans as absolute truth, quoting it comprehensively but resulting in mediocrity.

To overcome these limitations, high - quality prompt engineering and continuous personalized data accumulation become particularly important. We see a clear trend: The capabilities of models are moving towards "equal rights" - the basic large - model technology is becoming increasingly popular and equal under the impetus of open - source forces and giants. However, the value of data will never be "equal". Therefore, the key future direction is to integrate public knowledge, private data, and personal preferences to create highly personalized exclusive intelligent models. This will be the core competitiveness of the next - generation AI applications.

But in the process of promoting technology implementation, we often face a core dilemma: "Afraid of high costs when using a private cloud (the poor are trapped for a while), afraid of data leakage when using a public cloud (the timid are trapped for a lifetime)". How to solve this dilemma? This also makes us think about a more fundamental question: Will the future of the Chinese Internet ecosystem be closed or open? We know that Google pays Apple tens of billions of dollars every year to ensure it can get traffic from iPhones; Amazon also spends a lot of money to buy Google traffic.

In China, search engines and e - commerce are separated. Major APPs tend to build closed ecosystems, firmly locking users and traffic within their own systems. This closed - door tendency was pre - staged in the early days of the mobile Internet. Around 2011, we discussed with industry executives: 'What will be the entrance to the mobile Internet in the future? How will the form of mobile search evolve?'

At that time, I predicted that the app store would become the core entrance. Users would first search for and download apps in the app store and then conduct vertical searches within the apps. The role of general search engines (such as Baidu) would transform into a "backup provider" - mainly serving the long - tail needs not covered by the app store, and its importance would be far less than in the PC era. Later development has confirmed this prediction.

V. Ecosystem Openness: An Inevitable Choice for AI Development

So, in the face of the emerging intelligent Internet, that is, the AI era, should the industry be "closed" or "open"? We have also widely consulted the opinions in the industry. The mainstream view is that China doesn't need to completely copy the MCP protocol model based on the open Internet in the United States, but must find an open path suitable for its national conditions.

How to promote openness? We also have some initial ideas:

One is to build a general technology layer. Just like there was an industry - level message - pushing SDK in the mobile Internet era, we can launch an AI SDK to connect the underlying general - ability interfaces and lower the access threshold.

Two is to jointly build an open application ecosystem: Under the guidance and coordination of the competent department (such as the Ministry of Industry and Information Technology), promote the establishment of an open AI application and service ecosystem platform. Its core is not to force users to download and install but to provide a unified and convenient service discovery, access, and call mechanism. So I think the key in the future may be to optimize the protocol with necessary "additions" according to China's actual situation, and the industry competent department should take the lead in quickly establishing implementable technical standards or procedural specifications and keep them updated continuously. At the same time, industry companies, including us, should also be more active in contributing open - source code and basic modules to the industry to promote technology sharing and ecosystem prosperity.

Regarding how to better list relevant services, I also made a simple cooperation block diagram for everyone to see. Our core goal is to make it convenient for service providers to list their services. But a key question is: Who will review and govern the services? Relying on a single company is definitely not realistic. This requires us in the ecosystem to jointly explore an effective governance mechanism.

VI. Industry Symbiosis: Sense of Boundaries and Diversity

Not long ago, we jointly organized a deep - level closed - door seminar in Hangzhou with institutions such as the China Academy of Information and Communications Technology. The participants included diverse App representatives from leading to mid - tier companies. An interesting episode was that when old friends met, they joked with each other, asking, "Why haven't you 'died' yet?" - This precisely confirms that seemingly basic functions such as weather queries, calendar services, and message pushing can survive for ten, twenty years, or even longer, transcending market cycles. Behind this is the sustainable value formed by the ultimate satisfaction of users' rigid needs and continuous in - depth exploration. These "small and refined" services form a solid foundation for the prosperity of the ecosystem.

At that time, we mainly discussed the new form of traffic interaction from the mobile Internet to the intelligent Internet era. We proposed the concept of "Mocha" - just as conveniently as making coffee, seamlessly embed multi - modal service cards (Multi - modal Card) into the conversation flow. This led to a key difference:

One is the "all - powerful Agent" route: It advocates completing all operations in one - stop within the conversation interface to form a closed - loop.

One is the "look but don't act" route with a "sense of boundaries": The conversation interface should focus on efficiently presenting information and understanding intentions. When it comes to specific services (such as purchasing and booking), it should jump to the native application through "Mocha" to complete the operation. This not only ensures a smooth experience but also reserves a living space for professional service providers.

We firmly choose the latter. The reason is very practical: If all functions are "taken over" by large models, millions of APP developers will lose their value fulcrum. This does not conform to the principle of ecosystem symbiosis. The future intelligent ecosystem should be like a healthy business society, requiring clear value boundaries and a spirit of cooperation.

When the open - source storm of Deepseek hit at the beginning of the year, I made a judgment. I think