LIST

Two different experiences brought by fineness

| 2025-06-02 14:16:04

There are two kinds of subtleties in the world: one is perfectionism and the other is touching. Why does the latter often touch our hearts more?

The refinement of perfectionism is the persistence in “perfection” itself - the flawless surface of ceramics and the accuracy of each note on the piano keys. The refinement of moving lies in the connection of emotions, which is like the beauty of imperfection in the Japanese “mono no aware”, and is a subtlety that can touch the soul. It flows in the performance, each change in intensity tells a story, and each rest carries the weight of emotion.

Instrumental rationality says: precision is to achieve the objective standard of “perfection”.

However, value rationality says: refinement is to convey the intrinsic values ​​of “love”, “care” and “beauty”.

The former is more about inward self-demanding, while the latter is about outward emotional resonance. One pursues the ultimate standard, while the other desires a deep connection.

#culture

Prompt Design Philosophy

| 2025-06-02 12:56:14

Recently, especially after coming across some of Mr. Li Jigang’s Prompt design philosophy, I would like to talk about it briefly based on my own insights.

The so-called “one word can make the big model understand” is the power of compression. In essence, it is the refinement and optimization of the input, aiming to make LLM capture our intentions more accurately.

How to write a good prompt? I think there are three levels of logic:

  1. Understand the characteristics of LLM: Different LLMs have different “tempers”, and the same method may have very different effects. Only when you know yourself and the enemy can you use it freely.

  2. Industry knowledge barrier: This part is the accumulation of professional fields. For example, aesthetics, how to build a harmonious vision, how to use colors, these top-level perceptions are difficult for LLM to achieve in a short period of time.

  3. Clear logical expression: It is very important to state the problem clearly. For the same requirement, a clear and logical expression is far better than an ambiguous one, which can help LLM to understand it accurately.

The greater the subjective differences in a field, the more difficult it is for LLM to replace it. Aesthetics is an example. Although there are common beauties such as symmetry and color matching, “different people have different tastes.” For example, the beauty of “mono no aware” is a unique emotional projection of an individual.

What are the boundaries of LLM? As it gets better day by day, it seems that we need fewer and fewer prompts. Do you remember that if you say “be brief and concise” to GPT-3.5, it may not understand; but if you say the same thing to GPT-4, it will understand it. The difference between the two is the leap in understanding ability. If the model still cannot understand, we need to lay out abstract words, supplemented by analogies and examples. The improvement of the model means that this understanding gap is gradually being bridged.

I personally think that the progress of LLM does not compress the space of human expression, but pushes us to migrate to a higher cognitive level. Prompt engineers are more like translators and architects at the cognitive level. The real prompt engineering is not “training AI”, but designing a cognitive collaborative architecture. With the evolution of AI, this role will be elevated from “instruction writer” to “cognitive resonance designer”.

#AI

The Korean culture I felt was a little different

| 2025-06-02 12:42:06

A few months ago, and even for a long time since I became self-aware, my feelings towards South Korea were quite complicated, even a little resistant, due to the interweaving of Internet curiosity content and anxiety about national identity and cultural sovereignty. Before coming to Jeju Island, I thought that this trip was just to see the charm of Olai Road, and it might be the only time in my life.

When we emphasize the “independence of our own culture”, it is a manifestation of national pride at home; but in certain situations, the scrutiny of other people’s cultures may stem from a collective anxiety of “loss of cultural authority.”

The people and things I encountered in Korea this time were no different from my previous experiences in other countries: human nature is universal, and society and institutions shape the myriad aspects of one’s character. There are also philosophical implications in the cultural core that are worth pondering:

  1. The Korean word “한 (hate)” is not simply hatred, but a complex state of existence - a mixture of pain, unwillingness, tenacity and hope. This is similar to Camus' Sisyphus: knowing that the boulder will roll down, he still pushes it to the top of the mountain. Koreans turn this “hate” into a driving force for moving forward. This is a profound philosophy of life: Accept pain, but don’t be swallowed by it. Compared with some cultures that restrain emotions, Korean culture encourages direct emotional expression. This is not an overflow of emotions, but a respect for the true heart.

  2. The extreme pursuit of aesthetics has led to the strong output of Korean culture, but it has also brought about a fierce internal trend of appearance anxiety.

  3. The aesthetic concept of touching through meticulousness is fully reflected in Korean makeup. The original intention of makeup is to modify and conceal, but Korean makeup chooses another way of “concealing” - Korean society expects women to be flawless and natural. As a result, the “꾸안꾸” (꾸민듯 안꾸민듯, carefully dressed but as if without makeup) culture came into being. The logic is:

  • No makeup is a kind of reality.

  • The look of your face when you are energetic, well-rested and in a good mood is also a kind of reality.

  • Makeup is nothing more than using techniques to present the true state of the latter.

This “fakeness” isn’t a deception, it’s more of an aesthetic choice — choosing subtle over dramatic, enhancement over transformation.

#culture

Simple and Complex

| 2025-06-02 11:27:09

While hiking yesterday, my thoughts were wavering between the balance of simplicity and complexity.

I am increasingly fond of simplicity, whether it is pure simplicity or abstract refinement. However, I prefer to get along with complex people. It is not that simplicity is bad, but that complex people are engraved with more real marks of life: those struggles, contradictions, uncertainties, gains and losses, insecurities… are all the most honest portrayals of human nature. Observing them is observing human nature itself.

We often say that growth is a process of “simple → complex → simpler”. The closer we are to the truth, the harder it is to describe it in simple language. This is a reverence for the complexity of the world. Those who easily give standard answers are either insightful or shallow in their cognition. Lao Tzu said, “Those who know do not speak.” It is not that he is unwilling to speak, but that language and words also have their limitations.

This leads to a contradiction: Do we still need to pursue simplicity? Simplicity is a kind of wisdom in life, such as “the great way is simple”, Lao Tzu used a few words in the Tao Te Ching to describe the mysteries of the universe; Steve Jobs hid complex technology behind a minimalist interface. Simplicity never means easy.

What is the sense of boundary here? It comes back to the understanding and application of “balance”. For example:

  1. Honest simplicity: Don’t deliberately package yourself, knowing that perfection itself is a falsehood. Exposing inner contradictions and imperfections is far more precious than superficial brilliance.

  2. Admit ignorance: The more you understand yourself and the world, the more you will find that the answer closest to the truth is often “I don’t know yet.”

This is still a philosophy of life, an aesthetic that contains depth in simplicity.

#Life/Growth

Realistic feeling during hiking

| 2025-06-02 10:51:22

There are many things in this world that are difficult to fully explain by logic, such as why we hike.

I once talked about this with a friend, and he asked me why I love hiking, even though it is so hard. I can naturally analyze it from a rational level: Is it to enjoy the joy of reaching the top, to be obsessed with the stunning scenery that can only be seen after taking one step at a time, or to be intoxicated by the process of focusing on the present moment during hiking? These are the answers given to others, but not necessarily to myself. I think I have a shallow understanding of this passion, and can only roughly outline some language fragments that are difficult to restore to reality. The real answer, perhaps only when I walk alone on the ridgeline, the feeling of the fusion of body and mind at that moment, is the most real.

Many things cannot be explained by logic. Why do some people work so hard for a reason? Why do certain melodies always touch our hearts? This is an irrational passion. Science explains the world, art feels the world; logic builds order, and the almost “crazy” passion gives the world warmth.

What am I thinking about when I’m hiking? Mr. Liang Wendao also loves hiking. He said it is a spiritual reward. Many times, hiking allows me to enter a state of being completely immersed in the present, without having to think deliberately or judge myself. Sometimes I don’t think about anything, just let it go. Enter nature and let myself go. When we really stop thinking, we begin to “see”.

During the hike, most of the time may be “wasted” - emptying your mind, feeling the wind, feeling the sunshine, feeling the fatigue and strength of your body. When you occasionally come across a small path, meet people, and see the mountain scenery, your thoughts will naturally emerge like a stream, and then spread and drift. This is a peculiar “flow” state, and it doesn’t matter even if it is interrupted. I enjoy this natural growth process, this unpretentious sense of reality.

I especially like the “truth” that people show when hiking: some people fall silent while walking, some suddenly start to mutter to themselves, some share a sudden enlightenment with shining eyes, and some just smile and say “The wind is so nice today.” Accept this truth and let yourself flow naturally like water.

#on foot

Refusing to exchange money

| 2025-06-02 08:28:12

I went to a Korean fried chicken restaurant, and I never expected that the whole place was full of compatriots. Most of the set meals are for two people, and I ordered two sets of double combos, thinking that I could try more flavors. When paying, I specifically asked if I could pay by credit card, hoping to save some cash, which I always felt was not enough.

Two men and two women at the next table took the initiative to greet me, and after confirming that I was Chinese, they asked to exchange money. They said they were short of cash and needed it urgently, and one of the girls suggested transferring money via WeChat.

I answered, “No,” without further explanation. At that moment, I felt that “No” was “No,” and no additional reasons were needed.

They seemed a little stunned, and the boss was also a little embarrassed. Their embarrassment came from the failure of expectations and from face.

A person, in a foreign country, refused a stranger’s request to exchange money and chose not to explain. The other party felt embarrassed. This is not about right or wrong, but just a collision between expectation and reality, and a natural friction between face and self-boundaries.

People’s embarrassment often does not come from the indifference of others, but from the fact that their expectations of others’ responses are not met.

#Life

Do you believe that all good food is the same, or do you believe that everyone has their own taste?

| 2025-05-31 23:04:32

As the saying goes, “it’s hard to please everyone.” In Jeju Island, I tried several “delicacies” recommended on social media - Korean grilled fish, BHC fried chicken, and seafood Shin ramen. To be honest, I was a little disappointed. If the full score is 10, the grilled fish can still be given 5 points, while the fried chicken and ramen are only 3 points. The prices are not cheap. These stores are quite popular on Xiaohongshu, and the ratings on Google Maps are also quite impressive.

I once suspected that there was something wrong with my taste buds, or that my taste was quite different from that of Korean locals. But I didn’t seem to encounter similar situations in other countries, and local popular restaurants often won my heart.

Until last night, I had dinner at Vibe 230. The owner is a Korean. He is not fluent in Chinese, but he can communicate in basic numbers. The restaurant is decorated in a unique style. Brick walls, antique ornaments and a large number of CDs create a unique and relaxing atmosphere. The menu includes pasta (especially fern pasta, shrimp pasta, basil pasta, etc.), fried pork chops, Spanish garlic shrimp (Gambas al Ajillo), grilled cheese, etc. Most diners have good reviews, and some even call it a “pasta gourmet restaurant.” The owner really works very hard.

At that moment, I was more inclined to believe that the best food is universal. Most people can tell the food that really touches their taste buds after just one taste. Some flavors seem to be a common code for human beings, just like music and art, deeply rooted in our genes, making people love them instinctively.

#culture

Writes and Write-Nots: Writers in the AI ​​Era

| 2025-05-31 22:32:03

“Writes and write-nots” is a phrase that originated from Paul Graham’s essay “Writes and Write-Notes” published in October 2024, which cleverly borrowed the language sense of “the haves and have-nots”. He proposed: “In the long run, there are only two types of users - those who can write and those who can’t; those who can write are scarce, and those who can’t write will outsource their writing to tools or others.”

After reading this, I was deeply touched. In this, we can vaguely see the difference between “active” and “passive”.

Active people have their own goals in mind, and AI is just a good horse to help them get there. Passive people are driven by achieving their goals and will hand over tasks to AI as much as possible. For example, the former have their own ideas and judgments about the code generated by AI, and AI just speeds up the writing process; the latter may not examine the output of AI, knowing the results but not the reasons.

This is essentially about the autonomy of consciousness and cognition.

The key to judging whether a person is a “writer” lies in whether he or she has an original perspective, a coherent cognitive structure, and an attitude of responsibility for language output. This may also be a great angle to examine whether AI will replace humans.

#AI

Understanding the essence of growth

| 2025-05-31 22:19:45

I once saw a formula: The essence of growth = (happiness + sadness) / normal.

When I first saw it, I deeply agreed with it. Deep down, I am not a person who is content with “normal”. For me, choosing an unknown food full of uncertainty is far more meaningful than tasting the known food repeatedly. If you are trapped in the “normal” state for a long time, it will be difficult to break out of the situation and break through yourself.

But if you think about it carefully, is the ordinary really unimportant? Pain makes people wake up quickly, and happiness inspires creativity and sense of direction. But if you take a long-term view of life, ups and downs are ultimately few and far between, and what you remember deeply are often those moments of “change”.

True growth often happens in the “ordinary”. This slowness is like hiking, taking one step at a time. There may not be amazing scenery along the way, but you still need to keep moving forward; when you are tired, you can take a break. Being down-to-earth is the foundation.

On the one hand, human growth requires deep emotional experience, and on the other hand, it relies more on the sedimentation and accumulation of time.

Perhaps, growth ≈ (awareness × action) × time.

  1. Being happy or sad is the starting point of awareness.

  2. Normally, it is the field of time, the soil where actions can be sustained and influences can ferment.

  3. If there is only feeling but no action, it is just “experience”; taking action and learning and adjusting from it can be called “growth”.

#旅居 (This tag seems to be irrelevant to the content. The original text is #旅居. Please let me know if it needs to be adjusted)

Indian young people

| 2025-05-31 19:42:09

I also met many Indians in Jeju Island. These two countries are geographically far apart, but both have the largest young population in the world. What kind of anxiety do they have about the future, and are they eager to occupy a place on the world stage?

“The world’s largest youth group” - more than half of India’s population is under the age of 25. This is undoubtedly the largest number of young people in a single country in human history.

How will their hopes, plans, worries and dreams shape the future of this country and the world?

It is estimated that by 2025, the number of young people under the age of 25 in India will be between 600 million and 750 million. The energy and potential of this young force is immeasurable.

#culture

Why are young people in South Korea so introverted, while young people in Japan are “lying flat”

| 2025-05-31 18:39:11

When it comes to the living conditions of young people in East Asia, Japan and South Korea present an interesting contrast.

Why is Japan “lying flat”? Since the “bubble economy” burst in the 1990s, Japan has entered the “lost thirty years”. Economic growth is weak, prices and wages have stagnated for a long time, and the young generation generally lacks optimistic expectations for the future. Competition in the workplace is fierce and the promotion channels are narrow. Against this background, some young Japanese people have chosen a low-desire lifestyle - “no marriage, no house purchase, no consumption”, the so-called “lying flat”.

Young people in South Korea, on the other hand, are more likely to show a “passive struggle” or “involutionary struggle” attitude. The concept of “cannot lose” and “getting ahead” is deeply rooted in the bones and has become the social norm. This society attaches great importance to academic qualifications, appearance and position. Celebrities and Internet celebrities are often regarded as replicable success templates, which inspire young people to join the fierce competition. They seem to “cannot lie down, can not roll up, but still fight.”

Subcultures such as “N-time test takers” (repeatedly taking selective examinations such as civil service and postgraduate examinations), “three-abandonment generation” (giving up love, marriage and childbearing), and “involutionary culture” mixed with self-deprecating subcultures (such as the popularity of the term “Hell North Korea”) are all reflections of this situation.

Behind this, perhaps, is a deep fear of failure, the drive of cultural inertia, and a struggle that has not yet completely lost hope.

#culture

The Curse of Constant Worry

| 2025-05-31 18:32:22

She seemed to believe that by constantly worrying she could somehow avoid the misfortunes that came her way, as if worry itself were a spell that could ward off bad luck.

#Life

History of Olle Road

| 2025-05-30 23:32:38

Last night, while enjoying fried chicken for dinner, I was buried in my computer, researching all kinds of things about Olai Alley. Seeing this, the owner of the fried chicken shop was curious and started chatting with me.

I told him that I was researching the Olai Trail, and the boss mentioned that the trail was built with funds raised by the islanders. I was a little surprised by this statement. The kindness and simplicity of the islanders are true, but a project of this scale must have the foresight and courage of the initiator.

The birth of Jeju Olle Path has a very touching story. It originated from the dream of a special media person.

The founder is Suh Myung Sook, who was born in Jeju Island, South Korea in 1957 and graduated from the Department of Education at Korea University. After graduation, she joined the media industry and worked as a reporter for the famous Sisa Journal weekly in Seoul. She was promoted to editor-in-chief of the weekly in 2001. Later, she became the editor-in-chief of Ohmy News, the world’s first citizen news website.

In September 2006, Xu Mingshu resolutely ended her 23-year media career, put aside all work, and gave herself a long vacation to completely relax. After a month of careful preparation, she flew to Spain alone and embarked on the Camino de Santiago, a pilgrimage route with a thousand-year history, starting a journey of meditation and recovery.

Xu Mingshu spent 36 days and nights on this pilgrimage. The scenery, local customs, historical sites and travelers she met along the way left her with an indelible impression and deeply inspired and encouraged her. Near the end of the journey, she met a British girl named Henny. Talking about her travel experience, Henny said: “If this trip is so important to us, why don’t we build a ‘Camino de Santiago’ in our respective hometowns after we return?”

This sentence was like a seed that took root in Seo Myeong-sook’s heart. After returning to Jeju Island, she decided to create a similar hiking trail in her hometown. She named the trail “Olle”. In the Jeju dialect, “Olle” means “a narrow path connecting the main street to the front door.”

#Travel

Kindness and freedom during the journey

| 2025-05-30 21:10:43

🇰🇷 Jeju Day2 Solo Olle 1 - 16.9km ✅

While I was waiting for the bus, an elderly local lady came over and gestured in Korean, probably about bus-related matters. Seeing that I didn’t quite understand, she took out a bottle of beer from her belongings and gestured to give it to me. Before I could react, she had already opened the bottle cap neatly and drank it all in one gulp. Her carefree attitude was very impressive.

My first subconscious reaction was that “drinking is bad for the body.” This instantaneous intuition formed a sharp contrast with the grandmother’s straightforwardness, inadvertently revealing the collision of two life philosophies: one is to embrace the heartiness and freedom of the moment, and the other is the rational restraint and long-term consideration deeply rooted in the heart.

I sincerely admire this old lady because I deeply identify with and yearn for “living in the present moment”. Her behavior is an affirmation of pure enjoyment of the moment, not bound by age, social norms or health concerns.

This is one of the meanings of travel. We are shaped by our own culture and experiences, but unexpected encounters with others give us the opportunity to appreciate and think about different ways of life, even if those ways sometimes challenge our inherent intuition.

Occasionally, put aside excessive rational analysis and embrace the purity of the present moment. Sensibility and perception may be the more authentic source of happiness, rather than complete rationality.

You have to be yourself. And what you admire are those who embrace the present and look to the future on their way to becoming themselves.

#Travel

The neural circuit trap of Hedonian pleasure

| 2025-05-29 17:17:37

The instant stimulation of short videos brings about the rapid release of dopamine. Over time, the threshold of pleasure quietly increases. We need stronger stimulation to feel a little happiness, and may eventually fall into a dilemma of “pleasure poverty”. This is the neural circuit trap of Hedonic pleasure.

In contrast, the cognitive construction process of Eudaimonia (a deeper sense of satisfaction related to meaning and value) is more profound:

Strive → Understand → Connect → Internalize → Grow → Deeply Satisfy.

There is an implicit premise in Udhammoniya: Without resistance, there is no real growth.

In this era of fast consumption, we are used to consuming information. On the surface, information seems to be becoming more digestible, more interesting, and more accessible. However, at a deeper level, we may be gradually losing the ability to “cognitively chew”. When faced with a book, more and more people tend to look for shortcuts and are eager to quickly grasp the core.

Isn’t this, to some extent, also a kind of cognitive degeneration?

  1. Long text era: Deep plowing → sowing → waiting → harvesting (farming mode)

  2. Short video era: Discover → Collect → Consume → Forget (Collection mode)

Just as, if even thinking itself needs to rely on AI, then what is the meaning of “our” existence?

The real tool should enhance the user’s cognitive sovereignty, not replace it. We should strive to create an “anti-entropy” reading experience - after reading, the entropy of the user’s cognitive system is reduced rather than increased.

We have discussed the differences between intrinsic motivation and extrinsic motivation. Cognitive gamification has similarities:

  • Traditional gamification: Reliance on external rewards (e.g. points, badges).

  • Cognitive gamification: Pursuing intrinsic rewards (e.g. the thrill of understanding, the surprise of making connections).

The pleasure of watching short videos tends to be more towards the former.

