Next, I will go through the AI product manager trilogy – cognitive, business, and technology; There are about 6 articles, and I will share my experience on the transformation of AI product managers from three perspectives: AI cognition, AI technology, and AI application. I also hope to help all the production and transportation students who may be confused, anxious, or hope to embrace AI in this AI era.
Foreword: The closer the AI, the greater the anxiety; The faster the AI, the bigger the gap.
AI large models are developing too fast, and the wheels of artificial intelligence are rolling forward, making a deafening roar;
Some students will say, “No, right? AI is still far from us. ”；
Is it possible that you have been thrown too far away by the fast-moving “car” of AI, so you can’t hear the roar at all?
Today, I will not talk about the algorithms, training methods, architectures, etc. of large models; In this article, I hope that through my article, everyone can face up to AI and face up to this era.
However, AI is developing too fast, so what I am sharing now is to find out the certainty of AI belonging to this era in all the uncertainty, no matter when you read this article, you can understand it in time and “get on the bus”.
1. AI is developing at an unimaginable speed.
Let’s start with a brief review. We can find that all those well-known and popular products in human history have been used by users around the world for more than 100 million years.
From the original phone used for 75 years, to mobile phone use for 16 years.
With the development of the Internet, this speed is constantly accelerating. For example, the overseas version of TikTok reached the milestone of more than 100 million global users in just nine months.
In today’s AI era, ChatGPT only took two months, and DeepSeek only took seven days.
After 10 years of interaction design, why did I transfer to product manager?
After the real job transfer, I found that many jobs were still beyond my imagination. The work of a product manager is indeed more complicated. Theoretically, the work of a product manager includes all aspects of the product, from market research, user research, data analysis…
View details >
The whole world is accelerating, time is compressed, speed is accelerating. Therefore, in the age of AI, speed has become the most critical element, which greatly reduces the time it takes us to complete tasks.
ChatGPT has grown from 100 million weekly active users in 2023 to 400 million in 2025, and has now even exceeded 500 million, achieving fivefold growth in less than a year, and maintaining continuous rapid growth and increased market penetration. This shows that AI technology has had a direct impact on the global business world
In addition to its rapid development, another notable feature of AI is the universality of its application.
I don’t know if you remember, in the past, when we talked about the metaverse, digital twins, etc., it was actually more just a concept;
However, today, unlike in the past, new technologies have concepts first and then landed, and people’s cognition of AI now does not revolve around the concept of “large model”, but is reflected in various applications used in daily use, such as ChatGPT, Wenxin Yiyan, Doubao and a large number of large language models at home and abroad.
Therefore, in the era of AI, AI is not only a rapidly developing concept, but also a term that is deeply rooted in the hearts of the people and integrated into life.
With the release and open source of reasoning models led by Deepseek, the intelligence level of AI large models has reached an unprecedented level.
DeepSeek-R1 achieved an accuracy rate of 79.8% in the math competition (AIME 2024), surpassing 96.3% of human participants1331; 90.2% accuracy in MATH-500 testing; Codeforces Programming Competition Score 2029, has surpassed the level of professional programmers.
From a modal perspective, since the release of ChatGPT, we have only seen a simple conversational model.
Today, AI has evolved to help solve problems in many forms. In addition to models of various sizes and IQs in the cloud, models for writing code specifically (e.g., Cursor, Trae, codebuddy) have been developed;
Understand images, videos; models that generate images and videos (such as MidJourney, i.e., dreams);
In addition, there are models that can simulate human assistants, automatically operate mobile phone and PC interfaces, etc. (e.g. AutoGlm, Manus)
Especially in the past two years, 22-24 years, AI has experienced an explosive pace of change, which is unprecedented. AI technology allows us to achieve more progress in less time. Moreover, this progress has not stopped, and the iteration rate of AI is still very fast, so it can be expected that AI intelligence capabilities will continue to improve.
2. AI is rapidly integrating into our lives and work.
At the beginning, for example, OpenAI’s model was expensive due to technical barriers and high training costs, so the call cost of the GPT-3.5 horizontal model was:
In June 2023, the cost of calling an AI model that reaches the level of GPT-4 is 36USD/per million tokens;
However, with the optimization of model architecture, the reduction of hardware costs, and the rise of open source models like deepseek;
In January 25, the performance of the DeepSeek-R1 model was benchmarked against OpenAI’s o1, but the call cost was only 1.8%-3.7% of the latter (input token 1 yuan/million token, output token 16 yuan/million token)
Even with the growing and larger open source markets such as Wenxin large model and Tongyi Qianwen, more and more free large models have appeared.
