In today's technological wave, innovators and enterprises are facing severe challenges brought about by the fierce collision between technological beliefs and commercial laws, and are trapped in three major dilemmas: technological cliff, engineering gap, and commercial fog. So, what is the golden standard for identifying "high value-engineerable-strong closed loop" scenarios? How to achieve the "two-way rush" of AI and business based on platform capabilities such as LLMOps?
Recently, InfoQ's "Geek Appointment" X AICon live broadcast column specially invited Mars Radio Co-founder & CTO Xu Wenjian, Alibaba Senior Technical Expert Li Chenzhong, and AutoGame founder Zhang Haoyang to discuss the real path for the implementation of large-scale model products as the AICon Global Artificial Intelligence Development and Application Conference 2025 Beijing Station is about to be held.
Some of the highlights are as follows:
- In the future, big models will become public resources like water and electricity, and we should focus more on building private domain models, products and data flywheels.
- The training of general large models will absorb all available data, so only information that is truly in "data islands" can constitute a unique advantage.
- The product deliverable of the future will no longer be the code itself, but a model-driven capability.
- The real innovation will occur at the application layer: based on a general base and combining expertise in various fields, we will build rich vertical scenario applications.
- Traditional talents emphasize deep specialization in vertical fields and play the role of experts or executors in large organizations. The core value of future talents lies in the breadth of vision, such as developers understanding design, product and art logic at the same time.
The following content is based on the live broadcast shorthand and has been edited by InfoQ.
When did we "start doing AI"?
Xu Wenjian: When GPT first came out, what did you think? Was there a moment when you thought, “I have to do something right away”?
Li Chenzhong: When I was working on an intelligent customer service robot in 2012, I tried to use traditional methods to improve the results, but I could not break through the bottleneck. Whether it was sentence matching or exhaustive expressions, the results were stiff and mechanical. Even if I used language skills to try to imitate human answers, the content was often empty and specious, just playing word games. After that, when the industry experienced the AI wave in 2015 and 2016, I observed many cases, but I always felt that the hot AI concept at the time was still far from the effect I expected, until the emergence of GPT.
The first time I came into contact with GPT, I found that the large language model is completely different from previous AI technologies. It makes many things that were impossible in the past possible. It is no longer limited to rigidly handling specific vertical field tasks, but it shows a certain reasoning ability and can express in fluent and natural language, which is amazing. Therefore, shortly after the launch of GPT, 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 marks the opening of the door to a new world, just like the advent of the singularity moment, heralding the arrival 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 of its launch. The excitement brought by GPT kept me awake all night, and I explored its capabilities every day. At that time, OpenAI was not yet a giant, and rumors of cooperation with Microsoft, such as the news that Bing would connect to GPT, also caused a sensation. Less than a month later, ChatGPT broke through the 100 million user mark at the fastest speed in history, announcing the arrival of the generative AI revolution.
After the release of GPT 3.5, its intelligence level did bring epoch-making shock. Since then, because I focus on applications in the game field, I mainly focus on two major directions: one is NPC applications. The large language model has the ability to provide emotional value and solve problems, and can exist as a game NPC; the second is the field of AI programming. Shortly after the launch of GPT, 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 still a very novel tool in early 2023, realizing early AI-assisted programming, and then 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 is no clear turning point of "devoting myself to AI", but a process of quantitative change to qualitative change. GPT also shocked me when it first appeared, so in early 2023, I joined several doctoral friends from Beijing Normal University with a learning mentality to develop an AI personalized education product. Although the project ultimately failed due to insufficient personal 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 cutting-edge AI companies in China. During my time at Baichuan, I learned technologies such as Agent, and after leaving, I continued to invest in AI entrepreneurship. In the end, this accumulation triggered a qualitative change at some point-I suddenly realized that I had been deeply involved in the field of AI.
Xu Wenjian: What has been the biggest cognitive change in the past two years? Are there any memorable "stories" or experiences you can share?
