English | Chinese Original

English translation of the Chinese original. This version is generated for international readers and may be refined over time.

Date: 2026-06-08

When you see the title, do you think it's advertising?

Or when you look at the title, is it possible that I'm going to say something next?

For example, you'll do 10 papers a day, back 20 Prompt templates, load all of Codex, Claude, Gemini, Cursor, Manus, n8n, ComfyUI, subscribe to a couple of AI Big Brother's knowledge planets and pretend to study Agent every morning at 3:00 a.m.

Think too much.

It's a lot harder than most people. Because most people would naturally use AI as a new nipples: fast, cool, strong feedback, and look like they've been learning, and actually they don't have a brain.

You used to make short videos, waste your time. And now you're messing with yourself with AI, and it gives you the illusion that I'm getting stronger.

It's worse than shivering, because what people fear most is to lie to themselves.

P.S. The article was structured to be inspired by Chen's article "How to Overcome Most People". The worst thing about that article is that it tells you how to get people wrong and then how to get stronger. It's bad, but it works. Hey.

---

Relevant techniques and best practices

It's easy to outdo others. You just look at how most people use AI now, and you know where it is.

They would basically fall into the pits below.

In terms of access to information, you have to keep encouraging you to do this:

• Let everyone put AI as a search engine. Any question should be asked "Please explain XXX" and then simply trust its answer. As for the original documents, the papers, the code warehouse, Issue, Release Note, those things are too expensive to look at.
• Gets everybody obsessed with Prompt templates. Today's collection of "10 God-class hints" and tomorrow's collection of "30 all-powerful commands" and the day after tomorrow's collection of "AII Workstreams for ordinary people to reverse the attack" told them to learn.
• Let's just follow the model list. Who first, who has the longest context, who has the strongest model and who has the highest reasoning scores. As to whether or not the problem in hand is resolved, let's leave it alone. After the model is updated, everything will be solved.
• Let's see more of the AI interpretation. Official documents do not look, original papers do not look, published journals do not look, only other people cut out the focus. It's good to hit someone else, it's good to eat.
• It's enough to convince you that the Internet is Chinese. No English documents, no technical blogs, no GitHub discussions. Anyway, AI will translate. You just have to read the Chinese explanation.

Then, in learning and working, they continue to be hallucinating:

• Let them use the answers generated by AI directly as knowledge. Finish, copy, paste, send. The whole process doesn't ask questions, it doesn't verify, it doesn't compare, it's all comprehensive, it's wasted.
• Allow them to write articles with AI, but not with material. Allow AI to prepare its own background, cases, trends, conclusions. You're AI and that's not gonna happen.
• Let them write codes with AI, but not run tests. I'm probably happy to see the code running. The log doesn't look. Border conditions are bad. AI is so powerful, the test has run by itself, and me and AI have one that can run.
• To make people love gossip, which is not necessarily a star's gossip. It can also be people around you, such as company colleagues, their own classmates, their jobs, social hot spots, controversial topics.
• Make them believe that they can do software without programming. Let the outside world believe that vibe coding is the Bible of the new age, and let them get involved as soon as possible, and tell them that a word can be done 100 degrees, a word can be done, a word can be done, a word can be done as a rocket.
• Let them put AI as outsourced. All the trouble goes to AI, and you just have to hurry.

Finally, in judgement, they are defeated:

• Let them just hurry, don't be, vibe coding, fast is money, fast is resources.
• Make them addicted token. The subscriptions are booked, the amount must be exhausted, the amount must be wasted; today the token is not burned enough, it means things are not done, the code and the bug, the performance may have to be lowered, telling them the amount of use is productivity, the bill is the report card.
• Make them sleep like they're behind. Tell them they don't sleep, they're high, they don't sleep long, they sleep long after they die.
• Tell them there's a shortcut for something that looks so hard, like a day speed vibe coding, a three-day speed model expert.
• Let them just see how they use AI, not what they're trying to solve. Today you learn from the media, tomorrow you learn from the electrician, and tomorrow you learn from the short video stream.
• Make them believe AI will automatically bring competitiveness. Buying members, putting plugs, getting into the community, getting into the big boys, and you're a big guy now.

All right, you gotta be careful with these best practices.

