Survival Guide in the AI ​​Era: Why Your Experience Is More Valuable Than Knowledge?

01 AI Wall and Development Direction

The "scaling law" has always been the underlying rhythm of the rapid development of large models: as long as computing power and high-quality data are continuously invested, model performance can steadily climb on a logarithmic scale. Companies such as OpenAI, Anthropic, and Google still publicly state that they do not see the "ceiling" of this curve. However, although the curve has not yet reached the top, the slope is getting blunter - although we have not hit the physical limit, we have clearly hit the "economic-engineering wall" .

The first is computing power and energy consumption.  To train models with tens of billions or trillions of parameters, an increase of 1%–2% in quality requires multiple GPU clusters, massive amounts of electricity, and expensive cooling. Simply put: spending ten times the money only results in a few percentage points of improvement . The input-output curve begins to bend.

The second is data.  High-quality human corpora have almost been consumed, and what remains is "long-tail content" with high noise and high bias. Although synthetic or crawler corpora can "extend life", they are difficult to replace real and diverse human experiences, and may even amplify illusions.

The third is reasoning and memory.  Large models are good at statistical matching, but deep causal reasoning and cross-context decision-making still rely on "chain prompts" plug-ins; the longer the context, the higher the calculation and cost, and it is still difficult to achieve persistent and personalized memory.

To make big models go further, we cannot just rely on piling up computing power, but we need to let AI help factories, energy and materials become more efficient. AI can be used in new chip design, production line scheduling, wind power and photovoltaic power generation forecasting, helping the industry save electricity, materials and speed up; the industry will give back cheaper, stronger computing power and more real-time data, allowing the model to learn faster.

Only by forming a positive cycle of "AI helping industry upgrade, and industry feeding AI growth" can we break the current bottlenecks of expensive computing power, high electricity costs, and insufficient data.

Therefore, the current development of AI has basically formed three directions:

First, continue to optimize the underlying computing power and energy . OpenAI and Microsoft are jointly building the Stargate computing park, which uses liquid cooling and renewable energy; AWS is developing its own Trainium‑2 chip and conducting experiments in hydrogen energy data centers; Nvidia's Blackwell architecture continues to increase the performance per watt. The goal is straightforward: to provide more effective FLOPs within the same budget, so that the unit cost of training and inference can be reduced by another order of magnitude.

Second, use new architectures and algorithms to improve efficiency . Google DeepMind's Gemini 2 introduced sparse activation and retrieval enhancement; Anthropic combined long context with RAG 2.0 on Claude 3; Meta's Llama‑MoE uses expert routing to complete reasoning by activating only a few weights. The idea is the same: do not blindly expand parameters, but let the model spend computing power on the most needed parts.

Third, embed the model into a measurable business closed loop . OpenAI's Assistants API, Google Vertex AI Agents, and Tesla's autonomous driving system all directly use models to solve programming, customer service, driving, and other tasks, and use business KPIs (response time, failure rate, takeover mileage) to verify the effect. If the model performs well, more real-time data will be obtained; if the performance is poor, feedback will be given immediately and improvements will be made. In this way, training and application merge, which not only alleviates data shortages, but also avoids ineffective waste of computing power.

In addition, although the development directions of major companies are different, they also have a new consensus:

The question has shifted from “who has more model parameters” to “who can embed the model into real scenarios and accumulate user networks the fastest”. Whether making chips, adjusting models, or building cars, the same question must be answered in the end: What specific pain points does your AI help users solve?

Only when the scenarios are running smoothly, the data loop is closed, and the effects are measurable, will capital continue to be invested and the model will have nutrients to upgrade.

Therefore, in the next stage, AI agents developed by major companies will take over a large number of standardized scenarios and links. People who only have the ability to handle a single scenario are likely to be eliminated.

02 Knowledge, experience and scenarios

Nowadays, many people are talking about equal access to knowledge. Indeed, the Internet and AI have made knowledge within reach and are no longer an obstacle to human development.

But the competitiveness that knowledge brings to people will inevitably decline, because scarcity is the key to value.

At this time, what widens the gap between people?

It's experience.

Everyone can use the big model, and its training corpus and searchable data are all public information. But it does not know your capabilities, your customers, or your specific problems, which are the result of long-term personal accumulation and capabilities in specific scenarios.

For example, if you are in charge of purchasing for a manufacturing company and have visited hundreds of parts factories in the past ten years, you would know which one would secretly change materials during the peak season, which boss is most afraid of cash flow interruption, and which highway will inevitably lead to delivery delays once it is repaired.