So we will find that the token cost of AI large models has dropped by more than 280 times in 3 years!
What are the implications of such an outcome?
This will make it easier for AI to integrate into our lives and business worlds.
Add some side arguments,
In 2023, there will be a total of 190 purchases on AI, with a total amount of about 590 million yuan.
In 2024, the number of purchases will increase to 1,519, an increase of almost 8 times, and the total amount will reach 6.4 billion yuan, an increase of more than 10 times.
In the future, I believe there will be greater growth.
Well, at this time, some friends may want to ask, since AI is developing so rapidly and more down-to-earth, why doesn’t my current company feel too much AI atmosphere?
This starts with the three stages of AI development and enterprises:
1、The “basic model exploration period” led by Internet giants
At the beginning of 2023, at that time there were few models on the market to choose from. At that time, AI mainly provided basic models for commercial use, with high prices and high technical requirements, so it was inevitable that the first batch of attempts would be our so-called Internet giants.
2、The “industry model explosion period” in which medium-sized enterprises participate
From June to December 2023, with the emergence of more large models and the development of domestic large models (such as Alibaba’s Qianwen and Tencent’s Hunyuan), the portrait of AI companies has obviously spread from Internet giants to more medium-sized Internet companies or some companies with strong informatization capabilities. At that time, the value brought by the model had already begun.
3、The “inclusive landing period” of business-oriented companies
From the beginning of 2024 to April (DeepSeek released the open-source model R1), model capabilities are spreading to a wider group, including those with weak business-oriented, technical or informatization capabilities. AI agents (agents) and low-code platforms (such as Baidu App Builder) have emerged, and even non-technical companies can quickly build customized applications (such as intelligent customer service, report generation, etc.).
So, now you can think: Is your current business in the third stage? This is also the right time to apply for the transformation now, in fact, it is not too late.
3. AI is evolving towards AGI
What is AI and what is AGI?
From the literal English translation, it is only changed from Artificial Intelligence to Artificial General Intelligence; But is it really as simple as universal?
Next, I will list a few timelines so that everyone can better understand and comprehend.
- In 1948, Turing had already tried to let machines play chess against humans.
- In 1997, IBM launched the “Deep Blue”, a machine capable of playing chess against humans. It defeated the world champion of chess, Kasparov, for the first time.
- In 2016, it was generally believed that since the machine could quickly solve chess, it would also be able to solve Go next, but it was known that in the 16 years, that is, in the development process from “Deep Blue” to “Alpha Go”, people really solved the Go problem, and there was a gap of 20 years.
Why is it all chess, but it took AI 20 years to go from solving chess to Go?
We can observe chess, the complexity of chess itself is relatively low, and its grid is only 8 by 8.
At that time, machine learning was to defeat humans by giving machines existing chess rules and human experience, and then using violent calculations to complete more steps than the human brain’s thinking limit and imagine more feasibility.
The machine canViolent exhaustion combined with pruning algorithm(e.g., α-β pruning) can significantly reduce the amount of computation. For example, IBM’s “Deep Blue” can evaluate 200 million chess games per second through a dedicated chip, combined with a rule library and endgame database of human experts.
But how does AI solve the problem of Go?
Is it okay to still use violent algorithms like chess? Yes, but the problem becomes extremely complicated.
The chessboard of Go is 19×19, each possible move (search width) is as high as 250 moves, the search depth is about 150 steps, and the total search space reaches 10^170, far exceeding the total number of particles in chess and even the universe (about 10^80), which is a very huge amount of calculation for a machine.
So the complexity of Go forces the AI to turnmachine learningandSelf-chess。
In March 2016, AlphaGo played against Go world champion Lee Sedol for the first time;
AlphaGo, which no longer calculates directly through simple algorithms as it used to be; It is equivalent to using neural networks to train the endgame of 30 million chess games in human history, not relying on the rules set by humans, but by learning from these samples to refine the rules, and refining human experience into the model itself. Thus, in this way AlphaGo was born and it managed to overcome humanity.
In 2017, DeepMind released AlphaGo Zero again, which beat its predecessor, AlphaGo.