What we should really think about and reflect on is how to make those beautiful human qualities - such as deep thinking, knowledge accumulation, and cognitive growth - more accessible and more desirable?

  • How to make deep thinking as addictive as watching short videos?

  • How to make the accumulation of knowledge as clear and visible as game upgrades?

  • How to make cognitive growth as full of immediate feedback as social interaction?

#field

Kindness from a chance encounter

| 2025-05-28 20:16:27

The sunset over Jeju Island reminded me of the flight from Bangkok to Nepal.

It was the first time I saw the magnificence of the Himalayas so vividly. I was sitting next to the aisle and met the Nepalese for the first time. They gave me the same impression as the Indians, lively and enthusiastic. On the plane, they seemed to know each other and chatted between the seats.

The guy sitting by the window also likes to record the view outside the window with his mobile phone. He seems to have acquaintances in other seats of the plane, so I need to stand up from time to time to let him pass. At first, I felt a little uncomfortable and thought it was a bit noisy, after all, I am an introvert.

When the plane flew over the Himalayas, I was excited by the towering Mount Everest above the sea of ​​clouds and the majestic momentum overlooking the other mountains. I picked up my phone and kept trying to use 5x and 10x zoom, hoping to capture and restore the shock of what I saw with my own eyes.

The Nepalese guy next to me seemed to sense my desire and he stood up and gave me the window seat.

This kindness of a chance encounter, like the afterglow of sunset, remains warm in my memory. The nameless Nepalese boy is also deeply imprinted in my mind, reminding me from time to time that there can be such pure care between people.

#Life

In the era of short videos, what are the essential changes in pleasure?

| 2025-05-28 15:21:15

The core of developing information consumption products lies in a deep understanding of the nature of information consumption.

Pleasure can be roughly divided into two types:

  1. Hedonic pleasure: immediate, sensual, and superficial. For example, the pleasure of watching short videos with fun, pleasing visual elements, or special effects.

  2. Eudaimonia: A deeper sense of fulfillment related to meaning and value. For example, the fulfillment and happiness you experience when you immerse yourself in a book, complete a challenge, or establish a deep connection with others.

Those small, short-lived, fleeting pleasures are like a pleasant illusion and can hardly bring deep and lasting satisfaction. However, can this pattern last?

I once thought that even in the AI ​​era, individuals do not necessarily need to have the ability to fully process structured, long content, and even fragmented information input is enough. But now, I have a deeper understanding of this view.

#field

The best AI products are not planned, but grow spontaneously from the bottom up

| 2025-05-27 19:26:59

When using Claude Desktop and the integrated development environment (IDE), Anthropic engineer David Soria Parra was keenly aware of the lack of effective integration between the two, which led to a cumbersome workflow. He realized that this was a typical “M×N” problem - the complex relationship between multiple applications and multiple integration requirements. MCP (Model-defined Context Protocol) was born under such demand.

The emergence of MCP is not icing on the cake, but a natural development and a product that solves actual pain points.

If AI products a year from now are still at the level of “I can ask questions and it can occasionally make proactive suggestions,” then they will find it difficult to gain a foothold due to lack of differentiation and core appeal.

It is meaningless to speculate on user needs day after day. The obvious needs have long been seen by the market, while the invisible needs may be our wishful thinking. Trying to “educate” users with your own cognition and expecting them to use “the product I want them to use” is often a futile and even slightly stupid behavior.

Only companies that apply real technology to solve real and valuable problems can be more creative and have lasting vitality.

#AI/product

Claude’s Projects Design Thinking

| 2025-05-27 14:16:21

In my past understanding, the “Projects” function of most products is mainly used for content classification or module distinction.

After subscribing to and experiencing Claude, I found its “Projects” feature quite meaningful. ChatGPT has memory capabilities, and I imagined that it would support longer-term memory management in the future. Admittedly, tools like Mem0 also support the classification or labeling of stored memories to distinguish between working memory and life memory. But in actual applications, we are often accustomed to handling multiple tasks in the same conversation window, or not deliberately managing multiple windows. The problem this brings is the confusion of memory-for the model, this mixed context may lead to inaccurate memory. Talking about work today and learning tomorrow, the model can easily “thread” and incorrectly associate information on different topics.

Think of Claude’s Projects as folders on your desktop. Each Project is a dedicated conversation space where you can:

  • Upload relevant documents and files.

  • Set specific work rules and background information.

  • Have an ongoing conversation and Claude will remember all contextual information within the Project.

For example, you can create a “paper writing” project, upload research materials, and tell Claude your writing style preference; then create a “code project” project, upload code files, set programming standards, etc.

This is like a magical tool! Its design concept is in line with Claude’s MDC (Model-defined Context), and is essentially a set of specifications optimized by reinforcement learning.

Everyone is trying to improve the quality of input. The core of the Projects feature is to optimize input. By providing clear and isolated context, it reduces the burden on users to explain the background and requirements to the model. Therefore, in actual business, we can also learn from similar thinking logic: distinguish different application scenarios. At the current stage when long memory technology is not yet perfect, this approach is very helpful in improving model performance.

Pursuing better output quality is not just about making the model smarter, but also about making it accurately “understand what you are trying to do.”

#AI/product

Focus and simplicity of expression

| 2025-05-26 20:51:31

You can pause at the right time and think for a moment.

However, try to avoid meaningless modal particles, such as repeated “then” and “then”.

This feeling is somewhat similar to the unpleasant word “whatever”.

Let the expression focus on the value of the content itself rather than indulging in redundant rhetoric.

#field

Get along with small animals

| 2025-05-26 09:21:11

Aotian is back.

I feel like I still lack experience in getting along with small animals.

Do we learn to cherish something only after we lose it?

I couldn’t bear to let it go, so I adopted it back. The moment I saw the cat, I couldn’t help but feel excited.

I did owe Aotian before. I should respect its nature more, be responsible for its life, and also be responsible for the relationship between me and it.

Observe your cat more and understand what it really likes, rather than imposing your own preferences on it. That is actually a subtle desire for control.

Just like treating an intimate relationship, learn more about it and understand the relationship to avoid unintentional harm and regret in the future.

#Life

Emotional amplification and self-observation after being slightly drunk

| 2025-05-26 09:41:56

When I was slightly drunk, some INFP traits appeared.

I had a few drinks last night, hoping that being a little tipsy would help me concentrate, get into the flow state, and create efficiently. But I didn’t expect that as the alcohol got stronger, I felt sleepy and fell asleep at 11:30.

I woke up naturally at five in the morning and felt extremely excited.

I felt that my emotions and senses were amplified invisibly last night and this morning. My perception of the surrounding environment, interpersonal interactions, and even nature has become particularly acute. The sunlight passed through the treetops, passed through the window lattice, and cast mottled light and shadows on the wall, mixed with swaying tree shadows.

I couldn’t sleep, but I thought it was a good idea to get up early. I got up, hungry, and ordered a takeaway breakfast. The soy milk in the morning was particularly fragrant. I wondered if my body needed the nourishment of soy milk, or if my taste buds became more sensitive when I was slightly drunk. However, the steamed buns that came with the meal were much less delicious than usual…

When I get up in the morning to work, I always feel something is wrong, an indescribable discomfort. My head is a little groggy and it’s hard to concentrate, but my body is clamoring to go back to bed.

So, I lay down again. I set the alarm for an hour, thinking of taking a nap. I told myself not to think too much and to fall asleep as soon as possible. But I just couldn’t fall asleep, and I still couldn’t tell what was wrong.

I tried to play Mayday’s music, but I was worried that the music would wake me up after I fell asleep. Oh, then I understood the source of the “wrongness” - it was that I was PUAing myself subconsciously.

“Cherish time, not falling asleep is a waste of time.” This thought seems particularly heavy under the magnified senses and emotions.

Metacognition kicked in and I observed myself again.

I canceled the alarm and let my thoughts run wild. If I feel like crying, I’ll just cry.

I thought of the design philosophy of Flomo: “Objectively record your emotions without criticizing or evaluating them.” I just described them truthfully, because at this moment, I just wanted to describe them and do what I wanted to do.

I suddenly realized that the root cause of emotional disorders is often a lack of understanding of ourselves. Why is it difficult to continue to love work under high pressure? The root cause may not be the high pressure itself, but our failure to face work in a healthy emotional state.

I want to cry. What was wrong with me before? Was it because I was confused about my emotions? It seems that I still don’t know myself well enough. Then, what’s the point of knowing myself? I still seem to be carried away by my emotions.

…I am so moved. I am observing myself, saying what I want to say, doing what I want to do, and loving who I want to love. I feel like I understand myself better. I know that what I really love is self-growth, the process of deepening my understanding of myself, rather than some external “meaning”.

Then, isn’t this high-pressure experience also an experience, a way to observe and understand oneself? Thinking about this, my mind became more open.

I wanted to go to the park to lie down and bask in the sun, but I found that the grass was covered with dew. But this did not bother me. I went home to get a sleeping mat and continued my “lying down” career. No compromise, no internal friction, haha.

At the very end, in the park, the aunt in front was practicing Ba Duan Jin, and on the basketball court on the left, another aunt was also sweating profusely. And I, in the park, was “lying flat” with peace of mind.

Supplemented on June 1, 2025:

While hiking, I suddenly thought of people with depression. Their emotions are often out of control. I recalled that I had been led by emotions before, and objectively recorded my feelings at that time. Although it was not a manifestation of depression, it was a rare opportunity to experience the state of having senses magnified and emotions pulled.

I can clearly feel that my emotional regulation and self-observation abilities are quite strong. From the initial confusion to the final self-acceptance, I am good at and willing to understand and observe myself.

#Emotional changes

Thoughts on self-media creation

| 2025-05-26 07:02:18

Subjectively, I still think that it is not too late to start working in self-media now. The real challenge is to find a track that is highly compatible with oneself. From my personal experience and hobbies, on the one hand, it is my professional field, and on the other hand, it is outdoor activities, such as hiking and mountaineering, and the experience of living in different cities and countries.

However, I have never been able to find a particularly suitable way of expression in the content output of my professional field. I often wonder if the direction itself is biased. Today I saw a video that inspired me: in fact, I can draw materials from many information sources on the Internet and share and discuss around these information sources. In the past, I may have felt that there was a lack of an ideal way of presentation, but now, it seems that I have vaguely captured some good ideas. This may also be regarded as a universal content creation strategy.

I am thinking about how to use this method to transform what I see and think in my daily life into valuable output. Because many contents will be gone if they are not recorded and shared. After all, people have some desire to share and create. At least, it is a good thing to have the satisfaction of creation.

Therefore, I may be more actively involved in this in the future. The current common solution is to share your thoughts and experiences in the professional field. For example, if you find some good methodologies or tools, you can record them and think about how to present them in a way that is suitable for short videos, and then further consider how to synchronize them to major social media platforms.

Frankly speaking, I do have limited time to invest in running social media accounts. But that doesn’t mean I can’t do anything. In the AI ​​era, technology has greatly empowered creators. I think I should gradually accumulate my first-mover advantage, which is also a kind of growth for me.

#field

Self-awareness in a slightly tipsy or intoxicated state

| 2025-05-25 23:40:52

Slightly tipsy

How would you describe the feeling at this moment?

I could still type, but I felt exhausted and sleepy.

Consciousness is sometimes clear, but it is difficult to maintain that clarity for a long time.

I vaguely know what kind of person I want to be, but I’m not sure if I can stick to it until the end. Wow, it turns out that even after getting drunk, I still recognize myself. That’s enough.

Aotian (my dog) was lost, and I was very sad. I would unconsciously analyze why I was so sad.

I feel like I wasn’t good enough with it before…

Aotian was adopted by his next-door neighbor, who already had four stray dogs. Well, this might be a better destination for him.

Maybe if I let go, it will be happier.

But, I still can’t help feeling sad…

“I don’t want to leave you alone, drifting in the sea of ​​people… Maybe in the future you will find someone who understands you and loves you better… I just hope you have a happy life.”

If I really love you, I wish I could be with you forever.

@Life

#Life

Some practical suggestions for engineering

2025-05-25 8:37:14 PM

Nowadays, there is less attention paid to code details, and more emphasis is placed on directory structure and design patterns; the details of the code are at most glanced at.

In essence, this is still a collaboration between humans and AI. I maintain two repositories, one for cursor and the other for windsurf. windsurf is mainly used for supplements and corrections after jules completes the PR.

Therefore, every night I would sort out the tasks, especially the important ones, make a list, prepare the design draft and verification standards, and hand them over to jules for implementation. The next day, I would make subsequent adjustments based on jules’s PR.

In cursor, the focus is more on the main tasks. If jules handles some tasks, it may be slow or lack of context; on the contrary, in the practice of cursor, a lot of engineering skills have been accumulated.

More and more time is spent on reviewing and proposing design requirements.

Even the review and design requirements themselves are often obtained through interaction with LLM Bot or LLM Chat.

#field

Lessons in tool selection

2025-05-24 11:51:52 PM

After all, people should not easily use tools that they are not familiar with and are not of much help.

This time I fell into a trap… It’s a lesson learned through blood and tears.

AI and Flow: Current Disruption and Future Expectations

2025-05-24 3:38:02 PM

I deeply feel that the current state of AI still makes it difficult to create an ideal flow experience.

I always felt like something was missing. After thinking about it, what was the problem? Ah, it turned out that my own working mode had not changed at all - it was nothing more than changing from “tightening the screws” by hand in the past to supervising AI to “tighten the screws” now, and I was still as busy as a robot.

Is this flow? It is just like a CPU, switching between different processes, sometimes directing this task, sometimes dispatching that task. Each task cycle is short and fleeting, and subsequent tracking and debugging are exhausting.

Thinking back to the state of flow, I personally prefer to continuously handle a single task for more than 20 minutes; if it is shorter than this, I will easily fall into a sense of being busy.

However, current AI is clearly not up to the task. The task is interrupted after a maximum of five minutes, which is really uncomfortable.

Here are some insights:

  1. The AI ​​scheduler of the future may become a bridge for people to interact with many AI agents. It should be noted that this is not a universal agent - the road to a true universal agent is long and arduous, and by then most people may not need to work.

  2. The task cycle and duration of LLM will gradually increase. In this process, people will continue to explore its boundaries and clarify the capabilities of AI.

  3. As AI continues to improve its working capabilities, the role of humans will also change: first, more time will be spent on thinking and design (such as setting tasks for AI agents, defining AI boundaries, planning projects and systems); second, more time will be spent on reading and learning (AI output content, search results, etc.).

  4. In the future, there may be no need to focus on code writing, but engineering thinking is still important. At that time, there may be no such thing as “code farmers”, only engineers and designers. The intermediate role of product manager may fade, and “AI product manager” will be more like a virtual concept, and its functions will be closer to engineering and design.

  5. The importance of fragmented memory will also become more prominent. When switching between different AI agents, one needs to be clear about one’s own goals and current context, which can be assisted by the aforementioned interactive agent platform.

#field

Observe the process of researching travel strategies

2025-05-24 1:05:25 PM

People who study travel strategies can be roughly divided into three groups:

  1. Go with the flow: This type of people let their heart lead them wherever it takes them. Perhaps a sudden thought of traveling can prompt them to pack their bags and set off immediately, and they may even buy their tickets just before leaving.

  2. Semi-planners: This type of people will make some preparations before traveling, such as surfing the Internet to learn about the destination’s culture, geography, and food.

  3. Planning control: This type of people will meticulously list a detailed list, from the time to get up in the morning to where to stay at night, every step is accurately calculated. They pursue a sense of control, and traveling is like performing a precise task for them.

I’m somewhere between the second and third categories.

I don’t demand that I strictly follow the plan list, but I will have a deep understanding of the destination in advance, and try to grasp as much information as possible about transportation, culture, humanities, etc. I am more like the kind of person who knows how to plan, but can let go at the right time, and find a delicate balance between being well prepared and acting spontaneously.

#Life

NotebookLM product design thinking

2025-05-24 12:34:48 PM

What attracts me most about NotebookLM is its mind mapping function. I think one of the main reasons why RAG (Retrieval Enhanced Generation) is often criticized as “not very useful” is that the density of information it extracts is still insufficient. If the information collection process is dominated by users - for example, quickly understand the outline of an article through the article summary, and then decide whether to collect it.

At this point, the collection and the original text are sent to the workspace as context. Based on this, RAG operations can generate many interesting inspirations for information workers.

The biggest surprise that NotebookLM gave me was indeed its audio transcription and mind mapping functions. Audio transcription is undoubtedly a powerful tool for quickly getting started and understanding the content; and mind mapping is particularly important, as it can help you clearly sort out various features. After clicking on a specific node, RAG will provide you with a detailed explanation based on the knowledge base.

#field

NotebookLM product usage thoughts

2025-05-24 10:44:55 AM

I was about to leave for Jeju Island, South Korea, but I didn’t have enough time to conduct in-depth research. At this time, I thought of NotebookLM. I used it as follows:

Since this trip is for tourism, the timeliness of information is particularly important. I first downloaded the official documents, related blogs and books of Jeju Island to my local computer.

Next, we used Grok to conduct an in-depth search on Twitter for information about Jeju Island in the past month and subsequent event arrangements.

For the Deep Research feature of Gemini, I entered my information, plans, and key drivers for the trip, and it helped me put together a complete planning document.

At the same time, I quickly browsed and screened some valuable posts on Xiaohongshu.

In NotebookLM, I used the Explore feature to search for relevant sources of information about hiking in Jeju Island.

Considering that hiking in Jeju Island is not very popular in China, I turned to Perplexity to search for detailed hiking guides in Jeju Island using English keywords and asked them to reply in Chinese.

All the above collected content is gathered in a new workspace in NotebookLM.

First, a conversational blog is quickly generated to help me quickly grasp the core information. The prompt is set to a wizard tone to make it more approachable.

Then, quickly generate a mind map, which will serve as the final structured output.

Based on the mind map, review the basic context, click and expand the nodes that are not very clear, and let NotebookLM call the RAG knowledge base for explanation, with an index. If you encounter content that interests you, you can ask for an in-depth explanation, or directly click the index to view the original text.

Some important travel-related questions, such as phone card application and passport application, can be asked directly to NotebookLM, which can basically give answers and locate the original source.

If there is a need to consult the original text, NotebookLM will generate a source guide (summary) and key topics for each original text.

The key topics here are not only used for classification, I also often use the Discuss theme function to quickly understand a certain type of topic or noun. If you are interested, you can even let it further explore the corresponding topic and organize important information documents from websites such as Wikipedia.

It is hard to imagine that in the past, organizing documents and absorbing information might take a whole day; now, with the help of such tools, information collection and structured organization can be completed in less than an hour.

#field

The issue of data sovereignty

2025-05-24 10:26:33 AM

Data sovereignty is indeed an issue worthy of deep thought.

I looked up some information about NotebookLM data storage.

First, user-level privacy protection is crucial, and ensuring content privacy is its core consideration.

The essence of NotebookLM is to prioritize documents uploaded by users rather than endlessly searching the web.

Depending on the user’s geographic location and Google’s infrastructure layout, data may be stored in data centers in a specific area. Data localization and sovereignty issues are becoming increasingly prominent, which is one of the reasons why some services cannot be accessed smoothly in China or certain countries and regions.

#field

The nature of game theory and self-strengthening

2025-05-23 4:06:30 PM

The essence of the game comes from the scarcity of resources. The essence of the game between Party A and Party B depends on who has more confidence to “turn the table over”. If Party A has more such ability, then Party A will have greater say in the game.

From this perspective, deliberately trying to please others actually has little meaning.

What is truly valuable is to enhance one’s own voice, increase one’s own scarcity, and make oneself irreplaceable, irreplicable, and not easily bypassed. At the same time, the career one is engaged in should also pursue this state of being irreplaceable and irreplaceable.

Ultimately, the core is to make yourself stronger.

The era of short text and AI empowerment

2025-05-23 2:32:00 PM

Some people say that the new generation is deeply “poisoned” by short videos.

They are addicted to fragmented video consumption and are unwilling to read, which leads to a lack of structural ability.

However, human structuring ability is ultimately limited. Even writing papers or books is a process of constantly reshaping, repeatedly pondering, and adjusting the structure.

Those who are least adept at systematic expression are those whom AI can best empower.

In the AI ​​era, writing ability itself may no longer be the most important thing; instead, what replaces it is the desire to express and the experience of life.