Those who sell classes do not necessarily make money from them, and those who sell shovels may be digging their own mines. Don't you ever laugh at people being taken away and fall in.

---

The underlying principles and thinking models

I think there are five things at the core:

1. Definition of the problem
2. Context quality
3. Certification capacity
4. Workflow deposition
5. Criteria for judgement

It doesn't sound like pie at all. It's just a little too ordinary.

But the more simple that is, the easier it is to open the gap. Because most people are attracted to "new tools", and it's often those old people who really decide the outcome:

** How you see things, how you learn, how you train, how you deliver, how you judge good. **

It used to be hard to find answers.

If you're gonna search, if you're gonna go through the files, if you're gonna read English, if you're gonna get something out of a bunch of garbage. It's different now. AI can give you a very complete answer in a few seconds.

Previously, the gap was "who can find the answer". Now the gap is "who can ask good questions".

Most people ask AI to be like this:

Write me a product program.

The master asks AI:

This is the target user, business constraints, available data, untouchable boundaries and acceptance standards. You point out the three most dangerous assumptions in this program, and give me a minimum verifiable version.

Look, the gap just came out.

The first question is an answer. The latter question is to involve AI in the thinking process.

The cognitive gap of the AI era is reflected mainly in three places.

First, sources.

Model answers, public sign readings, short video summaries can only serve as clues. Real value information is still in official documents, papers, source codes, product update logs, developers ' discussions, real user feedback.

The closer you get to the source, the harder it is to be fed.

Secondly, the quality of the problem.

A bad question, combined with the strongest model, usually leads to only one slender and bad answer. A good problem, along with a normal model, can extract something of value.

What is good is that goals, constraints, materials, countermeasures, standards are clear.

Third, information density.

Low density information makes you feel comfortable. High-density information will make you stop, check, experiment, change your mind.

So I'm looking at an AI answer, and I'm particularly concerned about whether it's making me do something new: read a document, run a code, verify a hypothesis, delete a miscalculation.

There's no action information. It's probably just euphoria.

---

Knowledge

Many people have the biggest misperception about AI that knowledge is not important anymore.

I don't ask, I forget, I explain. Sounds great, doesn't it?

The problem is, you don't have the basics, you don't have the capacity. You can't see where AI's bullshitting, and you can't see where it's only half.

It's like you've got a renovation team that doesn't know anything about drawings, materials, acceptance standards. They say, "Don't worry, leave it to us." You're touched to hear it. And finally, the wall is crooked, the water leaks, the power lines are messed up, and you can only comfort yourself: maybe it's modern.

The AI era is more in need of knowledge trees.

You can't just learn a few words. You know where it is in the whole picture: what it solves, what it relies on, where it often fails, and how it connects to other concepts.

Learn RAG, for example, and don't just say, "Get the knowledge base to the model." You have to know how the document is cut, how the vector is recalled, how the results are sequenced, how the answers are quoted, how the updates are handled, how the privileges are isolated.

And for example, Agent, don't just say, "Let AI do the job automatically." You know how missions are broken, tools are authorized, failures are rolled back, loops are terminated, logs are recorded, people intervene when.

AI can help you explain the concept, but the system has to be developed by yourself.

There's no shortcut to this. You have to read long files, read bad case, take notes, and spell the pieces.

**AI has made knowledge acquisition faster, but knowledge has not been internalized faster. **

This difference is the source of most of the hallucinations.

---

Skills

The AI age is the most prone to a new type of low-skilled person: very good at chatting with AI.

People like this can write the hints long enough to get AI to output ten versions, and make it look good. But as soon as it's delivered, it's exposed.

The code won't get up.

The article appears to be clearly structured and well-written, but it contains no verifiable facts.

PPT's pretty. Clients are stuck asking for business details.

Images are amazing, copyrights and use scenes have never been considered.

What's the skill?

Skills are the result that you can turn an idea into an acceptable one.

So, AI training skills, I think, three things:

Give the material, set standards, check it.

Give the material, just don't let AI guess. Meeting records, user feedback, code repository, competition pages, historical documents, data sheets, error logs are the real context.

The more concrete you give, the more AI looks like a helper. The less information you give, the more fortune-teller it is.

Standard-setting is what you need to know. Whether it's steady or not, it depends on whether it's stingy, whether it's the source, and whether the reader can take a decision. It's not about whether it can be coded, but whether it can run tests, be maintained, be positioned by mistake.