Li Chenzhong: I am a person with a technological idealism, and I also like to watch science fiction movies, such as Jarvis in Iron Man and other artificial intelligence characters. Therefore, I am amazed by the rapid development of AI. But to say when there was a huge change in cognition, the process is actually not obvious.
The only significant change is that I find that many ideas that once existed only in imagination are becoming more and more feasible. I started with an earlier version of ChatGPT, and its effect in some scenarios was amazing. As the model is rapidly iterated and upgraded, its capabilities are constantly enhanced, making the realization path of the future more and more clear. In the past, the development of the Internet industry was changing with each passing day, but the speed of AI progress in the past two or three years is completely another order of magnitude. In the two or three years since the outbreak of AI, the development speed of the entire field has been so fast that it can even be said to be beyond 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 humans. Another idea I had at the time was: it has become possible to build a virtual world driven by AI. For me, the scenes depicted in science fiction films and many of my personal ideas are no longer fantasies, but have become feasible goals with clear paths to realization.
The emergence of new things often leads to two attitudes: one is to embrace them with excitement, see opportunities and actively try them; the other is to wait and see with caution. This means that when you try new things, you are often caught between these two attitudes. For example, when AI first emerged, I was eager to get involved, but many people were conservative, thinking that the time was not right or the technology was not yet mature.
During this period, I realized the necessity of persistence. When you put forward an innovative idea, it is rarely widely recognized. Instead, people are more likely to wait and see or even oppose it. In this case, it becomes difficult to push things forward - it is not easy to convince others. Therefore, you often face the situation of insisting on your own but lacking support, and this process is bound to be full of challenges.
Zhang Haoyang: I was still working at Tencent in 2023, and I left to start my own business at the end of the year. As a practitioner in a large company in the early days, my understanding was that it was necessary to develop large models as early as possible, rather than relying solely on external interfaces. This was almost an industry consensus in the first half of 2023. Major companies were fully committed to the development of large models, and Tencent also launched a mixed model.
The turning point came in mid-2023, when the LLaMA model was accidentally leaked and open-sourced. The Chinese large model community quickly became active, and startups that were later known as the "Six Little Dragons of AI" emerged. But by the first half of 2024, a profound cognitive shift was that many people gave up the idea of developing their own large models. The reason is that, both at home and abroad, the capabilities of various underlying models have repeatedly proved that large model training is not something that everyone can do, and the significance of training or fine-tuning from scratch is weakening. Especially after the emergence of RAG technology, excellent retrieval and sorting mechanisms can often achieve ideal results, which has become a general consensus in 2024.
2024 is called the "Year of AI Application", but the real explosion will occur in 2025. Take AI programming tools as an example: after the release of Claude 3.7 at the end of February 2025, products represented by Cursor have leapt from auxiliary programming to real AI programming. The improvement of model intelligence has completely changed the product and upstream and downstream capabilities. This has in turn given rise to a new understanding: self-developed small models will become necessary in the future.
AI Agent startups need to build the "product-data-model" iron triangle, and only by closely combining the three can barriers be established. For example: if there are only products and models, entrepreneurs need to rely on shallow work such as prompt word engineering. However, with the launch of plug-in functions by OpenAI, a large number of shell applications have been eliminated, proving that this type of model lacks barriers. Especially when models such as Claude 4 have reached the level of senior engineers or even architects, the upper-level application space is further compressed. Another type is the concept of "private domain big model" that was popular last year, such as integrating dirty data in the medical field to train exclusive models. If there is a lack of product entrances directly facing users, this type of work is easy to sink into the sea or become someone else's wedding dress-because it is impossible to form a closed data loop.