The big model can instantly list the national supplier list and capture the raw material price curve, but it cannot judge whether the other party’s warehouse is blocked based on the tone of a few phone conversations, or find signs of "wrong goods" in the details of the quotation like you do.

This is a scenario-based, warm experience, which cannot be searched or copied, and can only be developed through actual combat.

So what is the difference between knowledge and experience?

Knowledge can be learned, but experience can only be gained through practice.

The knowledge you learn is highly conceptual, abstract, and verbalized, which also means that you tend to lose sight of the details.

For example, if you want to make a product for college students, then you should talk to many college students and even conduct field research in universities.

But you are easily blinded by existing things. You think of college life as love, classes, postgraduate entrance exams, part-time jobs, canteens, dormitories, water and electricity, clubs, self-study, exams, etc. But these are highly conceptualized college lives. It is impossible to see new opportunities based on highly conceptualized cognition.

For example, at the end of June every year, there are piles of books and daily necessities that no one wants downstairs in the dormitory. Graduates are anxious to leave school and have no intention of running "Xianyu", but they are reluctant to throw them away directly, so these things are left next to the trash can downstairs, waiting for the school to clean them up. This scene does not have a fixed name, nor is it in any conceptual framework, but if someone quickly provides a service of "door-to-door classification, one-time settlement, and instant recycling", it is likely that a business can be made in an instant.

The common feature of these scenes is that they are outside the conceptualized and labeled university life. They have no fixed names and no one cares about them, but they are real needs.

Good opportunities come from the margins, from the unnamed place.

Moreover, experience cannot be acquired overnight; it requires time to be honed in real-life scenarios and requires one to personally bear the consequences of their decisions.

No matter how much knowledge you accumulate, you will not be able to solve small problems in the real world if you don't spend time, energy, brainpower and mental effort to accumulate effective experience.

Today is an era of oversupply, with oversupply of products, overcapacity, oversupply of information, and oversupply of knowledge. “I can solve math problems,” “I can provide knowledge services,” “I can complete XX task” are no longer very meaningful. The most important thing is to be able to deliver results.

So, where do the results come from?

Delivery from specific scenarios.

First, “specific” means personalization.

I see that many people regard the answers given by DeepSeek as information sources and arguments, which is actually the same as you regarding the content on TikTok and Xiaohongshu as the real world.

All the outputs of the current big model are the result of supervised fine-tuning, reinforcement learning, and prompt setting. It is just trying its best to provide answers that meet human expectations. It is your questions that allow the big model to temporarily generate various answers.

So, it makes no sense for you to ask DeepSeek to give you life advice, telling you what books to read or what to do next.

DeepSeek cannot see the details of your life, what you have done, how many times you have failed, your core, where your sense of existence is located, or your level of mental, energy, and brainpower.

You have to make your own way.

Second, "scene" represents where real problems occur. It means that you must step into a specific context and collide with real people, real needs, and real conflicts.

Only in real scenarios can you know what users really care about and what obstacles they really encounter. The pain points that users talk about and the pain points in actual actions are often not the same thing. If you chat with a few college students and ask them what they need, they may tell you "I want to improve my learning efficiency" or "I want to improve my social skills."

But what really makes them anxious and uncomfortable may be that they have no place to continue studying after the library closes at 11 o'clock in the evening, or that they cannot rest because their classmates snore too loudly after the lights in the dormitory are turned off.

This is the scene: only when you stand in it, can you know what difficulties and needs people really face. The scene is not those abstract keywords in your mind, but the real moment that is so specific that it can no longer be disassembled.

Third, "delivery" means that you actually provide value and truly solve that specific problem.

Behind delivery are concrete actions and concrete results, not knowledge or ideas themselves. Only when the results are clear and measurable can your value be truly proven and your experience be truly verified.

Therefore, the results come from the delivery of specific scenarios. This delivery must be based on your personalized experience, must be rooted in real scenarios, and must be able to truly create the value that users want.

Sequoia Capital’s AI Summit this year clearly stated that a true AI product is not about “whether it has capabilities” but “whether it has results”; it is not about “what you click it to do” but “what it accomplishes for you”.

Sam Altman made it clear in a conversation with Ben Thompson that he would rather have OpenAI become a website with a billion users than a state-of-the-art large model.

Therefore, the next stage of AI is to go deep into specific scenarios, solve problems, and deliver results.

03 Human Competitiveness

What does it mean when AI enters real-world scenarios, directly solves problems, and delivers results?