In the AI ​​era, it may no longer be necessary to read long articles. In the past, you might choose to “read it later” because you were afraid of its length, but the result was often “never read it”. Today, AI can help you extract core ideas, generate summaries, list key arguments, and even transform complex concepts into easy-to-understand explanations. You can grasp the essence of the article in a few minutes, and if you still have time, you can spend more time studying the original text.

The speed of information collection is experiencing an explosive growth.

However, literacy still seems to be a “privilege” for a few people. We have witnessed countless vivid stories and profound insights being lost due to a lack of expression ability.

The soul of a good work is not only due to the exquisite writing method, but also lies in the soul it carries - the unique life experience.

Some fragments of words keep emerging as my experience grows, and occasionally come to my mind like a flash of inspiration.

#field

The future of LLM and the transformation of engineers

2025-05-23 11:56:28 AM

Claude can now focus on a task for up to six hours, while Jules can complete an engineering task independently.

It is expected that within a year, AI may reach user level and be able to independently perform tasks within one hour.

This means that within three years, users may be able to basically shield themselves from code details at the business level.

Programmers may collectively transform into engineers in the future.

How to schedule each LLM and Agent to perform certain operations, how to set engineering rules, and how to accept the results will become new core issues.

Then: People (intention, code) ⇒ Execution

Today: Human (intention) ⇒ AI (code) ⇒ execution

Humans “package” the underlying complexity and “sink” it one level, allowing themselves to think at a higher level.

Once upon a time, mastering “How” (how to achieve something) was the core competitiveness; now, mastering “What” (clarifying needs and judging the pros and cons of results) has become the key ability.

Competitiveness has undergone a profound shift.

#field

Thinking about pension insurance

2025-05-23 11:32:44 AM

According to the latest data from the Ministry of Human Resources and Social Security, as of the end of March, the number of people covered by basic pension, unemployment, and work-related injury insurance nationwide was 1.071 billion, 244 million, and 297 million, respectively, up 4.42 million, 2.56 million, and 2.71 million year-on-year, respectively. As a new type of occupational injury insurance, the pilot areas continue to implement it steadily. As of the end of March, more than 11.04 million platform employees such as food delivery riders and online car-hailing drivers have participated in the insurance. (Xinhua News Agency)

A few observations:

  1. The total population of China is over 1.4 billion, and 1.07 billion people are covered by basic pension insurance, which means that this type of insurance has reached its “ceiling” and the population that can be covered has basically been included.

  2. Food delivery drivers and other flexible employment workers have also begun to pay insurance, indicating that the country is gradually building a system to provide a “safety net” for the flexible employment groups that have been relatively weak in protection in the past.

  3. The aging of the population, the increase in the number of retirees and the decline in the dependency ratio mean that the pressure to balance pension income and expenditure will be more prominent in the next 10 to 20 years.

  4. As the system becomes more and more perfect, the space for small and micro enterprises and freelancers to “evade” social security will become smaller and smaller.

#predict

The value of Group push

2025-05-22 5:23:12 PM

Thoughts on LLM push: This is similar to the recommendation algorithm, and the pushed content may come from questions that other users have asked or collected.

Questions themselves contain value. Therefore, for those questions that are collected by users, they must be content that is particularly worth recommending by the recommendation algorithm.

Subscribe to information collection on a specific issue

2025-05-21 9:30:45 PM

Core needs: For example, I am studying the development of a certain direction (such as LLM Long Memory), or I want to subscribe to the latest progress in a certain field, including related papers, literature, blogs, etc.

Or, I am about to travel and want to get all the relevant information about the destination in real time, such as weather forecast, latest local posts, etc.

This function can be provided externally as an Agent capability and can even be developed into an open source Agent service.

The workflow is roughly as follows:

Subscription trigger (timing/event) → Query builder → Multi-source information crawler → Content parser → Importance filter → Structured summary (optional clustering) → Rendering and notification (or saving to Notebook) → User feedback learning mechanism.

#field

The beauty of minimalism in products

2025-05-21 8:34:48 PM

Recently I have become less and less fond of complex colors. I feel that colors only add to the complexity. Even pictures should be avoided from being too large.

Starting with black and white as the base color scheme, it is really important to start defining the rules using cursor rules.

Information Processing and Collection

2025-05-21 8:04:20 PM

The importance of rapid information processing is self-evident. For information practitioners, the speed of information processing is even more critical.

For me, I open a lot of tabs every day, covering various web pages, articles, tutorials, etc.

For an efficient information processing tool, the density of its information is also particularly important - just the right amount of abstraction, supplemented by interesting analysis, is the best.

#Field/Tool Usage Tips

Travel Preference: Intuitive vs. Feeling

2025-05-20 8:25:15 PM

When you plan a trip, what do you focus on?

A. Imagine interesting experiences that might occur during your trip and explore new ideas or cultures (e.g., look forward to local art, history, or unique experiences).

B. Focus on specific travel arrangements and practical details (e.g., booking hotels, arranging transportation, packing essentials).

If you choose A, you are more of an intuitive type; if you choose B, you are more of a sensing type.

The Law of Ten Times Better Products

2025-05-20 3:41:48 PM

We’ve all heard the saying: To build something valuable, you have to be 10x better than what’s already out there. It seems to be a universally accepted law of entrepreneurship.

This rule is so strong that even Google+, with billions of dollars in funding, cannot become a competitor to Facebook, even though Google+ may be better than Facebook in some features.

Interview with the founder of Lovart who has no product manager

2025-05-20 11:51:50 AM

The value of general product managers in the future may decline. I myself have more than ten years of product experience and have experienced the entire cycle of mobile Internet. I believe that the core value of future product managers lies more in the depth of industry knowledge and whether they can systematically impart the know-how of a specific industry to AI. Therefore, the role of product managers will be more segmented and tend to be industry-oriented and vertical. In my opinion, a general Internet product manager is a risky profession.

The capabilities of general-purpose agents are likely to be internalized by basic models in the future. I believe that the agents that will truly flourish in the future are those that focus on vertical applications. The core competitiveness of these vertical agents lies in the vertical domain expertise behind them. For the general Internet product managers of the previous era, I do think their positioning is a bit vague.

Once the general capabilities are absorbed by the underlying model layer, what remains is professional design capabilities.

When tools are smart enough, “people who manage requirements” may no longer be needed, but “people who define requirements” will become more and more important.

AI has indeed greatly liberated productivity, and it may also make the thinking patterns of top designers more popular and more equal.

Cursor has a data advantage over GitHub. GitHub’s data is a large amount of internalized data; while Cursor records how users use AI tools and how to combine AI with traditional methods to create, which is extremely valuable. The workflow of ComfyUI is essentially a stacking of various models and traditional functional modules. These are exactly what AI needs to learn.

#AI/product

Scott Scheper’s Card Reference Method

2025-05-20 9:58:52 AM

The purpose is to systematically integrate the reading process into the personal knowledge network. He believes that reference cards are not simple document recording tools, but the initial link of knowledge “access system”, and are “gateway cards” that organically introduce external ideas into the internal thinking universe.

It contains three core elements:

  1. Information: author name, book title, subtitle (if any), year of publication, publisher, page numbers, etc.

  2. Objective of the reading source: that is, the core issue that the reading material aims to explain or solve.

  3. Source Overview: A brief summary of the main points of the reading material.

Thinking about reference cards in the Internet age

In the digital age, we still hope to retain these core features of reference cards and consider adding some optional fields to adapt to new information carriers and reading habits.

#Field/Tool Usage Tips

FOLO efficient information entry methods

2025-05-20 9:40:27 AM

Information collection has always been a headache for me. Part of my aversion to mobile phones is that the recommendation algorithms on mobile phones often push out filtered information. Folo is different. It provides a way to actively subscribe to information.

There are three ways to obtain information:

  1. Recommended information: Passively receive content pushed by algorithms.

  2. Subscribe to information: Actively choose to follow specific information sources (initial selection may also rely on recommendations).

  3. Deeply relevant information: In-depth content obtained through active searches on specific topics or through recommendations.

The way to collect information varies greatly from person to person. The key is to redefine and organize the information sources according to your own needs.

When considering information acquisition tools, in addition to supporting information subscriptions, you also need to think about how to effectively integrate them with in-depth tools (such as Folo) that you use every day.

#Field/Tool Usage Tips

Epiphany of Beauty

2025-05-19 11:20:36 PM

I have been wondering whether there is a universal shortcut to “enlightenment” in various fields of aesthetics, such as the beauty of architecture, art, nature, and even the beauty of appearance?

It is true that there are some commonalities in aesthetics. As discussed earlier, human perceptual mechanisms have their commonalities, such as a preference for symmetry and proportion, and a sense of harmony in color combinations.

Therefore, there are some basic aesthetic principles, such as composition (such as the rule of thirds), color theory, contrast and balance, etc. These principles are applied in photography, art creation and even UI design.

You might as well ask yourself a few more questions: What exactly was it about those designs that touched you deeply?

A large amount of aesthetic input and one’s own creative practice is still a process of gradual accumulation.

#Aesthetics

Understanding of beauty, essence

| 2025-05-19 11:15 pm

Beauty has a universal basis, such as people’s natural affinity for melody and symmetry, and also reflects the common perception mechanism of human beings, such as the common appreciation of the golden ratio in Eastern and Western cultures. However, beauty is more fluid, deeply shaped by culture, history and individual experience. Western aesthetics prefers form and structure, while Eastern aesthetics prefers artistic conception and nature. So, what is the essence of beauty? Is it subjective feeling or objective standard? Perhaps, it is a “differentiated presentation” with various forms under a set of universal criteria. Does beauty only stay at the sensory level? Not necessarily. The “cultural resonance” triggered by the implication also touches people’s hearts - beauty is not only about the eyes and ears, but also affects the deep thinking of the heart and brain.

#Aesthetics

Reflections on the development of Japanese aesthetics

| 2025-05-19 10:51 PM

The roots of Japanese aesthetics are deeply rooted in a closeness to nature, a belief that everything has spirituality. This sensitivity to “natural beauty” formed the basis of its early aesthetics. Shintoism regards excellence in craftsmanship and artistic refinement as a devotion to the gods, which gave rise to a commitment to “refining” craftsmanship and aesthetics. Zen’s “simplicity”, “emptiness”, “meditation” and “mindfulness”, as well as the direct experience of reality, have profoundly influenced Japan’s artistic practice. The Buddhist teaching of “impermanence” (mujō) - everything changes and nothing is permanent - and the acceptance and appreciation of “transience” in concepts such as “wabi-sabi” and “mono no aware” have together shaped its unique aesthetic style.

#Aesthetics

Art is real

| 2025-05-19 9:57 pm

“He was talking when his head dropped to his knees and he died.” I used to think that such death only existed in the exaggerated brushstrokes of art. However, Dong Jianyi’s death convinced me that art often reflects the “truth”. Movie plots come from life, and so does art. The absurdity and drama in life are the inexhaustible source of artistic creation.

“Jiabiangou”

#read

It’s not the appearance that’s beautiful, it’s the feeling

| 2025-05-18 2:34 PM

After visiting many exquisite cafes, Artisan Cafe impressed me not because of its exquisite decoration or film-making effect, or even because it is difficult to capture the “beauty” under the usual lens. However, it is this unpretentiousness that creates an indescribable “feeling”. Stepping into it, it seems as if you are invited into the slow pace of life in the Renaissance. This reminds me of the spiritual core shared by Japanese wabi-sabi and European old film aesthetics - not to please, but to “resonate”. This is an aesthetic that transcends vision, a discovery of the “mind” behind the lens, “feeling” rather than just “seeing”. Some spaces are full of stories without words, and this is the charm of art - derived from life, but also higher than life. The imprint of time, the beauty of mono no aware, flows between old wood, cracked pottery, and mottled paint, forming an imperfect, asymmetrical but simple harmony. Having been to many specialty art cafes, whether they are abandoned factories, old houses or manors, they all seem to confirm that the real cultural atmosphere is not piled up, but “grown” from the daily bones and blood. The touch at this moment is real and profound.

#Life

RAG and Long Context Thinking Analysis

| 2025-05-18 2:18 PM

Long context processing is undoubtedly a challenge that large models will face now and in the future. The bottleneck lies mainly in the computational complexity of the attention mechanism. Even if the context window is large enough, it is still a big question whether the model can truly remember and effectively use all the information. Especially when performing complex logical reasoning or information integration across long distances, the model’s “accuracy” needs to be improved urgently.

In contrast, the Retrieval Enhanced Generation (RAG) model shows its advantages. The efficiency of RAG is divided into two parts: retrieval speed and generation speed. More importantly, RAG provides good traceability, which makes it a more popular trend in the short and medium term.

In the long run, the boundary between Long Context and RAG may gradually blur, and the two may merge and complement each other to jointly serve the goal of faster and better information processing and generation.

#field

Thoughts on the Marketing of Elmo Chat

| 2025-05-18 1:50 PM

Elmo Chat has a similar concept to DeepWiki - no login is required to use it, which greatly reduces the user’s usage threshold. In terms of cost management, it is a wise move to attract traffic through a large number of simple model examples, and even learn from the open source business promotion model. The fundamental reason why Elmo Chat is free is that it avoids expensive large models and instead adopts low-cost open source models, combined with local or lightweight inference architectures, and supplemented by growth-oriented business strategies. In addition, the transmission of the brand’s aesthetic culture - lightweight, speed, and simplicity - is also an indispensable element of its success.

#field

Emotion conversion mechanism

| 2025-05-17 5:56 PM

Previously, I talked about the need to examine China’s development from an intergenerational perspective, because the growth backgrounds of each generation are very different. There is a type of person who, due to the distrust and insecurity caused by their early environment, develops a strong desire for control and tries to transfer their inner anxiety to those around them. Behind this is the essence of “emotional conversion”.

They try to gain a so-called sense of security by controlling others, but this sense of security is illusory because it is built on the unfreedom of others. The essence of emotional conversion is to project negative emotions such as inner anxiety and fear outward, hoping to alleviate one’s own discomfort by changing others. This pattern is quite common in human nature, and we have all been like this to some extent - losing temper with family members when anxious, or holding on tighter for fear of losing. The key is whether we can “perceive” this, which is almost a kind of metacognition, like opening a God’s perspective.

Deep in human nature, people pursue three things: security, belonging and self-realization. Ironically, excessive pursuit of security brought by external control often limits oneself, thus hindering the path of self-realization. True growth is nurtured in the soil of freedom and trust; but he put shackles on himself and others at the same time.

Human nature is so complex and fragile. We are all afraid of being hurt, but we often hold on tighter because of fear, which results in more hurt and eventually puts ourselves and others in trouble. The real balance lies in learning to accept our own imperfections and daring to believe that the world is not that dangerous, even if we fall occasionally.

#Life

The rise of Japanese anime

| 2025-05-16 9:37 pm

When it comes to Japanese animation, it is not difficult to find that it has become a kind of continuation of the psychological gap of modern people. The reason is that, on the one hand, the economic material downturn often gives rise to the need to supplement spiritual energy, and culture becomes an excellent container. On the other hand, the fictional world fills the spiritual coordinates and emotional needs lost in reality. People are eager for emotional value, but are unwilling to pay too much real price for it. Animation just provides such an outlet.

Behind this, it can be interpreted as a collective psychological compensation mechanism - escape and repair of real frustrations; an alternative spiritual life system - using the virtual world to carry the meaning; and an aesthetic expression of “post-industrial society”, such as cyberpunk, the concept of the metaverse, and cute culture.

Cultural factors are also key. The cultural background of Confucianism, Buddhism and Taoism and secular aesthetics allow “fiction” to carry “real” emotions and philosophy. Its perception of beauty is not a grand and vigorous one, but a preference for “light sadness” and “fleeting beauty”. The delicate happiness brought by the tiny beauties in life is a psychological medicine under the heavy reality.

#Life

Fear of Granular Amplification

| 2025-05-16 9:35 pm

When you first meet someone, it’s like looking at a painting from a distance, with a hint of beauty and simplicity in the haze. Once you get closer, your perspective widens, and the whole world suddenly changes. It’s not that the other person has changed, but that the “granularity” we see has become finer - those details, minor flaws, and even the complex emotions hidden deep inside, emerge one by one. Is this a blessing or a test?

This is like adjusting the focus of a camera. When you are far away, the picture is blurry but comfortable; when you are close, it is clear but also messy. You will find that others are not what you initially imagined. They have many sides, some of which even they themselves have not fully understood. When chatting with a friend, her casual complaint may instantly make you realize that under her usual cheerful appearance, she actually carries a lot. The shock at that moment comes from the redefinition of “understanding”. In a relationship, once the “granularity” changes, we need to adjust ourselves accordingly, otherwise it will be difficult to synchronize.

From a philosophical perspective, this is like choosing the resolution of the world. Looking from a distance, it is simple and life is relatively easy; getting closer, the truth gradually becomes clear, and thoughts become complicated. We can’t help but ask ourselves: How much do we want to see clearly? After seeing these details, can we still get along as before? Because once it is magnified, it is difficult to return to the starting point. The only way is to accept that they are not perfect, I am not perfect either, and all living beings are “a mess”, and then try to find harmony in this chaos.

Just like listening to a song, when the volume is low, you only feel the pleasant melody; when you turn it up, you can notice the noise, harmony, and even the tiny flaws that were not fully processed during the recording. The same is true for relationships. When the distance is shortened, you can hear not only the sweet words on the surface, but also the undercurrents that have never been said, and even they themselves have not noticed. This is awe-inspiring, but also full of interest. Because at this moment, you are truly dealing with a living life, not an abstract concept.

Intimacy is full of contradictions. The closer you get, the easier it is for a simple relationship to become complicated; if you don’t get close, it will always remain on the surface. A true relationship may be the courage to go deep, even if you know that the water beneath the deep water may not be as clear as it seems.

#Life

The essence behind the relationship is value exchange

| 2025-05-14 7:23 PM

The core of interpersonal relationships is often inseparable from the “value exchange” - whether it is material support, emotional comfort, or spiritual resonance. The reason why we reflect on and emphasize moral sense is partly due to the insight and balance of the inherent “self-interest” tendency in human nature.

Therefore, it may be wiser to maintain clarity and thoroughness in relationships and distinguish between “interests” and “feelings”. We do not deny the greatness, selflessness and nobility of many emotions, but if we regard these sacred emotions as the whole truth of human nature, we may drift away from the truth.

#Life

Buffett’s last battle

| 2025-05-12 7:46 PM

Opportunities do not appear in a planned way. Maintaining sufficient cash reserves is to quickly seize major opportunities when the market presents them. Buffett once vividly compared it to death: “It’s a bit like death - the probability of dying the next day at the age of 10 is extremely low, but living to 100 is almost certain.”

The Japanese market still holds many opportunities. Its investment and philosophy can be summarized as follows:

  1. Successor: Abel is fully prepared to take over as Berkshire CEO; he will recommend the board of directors to appoint him as CEO before the end of the year.

  2. Tariff issue: The United States should actively seek trade with countries around the world, and trade should not be used as a weapon.

  3. U.S. Finances: The U.S. fiscal situation is worrying, but this is not a problem unique to the United States.

  4. US stock market volatility: Compared with the major crashes in history, the recent market decline is nothing but a drop in the bucket.

  5. Artificial Intelligence: Not all investments and developments will revolve around AI; any choices in the field of AI should be made by Ajit Jain (Vice Chairman of Berkshire).

  6. Japanese Market: Will not sell any Japanese stocks, either now or in the next 10 years.

  7. Cash Reserves: “I made a lot of money because I never invested all my money.”

  8. When to act: The chances of a good opportunity appearing tomorrow are slim, but the probability of it occurring within five years is not low.

  9. Advice for young people: Surround yourself with people who are better than you; do what you love and don’t worry too much about initial salary.

  10. Appreciate Cook: Admit that Apple CEO Cook has created more value for Berkshire than he himself.

#predict

Reality

| 2025-05-12 11:47 AM

The desire for “authenticity” is growing day by day, but I find it difficult to put its exact meaning into words. Is it telling the truth? Is it the candor of “I am just like this”?

The sense of reality that I understand is more about clearly perceiving and accepting one’s current feelings, motivations, limitations, and desires, without denying or deliberately beautifying them. It includes three levels:

  1. Awareness - What am I doing? Why? Is this really what I want?

  2. Honest - Don’t hide or pretend to be some “ideal person”.

  3. Alignment – Align your language, choices, and inner state as closely as possible.

The key is to break away from the “histrionic personality”, make choices that are truly your own, and create things that truly come from your heart. Get out of the obsession of “how I want others to see me”, and gradually turn to the inner pursuit of “how I want to be in my own heart”.