Take Cursor as an example: it connects to models such as Claude, Gemini, and OpenAI, but if its users do not turn on privacy mode, the generated code may be used to train self-developed programming models. Cursor handles tasks such as file editing through a tool chain similar to the MCP mechanism, while accumulating user behavior data. When it becomes the entry point for user habits, its model may be better at programming than the underlying model. Such products may appear in various fields in the future, such as self-evolving games and social tools. Of course, we must also be wary of the monopoly of giants. The core value of self-developed private domain small models is that after combining domain knowledge with products to form a data flywheel, product-data-model forms a strongly bound ecological relationship, which makes it difficult for even large companies to enter, thus establishing barriers.
Another notable realization is that the cost of large models is decreasing at an exponential rate. In March 2023, the Stanford Town experiment cost thousands of dollars to run a night. But only half a year later, the model capabilities improved and the price plummeted. By May 2024, models such as DeepSeek were 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, thousands of players spend only tens of dollars on large model tokens per month. In the future, large models will become public resources like water and electricity, and we should focus more on building private domain models, products and data flywheels.
Xu Wenjian: Cursor can do this because it has become the undisputed leader in the industry and has enough data to support its training of small models and reduce costs. But for most companies, before becoming industry leaders, their data collection capabilities cannot compete with giant companies that train general large models.
Li Chenzhong: This also explains why everyone generally focused on model fine-tuning when big models first emerged. Because big models are iterated very quickly, their version upgrades are often overlay-based. In essence, only top teams or large companies with abundant resources can continue to invest in training big models. If ordinary teams invest in training, their results are likely to be overwritten by the progress of general-purpose big models in the short term.
I think the only valuable or threshold situation is that your data is an exclusive resource that cannot be obtained on the Internet, or it is highly personalized in a specific vertical field and extremely difficult for external companies to access. The training of the general large model will absorb all available data, so only those information that are truly in "data silos" may constitute a unique advantage. If the data does not have this nature, I think the rapid iteration ability of the general large model will quickly cover the training results of a specific team.
Zhang Haoyang: Do you two believe that there will be a truly universal agent in the future? Is it worth investing in products like Manus?
I personally don’t believe that universal agents will become a reality. This is because private domain data and models have unique value, which cannot be easily replicated by large companies with just massive amounts of 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.
Take our own games as an example: we designed a set of proprietary interfaces whose rules apply only to the game environment. Players drive AI to create new logic in the game through natural language instructions. We then use the generated data to fine-tune the proprietary small model, making it increasingly good at generating code for this product. Even if large companies obtain this kind of data, it is difficult to use it effectively because it is tightly bound to specific products. Therefore, it is still valuable to build private domain models for specific products. Of course, large companies may achieve breakthroughs in a certain vertical field, but their model results can ultimately only serve specific products, not omnipotent.
I believe that the product deliverables of 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 why I question the general agent: whether Manus is a similar product, its "universality" is limited, and its performance in specific scenarios is often inferior to the products of teams focusing on this field.
Zooming out, I don’t think it’s possible to invest huge manpower to fine-tune the user experience and data of all segments to deliver satisfactory results. At least in the next 3 to 5 years, this is not feasible. Unless AI technology evolves to the point where it can train itself - for example, through a GAN-like left-right fighting mechanism, it continuously optimizes in specific scenarios and eventually approaches the level of the leading products in the field. Even so, I have doubts whether this "universal agent" can still be called "universal".
Li Chenzhong: You are discussing more about deep application scenarios in vertical fields - these scenarios are unlikely to be covered by a single product. The general large model provides basic capabilities, such as stronger intelligence, better command compliance, faster response speed, and more complete reasoning chain.
When the base capability is applied to a vertical field, it needs to be customized for that field, which is similar to the division of labor among humans. The physiological basis of all people is essentially similar, but professional depth is formed through deep cultivation of the industry. Systems built in a specific vertical field will continue to accumulate and optimize around that field, thus becoming more in-depth and professional in that field. Although the underlying base model may be the same, the design of the application layer makes it focus on a specific direction.