This means that the single, standardized scenario experience that you have accumulated in the past through "skillfulness" or "diligence" will depreciate rapidly.

for example:

• You have worked in finance for ten years, doing accounting and tax filing every day, but you have never thought about how to optimize the company's cash flow;

• You have been working in customer service for five years, answering the same simple questions from users every day, but you have never thought about how to fundamentally optimize the customer experience process;

• You have been working in administration for seven years, filling out forms and organizing documents every day, but you have never thought about how to automate the process to make the entire department run more efficiently.

These experiences can be easily replaced by AI because they are all standardized scenarios. Your value lies only in your ability to skillfully repeat the action itself, rather than in your in-depth understanding and optimization of the logic and goals behind the action.

So, what kind of people will not be replaced, but will be able to take advantage of the situation to make a leap forward?

Specifically, people with three abilities are more competitive:

First, build a personalized knowledge and experience system

The significance of establishing the PMCDE knowledge and action system is to create a unique "personal data asset" by recording, organizing and iterating personal knowledge, experience and insights, laying the foundation for customized and scenario-based applications of AI.

Each note is a reflection of my personal thoughts, containing my unique taste and aesthetics. It is not simply copying and pasting information from the Internet.

Moreover, I also present a lot of my thinking logic in words, which can be tested by AI. For example, my understanding of human nature:

I asked GPT-o3 to understand and evaluate:

That way I can revise my ideas to better fit the real world.

Through each iteration, I can obtain a lot of private data and experience to make up for the shortcomings of AI's lack of scenario-based perception.

You have to know that everyone asks common questions, and every model can give common answers. What you need to do is to highly customize the large model. If one day the supervised fine-tuning and reinforcement learning of the model can be based on your data, through repeated interactions, you will be able to polish out a personal assistant that suits your aesthetic.

Second, deeply participate in scenario-based collaboration

Business value comes from specific scenarios. Through in-depth interaction with people, teams and the environment, we accumulate scenario-based experience, build relationship-based trust and influence, and become the "key nodes" in the scenarios. Such people are extremely valuable.

Because AI is good at processing standardized data, but cannot perceive the "temperature" of the scene - such as the hesitation in the customer's tone and the hidden conflict in the team meeting. These contextual insights can only be obtained through human interaction.

And as your accumulated user network grows, your personal influence will also grow, allowing you to quickly mobilize resources in complex scenarios and deal with uncertainty.

For example, McDonald's has solved the problems of food standardization, scale and efficiency very well, and has even standardized the "smile". Whether you are in Africa or Central Asia, you can see the McDonald's signature smile on the faces of McDonald's waiters. This is the pinnacle of industrialization.

At the same time, McDonald's has carried out in-depth localization and integrated global resources into the local ecology. This allows us to experience that McDonald's will make porridge, soy milk and fried dough sticks for Chinese users.

This is the coordinated development of resources.

The same is true for individuals. Remember these three steps:

① Make yourself “predictable” first, keep your words, and complete what you promised on time; when others come to you, it’s like ordering a meal at McDonald’s: you always know what you will get.

② Learn to "localize" and use different statements and solutions for different people and different occasions. For example, if a state-owned enterprise needs stability, you should talk about compliance; if a startup needs speed, you should talk about efficiency. Your core skills have not changed, but the packaging has become more appropriate.

③ Keep doing it and keep being remembered. If you cooperate once and others are reliable, they will come again. If you cooperate three times and others are reliable, you will become someone who "must be added to the group". The more scenarios you have, the faster you react and the easier it is to borrow resources.

Simply put: stable output + flexible response = key nodes in the scene .

If you do this, opportunities will naturally come to you - because everyone knows that with you around, troubles will no longer be a problem.

Third, let AI amplify its capabilities

AI can quickly process massive amounts of information, generate diverse solutions, and reduce trial and error costs, but the final decision and value judgment depend on humans.

This requires you to dig inward and tap into your own system capabilities.

You must proactively learn how to collaborate with AI and use it as a “superpower” rather than letting it become your adversary. Specifically:

• Let AI quickly collect, aggregate and organize information for you;

• Let AI generate multiple possible solutions for you, and make key decisions yourself;

• Let AI help you quickly experiment with new ideas and verify results at the lowest cost.

AI is a tool, not an answer. You need to be the one who controls the tool, not the one who is controlled by the tool.

The above three capabilities are the key to your ability to coexist, collaborate, and achieve transition with AI in the future.

On the general big model, focus on your own scenarios,

Actively build a closed loop of personal experience, quickly transfer experience to different scenarios, and use AI to amplify rather than replace yourself , and you will not only survive in the AI ​​era, but also live better and better.