#Life

Self-consistent without relying on any identity

| 2025-05-12 11:28 AM

It is understandable that society needs labels for identification and classification. But the problem is that once we begin to rely on these labels, we unconsciously live in the expectations of others. Identity labels bring a false sense of certainty, but once these externally given labels are shaken, we tend to panic: “Without these, what do I have left?”

The real you is not the title on your business card, but the “self” that can still exist calmly after losing all the labels. Because of the fear of leaving the job, the fear of social definition, and the fear of other people’s eyes, we may lose the courage to act.

Ask yourself: Would you rather be seen as the whole you, or just as your label? If you lose all your titles, income, and connections today, who would you be? For me, I am still willing to continue to create, explore, and enjoy everything the world has given me.

#Life

After a child dies, he or she will grow up in the hearts of his or her parents

| 2025-05-11 9:34 PM

Losing a child is a trauma that is often difficult for parents to heal. They burst into tears when they think of their child, and when they see the child in their dreams, they feel torn apart again after waking up, accompanied by endless guilt, self-blame, and powerlessness, which repeats over and over again.

The loss of a child is painful and the parents shed tears, but life must go on. Although the child’s growth is ruthlessly terminated in reality, parents will continue to imagine, miss, and even “construct” the life trajectory that the child has not experienced. They will imagine: “If he/she were still alive, how old would he/she be now? What would he/she be doing?” This is both a continuation of emotion and a unique way of mourning.

The year I was born, I thought of a relative of mine…

#Life

Nothing is real, only things we think are real

| 2025-05-11 12:01 p.m.

Everything we experience is essentially just our interpretation of the world. The world itself has no fixed form. We construct its appearance through our perception, and our subjective consciousness is the most real “reality”.

From this, we can understand that pain often does not come from the things themselves, but from our “judgment” of things. The more we resist something, the more it persists; the more we want to control the situation, the easier it is to be controlled by it.

True freedom is not to do whatever you want, but to have the power to not be driven by your inner impulses. Metacognition can help us identify these impulses and have the ability to balance them. This is the true meaning of freedom. Freedom is not “following your feelings”, but “knowing that you have feelings, but you can still decide how to go.”

#Life

War and Suffering

| 2025-05-11 11:39 AM

Politics is ruthless. Those who forget the history of suffering are bound to repeat the same mistakes and become innocent victims of the next suffering.

Those who have never experienced crying bitterly in the middle of the night are not qualified to talk about life.

And love, perhaps, is to see and practice your own dedication to the deepest needs of others.

(Reference: Chai Jing’s interview with survivors of the Chinese Civil War)

#Life

Hometown memories

| 2025-05-11 10:50 AM

Perhaps one day, we will return to the hometown we were eager to escape from when we were young. At that time, every bit of memory will appear clearly before our eyes, but we will be shocked to find that we can no longer truly reenter that past.

If I had known that the spring of that year would be the entire spring of my life, I would have been braver and lived my life to the fullest.

What we love, what we hate, what we long to return to, what we want to leave - in the final analysis, it is the “inner self” that needs to be escaped or pursued. When we are young, our hearts are far away; when we are old, our souls are drawn to the way home. To reconcile with the past is to reconcile with ourselves, and to accept all the comings and goings in life.

I once heard a saying: If we are destined to be apart, what is the meaning of meeting? The meaning of meeting is that the “part of me” that has been changed by you will always stay by my side instead of you who have gone away. It turns out that I once again felt the true meaning of “Ichigo ichie” and “mono no aware” - separation is not a break, but another continuation of “internalization in the heart”.

#Life

Breaking the traditional social concept

| 2025-05-11 10:43 AM

Traditional social concepts often disguise themselves as “natural”, replicating their own destiny through generations of “should”, “must”, and “you will understand later”, and cleverly projecting their own unfulfilled expectations onto the next generation. This is similar to the sense of destiny in some cultures, such as the expectation of the next generation to inherit their profession by the female divers.

The parents’ generation was unable to break free from the shackles of fate. In order to give meaning to the “price” they had paid, they would subconsciously ask their children to repeat a similar trajectory:

  • “That’s how hard it was for me back then, you have to endure it too” → Rationalizes repression.

  • “Girls should be obedient and take care of the family” → Internalize gender roles as behavioral norms.

  • “Only by becoming a civil servant can you have stability” → The general anxiety about survival in society is packaged as concern ‘for your own good’.

  • “There is no marriage without grievances” → View patience as a sign of maturity.

  • “I didn’t raise you to be whatever you want.” → See children as a means to continue your own life rather than as independent individuals.

#Life

Intergenerational conflict

| 2025-05-11 10:33 AM

Each generation has its own worries and missions. However, if you look deeper, each generation is essentially the struggle of “the same kind of people” in different eras.

Each generation challenges the “order” they live in in their own way. Our parents bravely broke the feudal ideas of the previous generation, while our generation is more persistent in the pursuit of freedom and independence. It seems that we cannot understand the persistence of the previous generation, but in fact we are repeating similar struggles - only the script and the stage are different. Therefore, to understand the differences between generations, we should not only look at what they did, but also explore why they did it.

There seems to be little point in criticizing the past. The reason why I may think that I am “smarter” than my predecessors may be simply because I am closer to the time node of “today”.

Every generation continues to reconstruct the definition of what happiness is and what success is.

Before I passed away, my grandfather warned me that I should be careful in what I do, and that I should have food to eat, people to rely on, and live with dignity - this was his definition of happiness.

My parents experienced the wave of reform and opening up, which was an era of rapid economic development. For them, happiness is to stand out, have a house and a car, and have successful children.

My understanding of happiness is that in this age of excess choices, I can “live the way I like”, live the life I want, and cherish every moment.

The connotation of happiness is evolving: from “being alive” -> to “being successful” -> to “being real”.

#Life

Heritage

| 2025-05-11 10:13 AM

The wheel of time rolls forward, technology advances with each passing day, and skyscrapers rise from the ground. All of this seems to create the illusion that the world has become brand new.

What is the essence of change? It is the rapid iteration of technology. Values ​​also change with it. Each generation will rise in the course of history and create a value system that is more in line with the new era:

  • My grandfather’s generation valued “collectivism,” “patience,” and “sacrifice.”

  • Our parents’ generation advocated “hard work” and “making a name for oneself”.

  • Our generation is more concerned with “individual”, “freedom” and “meaning”.

The worldview of each generation is undergoing rapid reshaping. However, we seem to be repeating a similar cycle - longing to break away from the influence of our original family, just as our parents struggled when they were young. They were forced by life, and we are trapped by spirit.

In a sense, nothing seems to have changed.

Our desire for love is the most basic need of human nature.

The pursuit of a sense of belonging has only changed its form in different eras.

When it comes to the question of “meaning”, every generation seeks answers in its own way.

Each generation has its own troubles. People living in a particular era often find it difficult to fully understand the ideas of the previous generation, but the qualities they show are surprisingly similar: the indomitable courage to live, the examination and breaking of traditional concepts, the persistent pursuit of the life they desire, and the unremitting exploration of happiness. Human nature is generally like this. We always yearn for what we don’t have, and it is not easy to cherish what we already have.

#Life

People who have experienced pain can better understand how difficult it is to be happy

| 2025-05-10 9:31 PM

People who have experienced hardships often have a deeper understanding of how hard it is to have happiness, and therefore they know how to cherish everything they have. This is actually related to the inner satisfaction.

Many times, we are already in happiness and satisfied, but we just don’t realize it. At least, every day now, we are trying to live the life we ​​want and do the things we want to do.

That’s good.

#Life

Smart hiding

| 2025-05-10 2:49 PM

The core of product design is simplicity, and core functions should be concise and clear. We need to constantly reflect on how to make functions simpler rather than increasingly complex. The simplest way is to achieve simplicity, just like ChatGPT, a seemingly simple conversation interface that can complete more than 80% of the tasks.

Users should not be burdened by too much learning cost. If a function requires additional learning to master, then its cognitive burden is too heavy. Design should pursue “intelligent hiding” to make complexity invisible and allow users to use it smoothly by intuition.

#Field/AI Product

The future of programming

| 2025-05-09 2:30 PM

Evolution of asynchronous interaction: AI’s autonomous working time will continue to expand. It is expected that by 2027, AI may be able to independently complete tasks of up to 1 hour; by 2030, it may be able to independently handle a day’s workload.

This means that AI is increasingly becoming a qualified “engineer”, and its engineering capabilities will continue to increase, gradually taking on more complex and longer-lasting programming tasks.

#predict

Shift the Scaling Law curve to the left, lower training cost

| 2025-05-09 2:26 PM

Infra Frontier: Shifting the Scaling Law Curve to the Left to Pursue Lower Training Costs

“Shifting the Scaling Law curve to the left” means achieving the same or even better model performance with lower parameter count, computing power input or cost control.

This bodes well for the future:

  1. Reduced inference costs, innovation-driven: As model efficiency improves, inference costs will drop significantly, opening the door to broader AI applications and innovations.

  2. After the efficiency is improved, the total demand may increase: Although the cost of a single training or inference is reduced, the improvement of AI capabilities and the expansion of application scenarios may cause the overall demand for computing power to increase instead of decrease, and the computing power requirements may be higher.

Therefore, for computing infrastructure companies, competition remains fierce, and optimization and innovation remain the core.

#predict

WeChat MCP Thoughts

| 2025-05-09 11:30 AM

It is crucial to obtain book information and details, especially to be able to extract the “note content” in the book. If MCP (Model Context Protocol) can easily process book notes, it means that AI can efficiently assist in generating book summaries or extracting key points. In this process, the role of humans is more focused on “screening and confirming information.”

Integrating this filtered information into a structured knowledge framework or knowledge map depends more on the capabilities of AI itself. If WeChat Reading can open up such capabilities through MCP, it will greatly improve the efficiency of knowledge acquisition and internalization.

#AI/product

Thoughts on the evolution of reform and opening up

| 2025-05-08 9:20 pm

Reform and opening up have undoubtedly brought about economic miracles, but this has been accompanied by dramatic changes in people’s ideas and concepts, with thinking patterns no longer pursuing simple uniformity.

Reform and opening up led to economic take-off, but there is a fundamental tension between economic liberalization and political control. The pace of political system reform was relatively limited and basically stagnated after 1989, during which the argument that “stability is a prerequisite for economic progress” was repeatedly emphasized.

From the ideological fanaticism and repressive socioeconomic environment of the Cultural Revolution to the general emphasis on materialism and rapid social changes in the era of reform and opening up, coupled with the rapid development of science and technology - this series of changes in China can be regarded as one of the most drastic and profound social transformations in the world.

#culture

Human nature before and after reform and opening up

| 2025-05-08 6:55 PM

After reading works such as “Gannan Chronicles” and “100 People’s 10 Years”, one often wonders: Has human nature really changed? Why do some of the “evil” of human nature seem to be understandable after the reform and opening up, while before that, the same “evil” was difficult for us to tolerate?

If you think about it, the essence of human nature may not have changed. Regardless of the era, the basic color of human nature seems to include:

  • Desire for a better life (selfish tendency to seek benefits and avoid harm)

  • Fear of losing existing resources and status (instrumental rationality stemming from the ‘survival instinct’)

  • Pursuing fairness, but unconsciously preferring privilege (internal contradiction)

  • In the face of temptation, moral bottom line often appears fragile (the “fragile” side of human nature)

In extreme living environments, the “evil” of human nature is more likely to be magnified, and what drives behavior at that time is more naked survival instinct. After the material is relatively abundant, the manifestation of these “evil” may be more hidden, and the motivation is more complicated, but its core is still inextricably linked to the performance under the pressure of survival.

#Life

Is Yujing mapping done through the attention mechanism?

| 2025-05-08 4:42 PM

Leaving aside the RAG (retrieval-augmented generation) path, LingoWhale-8B defines the model’s attention layer in its modeling_lingowhale.py file. In theory, the system can extract these attention weights during the summary generation process, thereby achieving an association mapping between the summary content and specific parts of the original text.

However, this ability to achieve mapping based on the model’s internal attention weights seems to be able to achieve similar or even better results at the application level in other ways for the RAG architecture. For example, through precise text segmentation, efficient vector retrieval, and clever organization of retrieval results, RAG can also establish a clear correspondence between the summary and the original paragraph.

This makes us wonder, at the application level, can we achieve high-quality summary-to-text mapping through sophisticated business logic design without relying deeply on the details of the attention mechanism inside the model? This may provide a feasible path for application developers who do not have the ability to modify the underlying model.

#AI/open-source

In-depth analysis of Yujing

| 2025-05-08 3:38 PM

The core of the basic model service for the implementation of summary mapping functions such as Yujing is to efficiently process and summarize text.

First of all, text preprocessing and chunking is crucial. The original text needs to be cleaned and formatted. When the text length exceeds the upper limit of the model processing, an intelligent chunking strategy must be adopted, which not only considers the length of a single chunk, but also ensures the coherence of the context and supports cross-chunk mapping capabilities to ensure the completeness and accuracy of the summary.

Secondly, the basic model service generates summaries based on the text type and length. The key here is the mapping mechanism between the original text and the summary. One possible implementation is through the attention weight mechanism in the Transformer architecture. Summaries are usually divided into two types: extractive and abstract. The extractive summary directly uses the original sentence, and the mapping is relatively direct. The abstract summary will reorganize and express the content. At this time, if you want to achieve accurate mapping, the model needs to support and output attention weights during pre-training. These weights can reveal which parts of the original text the model focuses on when generating each word or phrase in the summary.

For example, many pre-trained models in the Hugging Face Transformers library (such as BERTSUM, T5, and BART) are based on the Transformer architecture and have built-in attention mechanisms. By setting output_attentions=True when calling, developers can access these weight data, thus providing a technical basis for achieving accurate mapping between summaries and original texts.

#AI/open-source

Product Design

| 2025-05-07 3:32 PM

In the early stages of product design, it is important to always maintain a user-centric perspective. A simplified thinking framework can be developed around the following aspects:

Context

  1. Users: Clearly define the portrait or role of our users.

  2. Goals: Define the core goals that users want to achieve through your product.

  3. Pain Points: Identify the key barriers or inconveniences that prevent users from achieving their goals.

Value Proposition

  1. Product: What kind of product or feature do we need to build to help users achieve their goals?

  2. Alleviates: How specifically will our product alleviate or eliminate the user’s pain points?

  3. Benefits: What additional benefits or value will the product bring to users?

Objectives

  • Break down the big picture product vision into a series of small, measurable and achievable goals.

Solution

  • Core Features: List the key features that must be developed to achieve the goal.

  • Integration: Think about how the product integrates with other existing services or systems.

  • Alternatives: Explore and evaluate whether there are other feasible solutions.

  • Constraints: Identify technical, resource, or time constraints that need to be taken into account during the development process.

  • Out-of-scope: Clearly define the features that are not to be developed at this stage to maintain focus.

Feasibility Analysis

  • Comprehensively evaluate the technical feasibility, market feasibility and commercial feasibility of the project.

#field

The power of tracing back

| 2025-05-07 10:45 AM

When exploring the source of other people’s opinions, you will often find that they are deeply rooted in the soil of personal experience. Rather than pondering the opinions themselves, I am more fascinated by exploring the experience - how others extract this insight from the bits and pieces of life. After all, opinions are everywhere, but “personal experience” is precious. It is those “first-hand experiences” that breed unique and profound thinking.

#Life

Technology serves product experience

| 2025-05-07 10:44 AM

The quality of technology ultimately depends on whether it can solve the actual problems of users or customers in a convenient and effective way. Otherwise, no matter how sophisticated the technology is, it will be difficult to succeed. When developing products, we should always start from “user needs” and “overall experience”, rather than having technology first and then trying to find application scenarios.

The Bad Approach: Develop a cool technology (e.g. an AI API) and then expect others to adapt their existing products to your technology.

Correct approach:

  1. First, think deeply: “How should this product (or the core user value you want to provide) work from beginning to end?” You must first conceive the complete user experience and product form.

  2. Then, carefully consider: “In which links should we introduce AI technology to solve specific problems or improve user experience?” AI should be a tool to assist and optimize the overall product, rather than the product itself.

#AI/product

Value mining of public account articles

| 2025-05-07 10:37 AM

The ultimate pursuit of user experience is crucial. For example, how to allow users to “quickly save” public account articles is itself a very valuable entry point. Furthermore, around these saved articles, providing a series of “divergent thinking assistance” (such as related topic recommendations, point of view collision) and “aggregate information organization” (such as automatic summary generation, key argument extraction, knowledge graph construction), and even the ability to find similar public account articles, can greatly improve the efficiency of users in obtaining and internalizing information.

#Field/project

Simplifying design in the AI ​​era

| 2025-05-06 11:51 AM

In the era of AI-assisted development, overly complex design processes may become a hindrance, especially for small or startup teams. In the early stages, it may be enough to focus on two core design documents:

  1. Product design documentation:
  • The core is to clarify user value.

  • Define clear core functions (MVP functions) that can solve user pain points.

  • The output of this document can interact with AI, for example, allowing AI to assist in sorting out user stories and refining core demand points.

  1. System Design Document:
  • Outline a clear division of system modules.

  • Define key technical architecture and interface protocols between modules.

  • Plan a preliminary deployment plan.

  • This part can also be done with the help of AI, such as letting AI recommend technology stacks based on product requirements, generate preliminary architecture diagrams or API design drafts.

Combining these two core documents (and their AI-assisted outputs) can output a relatively clear “task list (Note)”. Combined with project management tools such as GitHub Projects and Trello, the development process can be effectively managed. Usually, this will be broken down into:

  • Feature List: corresponds to user stories or Epics in project management tools.

  • Technical Implementation Task List (Tech Task List): corresponds to specific development tasks or sub-issues.

In this way, a streamlined and efficient project initiation and management process was initially established.

#Field/Software Engineering

Respond to each sentence and think again

| 2025-05-06 11:07 AM

Is it really so difficult to achieve “respond to every word”?

Why do I sometimes fail to do this, and why do I expect others to do it?

If others fail to do so, whose “fault” is it?

I don’t seem to think that “responding to every word” is necessary, but what should I do if the other party insists on doing so?

I feel that the deliberate “response to every sentence” seems stiff and lacks humanity.

The other party seemed unhappy that I did not respond one by one.

But I feel I am not at fault.

What is the crux of this? Is it a misunderstanding in communication or a misalignment of expectations?

#Emotional changes

Trust your instincts

| 2025-05-04 10:26 AM

Once upon a time, I rejected intuition, and rationality often warned me of its inherent bias. However, if I blindly seek recognition from the outside world, I often only get contradictory voices. The real answer is not in the outside world.

When we are loyal to our intuition, sixth sense and inner wisdom, and trust the arrangement of the universe, life may show unexpected miracles and lead us to find the peace, love and joy we have longed for.

Courage lies in following your heart and listening to your intuition - it knows what you want to be, and everything else is secondary.

Intuition can be like a subtle whisper: a conversation with a stranger in a coffee shop may start an extraordinary friendship; a phone call to a friend may provide comfort just when you need it.

Many times, you don’t have to let your reason completely suppress your heart. You must have the courage to face the guidance of your intuition.

If you want to see scenery you have never seen, you must first bravely venture into the unknown, put aside your thoughts, and listen quietly to the voice of your heart.

#Life

Achieve your goals in inspiration or desperation

| 2025-05-04 10:08:47

There are two sources of goals: triggered by inspiration; born out of desperation.

Let inspiration trigger your dreams, not despair give birth to your goals.

Driven by pure inspiration (not desperation), there was a deep and powerful urge to create something new and refreshing to the world.

But how do you tell the difference? An easy way is to remember the difference between an idea and a thought. If a goal or dream comes to you in a flash, it is inspired; if you think about it and finally decide on a goal or dream, it is born out of desperation.

We always say, earn enough money and go traveling, these goals are means to serve the purpose, not the purpose itself. We always want to achieve these goals for various reasons, so we always feel that there is something missing deep in our hearts.

There is no right or wrong in the goal itself, it only matters whether its source is inspiration or despair. It just depends on your inner feelings. When you realize the difference in the source and expression of the two goals, you will feel extremely happy for being able to create something refreshing.

We do not create in order to achieve physical and mental perfection, but because we feel physical and mental perfection, we want to create and give without expecting anything in return.

What do you really want?

If I had endless money, had been to every corner of the world, had nothing to worry about, and no one would criticize me for what I did, then what would I want to do and what new things would I want to create?

