From GPT to Agent, how can technology and business "meet each other halfway"?

In today's technological wave, innovators and enterprises are facing severe challenges brought about by the intense collision between technological beliefs and business imperatives. They are deeply trapped in three dilemmas: the technological cliff, the engineering gap, and the business fog. So, what is the golden yardstick for identifying "high-value - engineerable - strongly closed-loop" scenarios? How can we achieve the "mutual pursuit" of AI and business based on platform capabilities such as LLMOps?

Recently, the live - streaming program "Geek's Appointment" on InfoQ, in collaboration with AICon, specially invited Xu Wenjian, Co - founder & CTO of Mars Radio Wave, Li Chenzhong, a senior technical expert at Alibaba, and Zhang Haoyang, the founder of AutoGame. As the AICon Global Artificial Intelligence Development and Application Conference 2025 in Beijing is approaching, they jointly explored the real path for the implementation of large - model products.

Some of the wonderful viewpoints are as follows:

• In the future, large models will become public resources like water and electricity. More focus should be placed on building private - domain models, products, and data flywheels.
• The training of general large models will absorb all available data. Therefore, only the information truly in a "data island" can potentially constitute a unique advantage.
• In the future, the product deliverables will no longer be the code itself, but a model - driven capability.
• True innovation will occur at the application layer: Based on the general foundation, combined with professional knowledge in various fields, a rich variety of vertical - scenario applications will be built.
• Traditional talents emphasize in - depth specialization in vertical fields and play the roles of experts or executors in large organizations. The core value of future talents lies in the breadth of vision. For example, developers understand the logic of design, products, and art simultaneously.

The following content is based on the live - streaming transcript and has been abridged by InfoQ.

When did we "start doing AI"?

Xu Wenjian: When GPT first came out, what were your thoughts? Was there a moment when you thought, "I have to do something right away"?

Li Chenzhong: In 2012, when I was working on an intelligent customer - service robot, I tried to improve the effect using traditional methods, but I always struggled to break through the bottleneck. Whether it was sentence - pattern matching or exhaustive expressions, the results were always rigid and mechanical. Even when I used language skills to imitate human responses, the content was often empty and ambiguous, just playing word games. Later, during the AI wave in the industry in 2015 and 2016, I observed many cases, but I always felt that the popular AI concepts at that time were still far from the effects I expected, until the emergence of GPT.

When I first encountered GPT, I found that large - language models were completely different from previous AI technologies. They made many things that were previously impossible possible. They were no longer limited to rigidly handling specific vertical - field tasks but showed certain reasoning abilities and could express themselves in fluent and natural language, which was amazing. Therefore, shortly after GPT was launched, I began to try to apply it to various scenarios, gradually expanding from content summarization to tasks such as reasoning and judgment. In my opinion, the emergence of GPT marked the opening of the door to a new world, like the arrival of a singularity moment, indicating the coming of a completely different era.

Zhang Haoyang: As an early user of GPT, I remember that I finally got an account on the fifth day after its launch. The excitement brought by GPT at that time kept me awake all night, and I explored its capabilities every day. At that time, OpenAI was not yet a giant, and there were rumors of a cooperation with Microsoft - for example, the news that Bing would integrate GPT also caused a sensation. Then, in less than a month, ChatGPT reached 100 million users at the fastest speed in history, announcing the arrival of the generative AI revolution.

After the release of GPT 3.5, its intelligence truly brought an epoch - making shock. Since then, because I focus on applications in the game field, I mainly pay attention to two directions: One is the application of NPCs. Large - language models already have the ability to provide emotional value and solve problems, and can exist as game NPCs. The other is the field of AI programming. Shortly after GPT was launched, Cursor became one of the first products to integrate large - model capabilities into the VS Code editor. Although Cursor is well - known today, it was a very novel tool in early 2023, achieving early AI - assisted programming, and later gradually developed the concept of Agent programming, until Vibe Coding in 2025.

Xu Wenjian: My experience is slightly different from that of the two teachers. There was no clear turning point of "diving into AI", but a process of quantitative change to qualitative change. GPT also shocked me when it first appeared. So, in early 2023, with a learning attitude, I joined hands with several doctoral friends from Beijing Normal University to develop an AI - personalized education product. Although the project ultimately failed due to my personal lack of ability, it planted a seed in my heart: AI can indeed change many things and may be a precious opportunity for our generation. After the project failed, I joined Baichuan Intelligence, hoping to understand the understanding and practice of technology by leading domestic AI enterprises. During my time at Baichuan, I learned technologies such as Agent. After leaving, I continuously invested in AI entrepreneurship. Eventually, this accumulation led to a qualitative change at a certain moment - I suddenly realized that I had been deeply involved in the AI field.

Xu Wenjian: What was the biggest cognitive change in the past two years? Are there any impressive "stories" or experiences you can share?