The reason why I believe that universal agents are feasible is not that a single agent can handle all domain tasks without adaptation, but that such agents have a basic capability framework similar to that of humans - they have completed the evolution from "monkey to human" and have planning capabilities, logical reasoning capabilities based on background information, and tool calling capabilities. Its application in different fields will have different performances. The reasons are: first, the domain knowledge base configured for it is different, and the knowledge retrieved through RAG will also be different; second, the configured domain tools are different. No agent will be equipped with all MCP tools. Even Manus will only configure the corresponding tools according to specific domain needs. For example, Agents in the game field will be equipped with MCP tools combined with game products, while Agents in medical or other fields will be equipped with tool sets related to the field.
Therefore, the logic behind the establishment of Manus is that its core base has universal capabilities. When the corresponding input source, dedicated tool chain and domain knowledge base are mounted on it in the target field, it can be transformed into an effective application system in the field. This is the meaning of the existence of universal agents.
Xu Wenjian: Universal Agent does not need to take on all tasks, nor does it need to be an "all-round" product. It can serve as a core backbone, focusing on handling general capabilities, while leaving the professional knowledge in vertical fields to various vertical Agents mounted on it. The core value of Universal Agent is to act as an entry or integration hub to connect different capability modules.
Zhang Haoyang: OpenAI proposed the Function Calling mechanism a long time ago, and the open source community has also been exploring similar functions. However, this type of product did not really explode until this year. I think the fundamental reason lies in the leap in the capabilities of the underlying model. Function Calling is essentially not much different from the tool calling protocol, and the core of both is to execute external instructions. Some people have previously tried to implement similar functions by calling command lines or tool chains through Function Calling, but the effect was limited. The popularity of Manus coincided with the release of Claude 3.7, which confirms that the key driver is that the model base capability reaches a critical point. In addition, if the giants come out to formulate unified tool standards and open up their powerful base models, users can directly develop segmented applications based on the underlying capabilities. At this time, the value of encapsulation layers such as Manus is questionable.
Xu Wenjian: Mr. Haoyang, the core difference in our cognition may lie in the judgment of Agent barriers. In the "product-data-model" triangle you proposed, I particularly think that Agent itself has a unique barrier - its value is rooted in the professional knowledge of the vertical field. The advantage of Agent is that it can be continuously iterated: relying on the upgrade of the general large model, it is always "one step ahead" of the general model in a specific field. This advantage comes from the deep understanding and optimization of the vertical field. Therefore, no matter how the general model evolves, Agents with domain expertise can always maintain local advantages.
Li Chenzhong: In essence, there is not much difference between Function Calling and MCP. The value of MCP lies in promoting industry standardization and making tool calls more centralized and standardized. However, the core difficulty in building an excellent Agent lies in its planning ability, and the key to Manus's future lies in this: how to effectively decompose the task steps around the goal and execute them in place? Although this kind of task planning will be different depending on the specific field, it will also have a basic level that is unrelated to the vertical domain. Similar to putting aside professional attributes, everyone has the ability to plan daily affairs such as shopping for groceries and traveling. After adding professional attributes, such as programmers, they have the ability to design system architecture, write code, and use various exclusive tools. Then, around the goal of landing a product, they will carry out a series of task planning and decomposition, and use the above capabilities. Therefore, Agents will differentiate in vertical fields, but their base will have basic planning capabilities that are unrelated to vertical domains.
My view on the future product form may be more radical: perhaps the final form will evolve into "Data-to-Agent". All interactive interfaces can be dynamically generated based on the current scenario, with the cost approaching zero - just like Jarvis in "Iron Man", who can generate interfaces in real time according to needs, mobilize tools, and provide dynamic responses in combination with knowledge bases.
Regarding the Agent ecosystem, I think there will only be a limited number of general-purpose Agent frameworks in the future, with core planning/thinking/perception/tool calling capabilities. Based on these general-purpose bases, combined with professional knowledge and scenarios in various fields, a series of rich vertical applications can be built. There is no need for a large number of general-purpose Agent frameworks, but there will be many upper-level applications.