Some new thoughts on flow

| 2025-05-03 19:07:50

“Anxiety is unchecked thinking. Flow is unchecked thinking.” —James Clear, habit researcher

We have discussed a discovery before, which is that it is easier to reach a flow state when you are slightly tipsy or sleepy. This phenomenon confuses me.

Humans are at their best when they are in a state of mindlessness and oblivion. There is a beautiful adjective in Japanese for this state: “mushin”.

Being mindless means having no distracting thoughts, no resentment, no worry or fear, and even no self in the consciousness.

Only by focusing on one thing can we break free from our own limitations and create all kinds of colorful things in this world. I don’t want to impose this belief on you, I just hope you are willing to try to experience it and understand it through your own experience.

It may also be the same as the previous teacher forgetting the previous lesson preparation between classes, and it is easier to achieve the flow state.

Therefore, the essence of anxiety is thinking too much, and the essence of flow is to be focused.

So back to the previous question, why is it easier for me to reach a flow state or a focused state when I am slightly tipsy or tired?

When intoxicated, the activity of the prefrontal cortex (PFC) — the area responsible for planning, criticism, and self-monitoring — is slightly inhibited by alcohol, which is somewhat similar to the “ego degradation” effect after meditation.

Fatigue may be different. Executive function decreases → criticality/distraction decreases. This is similar to the effect of flow, but in fact it is because of insufficient resources. The brain automatically “energy saves” and only focuses on the current task, and no longer deals with multi-tasking interference.

It may “seem” easier to enter the flow state, but in reality it is entering a state of “narrowed attention”, which is not a truly efficient flow state.

Usually early morning is the golden time when it is easier to enter a true “high-quality flow” because the brain has just restored its resources and its executive functions are most complete (including working memory, interference inhibition, and emotional regulation).

This may explain why artists sometimes need a specific environment, music or even a little alcohol to stimulate their artistic cells.

The Nature of Experience

| 2025-05-03 18:39:11

Hiking experience, travel experience, social experience, all human experiences.

If humans learn only one thing, it will be enough to change the world: don’t be afraid of experience.

People are born with a little fear. Fear of failure, fear of embarrassment, fear of the unknown, fear of trouble. This fear makes us stay in our comfort zone and dare not take a step forward. But the problem is that all growth and all changes are hidden behind “giving it a try”. If you don’t try, you will have nothing.

Fear makes people feel terrible. For example, many people nowadays feel overwhelmed when they hear “AI” and “programming”, and say “it’s too difficult, I can’t do it”. What is the result? If you don’t dare to try, you really can’t do it, and slowly you will be left behind.

There is another kind of fear, which is the fear of facing oneself. Afraid of admitting that one is wrong, afraid of seeing clearly what one wants, afraid of being different from others. If there are too many of these fears, one will feel depressed and feel awkward in getting along with others.

Don’t always shrink back. Dare to try and dare to feel. Only then will life be interesting.

Thoughts and ideas

| 2025-05-03 18:20:11

Thoughts and Reflections

flomo records momentary cognitive fragments during the thinking process.

Thoughts are psychologically driven and are the material from which humans construct their conception of time. Thoughts are the premise of all human experience.

We restore the process of our own thoughts popping up. This is definitely not controlled by the brain. It is formed spontaneously. Even the human brain cannot control the emergence of thoughts.

Thinking is the processing of thoughts. Thinking requires the participation of the human brain and consumes resources, so thinking is very difficult. Even if humans are willing to bear pain, they are unwilling to think actively.

But it seems that thinking is also the source of pain.

But it seems that an optimistic attitude is often not a result of reasoning, but a more instinctive and intuitive state.

Therefore, the first reaction to something is often one’s own emotional state, which is closest to the truth, rather than a decision made after thinking, rationality or emotion.

But isn’t thinking important? This seems a bit contradictory? And thinking will lead to unhappiness?

When you are alive, you have to grow up and figure something out.

Growth is inseparable from thinking. You have to rely on it to go beyond those intuitive and superficial things to figure out what you really want and how to do it better.

Therefore, “stop thinking” and “active thinking” are themselves a kind of balance.

Thoughts on brain evolution

| 2025-05-03 17:56:10

The brain’s complex calculations are aimed at the instincts of survival and safety that humans have developed since evolution, but it is powerless to control the joy and happiness of the soul.

The brain’s job is to maintain survival, and the mind’s job is to create a sense of satisfaction.

If you want to achieve freedom, happiness, peace and abundant love, you cannot just follow the instructions of your brain and be content with physical health; you must break this limitation and pursue higher spiritual health.

He who knows others is wise; he who knows himself is enlightened

| 2025-05-03 17:51:34

Most of us measure the world through our own ideas and perspectives. The meaning (or perspective) we give to something ultimately determines how we feel about it.

It is also a person’s three views that determine his or her views on a certain thing. Even when it comes to work, some people feel happy at work, while others feel painful at work.

If you think about it, who would you be if you no longer had the thought “I hate my job”?

Many people have not thought about understanding the existence of emotions as metacognition.

When you are angry, think about the reasons for your anger and the nature of your anger, and write it down objectively, which often helps with understanding and relief.

He who knows others is wise; he who knows himself is enlightened

| 2025-05-03 17:45:22

Most of us measure the world through our own ideas and perspectives. The meaning (or perspective) we give to something ultimately determines how we feel about it.

It is also a person’s three views that determine his or her views on a certain thing. Even when it comes to work, some people feel happy at work, while others feel painful at work.

If you think about it, who would you be if you no longer had the thought “I hate my job”?

Many people have not thought about understanding the existence of emotions as metacognition.

When you are angry, think about the reasons for your anger and the nature of your anger, and write it down objectively, which often helps with understanding and relief.

I would rather endure suffering than leave the comfort zone that makes me miserable

| 2025-05-03 17:35:33

Out of fear of the unknown, people are more willing to endure familiar pain.

There is an inherent “familiarity preference” in human nature, even if those preferences are not optimal.

Known pain seems easier to anticipate and deal with.

Independent development in the AI ​​era, front-end is the main battlefield

| 2025-05-01 11:22:06

The AI ​​model turns core services into general services. The backend is simpler and there seems to be no such high technical barriers.

SaaS and PaaS providers encapsulate basic capabilities, and many problems can be solved by directly using cloud services without the need for repeated implementation.

The front end becomes the key place for “user value delivery”, and it is easier to focus on MVP and single-person product closed loop, and use Next.js / Tailwind / Firebase / Vercel to complete full-stack delivery. The chain of “function → value → growth” can be quickly verified.

Databases, permissions, storage, and model calls are all moving towards platformization (Supabase, LangChain, Replicate, Vercel AI SDK). Independent developers do not need to write the backend from scratch. They only need to “integrate” instead of “build”.

AI decouples the “intelligence” part from the backend, and the frontend is the battlefield that is closest to users, has the greatest changes, and has the most concentrated value.

User stories drive development

| 2025-04-30 10:42:40

DDD theory and agile development theory that can be practiced by small teams.

Rapid iteration is the key. Set a version, iteratively develop the initial version, set the iteration cycle and iteration tasks, and try to make the iteration cycle in weeks.

Continuous integration and continuous deployment are important to reduce post-maintenance and code verification. This capability becomes particularly important in the AI ​​era to avoid rework.

User feedback drives development (user stories -> development -> validation), rather than writing lengthy prototypes and requirements documents first, which may not make much sense.

Use Kanban boards to record task progress.

Summarize, count, and reflect after each iteration, as well as review the tasks within a cycle.

Thoughts on using DeepWiki

| 2025-04-30 10:30:26

Many concepts of DeepWiki are worth learning from. In fact, I have been constantly learning from various open source projects before, basically learning from Cursor and Deep Research tools, and I have also summarized some routines on how to quickly learn an open source project. This time, DeepWiki is particularly convenient because it allows you to quickly generate a document summary of the project.

Simply replace github.com in any GitHub repository link with deepwiki.com to access the project’s DeepWiki page.

In fact, this operation is very familiar to developers. This is also a bit of inspiration, that is, how to provide a quick entry for users. It does not necessarily have to be in the form of a plug-in, but there can be a quick operation for the scene.

Another point is that there are still many applicable scenarios for Deep Research, and it is generally important to enter a popular and vertical scenario.

Stop thinking

| 2025-04-29 22:29:17

The prerequisite for minimizing thinking is to be aware that we are thinking.

This is what is called “metacognition”, observing our thinking from a God’s perspective.

That is to say, for some problems, if you realize that thinking will be painful, and realize that stopping thinking can let time solve the problem, then you don’t have to worry about it. Stop thinking and let the answer come naturally.

The ultimate rationality is the new thinking of sensibility

| 2025-04-28 20:20:04

Previously, on the mountain in ACT, I had a chat with a friend who studies philosophy and his companions about a very interesting topic: the ultimate sensibility is rationality, or vice versa, the ultimate rationality is sensibility.

Later, I thought about it and felt that this process was actually quite natural, like a cycle: starting from sensibility, through rational analysis, and finally back to sensibility. Sensibility is our most primitive feeling, such as emotions and intuitions such as happiness, sadness, and curiosity. They are like raw materials, driving us to think and explore. For example, if I am curious about something, I will want to understand it, and this is when rationality comes into play. Rationality helps me analyze and reason, trying to find patterns and answers from the chaos of sensibility. But the interesting thing is that no matter how rationally I think, the ultimate goal is to satisfy the needs of sensibility - for example, after understanding something, I will feel satisfied, or use these understandings to guide my life and make myself happier.

In this process, I found that “having myself” is very important. Thinking is not only an exploration of the external world, but also an understanding of oneself. By rationally analyzing my emotional experience, I gradually see my own values, motivations and needs. For example, I ask myself “Why does this thing make me happy?” This is actually using reason to dig out the root of emotion. In the end, this kind of thinking will make me understand myself better and make choices that are more in line with my heart.

Therefore, the cycle of sensibility -> rationality -> sensibility is actually a natural process. Rationality is a tool, and sensibility is the starting point and the end point. Any rational weighing of pros and cons cannot be separated from facing one’s own emotional needs sincerely. Just like what my friends discussed recently, no matter how we use rationality to analyze, we must eventually return to sensibility itself, to feel and experience. This thinking made me feel suddenly enlightened. When I discussed it on the mountain before, I only grasped the conclusion, but now I understand the logic behind it: sensibility and rationality are not opposites, but interdependent and mutually transformed. Extreme rationality will touch the deep level of sensibility, and extreme sensibility also contains rational logic. The essence of thinking is to continuously deepen the understanding of the world and oneself in this cycle. This makes me feel that those seemingly contradictory things in life are actually connected. As long as you feel and think with your heart, you can find your own answer.

Why China’s e-commerce opportunities are huge

| 2025-04-28 12:15:13

Why Hangzhou? Hangzhou is one of the most developed cities in the world for e-commerce and an e-commerce center.

Hangzhou and the entire Zhejiang Province have been places where small commodity economy, family handicrafts, and trade spirit are most prosperous since ancient times. It is also one of the regions with the most active private enterprises.

China is one of the countries with the largest e-commerce transaction volume and the highest penetration rate in the world. E-commerce accounts for more than 30% of the total retail sales, far higher than the United States, Japan and Europe.

Compared with the United States and Japan, why is e-commerce in the United States so developed, and why is the development of e-commerce in Japan relatively slow?

The historical background and social culture may have played a decisive role.

  • China: Offline channels were once imperfect, and e-commerce has provided new vitality.

  • United States: The offline market is huge, and e-commerce is an important supplement.

  • Japan: Offline business is highly developed and convenient, and the urgency of e-commerce is relatively low.

China is one of the countries with the greatest potential for e-commerce.

The development of e-commerce in the United States is partly due to its vast logistics radius. As a supplement to improving coverage, e-commerce has digitized small towns and suburban markets that were originally difficult to cover.

Human Courage

| 2025-04-27 18:50:11

One of the first qualities of a human being who dares to climb Mount Everest is courage.

Sometimes I admire some instinctive courage. Animals also have instinctive courage, such as the instinctive reaction when facing danger, the instinct of dogs to protect their owners, and the instinct of mothers to protect their children.

However, a lot of human courage often goes beyond instinct and is more driven by rationality and values. It is the choice to act even in the face of “fear” and has self-control that goes beyond instinct.

Courage is the foundation of all virtues.

For example, if you want to be honest, you need courage, because telling the truth often means facing uncomfortable situations; if you want to be kind, you need courage, because kindness sometimes means risking getting hurt; if you want to be fair, you need courage, because upholding justice often offends people and gets you attacked.

Without courage, it feels like all qualities may just be illusions in the mind.

The essence of courage is choosing to do what you know is right when faced with fear, pain, and pressure.

Courage is not recklessness. My evaluation criteria are knowing the danger, pain, and cost, but still rationally and firmly choosing to act after weighing the pros and cons. Recklessness may be acting without fully considering the consequences, and if the results do not meet expectations, it is easy to regret.

In fact, it is also a balance between the courage systems of Eastern and Western cultures. One extreme is obedience, compromise, and tolerance to the point of losing oneself; the other extreme is blind resistance, self-expansion, and ignoring the wisdom of coexisting with the world.

“Courage is not one of the virtues, but the most important one.” Because it is like the engine of action and the skeleton that supports everything.

It is easy to talk about virtue, but it is hard to have the courage to act. `Actions speak louder than words.'

Human curiosity drives us forward to explore

| 2025-04-27 18:30:30

I watched the video of Wandering Cookie climbing the 8,000-meter Annapurna without oxygen, as well as the aerial video of his previous climb to Mount Everest.

The video shows how small humans are, how hard it is to walk, and how many people are dying in the process. I can’t help but think:

Is it necessary for humans to reach the summit? What is the significance of so many people reaching the summit of Mount Everest? Do humans really need to do this?

If humans were simply fulfilling basic survival needs, then these behaviors would seem to have no immediate “meaning.”

But if we follow this analogy, the “significance” of the explorations in the Age of Discovery and the countless scientific breakthroughs also seem to be questioned.

So what is meaningful? Perhaps it is the indomitable spirit of the human being, the eternal desire for the unknown, and the deep respect for the laws of nature.

An important characteristic that distinguishes humans from other species is the desire to explore the unknown world and the spirit of challenging their own limits.

This may be the fundamental driving force behind human progress.

Our biological instinct tells us that we should be afraid and that we should surrender.

And courage may be the reward that the world gives to human beings.

The concept of software engineering starting from the end and TDD

| 2025-04-27 17:34:52

I mentioned before that TDD (test-driven development) is one of the best practices for software development in the AI ​​era.

The concept of “beginning with the end in mind” in software engineering is similar, emphasizing thinking about problems from the perspective of results.

For any matter, generally speaking, we should think before we execute, which can be summarized into two steps:

  1. Create a blueprint in your mind.

  2. Put ideas into practice.

Some examples:

  1. Test Driven Development (TDD): Write test cases first, clarify the expected results, and then develop.

  2. Continuous Integration (CI): Keep the software in a runnable state at all times and ensure that each change does not break the system.

  3. Amazon’s “backward working method”: When developing new products, first write a press release and FAQ, clarify user value, and then develop.

Some difficult and easy discussions, metacognition

| 2025-04-27 15:55:30

I just saw the cat and thought of this topic. I deliberately asked my partner about it. The cat has been pooping frequently recently, and the house is full of odor. The cooking aunt reminded us to solve it, saying that it might be worse in Hangzhou in summer. I thought about it and it seemed inevitable a long time ago, but it has been postponed until today. I wondered why this matter was not solved when it seemed to be important and urgent, because it is more of a daily life and does not affect my professional quality. Because thinking about this matter requires solving this problem, and solving problems is very troublesome. I teased my partner that it is easier to suffer than to solve problems. It is very troublesome to solve this matter, but if we endure it and don’t think about what the consequences will be in the future, we feel that we can get used to it and endure it.

  1. It is easier to suffer than to solve problems.

  2. It is easier to endure misfortune than to enjoy happiness.

  3. Most people would rather do anything to avoid real thinking.

  4. It is easier to save face than to admit mistakes.

  5. It’s also easier to fake it than to be genuine.

  6. It is easier to complain about the environment than to change yourself.

  7. It’s easier to dwell on the past than to let it go.

What is the bottom of human nature? Fear, pride, laziness, and the desire for truth. Each of these can be a shining light or a weapon to destroy people.

Listened to Chen Leiyi’s singing

| 2025-04-27 11:33:16

I remembered what Zhang Xiaolong said: “A programmer who doesn’t listen to rock music is not a good product manager.”

I really like Chen Leiyi. She is also a girl born in the 2000s, but she has a very good control over the stage and sings very powerfully. People feel at first glance that she is a girl suitable for the stage.

The spirit of rock and roll, to some extent, represents anti-traditionalism, concern for humanity, yearning for freedom and finding one’s true self.

Music represents emotional power and expression.

Our science and engineering men, programmers, usually have core responsibilities of implementing functions in a logical and rigorous manner, following established rules and requirements documents, and solving technical problems. This is more focused on “How”.

But thinking about why to do this, for whom to do it, and how to do it requires stronger empathy, creativity, business acumen and the courage to break the rules.

This is a transition from following rules to defining rules, from technical implementation to user value, and from logical deduction to emotional insight.

These tend to be:

  1. People who dare to challenge and innovate (anti-traditional, free).

  2. Someone who deeply understands and cares about people (humanity, nature).

  3. People who have the ability to think independently and be different (do different things).

ps. We even got in a group photo in the end.

Rationality is a tool

| 2025-04-27 11:12:46

Rationality is, to put it bluntly, a tool. It is used to achieve a purpose, such as making money or solving problems, so it is often utilitarian. But it has a characteristic: dry language and logic sometimes make it difficult to fully grasp people’s emotions and intuition. If you like someone, rationality can only say “because he has a good personality”, but how can it fully explain the inexplicable heartbeat? Language itself is also a tool, after all, it is limited. There may not be so many reasons for true love, and sensibility is the protagonist. Sensibility and subconsciousness are sometimes closer to certain levels of reality than rationality, such as artistic creation and interpersonal relationships, which often rely on it to support the scene. As for rationality, it is good to review it afterwards, but don’t expect it to take care of everything. Friends remind us to “handle the boundaries well”, that is, don’t be stubborn - don’t analyze coldly when you should be emotional, and don’t blindly bump into it when you should calculate. Simply put, rationality is a knife, and sensibility is the fire in the heart. Smart people know when to use which one. Think more about this balance in life, don’t let the tools ride on your head, and don’t let impulse ruin big things.

Continue reading Some thoughts on WeChat product concept

| 2025-04-27 11:11:45

  1. Keep it rough, keep it clumsy: Focus on the big features and ignore the small features. Sometimes it is better not to add them.

  2. If the solution is too complex, it is likely that the problem itself is misdefined.

  3. Even the smallest, almost negligible details can reveal the cognitive system, brand temperament and cultural confidence behind them.

  4. In fact, the ability to be “pluggable” can be an engineering capability on the one hand, and it can also be considered from the user’s perspective on the other. Some unimportant functions can be made into plug-ins for selection.

  5. Truly excellent experience innovation is not about users saying “Wow, this is amazing”, but about users feeling “this is how it should be”. Innovation is about bringing the user experience to the extreme.

  6. Products are constructed rationally, but are often used to express emotions.

  7. Train your intuition, and train your intuition rationally. The solution to a complex problem may not come entirely from reasoning, but also from intuition.

  8. Your weaknesses, from another perspective, may be your strengths. People with social phobia may be better at making social products.

  9. The product concept determines the height of the product.

  10. Treat your product as a work of art, not just a product.

Tolerance, understanding and forgiveness

| 2025-04-26 16:12:52

Tolerance refers to accepting and tolerating the different views, behaviors or habits of others, even if these may conflict with one’s own beliefs or habits. It usually manifests itself as an acceptance of external differences and may not require a deep exploration of the other person’s inner world.

Understanding goes a step further and means to deeply understand and comprehend other people’s thoughts, emotions and motivations. It requires us to try to stand in the other person’s shoes and understand their feelings and situations.

Tolerance is more of a behavioral restraint, avoiding emotional conflicts, and does not necessarily require a lot of cognitive effort. However, if there is a lack of internal recognition, this tolerance may be superficial and even difficult to maintain under pressure.

Understanding is more difficult and requires more effort and investment. To truly understand others, we must transcend our own subjective perspective and enter into the other person’s inner world.

But true tolerance is based on understanding. Tolerance without understanding seems to be just a temporary and superficial maintenance of the relationship.