Li Chenzhong: I am a person with a technological idealist complex and usually like to watch science - fiction movies, such as the artificial - intelligence characters like Jarvis in "Iron Man". Therefore, I am amazed at the rapid development of AI. But if we talk about the specific moment when my cognition changed greatly, this process is actually not obvious.

The only significant change is that I found that many concepts that used to exist only in imagination are becoming more and more feasible. Starting from the earlier versions of ChatGPT, its performance in some scenarios was amazing. As the models were rapidly iterated and upgraded, their capabilities continued to increase, making the implementation path of future concepts clearer. The development of the Internet industry used to be changing rapidly, but the progress speed of AI in the past two or three years is on a completely different scale compared to it. Since the explosion of AI in the past two or three years, the development speed of the entire field has been so fast that it has even exceeded my imagination.

There is a sentence that can describe my mood at that time: I suddenly found that I had found a "free labor force" that could handle many things like a human being. Another thought at that time was: It has become possible to build a virtual world driven by AI. For me, the scenarios depicted in science - fiction movies and many of my personal ideas are no longer fantasies but have become feasible goals with clear implementation paths.

The emergence of new things often triggers two attitudes: One is to embrace them excitedly, see the opportunities, and actively try. The other is to watch cautiously. This means that when you try new things, you are often in the middle of these two attitudes. For example, when AI first emerged, I was eager to get involved, but many people held a conservative attitude, believing that the time was not right or the technology was not yet mature.

During this period, a more profound experience is the necessity of perseverance. When you put forward an innovative concept, it is rarely widely recognized. Instead, there is more waiting and even opposition. In this situation, it becomes difficult to push things forward - it is not easy to persuade others. Therefore, you often face the situation of persevering alone without support, and this process is bound to be full of challenges.

Zhang Haoyang: In 2023, I was still working at Tencent and only left to start a business at the end of the year. Initially, as an employee of a large enterprise, my perception was that it was necessary to self - develop large models as early as possible, rather than relying solely on external interfaces. This was almost the industry consensus in the first half of 2023, and all major enterprises were fully investing in large - model R & D. Tencent also launched the Hunyuan model.

The turning point occurred in the middle of 2023. After the LLaMA model was accidentally leaked and open - sourced, the Chinese large - model community became very active, and several startup companies later known as the "Six Little Dragons of AI" emerged. However, in the first half of 2024, there was a profound cognitive change: Many people abandoned the idea of self - developing large models. The reason is that, both at home and abroad, the capabilities of various underlying models have repeatedly proven that large - model training is not something that everyone can do, and the significance of training from scratch or fine - tuning is decreasing. Especially after the emergence of RAG technology, an excellent retrieval and sorting mechanism can often achieve ideal results, which has become the general consensus in 2024.

2024 was called the "Year of AI Applications", but the real explosion occurred in 2025. Taking AI programming tools as an example: After the release of Claude 3.7 at the end of February 2025, products represented by Cursor jumped from assisted programming to real AI programming. The improvement of model intelligence has completely changed the product and the capabilities of the upstream and downstream. This has instead given rise to a new perception: It has become necessary to self - develop small models in the future.

AI Agent entrepreneurship needs to build an "iron triangle" of "product - data - model". Only when the three are closely combined can a barrier be established. For example: If there are only products and models, entrepreneurs need to rely on shallow work such as prompt - engineering. However, with the launch of the plugin function by OpenAI, a large number of shell - wrapped applications were eliminated, proving that such models lack barriers. Especially when models like Claude 4 already have the level of senior engineers or even architects, the space for upper - layer applications is further compressed. Another type is the concept of "private - domain large models" that was popular last year, such as training exclusive models by integrating dirty data in the medical field. If there is no product entry directly facing users, such work is likely to sink without a trace or become a wedding dress for others - because a data closed - loop cannot be formed.

Taking Cursor as an example: It integrates models such as Claude, Gemini, and OpenAI. However, if its users do not enable the privacy mode, the generated code may be used to train a self - developed small programming model. Cursor uses a toolchain similar to the MCP mechanism to handle tasks such as file editing and accumulates user - behavior data at the same time. When it becomes the user's habitual entry point, its model may be better at programming than the underlying models. In the future, such products may appear in various fields, such as self - evolving games and social tools. Of course, we also need to be vigilant against the monopolization of giants.

The core value of self - developing private - domain small models lies in: After combining domain knowledge with the product to form a data flywheel, the product - data - model forms a strongly - bound ecological relationship, which is difficult for large enterprises to penetrate, thus establishing a barrier.