Audience: What do you think of Manus-style entrepreneurial opportunities?
Li Chenzhong: It cannot be said that there are no opportunities for starting a business in a vertical field based on general agents. But it depends on whether the development speed, innovation, and effect can surpass existing players - just like the emergence of DeepSeek. If you do have some unique methods, such as achieving significant innovation in key points such as planning capabilities, tool calls, and intent recognition, I think it is worth trying.
Zhang Haoyang: I don’t think the chances are great. I have had contact with the founding team of Manus and witnessed its journey from obscurity to overnight fame. From a venture capital perspective, Manus’ success is more like a “self-media carnival”: influential domestic self-media have vigorously promoted its popularity, and its strategy of initially targeting overseas users rather than the domestic market is also worth pondering. After it became popular, it quickly received investment from Tencent, and its valuation climbed to billions of dollars. But I think this is more of a “show” for the capital market, and the actual product value has not yet reached the ideal state of a general agent: early trial users’ feedback was lower than expected, the current popularity after opening has also subsided, and the user experience is not perfect. In many segmented scenarios, combining specific workflows with large model capabilities can solve problems more effectively.
Therefore, there is little chance of starting a business like Manus. This is not to question the team's ability, but this matter has a "first cause effect": the first to do it is genius, the second is futile. It is difficult for capital to buy in this narrative again, especially when Manus's brand effect has been formed and its technical barriers are often questioned. Subsequent imitators will face the question of "Manus has received huge funds, how can you surpass it?"
In addition, I always have doubts about the feasibility of "universal agent products" - I don't believe that one team can solve all application scenarios. Unless you are positioned to provide the Agent "base" itself. But at the current stage, this still requires a professional team to encapsulate the technology. In the future, large models may evolve into underlying infrastructure: users only need to make demands on the Agent around them, and it will be able to independently write code, build frameworks, and return results. By then, teams focusing on upper-level frameworks will no longer be necessary, and core capabilities will be attributed to the infrastructure layer (Infra).
What is the most difficult part in developing AI products from 0 to 1?
Xu Wenjian: We have all worked on AI projects from scratch. In your experience, what is the most difficult part? Model capability? User closed loop? Or something else?
Li Chenzhong: In practice, model capabilities have indeed been a key bottleneck, but this is relative. After the scene is selected, the model effect is often unstable or lower than expected, which is often hindered in the early stage. This needs to be gradually improved as the model is upgraded. However, during the project execution, I found a workaround: when the model reaches a basic level, the model itself does not necessarily become an absolute bottleneck.
Observing user usage habits, a common misunderstanding is to directly input long prompt words (hundreds or even thousands of words) containing multiple complex tasks into the model, resulting in unstable output. At this time, it is easy to blame the model for insufficient capabilities. The improvement method is to break down complex tasks into multiple simple subtasks and let the current model execute them step by step. Practice has shown that the model responds more stably to simple tasks. As the model is upgraded, the overall effect will be further improved on the original basis. This is similar to management: when faced with complex tasks, replacing smarter people (model upgrade) is one of the solutions, but breaking down tasks and letting ordinary people (current model) execute is also feasible.
Therefore, as long as the model reaches the basic capability line, the bottlenecks in most scenarios can be solved. It is necessary to distinguish between model capability and commercial feasibility: the model effect is the basic support, but the success or failure of the project depends more on whether the business model itself is established. On an invalid model, no matter how good the model effect is, it is futile.
Zhang Haoyang: Mr. Li pointed out the key point, which is the fit between product and technology (TPF). I have personally experienced the application of large models for four years, and my biggest feeling is that the rapid improvement of model capabilities is gradually eliminating a lot of dirty work. The Manus team also shared that after the model base capabilities are enhanced, many prompt word projects are no longer necessary, and its ability to follow instructions, reflect and plan has been significantly improved.