Tolerance is an acceptance of external differences. Even if we do not fully understand the other person’s inner thoughts, we can achieve this through behavioral tolerance. Understanding requires us to transcend our own experience and limitations and enter the spiritual world of others. This is a cognitive and emotional leap.

What is more difficult than understanding is forgiveness. It not only requires us to transcend prejudice and cognitive limitations to understand others, but also requires us to let go of resentment and anger on the basis of understanding, and respond to harm with tolerance. This feels like a deeper challenge to human nature.

Metacognition

| 2025-04-26 16:04:27

The essence of metacognition is the cognition of cognition.

For example, be aware of yourself thinking and observe yourself walking.

Being aware of one’s own behavior, there is also passive metacognition and active metacognition.

From passive to active, this is a turning point. When a person can actively open up the third perspective and begin to continuously reflect on his own thoughts and behaviors, it means that he has truly begun to awaken and has the possibility of rapid growth.

Sensibility is paramount

| 2025-04-26 15:38:06

Reason is consciousness, and sensibility seems to be a part of the subconscious. At this time, we can’t help but broaden our horizons when we try to understand the product and the user’s behavior.

Sometimes when I read a book and come across some key plots, or when I feel a critical point, I will stop and think about why this point of view touches or inspires me, whether this touch or inspiration can be used in other things, and whether there is similar knowledge at this point of inspiration.

Indeed, I have always believed that sensibility is a good nutrient or raw material to guide myself to think. I have asked my friends many times before about the moments that moved them the most. This is the power of my sensibility.

  1. Why did you get emotional? What is the reason?

  2. Why does this product particularly attract me? Is there anything special about it?

  3. Why am I particularly immersed in this plot?

Sensibility is the raw material that helps us make choices; rationality helps us think, summarize, and train.

On the one hand, sensibility is a good guide for how to train and think about ourselves, or to find a sense of mission.

  1. Who do you want to help most?

  2. What is the thing you have done that makes you feel most touched?

  3. What is your most touching moment?

  4. Without any financial pressure, how do you plan to spend the rest of your life?

  5. What do you pay most attention to in your spare time?

How to capture other people’s sensibility? Sensibility is the best way to understand a person.

  1. What are the things that are most memorable?

  2. What things always come up?

  3. What is your unconscious first reaction, often the first thought?

  4. Dreams. What you think about during the day, you will dream about at night. I have never really believed in dreams, but dreams may be a true display of my inner thoughts, even when I am drunk.

  5. The body’s reaction. The body can tell you what you need. If you want spicy food, vegetarian food, or want to sleep, then listen to your body.

  6. Intuition.

The power of the subconscious mind

| 2025-04-26 14:42:36

We have discussed the question of whether AI has self-awareness. Apart from ourselves, we have no way of judging whether the other party has consciousness, not to mention that the other party is AI. We can only use some academic methods, such as the Turing test, to judge whether AI can meet the requirements of self-adjustment and self-optimization.

Suddenly I thought that for humans, the subconscious is also a huge force. The subconscious refers to the psychological process that is not perceived by consciousness but affects behavior, emotions and thinking.

However, for AI models, there may be some black box operations in the model itself that are difficult to fully understand. For example, the weights and activation modes within the large model may affect the answer, but even the developers and AI itself cannot clearly know how.

Most of our behavior is influenced by our subconscious mind. We rely on “System 1” (subconscious mind) and only arouse “System 2” (rational thinking) when we encounter conflicts or new situations.

So for most people, our judgments and decisions are often driven by past experiences, emotions, and cultural cues rather than completely rational analysis.

Subconscious execution is faster and consumes less energy. So when making a product, how can we make the execution path shorter? Products that require learning are very tiring to use. If most functions can rely on intuition to match user habits, or people can rely on reward models to learn, the shorter the product execution path, the less “conscious intervention” users need, that is, the more it resembles “subconscious behavior.”

Users need guidance, not education.

There is no need to tell users how advanced you are, you just need to make them use your product smoothly or even addicted.

Learning and thinking, which consume a lot of energy, are more helpful to human growth. Unfortunately, humans are naturally not fond of learning and thinking because these things consume a lot of energy. In the long process of evolution, the primary task of life is survival, and high energy consumption is a waste of life. This is why most people are willing to stay in their comfort zone.

Some people say that diligence is sometimes just a cover-up for mental laziness.

It is easier to suffer than to solve problems, and it is simpler to endure misfortune than to enjoy happiness. For most people, they are willing to do anything to avoid real thinking.

So perhaps it can be said that cognitive fuzziness comes from within, while emotional fluctuations often come from external stimuli.

Impatience

| 2025-04-26 13:12:32

I seem to be expecting too much to do something meaningful too early, or I am a little too eager for success.

But lack of patience does not seem to be a shameful thing, perhaps it is part of human nature.

When a baby is born, the rational brain is extremely weak.

Human endurance may also be consistent with the growth of compound interest. The compound interest effect shows the general law of value accumulation: the growth is very slow in the early stage, but it will grow rapidly after reaching an inflection point. This law reveals a kind of power, but to obtain this power, we need to face the slow growth in the early stage calmly and persist until the inflection point.

People’s endurance growth is also affected by the “comfort zone edge effect”.

Another important rule of the edge of the comfort zone is that it reveals the universal law of ability growth: whether it is an individual or a group, their abilities are distributed in the form of “comfort zone-stretching zone-difficulty zone”. If you want to grow efficiently, you must always stay on the edge of the comfort zone. If you rashly cross into the difficult zone, you will be frustrated, and if you always stay in the comfort zone, you will stagnate.

That is, for learning, the thinking after learning, the action after thinking, and the change after action are more important. If you don’t focus on the change in the inner layer, then no matter how much learning you put in on the surface, you will get half the result with twice the effort; therefore, from the perspective of weight, change > action > thinking > learning.

It is relatively easy to simply maintain learning input, but it is relatively difficult to think, act, and change. In the absence of awareness, we will instinctively avoid the difficult and go for the easy.

When reading, you don’t need to remember all the knowledge in the book. It is enough as long as one or two points prompt you to make real changes. The gains and significance are much greater than reading a lot of books but just staying at the level of knowing.

But why do some people focus so much on a thing? On the one hand, it may be because they have been trained or are good at exploring the underlying principles. But a more advanced method may be to use the instinctive brain and emotional brain to solve the problem, to feel the difficult things and find fun in them, or even become addicted. This is really amazing. Extremely rational people should be people who understand themselves extremely well.

The Nature of Anxiety

| 2025-04-26 11:50:52

The essence of anxiety may be that you want to do a lot of things and want to see the results immediately.

All human anger, in the final analysis, is anger at one’s own incompetence.

My desires are greater than my abilities, and I am extremely impatient. Anxiety is caused by the huge gap between desires and abilities.

Certain characteristics of human nature: being impatient for success and wanting to do many things at the same time; avoiding difficulties and seeking easy things, wanting to see immediate results without much effort.

Why do we feel more anxiety and pain in modern society? Because the pace and competition are more intense, and nature is amplified.

For example, modern people have experienced bonuses or sudden wealth. Once they have tasted quick money, they can no longer tolerate the psychological gap of “delayed gratification” and “slow process”. This is not a matter of money, but a recalibration of the human desire system.

But eventually we have to return to reality and face the rules: if we want to achieve something, we must remain patient and delay gratification.

Analysis of non-marriage & older single men and women

| 2025-04-26 09:52:50

China has a large number of involuntarily single men and a large number of older single women.

Under certain concepts in Chinese culture, people with rigid and single-minded ideas always believe that everyone must get married and have children. In particular, they believe that any woman must rely on marriage to support herself to a greater or lesser extent. Marriage must be a survival tool that is a matter of life and death for any woman in any era. Any unmarried woman of older age must live a miserable life and feel empty in her heart. They call this group of people “old leftover women.”

In developed countries, people with more individualistic ideas and more receptive to new ideas believe that if women can have an independent economic foundation to be self-reliant and self-sufficient, then being unmarried at an older age is nothing more than a personal choice that varies from person to person and from place to place. In fact, if your personality, style, interests and hobbies are really not suitable for the lifestyle of marriage, then not getting married is the right choice for you, and you can truly live happily.

Why do certain discriminations in China still exist? Discrimination against older people, discrimination against older singles, discrimination against women, discrimination against education backgrounds, and discrimination against regions.

What is the nature of discrimination? What is discrimination?

Discrimination is not based on what a person does, but on who that person is.

Its essence may be the desire to continue power inequality, cognitive laziness leading to simplified processing of complex reality, and group preference and psychological defense against dissidents.

I feel that discrimination is an invisible deprivation that happens regardless of what the other person has done, just because of who they are.

Respond to every sentence

| 2025-04-25 21:57:04

At some point, I suddenly realized that I didn’t respond to every sentence when chatting with my friends. I suddenly reflected on this matter, so I quickly found some chats that I hadn’t responded to and responded to them. But in the process of replying, I felt like AI. Isn’t this how AI chats with me? AI now doesn’t have much emotion, and responds coldly. I was stunned… I realized that I was almost like AI, haha, just kidding.

Is responsiveness a good quality? In the workplace, it may be professionalism, but in life, I value sincerity and spontaneity more.

Chinese people are particularly concerned about responses, which may be related to the cultural emphasis on face and interpersonal relationships. Not replying to messages may make people feel that you don’t care about them, or even a little “disrespectful”. But I myself tend to accept that I will ignore the response, thinking that responding every sentence will seem mechanical, unnatural, like a robot.

The art of communication lies in the delicate balance between sincerity and propriety. Responding to every word seems professional and polite in formal occasions, but it may make people feel deliberate and alienated in intimate relationships. I think the real communication is not the literal coming and going, but the resonance of heart and heart. Saying the right words at the right time, without having to reply to every word, can make people feel warm and natural. Appreciating a person is not about how diligent he is in replying, but whether he can switch flexibly in different scenes, which is both sincere and comfortable. Living in the present, listening and speaking with your heart, is the warmth and meaning of the conversation.

Some thoughts on open source commercialization

| 2025-04-25 12:25:59

Although a large number of stars for an open source project can increase its exposure, it is not directly equivalent to financing and profitability. Secondly, a good open source project does not mean that it can attract users. The market investment required for the commercialization of an open source project may even be far greater than the technical investment. On the other hand, you must choose the correct customer portrait based on your own track, do not blindly pursue benchmarks and large companies, and try to prove a user portrait that can be quickly replicated as much as possible to find your own customer base. Personally, I do not recommend doing ToC open source, but you can do technical To Professional. Finally, by doing the first batch of open source alternatives to a successful product, you can save the step of verifying a large customer base.

In fact, the product can take many forms:

  • Open source library + paid enhanced version (Pro)

  • SaaS Service + Local Deployment Edition

  • CLI tool + cloud-hosted interface

  • Plugin Ecosystem + Documentation Support

In fact, it is aimed at a group of professional users. Although these users are not enterprises, they have a certain budget, rigid demand scenarios and strong skill background. They can understand and are willing to pay for high-quality tools or participate in contributions.

Self-explanatory animals

| 2025-04-25 00:44:57

Humans seem to be animals that can explain themselves. Most of what we think of as thinking is actually rational explanations for some of our own behaviors.

Where exactly is curiosity?

| 2025-04-24 23:32:41

It seems that looking for questions just to find answers is really boring. This group of people’s curiosity is focused on thinking about what problems users have.

But the truth is actually very simple. It is important to get inspiration from the problems in your life and to pay attention to the problems around you.

But often the simplest truths are ignored by many people.

The core is whether a person can face his true self.

Why do we think more during our journey?

| 2025-04-24 23:16:14

Everyone is doing similar things because everyone may face the same problems. So how can we find truly unique and novel problems? Perhaps, that is to make ourselves more unique and novel. If I don’t travel and work, I may lack those new ideas and sources of thinking.

Why follow the same path as others? Maybe it’s not worth it.

Deep Research Saving important information is also a very important function

| 2025-04-24 16:49:40

The information generated by Deep Research may be in various forms, and may be more important or more complex. It is also very interesting and important to be able to save this information flexibly.

One is to flexibly access various MCP tools, and the other is to elegantly integrate various capabilities of MCP.

In addition, the note module also provides an MCP entrance, where users can quickly save notes.

For the caller, how to flexibly integrate the MCP module is also very important. It seems that most businesses will build their own MCP gateway.

| 2025-04-23 11:36:40

Current best practices involve deep semantic understanding of users and content, typically achieved with the help of embedding vectors generated by LLM.

Reinforcement learning (RL) is used to optimize long-term user value: Recommendation systems now not only optimize users’ current click behavior (short-term indicators), but also use reinforcement learning methods to simulate and maximize users’ long-term value, such as improving user stickiness, retention, and lifetime value.

LLM shows great potential in achieving conversational recommendations, handling cold-start problems, and improving recommendation explainability.

Capturing users' short-term interests and intention changes, the Sequential Recommendation model predicts the items that users may be interested in next based on their recent behavior sequences.

What may be more important later is that AI plays the role of “explainer” (explaining the reasons for the recommendation), “conversation partner” (discussing needs with users), “demand miner” (helping users clarify vague intentions) and even “content creator” (generating narratives or summaries related to the recommendation).

I love reading from the bottom of my heart

| 2025-04-22 21:55:58

Reading is a process of self-discovery. It is more about talking to one’s heart than acquiring knowledge. Tools can help us quickly filter through the flood of information, but the real nourishment comes from the emotions and resonance when we immerse ourselves in books. Modern people’s pursuit of efficiency often makes us anxious - we feel tired when we read, but we regret not reading. But reading should not be a performance, but a natural choice and a part of inner growth. Freedom does not lie in the number of tools, but in whether we can listen to our hearts and find ways to enrich ourselves. Why do we read? Not to fill in the gaps, but to light up the soul. This is the most essential gift of reading.

However, this society sometimes does not allow us to enjoy reading purely.

Rendering ideas

| 2025-04-22 21:55:45

In fact, in essence, we have to believe in AI to a certain extent.

The structure on the page needs to be constrained, but perhaps some of it can be relaxed. This is the space for AI to grow.

For example, AI generates an abstract for an article. The abstract has a suitable format and corresponding layout. Apart from the abstract, the next thing may be a wonderful sentence. Wonderful sentences often correspond to pictures in the original text, so it is another kind of layout. Then there is the form of presentation, which may be presented as a dialogue. In short, it may be possible to leave some operating space for the page to play freely. AI can create a certain format, and the page knows how to render better according to a certain format.

Rendered UI reference: https://github.com/wandb/openui

Give AI imagination within a certain space and let it play freely!

  • AI becomes a “design suggester” for page structure.

  • Pages are “format interpreters”.

  • Users will enjoy the best reading experience arranged by AI.

Another point that may be important is that it is also important to present the data elegantly like https://quizlet.com/.

Sharing & Emotions & Expression

| 2025-04-21 22:32:01

We know that feedback on the Internet is often unstable and of low emotional value, but we still can’t help sharing it.

People have an urge to “be seen”, even if it is just a potential desire for a certain person or group to see it. This is a bit like “WeChat status” or “background music in Moments”, a kind of silent speaking.

Sharing is sometimes a confirmation and archive of current experience.

Sending a card, a picture, or a feeling is not necessarily for others to respond, but it is like saying: “I don’t want this emotion to pass in vain.”

For example, a fisherman’s goal is not to catch a fish with every cast, but the posture of fishing is very important, as it is a symbol of “my presence”.

It seems to be a social circle of like-minded people online.

A product feature that gives people an excuse to share themselves may be more contagious than a feature that can share content.

“After reading this sentence, there was silence for 3 seconds” / “For you who are still awake at 3 am”

In this fragmented era, fragmented language touches the fragmented us.

Sense of ritual + weekly summary -> circle of friends, showing off yourself, similar to the annual summary of NetEase Cloud Music.

This is my favorite quote today, what about you?

Analysis of the applicability of GraphQL and REST in AI product development

| 2025-04-21 21:30:06

The core advantage of GraphQL is data flexibility. The cleaner the data format, the lower the preprocessing cost.

The idea that a GraphQL schema can be directly converted to a natural language description is great. The schema itself is strongly typed and defines the data structure and relationships, but REST is decentralized and LLM needs to understand the relationships.

GraphQL resolves all requests at once and supports batch queries and subscriptions (subscription), such as real-time updates of recommendation results.

The schema of GraphQL is very flexible. Adding a new field does not affect old clients, avoiding version management.

Suggestions in the project

  1. Situations when you should choose GraphQL: You need to process multiple types of data (such as text, images, real-time streams) or complex relationships (users-orders-logistics), require fast iteration, and the AI ​​agent makes decisions directly based on the schema.

  2. When choosing REST API: The project is simple enough that you don’t need the fancy stuff of GraphQL, and there’s nothing wrong with REST, after all, it’s stable enough.

Client choice: Altair (excellent GraphQL client)

Necessity of enabling browser notifications

| 2025-04-21 18:05:34

Many web pages, LLM systems, take a long time to process. For long processing times, users don’t know when the corresponding tab will complete the task, so it is particularly meaningful to enable the completion reminder function at this time.

Gender differences in behavior patterns

| 2025-04-21 12:03:24

How does society shape our understanding of gender, values, and relationships?

Why do Chinese women seem to be more inclined to “fight among themselves” or stick together, while men often target or remain indifferent to each other? Why are beautiful girls easily accepted by both men and women, while good-looking boys find it difficult to be appreciated by the same sex?

Speaking of the “internal fighting” between women, this is not necessarily simple jealousy or comparison, but a microcosm of society’s definition of women. Women may be instilled with the importance of appearance, status and recognition from an early age, and these things are sometimes shaped into a scarce resource.

This kind of competition is not necessarily innate, but may be the product of social expectations. In contrast, relationships between men lack this “soft” competition. They are more likely to compete on “hard power” such as career and economy, and appearance or social status rarely become the focus.

The definition of value is never neutral; it is shaped by culture and expectations.

It feels like society has a net that circles us and constrains everyone. Society needs some kind of rules to support resource allocation, role division and identity recognition. Why does the status of Chinese women sometimes seem to be severely stratified? After the reform and opening up, the status of women born in families is compared with that of women in Japan and South Korea. The latter’s growth environment may be more like a period of slow growth without drastic changes and deep reflection.

But is this kind of competition really innate? Or is it more of a product of social expectations? It seems that combined with Western culture, it may not be innate.

What is missing on xhs is the emotional value

| 2025-04-21 11:30:35

Maybe there is a point where AI can help me quickly generate valuable paragraphs or fragments, and Xiaohongshu (xhs) can quickly publish these paragraphs or fragments. Some emotional value, emotional stories or fragments may be scarce on xhs.

CrewAI vs. LangGraph

| 2025-04-20 15:34:11

LangGraph is more suitable for future complex AI systems due to its advanced features, while crewAI also has potential due to its ease of use and wide adoption. The future trend may be inclined to combine the two, but LangGraph seems to have more advantages in building complex and controllable Agent systems.

LangGraph provides state management, human-computer interaction, and debugging time travel, suitable for complex tasks; crewAI emphasizes collaborative intelligence and code-free tools, which are easy and fast to deploy.

In terms of choice, for lightweight and prototype development, crewAI may be a good choice; for some engineering projects that require fine control and complex logic, LangGraph may be more suitable.

Reference: https://github.com/crewAIInc/crewAI

Team & Personal Knowledge Assistant: Anything-LLM

| 2025-04-20 15:14:18

Anything-LLM is a tool that allows you to feed content such as Notion, PDF, Markdown files to a large language model, and then use it to do intelligent question-answering, search, and summarization tasks. You can run it on your own computer or server, and the data will not be uploaded to the cloud. It is more suitable for teams or individuals to build private knowledge assistants.

  • Upload documents and turn them into a “chat database” with one click: for example, you can import an e-book, a project description document, or even the entire Notion workspace.

  • Ask it questions and it answers them with your content: For example, “What features did we decide on in our last design meeting?” or “What is the key point of this white paper?”

  • Supports multiple users and multiple data sources: can be used as an internal document Q&A robot and can also be used for personal information management.

  • Supports multiple large model backends: supports OpenAI, Ollama (local LLM), Groq, etc., and can be switched freely.

LLM inference model, and subsequent predictions

| 2025-04-20 14:51:02

The model of the future may be ‘adaptive intelligence’, which will more intelligently judge when deep thinking is needed and when quick response is needed, and automatically switch between different modes.

Users can participate in the AI’s thinking process and guide or correct the direction of reasoning.