Zero no longer relies on any human knowledge, only learning and accumulating experience by playing chess by itself. In three days, it played nearly millions of games, and its power has exceeded that of AlphaGo, the accumulation of thousands of years of Go, and it took only three days to achieve this achievement.
This shows the amazing capabilities of machine learning,
Machine learning basically goes through three stages: from basic learning→ self-playing→ and self-learning.
So let’s answer the above question, AI and AGI, the difference is one General, what is the difference?
Personally, I think: AGI is like a “digital version of the all-powerful human brain”– It not only completes specific tasks, but also acts like a humanFlexibility to learn any new skillmoreoverThink independently, draw inferences from one example, and solve problems you have never seen before。 We are still exploring the road of AGI.
I think we can refer to the definition of AGI released by OpenAI.
It divides AGI into five levels, from one to five, representing different stages on the path to AGI.
First, let’s understand what these five levels mean?
- Level 1：The core capabilities of chatbots are natural language understanding and generation, but they are limited to information interaction and lack deep reasoning and action capabilities。
- Level 2：ReasonersAI can solve human problems, begin to understand the complex logic of the human world and gain reasoning skills.
- Level 3：AgentsAI began to learn and be able to use tools to solve some problems in the physical world that were originally humans.
- Level 4：InnovatorsAI has been able to create and learn independently, and has the ability to think for self-improvement.
- Level 5：OrganizationsAI has become comparable to or even surpasses humans, and can become a member of an organization like a human, exploring the world and completing tasks alone.
At present, the word Agent is often heard on the Internet and at work, does that mean that today’s artificial intelligence has reached the L3 level?
In fact, I think the current artificial intelligence technology is in the transition stage from Level 1 to Level 2;
Although the current agent can divide labor and collaborate, in fact, you will find that most agents are still in the chat stage.
There are even more problems such as model hallucinations, information redundancy and coordination failures.
The real L3 requires the intelligent body to perform cross-application tasks autonomously, which is still a long way for artificial intelligence to go in terms of environmental perception and real-time interaction.
4. AI has had a disruptive impact on the product paradigm.
Did artificial intelligence only appear in these two years?
Of course not, since the mobile Internet era in 2014, everyone has actually subtly begun to come into contact with AI, such as the search engine, shopping recommendation ranking, music and video recommendations, and search itself have begun to use AI technology.
However, at that time, AI was often called vertical artificial intelligence, that is, it was not universal, and it was backed by a large number of algorithmic tools (such as clustering algorithms, collaborative filtering, etc.).
What does today’s big model look like?
Today’s AI generally uses deep neural networks, reinforcement learning, and multimodal fusion; Now AI is a general artificial intelligence, and with the reduction of model training costs, in this context, the threshold and cost of AI use have been significantly reduced. The scene does not need to be specialized, the threshold for use is simplified, and AI has almost become a basic production factor and is no longer a niche high-end technology.
I am now watching AI news from all walks of life every day, including the automotive industry, medical industry, terminal industry, education industry, consumer manufacturing industry, game industry, cultural tourism industry, etc., there is a steady stream of AI applications and inventions every day.
If AI becomes our basic factor of production, it will have a profound impact on the whole world.
From the agricultural era, the industrial age to the information age, every change is triggered by changes in basic production factors. And now, in the era of AI, if the basic elements change, it will mean that our upper business world will be completely reconstructed, which is undoubtedly a huge change.
everybodyWhen AI becomes popular and people believe that AI has the ability to reshape business, do you think AI will definitely disrupt our product paradigm?
What is a product paradigm? The product paradigm is essentially a human-computer interaction paradigm, that is, how to talk to the machine, give instructions, and how the machine feeds.
The key to solving this problem lies in the cost of machine understanding, that is, the cost of learning, which are two core factors.
The transformation of human-computer interaction in history
- [The earliest computer age]: The bandwidth of human and machine communication is limited, and the communication methods are extremely limited, limited to dozens of statements and instructions.
- 【PC Era】: Humans began to interact with the mouse. This period saw a significant advancement in interfaces from the command line to graphical user interfaces. It was found that the cost of learning was greatly reduced, and the interaction bandwidth was also improved.
- 【Mobile Era】: Mobile Internet further reduces learning costs and increases interaction bandwidth, and mobile devices also add more modalities, such as videography, photography, geolocation information, etc.