The core of TPF is that when the model capability is lower than the user's expected delivery level, the gap needs to be filled manually. Today, the model capability is infinitely close to or even exceeds user needs, and the challenge shifts to higher-level control and architecture design. For example, AI programming tools perform well at the code execution level, but still have shortcomings in system architecture design.
Another thorny issue is the memory mechanism. Even if the context window is expanded to millions of tokens, the model still has problems such as uneven attention distribution and forgetting of mid-segment information. The current optimal solution is still RAG and its variants. In the future, if the model can make breakthroughs in retrieval ranking (Rerank) and memory capabilities, it will greatly simplify the underlying optimization work. Despite the challenges, I am generally cautiously optimistic. Even if the model is regarded as a "super genius without memory", its value is significant enough.
Li Chenzhong: I have been thinking about a question recently: Why do models sometimes produce unexpected results after their capabilities have been significantly improved? I found a key point - the model may not be not smart enough, but because of its extensive knowledge, it can conceive of multiple solutions when faced with a problem, while we humans are limited by our own knowledge and often only anticipate one or a few solutions. When the model outputs ideas that we have not foreseen, it is easy to feel biased.
The root of this difference may lie in the lack of constraints. There are invisible constraints when humans make decisions: individual knowledge boundaries and environmental information (visual, auditory, social cues, etc.) jointly shape the expected path. However, the model only relies on input text prompts and lacks these potential situational constraints, which may lead to the output of seemingly "strange" but logically self-consistent solutions. Therefore, sometimes it is necessary to provide additional explicit constraint descriptions or requirements.
Even if the model is more advanced in the future, if this information asymmetry cannot be resolved, there will still be two tendencies: either overfitting to a specific scenario and losing flexibility, or using technical means to simulate the implicit constraint information received by humans. This is an issue worthy of attention in current model interaction.
Zhang Haoyang: This highlights the importance of multimodal capabilities.
Li Chenzhong: Yes. For the model to accurately match the user's contextual expectations, it must obtain input dimensions similar to those of humans. Otherwise, just like two people guessing each other's intentions, it is inevitable that there will be deviations. It is even more difficult for the model to fully match unspoken expectations. This will be the direction to be overcome in the future.
Xu Wenjian: In essence, we need to extend AI’s perception to the real world—establishing closer connections in vision, hearing, touch, and other dimensions. Only in this way can the model understand and respond to implicit conditions in complex situations.
Are talent standards being rewritten?
Audience: Has the type of talent most needed to develop LLM products changed? What kind of people will you look for first when building a team?
Zhang Haoyang: As an entrepreneur, I personally experience and practice the transformation of organizational forms. I strongly agree with the concept of "super individuals" - when I worked in a large company, I tried to play this role and deeply used AI to assist my work. With the improvement of Agent capabilities, I now work like a team of digital employees. After using our self-developed Agent tools and Cursor, the average effective code output of engineers in the team jumped from about 1,500 lines per week to 30,000 lines, achieving a 20-fold increase in production capacity.
I think this shift will become a universal phenomenon. Accordingly, the talent structure will inevitably change: individuals need to transform from tactical executors to strategic architects. The core capability of the future is management - but the management object will change from people to AI agents. The key is to learn to dispatch various capability resources to solve problems: when you can clearly express your needs and use the Agent's increasingly enhanced ability to follow instructions to break down complex tasks into executable subtasks, you can drive the Agent team to collaborate efficiently.
What Agents lack the most at present is global architecture capabilities. They are good at solving problems in specific aspects, but it is difficult for them to coordinate the overall solution. Therefore, what is needed in the future is compound talents - they are not required to be proficient in everything, but they must have a cross-domain vision and the ability to ask precise questions. Take me as an example: as a game engineer, although I am mainly good at front-end development, I understand server-side knowledge. By guiding AI through professional questions and using my own inspection results and cross-end integration capabilities, I can use Agent to complete the full-link work.