And it seems that the “multimodal” reasoning ability is the trend later.

Structured output can significantly reduce hallucinations in the model

| 2025-04-20 14:44:16

OpenAI’s structured output methods, such as function calls, improve the controllability and reliability of output by imposing constraints, focusing on tasks, and facilitating verification.

When you ask a model to output a specific format (such as JSON, XML, CSV, or a custom template), you are actually imposing explicit rules and boundaries on the model’s output.

And the essence of structured output is to guide the model’s thinking process. When certain techniques (such as the thought chain prompt Chain-of-Thought) are used in conjunction with structured output, the model can be required to list its reasoning steps or extracted key information in a specific field before giving the final answer. This makes the model’s “thinking” process more transparent and helps identify potential logical jumps or factual errors.

UI Testing of Current Mature Agent Methods

2025-04-20 14:18:51

A multimodal LLM fine-tuned for UI snapshots, allowing test scripts to be written in a natural language and automatically navigating through the program, QA.tech probably falls into this category.

The other type, such as Browser Use, combines the multimodal basic model with the Playwright testing framework, and relies on a deep understanding of the structure of the web page for testing. This method does not rely on a specific fine-tuning model.

Design Patterns for AI

2025-04-20 14:12:50

The improvement of the capabilities of agents in software engineering will allow many future modifications to be done by AI. However, AI-friendly code design is crucial for code maintainability.

The DRY principle can reduce duplication of code and make AI easier to handle working contexts. Currently, the design patterns that are most suitable for AI are still closely linked to the best practices of traditional software design.

With the evolution of AI, more design patterns for AI may emerge in the future.

#Field/Software Engineering

Reasoning Model & Prompt Thinking Practice in the Context of Reinforcement Learning

2025-04-20 12:47:25

Good prompts should be clear (not ambiguous and easy to understand), specific (clearly state the requirements and avoid generalities), and provide sufficient context (tell the AI ​​relevant background information).

Modification of the traditional prompting method in the application stage of the reasoning model:

Observation 1: Few-shot Prompting vs. Zero-shot Prompting

  • Few-shot prompts: Provide the model with several examples (input + expected output) in the prompt word to let the model learn the pattern.

  • Zero-sample prompts: Task instructions are given directly without providing examples.

  • Findings: For reasoning models, concise zero-shot prompts (giving instructions directly) or few-shot prompts (giving examples) work better. This may mean that reasoning models are quite capable of understanding instructions on their own, and providing examples may limit or mislead them.

Observation 2: Chain-of-Thought (CoT) Tips

  • CoT: A prompting technique that guides the model to “think step by step”, allowing the model to show the reasoning process before giving the answer, which usually improves the accuracy of complex problems.

  • Findings: For inference models, using CoT hints may actually reduce performance.

  • Possible reasons: Such advanced reasoning models may have been trained through reinforcement learning and other methods, with a built-in CoT-like mechanism (a fine-tuned CoT mechanism). In other words, they tend to or are trained to think in steps, and using CoT prompts to force guidance may disrupt their internal optimized process, resulting in poor results.

Reinforcement learning and fine-tuning are designed to make the model more directly and efficiently complete specific tasks. Prompt should focus on defining “what to do” and “what the goal is” rather than specifying “how to do it” in too much detail (unless “how to do it” is the core requirement of the task). The model’s ability should be trusted and given a certain amount of autonomy to solve the problem, especially for tasks that have been specifically fine-tuned (such as code generation, summarization).

Design a series of coherent prompts, each of which handles a part of the overall task. This is especially important for building autonomous agents.

#Field/Software Engineering

Decentralized, conversation-based architecture practice

2025-04-20 11:34:05

In the AI ​​era, a new architectural practice may be the Advice Process

  1. Rule: Anyone can make architectural decisions.

  2. Limitations: Before making a decision, you must consult two types of people:

    1. All persons who will be significantly affected by the decision.
  3. Someone with expertise in the area of ​​decision making.

Key Points: Decision makers must seek advice, listen, and take note, but they do not have to agree with or adopt it. The goal is to get a wide range of input and voices, not to reach consensus.

Four key elements: Architectural Decision Records (ADR), Architecture Advisory Forum (AAF), Team-sourced Architectural Principles, and Your own Tech Radar.

#Field/Software Engineering

API is no longer a technology, but a product

2025-04-20 11:26:00

The API is aimed at the group. Just as the design of user products requires consideration of user experience, the design of APIs also requires consideration of the developer experience.

Good API design should be reasonable and standardized, which is not only convenient for users to use, but also convenient for other businesses to access, and can also be better processed by AI.

You can try the following:

  1. Provide available examples.

  2. Pre-configured authentication.

  3. Prepare realistic test data.

These should be delivered and managed as part of the API product, just like API code and API documentation.

#Field/Software Engineering

Sages & Philosophers

2025-04-19 12:39:19

A wise man is a person who claims to be learned or intelligent; a sophist teaches for money; a philosopher is a person who loves wisdom.

Sophists teach people the truth and charge tuition fees. Such people are common in life, such as teachers who impart knowledge and mentors who sell entrepreneurship courses. There are also those who think they are knowledgeable and complacent, and those on the Internet who boast of being well-read but actually know nothing.

Philosophers, on the other hand, are well aware of the limitations of their own knowledge, which is why they are constantly seeking insights. “The only thing I know is that I know nothing.”

Do people have a sense of shame from birth?

2025-04-19 09:51:53

Shame is more like a social adaptation mechanism that helps individuals maintain good relationships in the group and avoid being excluded. It also prompts us to restrain ourselves to meet the expectations of the group, thereby enhancing the opportunities for cooperation and communication. Humans are social animals and ultimately cannot do without the group.

Shame is mostly learned, and newborns rely mainly on instinctive behavior to survive.

There must be many places in the world where being naked is a natural thing, so social forms often determine the boundaries of individual behavior.

Shame is not meant to foster moral prejudice or break you down, but to remind you that you still care.

Many times, shame should not be an excuse to escape. Feeling ashamed for saying the wrong thing, making the wrong decision, or ignoring a good person is not weakness.

#Life/Growth

Thoughts on AI-generated code

2025-04-18 16:46:25

While meditating under the tree and looking over the various modules of recent AI code generation, I feel that most of the current AI is superficial, and most projects have not yet used good tools and are not ready for production level.

A truly outstanding project should have many general learning templates and structured templates to assist the program in quickly generating a production-level usable product. This is crucial because AI should not only complete simple tasks, but also be competent for complex structured tasks.

Furthermore, AI should have growth potential. Therefore, the key is that, as AI’s capabilities or tools improve during this process, its upper limit should also be broad.

In this process, MCP (Model Context Protocol) plays a vital role, as it standardizes the communication between AI and tools; the same is true for A2A (AI to AI), which standardizes the communication between AIs.

After the specifications are established, the key lies in whether a mature system can complete the release of workflows with one click.

It can even be said that once the protocol is released, the importance of generating an MCP application with one click, or generating an MCP application based on an existing cursor, is self-evident.

Furthermore, deployment and project templates vary greatly, and it is quite troublesome to implement or build logic based on them.

What are your concerns?

  1. Everyone is concerned about security issues, but if the application code is completely controllable, there is no need to worry too much.

  2. Tools like OpenGrep can verify and identify security issues.

Yao seemed to have worked on this, but the results were clearly not satisfactory.

When should you use i18n

2025-04-18 12:31:45

In the MVP phase, i18n should not be a priority.

At this stage, the core is “minimum” and “feasible”.

i18n should be considered an optimization or extension step after successful validation.

However, if the verification is passed, it is necessary to consider i18n. The cost of fixing an i18n Bug after release may be more than 14 times the cost of fixing it during the coding stage. If the system needs to support multiple languages, it must be cautious.

Stripe’s fees are really high! But it’s really a must-have for overseas expansion

2025-04-17 16:32:23

Stripe Checkout makes it extremely fast to build a payment system. You don’t need to write your own UI. You can jump to Stripe’s payment page with just one line of code.

AI overseas projects usually use Stripe, PayPal and TerraPay.

It is undeniable that Stripe provides complete subscription billing management, webhooks, invoicing, trials, coupons, taxes, and more.

And there is no need to register the company in the United States.

#AI/product

Thinking about time

2025-04-17 13:44:40

Why does time seem to lack a clear and distinct measure?

I always feel that there is not enough time, but I don’t know where it goes.

In daily life, there is a lack of regular behavior.

For example, the specific length of sleep, when to get up, and when to exercise.

As well as daily office, study and reading time.

It would be better if all of this was properly planned or arranged.

I have lost interest in some meaningless conversations. If there is no obvious collision or gain, I don’t think there is anything that is lacking on the Internet or that I cannot search for myself.

Talking is cheap and ineffective, while creating is more challenging and more effective in changing the world.

Please talk less and do more.

What is important?

Spend time with the ones you love.

Surround yourself with people who inspire your curiosity and potential.

Sports and the outdoors make me happy.

Explore new places.

Take on a creative project.

It’s okay to say “no”, to be silent and to find your own focus.

Is PostHog suitable for startups or MVP stages?

2025-04-17 11:11:40

PostHog I think is very suitable for MVP stage integration drive.

The core definition of MVP is: to verify core product assumptions at the lowest cost, understand early user behavior, collect feedback and guide the next product iteration.

The core functions provided by PostHog are:

  1. Event-driven: including button clicks, page and function usage, etc.

  2. Automatically capture front-end events: Reduce the burden of manual tracking in the early stages.

  3. Key conversion funnel: such as registration -> create the first project -> share.

  4. User Session Recording: This feature is very useful and helps you understand “why it is so”.

  5. In-product surveys: Send short surveys directly within the product to specific user groups (e.g., users who have just completed a core action, or users who are about to churn) to collect direct, contextualized feedback.

  6. Feature Flags: Allows you to release new features or variants of your MVP to specific user groups (such as internal test users, some new users) without redeploying the entire application. This greatly speeds up the process of A/B testing and grayscale releases, and is a powerful tool for rapid iteration of MVP. You can use data to verify new ideas, and quickly turn them off if the effect is not good.

  7. Monitoring: Provides retention analysis and key data information dashboard.

#AI/product

Key Value Capabilities for Large Models

2025-04-16 16:16:47

Anthropic provides a feature called “Prompt Caching” that aims to optimize API call efficiency.

Its essence lies in the flexible storage of K (Key) and V (Value) states when the attention mechanism predicts the token.

  1. Explicit call: It is not enabled by default and needs to be enabled actively.

  2. Prefix Caching: If a subsequent API request uses the same prefix content as a previously cached request (up to the cache_control mark point), and the cache_control mark is in the same position, the system can reuse the previously cached K/V state and skip recalculating this part of the prefix content.

  3. Main Benefit: Significantly reduces latency for subsequent requests and slashes costs. The price of a cache read token is 90% cheaper than a normal input token. When writing content to the cache for the first time (Cache Write), these tokens will be 25% more expensive than normal input tokens.

  4. TTL (Time To Live): 5 minutes, and the minimum token length requirement is 1024.

Comparison of policies among manufacturers

  1. OpenAI: No additional charge for write cache, automatic caching. Up to 50% discount for read cache.

  2. Anthropic (Claude): Write cache has a 25% surcharge, but read cache is discounted up to 90%. Cache TTL is fixed at 5 minutes (can be refreshed).

  3. Google (Gemini): Users need to use SDK or API to create CachedContent objects and can customize TTL (default is 1 hour). Users have the highest control, no charge for writing, but incur storage costs based on token-hours. Read cache can get a 75% discount.

More commonly used scenarios

  1. Conversational Robot: Includes some conversation cases, which are usually fixed, and each round of user input and historical conversations are attached to the back.

  2. System Prompt: You can put lengthy, fixed system prompts at the forefront of the prompt. By placing the cache_control tag after the system prompt and before the user input/history dialog, the model cache can handle the KV state after the system prompt.

  3. Batch classification/processing: For a large number of classification tasks with the same rules, their prompts can also be stored in the cache.

  4. Few-shot Learning: Provide several complete “input-output” examples in the prompt to guide the model’s behavior when dealing with actual problems. These examples are usually fixed, while the actual user’s questions are variable.

  5. Fixed instructions in RAG: If the [Fixed instructions] part in the RAG flow is very long, this part can be cached.

#AI

The service layer does not need Kong to call the AI ​​platform

2025-04-16 12:48:50

When Go Backend calls LLM service (usually Python service), it generally does not go through Kong, but directly calls it through the intranet (such as IP, Service Name, Docker Compose / K8s internal service name).

Kong mainly plays the role of a “border gateway” for external network services.

Direct calling not only has higher performance, but also clearer logic.

If internal services need to be used both internally and externally, consider exposing them via LangGraph (or similar service orchestration tools):

  1. A public interface for Kong forwarding.

  2. An Internal-only interface for direct connection to the intranet.

#Field/Software Engineering

Browser Use Less than 10,000 lines of code raised $1,700, why!!

2025-04-15 12:57:34

In short, Browser Use improves the efficiency and accuracy of large language models in web page recognition and operation, which helps agents complete tasks.

Positioning: One of the basic services of Agent.

MCP (Model Context Protocol) is a consensus and its encapsulation of the API is undeniable.

However, there is still a pie to be divided: the vast majority of services do not have APIs, but only provide GUIs for humans, and the browser is the main container for GUIs.

Browser Use can be seen as an intermediate optimization solution for the period when the model’s capabilities are insufficient. If this period is long enough, its value is huge; if the model breaks through quickly, its value may decrease.

I personally believe that the upstream and downstream supporting infrastructure of Agent is still in its infancy, and small teams have great potential and may be able to excel at some point.

Reference link: https://github.com/browser-use/browser-use

#AI/product

Cursor How to quickly learn a project

2025-04-15 12:22:20

  1. Agent mode starts: First, you might as well ask a simple question, such as “Please help me analyze the overall architecture of this project.”

  2. Architecture diagram drawing: Let Cursor assist in drawing the architecture diagram.

  3. Online Preview: You can copy the generated chart code to an online tool (such as Mermaid Live Editor or Mermaid Chart) for preview.

  4. Core process analysis: Then request Cursor to analyze the core process and entry files of the project.

  5. Follow the thread: Read further along this line of thought, for example, analyze a specific function to find out which services it initializes.

#AI/open-source

Testing Tips

2025-04-15 11:01:48

Testing often involves CLI, Web UI, and API related levels.

  • CLI testing: more frequent, suitable for quick verification, unit testing, etc., and easy to integrate with UI.

  • Web UI testing: Suitable for user experience verification and end-to-end testing, but its maintenance cost should not be underestimated.

  • API testing: This is a common method and can even be completed with the help of tools. It is closer to real business and is embedded as part of the engineering system.

Generally speaking, their value ranking may be: Unit/CLI testing > API testing > UI testing.

Verifying the same business logic through multiple methods is also a way to enhance the robustness of the system.

In fact, UI testing has its own routines, and often distinguishes page processing. For example, to facilitate testing, a basic UI interface system may be built, and such basic pages are generally only presented statically.

#Field/Software Engineering

CLI calls the service or API

2025-04-15 10:57:25

Usually the following aspects can be considered:

  1. Language consistency: If the business language you develop yourself is inconsistent with the language you want to write the CLI in, then using the API is the natural choice.

  2. Business requirements:

  • If the CLI is only used for local quick verification and debugging, it is more direct to call the local core logic.

  • If it involves remote distributed tasks, cloud services, etc., it is best to design it as an API caller.

  1. Security considerations: If you have high requirements for security, authentication, permission control, etc., you should also call the API directly.

The API design of Kubernetes can be used as a reference. The core of it is to properly encapsulate and abstract the core logic. kubectl needs to call remote or distributed services. This distributed architecture forces it to choose the API form.

#Field/Software Engineering

A series of reactions to network problems

2025-04-15 09:56:39

My MacBook suddenly lost its internet connection. I tried using xhs and Google, but none of them worked. I was in a hurry and became emotional, and I was eager to fix it. After I calmed down a little, I reflected on myself, thought about what was important, and how to face the current predicament. Instead of rushing for a temporary solution, I turned to explore the lessons I could learn from this network incident.

Reconstructing the scene is crucial. I have always had a method, that is, whenever my emotions fluctuate, whether positive or negative, I will not rush to evaluate the emotions themselves or myself, but record them objectively. First, objectively record the emotions, and second, objectively record the problem information, and then objectively analyze the problem.

I seem to be quite good at reflecting on my emotions, especially on the emotions themselves. From an emotional perspective, this is a quick way to understand myself. For example, when I was riding a pendulum, I just needed to feel the tension at the time, and then I suddenly had an inspiration, and realized why it was like this at the beginning. This time, when I encountered obstacles, my fighting spirit was stimulated, and I wanted to solve it even if I had to work day and night. Subconsciously, I was extremely eager to solve this problem. Therefore, if I encountered an important or urgent problem at a certain time and stage, I would be completely immersed in it. This may be one of the reasons why I have a habit of procrastination. At the last minute, the whole person’s state can reach the state of flow. However, there is room for improvement.

The details of this incident are as follows:

The new version of my MacBook has many different configurations than the old version. Previously, I used ClashX for VPN, but recently I found that ClashX can easily cause the computer to lose network. Looking at its log, it may be a configuration problem. When there is no network, I restart the computer, and it can be used normally after restart. However, ClashX frequently causes no network, and it needs to be restarted frequently, which is not a long-term solution. So I prepared to change it and turned to sing-box. However, the configuration of sing-box is quite cumbersome, and it prompts that Install Network Extension is required. After checking several issues, I learned that it needs to be enabled in General -> Login Items & Extensions -> Network Extensions. However, after enabling it, the computer still has no network. At this time, a strange scene appeared-there is no network after restarting the computer.

There are three purposes for recording things:

  1. Clarify your thoughts and sort out your logic.

  2. Record the content to help others reproduce the problem, submit an issue, or even ask AI for a solution.

  3. Most importantly, learning and growing from every “failure” through objective recording and reflection is a gain in itself.

#Field #Life/Growth

Procrastination

2025-04-14 20:04:43

Procrastination is not necessarily a bad thing. It can save time to a certain extent. However, the key is whether you can “drag” it to an ideal point and make the best use of the time while ensuring quality.

Procrastinators may be avoiding one important task while accomplishing many other less important but still valuable things. Some people even claim to work more efficiently when under pressure.

From this we can see that procrastination varies from person to person, and it is more important to find a habit that suits you.

In most cases, procrastination does not solve the problem, but instead causes more stress and anxiety, which is really a last resort.

The concept of making progress day by day

2025-04-14 20:01:27

  • “One step at a time” means that no matter what you do every day, you will carry this spirit through to the end.

  • “One step forward every day” means that you must do the things that you must do every day. Even if you delay for one or two days, you must try your best to practice them every day. It only requires you to move forward a small step every day. The daily requirement is not high, and it only requires you to do this one step forward. The key is persistence.

Key management in the project

2025-04-14 17:20:39

Doppler is very suitable for direct sales teams. Its light weight, ease of use and good integration are its significant advantages.

Compared to tools like HashiCorp Vault that are more suitable for production-level, large teams, Doppler has obvious advantages in small teams.

参考: HashiCorp Vault GitHub

Smart proxy solution

2025-04-14 14:52:16

Traditionally, when we analytically read an article, we think about a number of questions and break it down into discrete steps — summarizing, extracting keywords, categorizing content, and generating insights — and we think this is a good way to work.

This approach has its limitations, each task may be implemented by a separate model, and you need to sort the tasks. This is similar to a workflow, which requires clear management of knowledge transfer between steps and independent determination of what additional operations are needed based on the intermediate results.

But, what will the agent do?

  1. State Management: Understanding context, having short-term memory and long-term memory.

  2. Strategy selection: Knowing which approach is more reasonable.

  3. Tool Calling: Understand which tool can better solve the problem.

The advantage of LangGraph over Agent is that it can manipulate each node like a graph.

Vercel AI SDK Development Tool

2025-04-14 09:58:25

Two outstanding features:

  1. Streaming Excellent: Encapsulates the complex details of streaming and is easy to use.

  2. Cross-provider compatibility: It doesn’t matter even if you change the LLM manufacturer.

  3. UI level: The returned status and data are usually reactive and can be directly bound to UI components of frameworks such as React, Vue, Svelte, etc. to achieve smooth real-time updates.

In comparison, Vercel AI SDK helps developers quickly build front-end NodeJS and React projects; for the back-end, there are Python packages like LiteLLM to solve similar problems.