- [AI Era]: AI’s expressive ability has become extremely rich, and any imaginable and expressive content can be input. At the same time, the interaction process is extremely concise and only requires dialogue. It lowers the barrier to learning and enhances expressiveness, making it easy for anyone to achieve the desired results, achieving the first fusion of expressiveness and ease of use.
As interactive capabilities and expressions evolve, our products may be given new forms and meanings, enabling reinvention。
Like what:
- In the past, in the workplace, we had to write meeting minutes, PPT, and Excel; Now AI can automatically generate meeting minutes through meeting content, provide corresponding data and content, and generate report documents in 1 minute.
- In the past, when we were studying, we needed to take different classes and do question banks; Now AI can capture students’ micro-expressions and handwriting pressure when solving problems through cameras, judge the state of “pseudo-understanding” and strengthen weak links in a targeted manner.
- In the past, when playing games, fixed NPCs and fixed plots; Now that NPCs have a memory chain that continues to evolve, the player’s choices a few years ago will still affect the current story branch. AI renders narrative lines in real-time that align with collective consciousness tendencies.
Etcetera. We are only in the infancy;
In the past, the skills of product managers often limited our creativity. We may have great ideas, but they are difficult to achieve because we do not have skills such as programming or painting. However, the emergence of large models has eliminated this ability limitation and provided us with a way to bring our ideas to life.
5. AI, it is not a tool, it is the “person” you need in your mind.
The initial form of the agent, which assists in collecting and processing data only at the information level. So in fact, everyone can integrate this ability into their own business.
As AI capabilities continue to expand, cross-application automation is already doing an excellent job. OpenAI has released its Agent suite, significantly lowering the threshold for using agents; The increasingly rich MCP protocols have made it extremely standardized for AI to connect to external services.
Therefore, we can no longer just look at AI as a tool. I think it will eventually translate into productivity that will deliver quantifiable business outcomes.
So what should we do? Many friends don’t know how to take the first step;
first: It is crucial for the understanding and selection of models, and only by understanding the principles of large models and choosing a smart enough model will the learning speed be fast. A reliable person and an unreliable person go to the company to apply for employment, and the reliable person may be competent in two weeks, while the unreliable person may not be able to do well even if he has worked in the company for two years. Therefore, I think in the first step, we need to know enough about the big model, its future and present life, and its way of life.
second: Push back your business to make solid data preparations. The core logic of AI models (especially machine learning and deep learning) is to extract patterns and use them for prediction or decision-making through learning from data. Only by transforming business logic into high-quality, standardized, and dynamically updated data assets can AI truly understand business needs and solve practical problems.
third: Learn about prompt word engineering. Prompts are “translators” that connect user intent with AI capabilities; This requires us to accurately disassemble the semantic structure of user needs; It can even promote the implementation of low-code/no-code AI functions; A good AI product manager must be a bridge between AI and business needs.
On the second point; Let me add another explanation:
In the past, we only thought of structured data as a data asset because machines could only process structured data. However, machines are now able to process data that goes far beyond structured data.
The advantages of unstructured data over structured data are reflected in two aspects.
First, unstructured data is an order of magnitude far greater than structured data, which is the norm. As a result, a small portion of your data assets can now be leveraged in their entirety.
In some cases, structured data may have been tainted by human subjective judgment. Because they are annotated according to specific tasks, resulting in missing contextual information, many details that have been lost or artificially subjectiveized. Therefore, in the age of AI, AI can make extensive use of unstructured data.
Finally, product managers need to revisit the essential question: Are we building tools or creating “digital lifeforms”?
Traditional product design follows the progressive path of “function→ experience→ emotion”;
The demand of the AI era presents a reverse structure of “cognitive resonance→ emotional dependence→ and functional realization”. Users need to be understood first, and then to be satisfied; Therefore, AI product managers must understand and empathize with user needs, and I believe that only in this way will the AI agents designed by everyone be recognized by the business and users in the end.
Notice
Next, I will bring you the AI product manager transformation trilogy – cognition (2) “Deconstructing the Essence of AI” as soon as possible;
In this article, I will focus on sharing with you my in-depth understanding of AI technology principles, although AI product managers do not need to be experts in algorithm and model training, but they need to master the framework and application boundaries of technical principles in order to effectively communicate needs, design products and promote implementation.