Traditional talents emphasize deep expertise in vertical fields and play the role of experts or executors in large organizations. The core value of future talents lies in the breadth of vision: for example, developers understand design, product and art logic at the same time. Such people with cross-domain cognition and agent collaboration capabilities will become true super individuals.
Li Chenzhong: Looking back at the early days of the Internet, excellent programmers were usually able to complete projects independently and had full-stack capabilities. As the industry scaled up, division of labor (front-end, back-end, data, algorithms, etc.) began to emerge with the goal of improving efficiency. However, division of labor should not be equated with profession. I am particularly opposed to treating division of labor (such as "I am the back-end", "I am the front-end") as a profession, because this may cause individuals to lose the ability to independently complete closed-loop tasks.
The development of AI is accelerating its return to this essence. In the traditional model, over-emphasizing the division of labor and specialization may weaken the ability of individuals to solve problems. Although AI cannot completely replace humans in handling all complex matters, it has greatly lowered the threshold for execution. The core of future talents will be to become "AI engineers" - this means that you need to have basic understanding and operational capabilities throughout the entire process of landing a goal or product. It will be difficult to adapt to the future if you stick to a single division of labor thinking.
As AI capabilities improve, the necessity of specific execution positions (such as pure test engineers) will decrease. The key is: someone is needed to be able to plan products globally - from business models, product definition to implementation paths - and effectively direct AI tools or "digital employees" to collaborate and execute. Talents who lack this global planning and integration ability and are only good at a certain technical link (such as back-end development or algorithms) will face challenges in their competitiveness.
Therefore, in talent selection, I value the underlying potential more: strong internal drive, desire to explore, the courage and tenacity to solve problems. Technical skills can be cultivated, but the ability to independently find topics, delve deeply into research, and integrate resources to achieve goals in the face of unknown areas is the core. As long as you have these basic qualities and a normal level of intelligence, it is only a matter of time before you master the required upper-level skills.
Audience: What are the future career prospects of artificial intelligence? Is the threshold very high?
Li Chenzhong: The first question is: Do you really love this field? If you love it, then you don’t need to worry too much about the future - you will definitely be able to surpass most people in this field and naturally do very well. But if you don’t really love it, you will have to compete with most people.
Zhang Haoyang: According to a recent survey in the United States, the major with the highest unemployment rate is computer science, which is in stark contrast to the grand occasion of graduates of this major "walking sideways" more than a decade ago. In a sense, the development of large models is a manifestation of "programmers killing programmers": when the intelligence of human control of machines evolves to a new stage, a large number of low-end practitioners may face the risk of elimination. In particular, AI has been able to replace the capabilities of low-end positions in many industries.
There is a cruel reality about the prospects of students majoring in artificial intelligence: only the top 5% or even higher geniuses in the field will have significant competitiveness. Because in the future you need to be smarter than AI to become the core force driving the evolution of AI. When AI capabilities surpass years or even decades of professional accumulation of humans, if you cannot transform into an agent manager or have interdisciplinary thinking, you are doomed to be eliminated. AI will not replace people, but people who can use AI will replace those who cannot use AI. The advantage of studying artificial intelligence is that systematic theories can deepen your understanding of the essence of technology. The key is - either become a top minority in the field, or empower AI capabilities in other fields. This deep understanding will enable you to go further than others.
Li Chenzhong: I divide the coding community into two categories: programmers (creators with innovative ability, exploratory spirit, and the ability to produce frameworks) and coders (mechanical workers who perform coding tasks). In the AI era, top programmers still have broad prospects; but coder-type roles will be quickly replaced by AI. Because although AI is creative, humans can only maintain their value by surpassing AI in more groundbreaking creative fields.
Xu Wenjian: Don’t be limited by technical details, but expand the boundaries of your vision, actively embrace new trends and new capabilities, and actively adapt to the changes of the era driven by AI.
This article comes from the WeChat public account "InfoQ" (ID: infoqchina) , author: AICon, and is authorized to be published by 36氪.