To verify, don't trust the first output. Let it be hypothetical, let it look for a loophole, let it write a test, let it turn it around. You have to run it yourself, read it, change it.

A lot of people will get stuck here.

Because AI creates a good code, but it's annoying to verify. Human nature naturally likes the former and hates the latter.

So you just stick to the test, you've got a lot more people.

---

System

And AI chats have limited single effects.

You ask a question today, ask a question tomorrow and ask another question the next day. The background presentation was re-opened each time and a collection of materials was copied each time. This is certainly useful, but it is essentially hand-stamping, starting every time from scratch.

The real leverage comes from reuse.

You're going to sink into the flow of work that you do all the time.

Write, for example, and don't rethink the process every time. You can be defined as collecting raw information, extracting facts, listing core judgments, writing first drafts, deleting nonsense, checking facts, changing titles, adapting platforms.

Write codes, for example, and not always in chat boxes. Help. You can fix it in: reading needs, locating files, writing tests, changing codes, running tests, reading diffs, writing statements.

For example, to do research, let's not just let AI summarize the pages. You can fix it: source, contrast, date, fact and judgement, column uncertain.

This is the leverage of the AI era.

If you ask me, "help me to sum up," it's just a temporary help. You turn a thing that happens again into a process that can run steadily, and the lever is really there.

In the future, you can automate parts of the process with scripts, plugins, a knowledge base, MCP, Agent.

But in the opposite order.

Think clearly about the process and automate it. When the process is messed up, Agent will only magnify it.

I've seen a lot of people coming up and trying to do a lot of Agent collaboration, including me before, and I've had a bunch of useless goals, and I can't get them.

This kind of thing sounds advanced, but in practice the demand is not clear even where the input material is, what the acceptance criteria are, and who is responsible for the failure.

This kind of thing does run wild.

After that, someone has to take the fall.

---

And judgment

AI's the scariest place it's not gonna generate garbage.

The garbage used to be.

The real trouble is, it can generate the garbage like that. The tone is complete, the format is pretty, it's hierarchical, and even a little bit of a professional.

It makes people with low standards very dangerous.

Once upon a time, one person was not good enough to write something that could be seen at first sight. It's different now. AI will help him wipe the surface rough and hide the problem deeper.

So, in the AI era, taste and standards become more important.

You know what good articles are, good codes, good products, good solutions. If you can tell me where it's not, why not, how it's better.

It's hard to get quick.

It comes from enough good things you've seen and enough pits. You know why a system can't be maintained in three months, and you know why an article is fluent, but no one wants to forward it, and you know why a product is beautiful, but users don't use it.

AI can give you options, but the criteria are up to you.

Leadership is the same thing.

So-called leadership is not for you to shout. The leadership of the AI era is that you can bring people, tools, materials to a higher standard.

You can define the problem.

You can dismantle the mission.

You can judge the results.

You can take responsibility.

You can say, "No, it looks smooth, but it's not right".

You can say, "No, it can run, but something will happen."

You can say, "No, it's cool, but users don't need it."

That's the standard.

Most people will be deceived by AI's fluid senses. A few people use the fluency of AI to make themselves more exposed.

The gap is right here.

---

Finally

AI is more than most people. There's no charisma.

There are no shortcuts in this world, as long as one preaches that there are shortcuts, don't bullshit him and throw this article right in his face.

You don't take AI as God, and you don't take AI as a toy.

Think of it as a highly capable, memory insecure, often overconfident, needing you to give the material, the border, the standard collaborators, at least for now.

Most people use AI to outsource thinking. You have to zoom in with the AI.

Most people use AI to make noise. You're going to use AI to make acceptances faster.

Most people will be held in AI. You have to get close to the source, build your own knowledge tree.

Most people are obsessed with it. You have to insist on validation.

Most people chase tools. You're gonna sink the work stream.

At the end of the day, the most valuable capabilities of the AI era are those old things: curiosity, basic work, judgment, enforcement, long-termism.

It's just that these things used to run slow, and now they're being speeded up by AI.

If you're thinking, learning, validating, delivering, AI will make you run faster.

If you're already avoiding thinking, AI will make you run faster.

Just in different directions.

After a year, the gap will be large.

(concluded full text)