Another significant perception is that the cost of large models is decreasing exponentially. In March 2023, the Stanford town experiment cost thousands of dollars to run overnight. But only half a year later, while the model capabilities improved, the price dropped significantly. By May 2024, models such as DeepSeek had effects comparable to GPT - 4, but the cost was only 10% of it. This year, the cost is almost free - taking the AI NPC interaction in our game as an example, the monthly large - model token consumption of thousands of players is only dozens of dollars. In the future, large models will become public resources like water and electricity, and more focus should be placed on building private - domain models, products, and data flywheels.

Xu Wenjian: Cursor can do this because it has become an undisputed leader in the industry, with enough data to support the training of its small model and reduce costs. But for the vast majority of companies, before becoming the industry leader, their data - collection ability cannot compete with that of the giants training general large models at all.

Li Chenzhong: This also explains why at the beginning of the rise of large models, everyone generally focused on model fine - tuning. Because the iteration speed of large models is extremely fast, and their version upgrades are often over - covering. In essence, only top - notch teams or large enterprises with abundant resources can continuously invest in large - model training. If an ordinary team invests in training, its results are likely to be covered by the progress of general large models in the short term.

I think the only valuable or threshold - setting situation is: Your data is exclusive resources that cannot be obtained on the Internet, or highly personalized data in a specific vertical field that is extremely difficult for external companies to access. The training of general large models will absorb all available data. Therefore, only the information truly in a "data island" can potentially constitute a unique advantage. If the data does not have this characteristic, I think the rapid iteration ability of general large models will quickly cover the training results of a specific team.

Zhang Haoyang: Do the two teachers believe that a truly "general Agent" will appear in the future? Is a product direction like Manus worth investing in?

I personally don't believe that a general Agent will become a reality. Because private - domain data and models have unique value, and this value cannot be easily replicated by large enterprises with only massive data. It often requires in - depth industry knowledge (know - how). The core is that data and products must be strongly coupled to form a real barrier.

Taking our own game as an example: We designed a set of exclusive interfaces, and its rules only apply to this game environment. Players use natural - language instructions to drive AI to create new logic in the game. We then use the generated data to fine - tune an exclusive small model, making it more and more proficient in generating code for this product. Even if large enterprises obtain this kind of data, it is difficult for them to use it effectively because it is closely bound to a specific product. Therefore, building a private - domain model for a specific product still has value. Of course, large enterprises may achieve a breakthrough in a certain vertical field, but their model results can only serve a specific product in the end, not be omnipotent.

I think the product deliverables in the future will no longer be the code itself, but a model - driven capability, which can be understood as MAAS (Model as a Service). Products will be driven by large models to achieve self - iteration and evolution. This leads to the core reason for my doubt about the general Agent: Whether it is Manus or similar products, their "generality" is limited, and their performance in specific scenarios is often inferior to that of products from teams focusing on that field.

Looking at it from a broader perspective, I think it is impossible to invest a huge amount of manpower to fine - tune the user experience and data in all segmented fields to deliver satisfactory results. At least in the next 3 - 5 years, this is not feasible. Unless AI technology evolves to be able to self - train - for example, through a mechanism similar to the left - right mutual - combat of GAN, continuously optimizing in a specific scenario and finally approaching the level of the leading product in that field. Even so, I still have doubts about whether this "general Agent" can still be called "general".

Li Chenzhong: You are discussing more about the in - depth application scenarios in vertical fields - these scenarios are unlikely to be covered by a single product. And general large models provide basic capabilities, such as stronger intelligence, better instruction - following, faster response speed, and more complete reasoning chains.

When the base capabilities are applied to vertical fields, they need to be customized and polished for that field, which is similar to the division of labor among humans. The physiological basis of all people is essentially similar, but through in - depth industry experience, professional depth is formed. The systems built in specific vertical fields will continuously accumulate and optimize around that field, thus performing more deeply and professionally in that field. Although their underlying base models may be the same, the design of the application layer makes them focus on a specific direction.

The reason why I believe that a general Agent is feasible is not that a single Agent can handle tasks in all fields without adaptation, but that this kind of Agent has a basic - ability framework similar to that of humans - it has completed the "evolution from monkeys to humans", with planning ability, logical - reasoning ability based on background information, and tool - calling ability. Its performance in different fields will be different for two reasons: One is that the domain knowledge bases configured for it are different, and the knowledge retrieved through RAG will also be different. The other is that the domain tools configured are different. No single Agent will be equipped with all MCP tools. Even Manus will only configure corresponding tools according to specific field requirements. For example, an Agent in the game field will be configured with MCP tools combined with game products, while an Agent in the medical or other fields will be equipped with a tool set related to that field.

Therefore, the logic for the existence of a form like Manus is that its core base has general capabilities. When the corresponding input sources, dedicated tool chains, and domain knowledge bases are mounted for it in the target field, it can be transformed into an effective application system in that field. This is the meaning of the existence of a general Agent.

Xu Wenjian: A general Agent does not need to handle all tasks or be an