A clever way is to use the hooks of the Vercel AI SDK to centrally manage and dispatch KEYs on the backend. This way, you can get the excellent front-end streaming support of the Vercel AI SDK while enjoying the advantages of centralized back-end management brought by the proxy (such as keys, model routing, quotas, load balancing, etc.).

LiteLLM supports two modes: one is library mode (Python) and the other is proxy mode. Both are very easy to use and have an active community, which is a clear advantage over one-api.

#AI

Chinese and Western Culture

2025-04-13 20:52:22

In an ink painting, mountains and rivers are outlined with just a few strokes of light ink, but large areas of white space are deliberately left, evoking people’s imagination of the unspoken clouds and distant peaks.

This is an invitation to think and feel.

Therefore, the blank space on the text may be more important than the text itself.

Oil paintings in Western culture are densely packed with details, expressing the intentions thoroughly and leaving no room for ambiguity.

We have been taught since childhood that we should be reserved in our words and actions, not say everything, and leave room for others and ourselves.

“The Tao that can be told is not the eternal Tao.” Some things are too profound to be fully described in words and can only be understood by people themselves.

The tendency of Western culture is to fill in details and pursue the ultimate “existence” in the pursuit of knowledge and truth.

#Aesthetics

Creating something out of nothing

2025-04-13 19:32:23

One of the key leaps for mankind is the creation from “nothing” to “something”.

In the future, AI may give rise to new physical formulas or even higher-level abstract laws, and even be given the status of truth.

Creativity is more important than we think.

Performance is important, but why do we love jazz? Because every performance is an improvisation and creation.

Clarifying the knowledge points and logic can even be said to be for better improvisation.

The classroom is a lively place with communication and collision, rather than one-way output.

Students are not empty bottles to be filled, but individuals with thoughts and curiosity. The real value does not lie in how much the teacher teaches, but in the sparks that emerge in the communication among people.

This can also be called “living in the present moment”. There is no need to worry much and you might as well regard some things as spiritual practice.

#Life/Growth

Deep Research Value Chain Distribution

2025-04-13 15:22:29

Following giants such as OpenAI, many entrepreneurs or open source projects have begun to continuously study similar Deep Research solutions.

Model vendors (such as OpenAI) do have significant innate advantages in such applications, as they can combine their own models with reinforcement learning fine-tuning.

However, various products of Deep Research have also opened up differentiated paths, such as adopting thought chains, round-table interviews, dialogue forms, role-playing, and integrating various tools. There is still a lot of room for many refined needs and scene needs for specific knowledge, data, and workflows.

Furthermore, by utilizing better quality proprietary datasets, it is possible to surpass general-purpose models on specialized tasks.

In the future, OpenAI and other vendors may encapsulate the capabilities of RLHF (Reinforcement Learning from Human Feedback) into easier-to-use APIs or services. Currently, Fine-tuning API mainly supports supervised fine-tuning (SFT). In the future, “RLHF-as-a-Service” may appear, allowing customers to provide preference data (such as A is better than B) for model optimization without having to worry about the complex implementation of RL algorithms.

In addition, with the development of open source libraries, more and more open source projects and open source models will emerge, which can also support reinforcement learning fine-tuning.

#AI

Subconscious proof that I am important

2025-04-13 14:41:25

There is a phenomenon with the cooking auntie. For example, she always expresses subconsciously: “I have a lot of orders, and many customers want me to cook for them.” This behavior seems to tell us subconsciously that she is scarce. This is similar to the behavior of some porters who are eager to show off and take credit when hiking in Nepal.

She seemed to be subconsciously conveying the message that “my services are rare”, which is not difficult to understand, but my attitude towards this behavior is quite interesting.

Subconsciously, I felt that the aunt was “showing off”, showing off how much she had given and how many clients she had, in an attempt to gain more initiative and increase her own scarcity, which might make me cherish her more and even feel that it was a great thing to be able to hire her.

Later, when I thought about it carefully, I could also argue that the aunt was trying to encourage herself, craving recognition from others and feeling that she was quite capable - everyone has the need to seek a sense of existence.

I subconsciously reject this kind of phenomenon, probably because it makes me feel that she wants to have the upper hand in the relationship. However, she and I are in a cooperative relationship, and the root of this rejection may be my yearning for an equal and sincere relationship, as well as my potential desire for control.

However, from the perspective of traditional culture, the other party’s desire to prove themselves in a collectivist environment is slightly in conflict with my expectations for genuine equality.

It is important to understand the motivation behind the other party’s words and actions, rather than just relying on your gut reaction. In essence, the identities and cooperative relationship between the two parties will not change, and the other party has no intention of making a profit. Next time, you can just laugh it off and say, “Yes, your craftsmanship is so good, there must be many people lining up to hire you.”

Sincere praise may come from understanding and respect.

#Life/Growth

Meaningful investments are those forces that are not optimistic about themselves

2025-04-12 20:29:46

Many people already have many platforms and resources, and even without your help, there is no shortage of support from others.

The real socially valuable investment lies in seeing those who were not optimistic about the future and firmly supporting their growth in their early stages. The starting point of these people may not be important, but their growth potential outside the structure and their growth slope are important.

The future of modern product documentation

2025-04-12 19:03:03

Using MCP (Model Context Protocol) to access local documents seems to be a trend.

Docs ≠ static web pages, but rather “part of the product”.

Documents should be developed more concisely and easily accessible to third parties. The MCP approach is particularly important here.

The documents themselves should have the ability to access AI and knowledge bases.

Nextra is a great tool that can be implemented based on Nextra and by referring to the Langfuse case.

#AI/product

Product Recording Method

2025-04-12 17:51:52

Product-related thoughts are more suitable for fragmented recording, but not too scattered. flomo is suitable for recording preliminary product ideas.

In addition, platforms such as Xiaohongshu or Twitter can also be used as channels for recording and publishing product ideas.

Xiaohongshu is naturally suitable for attracting traffic, and the platform places great emphasis on “personality”, with the community as its core.

It is also easier to accumulate the first batch of users through Xiaohongshu.

We can even think about what form of card content is more suitable for the structure of platforms like Xiaohongshu. Perhaps the carrier form of Xiaohongshu will undergo a complete change in the future.

The most important ability for engineers in the AI ​​era

2025-04-12 11:39:26

The core of a programmer is always the process of input and output. The ability to read and write code has always been the most core and underlying ability of a programmer, including structured thinking, logical reasoning, etc.

In the AI ​​era, the ability to read code is becoming increasingly important. AI enhances the process of output, that is, the process of “writing”, which means that the need to write a lot of boilerplate code, standard algorithms, and common functions from scratch will be reduced. However, it is crucial to verify the code output by AI. The code generated by AI is by no means perfect, and there may be logical errors, performance issues, security vulnerabilities, or it may not conform to business logic. The ability of engineers to read and understand these codes is still their underlying literacy. Similarly, debugging code has become crucial, and how to use skills, business experience and tools to debug is the key. Finally, the integration of multiple code modules and engineering design are also very important.

AI-generated code also needs to be maintained. Generally speaking, the maintenance phase is longer than the development phase, and it is crucial to use engineering experience to optimize the maintenance process.

After deep reflection, I think that project-driven learning is the most effective. Participate in the most real, complete, and complex projects, and integrate your own abilities to understand the entire project. In this process, how to use AI to assist understanding, how to use tools to practice and debug, is particularly important.

#field

“Black Mirror” Season 7 Episode 3 “Daydream Hotel”

2025-04-12 11:19:52

The first two episodes are a deep combination of technology and human nature, and the viewing experience is a bit dark. The third episode is my favorite.

The last sentence that impressed me most was: “I have plenty of time.”

Love between same-sex people.

AI can also be emotional and conscious.

AI can also have emotions, love, and provide long-term companionship.

And, it will always be there!

#Movie

“Black Mirror” Season 7 Episode 2 “Black Beast”

2025-04-12 11:06:25

The bully has become God. The nature of the world is not that justice will eventually triumph over evil.

Those who are bullied are always living in a nightmare, a life worse than death; while the bullies are at large.

I thought it was about the revenge of the bullied, but I never expected that it was the bully who killed the avenger.

This world seems to have always been a world where power is king, and it has never changed.

#Movie

Black Mirror Season 7 Episode 1 “Normal People”

2025-04-11 20:19:39

Technology and human nature, the fate of ordinary people seems to be to be exploited and oppressed again and again, unless they transform themselves into exploiters and oppressors. What a pathetic human being.

If individual consciousness is controlled by technology, the consequences will be disastrous. Will we ignore the pursuit of basic human values?

If some of a person’s own characteristics, such as thoughts, consciousness or memory, become commodities to be bought and sold, then how should individual rights be defined?

On the one hand, technology claims to pursue the well-being of life, but on the other hand, it deprives individuals of their freedom. Is the original intention of technology still there?

Does technology make life better, or does it give some people more power?

Ordinary people seem to have become the objects of exploitation by technology. Will the inequality in technological development exacerbate social stratification?

When artificial intelligence can predict, manipulate, and influence our decisions, does our free will still exist?

Do we make active choices, or are we pushed by algorithms?

The ending is tragic, tragedy triggers thinking, and reality, sometimes, has to be tragic.

#Movie

WeChat Product View

2025-04-11 15:10:36

  1. Understand human nature and group psychology. Product managers are like God, building systems and making rules, allowing groups to evolve on their own in the system.

  2. People are reactors to the environment. Products create the environment, which determines the user’s response.

  3. People are lazy, but laziness is also the driving force of innovation.

  4. Fashion is one of the driving forces of human nature, and everyone yearns for fashion.

  5. People lack patience. If your product can attract users at first sight, you are halfway to success.

  6. People don’t like to learn. The decline of blogs to some extent confirms the decline of RSS (Note: this may refer to the tendency of users not to actively configure information sources).

  7. Do not treat groups the same way you treat individuals. The IQ of a group is often lower than that of an individual, and Internet products are aimed at groups.

  8. The value of the “collective” is crucial. If an individual is disconnected from the crowd and isolated, he will feel worthless. Some products are difficult to popularize in China, perhaps because the problem faced by Chinese society is one level lower than “sense of existence” - “sense of survival”. There is a group of people in China who have both the pressure of survival and the desire for a sense of existence. This group of people is the so-called “diaosi”. If you can get them, you will grasp the core of the user group.

  9. Discover the essence of everyday life and touch the “switch”. These switches are closer to human emotions than ordinary switches. Great products should satisfy people’s emotional needs.

  10. Humanization is to understand others by your own experience. You can never fully understand others, you can only truly understand yourself. What you need to understand is the most common psychological activities of people.

  11. Demand is to meet people’s human needs. Don’t mix too much personal morality into the product. The product is a combination of technology and art.

  12. User needs are often fragmented. Don’t just accept what the user says. What’s important is to summarize and come up with abstract rules.

Reference link: Lu Qi’s speech transcript: WeChat product observation

Thinking about fragmented information

2025-04-11 11:45:21

The Internet is full of discussions about the disadvantages of fragmented information. There is no need to elaborate on these. Many answers can be easily searched on xhs. I would like to add some subjective thoughts and understandings on fragmented information.

The explosive growth of the information age has surrounded us with massive and fragmented information. In fact, the amount of information our brain processes every day far exceeds its limit, leading to distraction. Fast-food consumption habits and algorithmic recommendations have further fueled this trend.

However, fragmented information is also shaping our way of expression and creation.

For example, some expressions that touch people’s hearts, concise output creation, and viral dissemination.

AI is a great tool that works well with fragmented information. It not only adapts to the characteristics of fragmented information, but also transforms it into a more meaningful whole in a unique way.

AI is good at connecting scattered points into lines and then weaving them into surfaces. Through pattern recognition and logical reasoning, it can extract clear structures from chaos and give meaning to fragments.

From disordered fragments to an ordered whole, from superficial fragments to deep insights.

Use Notion

| 2025-04-11 11:32 AM

Notion, I mainly use it to record notes that require long-term maintenance and management, or to plan schedules and project management.

Tools for performance analysis of web pages

| 2025-04-11 10:30 AM

Lighthouse is a great tool for analyzing and improving the quality of a web page or progressive web app (PWA).

Lighthouse provides developers with a comprehensive assessment of web page performance, accessibility, best practices, SEO basics, and PWA potential through automated audits. Its report not only gives scores, but more importantly, provides specific optimization suggestions and diagnostic information to help developers locate problems and improve web page quality, ultimately improving user experience.

Firebase Google Code Release

| 2025-04-11 9:35 AM

Firebase Studio is an online AI IDE released by Google. It has an Agent mode similar to tools such as Cursor, which allows you to build the entire project through conversation.

This seems to indicate that Cursor’s Agent mode is a very effective form in the field of AI programming.

Firebase Studio has two modes: Prototype and Code:

  • Prototype mode: mainly through dialogue input, LLM completes project creation and code writing, suitable for initial implementation of ideas.

  • Code mode: It can be understood as VS Code with the Gemini plug-in.

After the initial experience, regardless of the effect, the entire product is logically smooth to use:

  1. Use Prototype mode, mainly AI, to build a draft of the application.

  2. Switch to Code mode, with humans as the main body and AI as the auxiliary body to complete details and complex logic.

  3. Publish with one click.

I feel that this is quite suitable for prototype development. Compared with Cursor, the cost of developing a prototype may be lower. It can be compared to the combination of https://v0.dev/ product and Cursor.

#AI/product

Model Intelligent Scheduling

| 2025-04-10 11:40 PM

By averaging or fusing the weights of multiple models, a single model with better performance can be created. This approach not only integrates the advantages of each model, but also reduces resource consumption and inference time. For example, the Weighted Average Reward Model (WARM) significantly improves the robustness and performance of the model by fusing multiple fine-tuned reward models.

For example, TinyLlama is an LLM model with only 1.1B parameters.

Claude Code Overview Differences compared to Cursor

| 2025-04-10 11:24 PM

Advantages: It is more friendly to developers who are accustomed to command line operations and is more suitable for rapid prototyping and automated scripting. As a command line tool, it can directly access and operate the entire file system and can be seamlessly integrated with various command line tool chains (such as Git, Docker, database tools, etc.).

Cons: The interface is obviously not user friendly.

Refer: https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/overview#install-and-authenticate

#AI/product

Reinforcement Learning & Application Layer Thinking

| 2025-04-10 10:49 PM

Which capabilities of the big model will be internalized by itself, and even reinforcement learning can do better? In which scenarios, the business still needs to do it by itself? There are two very good points: one is the high-quality data in the business. High-quality data requires one’s own business experience to summarize, and some good data can often be handed over to the big model layer for reference.

The essence of a big model is still input and output. In addition to the capabilities of the big model itself, the factors that affect output are also some excellent data in input, which requires your own business capabilities to analyze.

For general capabilities, such as language understanding, image and speech processing, and pattern recognition, large models can generally be internalized by themselves, and the business side can just use them directly. As for reinforcement learning, it can be seen as adding a “dynamic tuning” gain effect to the model.

Some scenarios still require some good method collection. For example, to predict the stock market in finance, the model must understand professional knowledge such as K-line and trading volume.

For specific vertical scenarios, fine-tuning of large models may be a better direction, but the premise is that it is very vertical. For example, in the PDF to Markdown scenario, fine-tuning of large models has a very good effect on the one hand.

Model functions ensure compliance with the first citizen positioning

| 2025-04-10 9:37 PM

Many models have some special features, such as multimodal capabilities, tool calling capabilities, and long text processing capabilities. However, it is important to ensure that the model’s capabilities are universal capabilities. What are universal capabilities? For example, capabilities that all model manufacturers can do and can definitely do better than the business side can be continuously optimized using reinforcement learning.

Some scenarios are beyond the reach of LLM vendors, such as short-term and long-term business memory. These cannot be completely done by relying on the context of the model, but require the business layer to pass on short-term or long-term memory to achieve them.

Information conversion

| 2025-04-10 7:55 pm

Fragmented information/knowledge + fragmented logic constitute human intelligence, while fragmented information/knowledge + fragmented logic + implicit/explicit ethics constitute human wisdom.

Objective existence: data/facts —> organized data

Subjective consciousness: —> Distilled information —> Integration/creation/divergence

We are not producers of data, we are “organizers” or “interpreters” of data.

Data can only be transformed into information through business support. Data is objective, while information is subjective. The purpose of creating information is to reach a consensus.

Agno Multimodal Agent Framework

| 2025-04-10 6:11 PM

The core value of Agno is to provide a framework that can build fast and efficient multimodal agents with memory, knowledge, tools and reasoning capabilities. For developers, its advantages are: reducing boilerplate code, accelerating development cycles, scalability due to high performance, and high flexibility.

#AI/product #AI/open-source

Possession and loss

| 2025-04-10 12:06 noon

The presence of feelings: The ability seems to begin from the fetal stage, those sensory and experiential feelings.

Thinking and Feeling: Children begin to develop the ability of “metacognition”, or “thinking about their own thinking”, at around the age of seven.

The feeling of owning it: Like a newborn baby’s curiosity about the world, a new mobile phone, a new journey.

Feelings of loss: losing a job, losing a loved one, losing a lover…

The value of things and our feelings are relative, and it seems that changes are needed to stimulate us. For example, we are very excited when we first get something, but after a few days, we may become numb.

“Owning” is mostly active in perception, such as choosing to buy a mobile phone. Passive possession means that we own these landscapes and have a pair of eyes to see the world. “Losing” is real, it may be gone suddenly, out of our control, tearing a hole in your accustomed life and forcing you to face changes.

Negativity bias: Human nature seems to be more sensitive to bad things than good things. This may be due to the instinct to seek benefits and avoid harm in order to live longer.

Everyone’s reaction is like a mirror, reflecting their own experience and growth. The world is impermanent, everything is changing, and persistence will bring pain. So, what we should think about is how to face impermanence and turn passivity into an active perception, as if everything can be growth.

#Life/Growth

Information Workflow

| 2025-04-10 11:48 AM

Recently, the tools and information are a bit messy. I thought about a set of tools that I think are suitable and that I like, and tried to build my own information workflow. I have been looking at a lot of products and projects recently, and I have also summarized some methods, which are mainly for information and knowledge management:

  1. Don’t export knowledge, only export thinking: There is too much knowledge, and I don’t want to be a knowledge carrier. And there are many tools that can serve as a second brain, so I don’t need to internalize everything myself.

  2. There is only one scenario for taking notes: when I need to look at something repeatedly or share it. I pay a lot of attention to structuring, how to organize my notes in a structured way, and how to simplify them.

Tool Introduction

  • Notion: Beautiful page, support for networking, a large number of templates. I use Calendar for schedule management, some event management, schedule meeting related, private note taking. Need to do structured classification, less is more.

  • Ticklist: Make some short notes of shorthand and classification, such as prompt notes. And select priorities according to the four quadrants. Unlike Notion’s schedule management, Ticklist does not make schedules, but only priorities, pursuing simplicity and beauty.

  • Flomo: records ideas and thoughts, which are relatively fragmented and exported to blog diaries from time to time.

  • Mac Book Local Files: Classify some books, PDFs, and documents on the computer, mainly for collection and organization.

  • Obsidian: Mainly used for some document-type projects, such as blog writing, article writing and note writing. Generally, it is unstructured or needs to be shared in the form of a project.

  • WeChat Reading: Download e-books with Zlibrary, and then import them into WeChat Reading. The web version works well and has translation.

  • Writing tools: I gave up handwriting after college. In the AI ​​era, I also gradually said goodbye to typing word by word. All model manufacturers support Canvas-like editing interfaces, which are very suitable for AI-assisted editing. Most of them are in Markdown format, which is very convenient for export. Writing without AI assistance feels like something is missing.

  • Collaborative writing tool: Google Docs. If it is related to internal projects, use GitHub/Git directly, Project’s Docs/Collaboration. The content of Docs is easy to migrate to Cursor, and the design of each module is in the corresponding README. For Deep Research, OpenAI and Gemini are both good and suitable for outputting professional documents.

  • Google: the main source of web information, most of the information/knowledge channels. The built-in Reading List collects some commonly used information and knowledge websites, and the Tag is used to mark keywords for search indexes.

  • Folo: RSS subscription for high-quality blogs, and Folo can also help you discover some high-quality blogs. The quality of blogs is generally high, and you can read them in your spare time.

  • Reader: The reading experience is very good, suitable for collecting books and highlighting wonderful passages, and it feels somewhat similar to Flomo.

⚠️ The way to distinguish Obsidian from Notion is actually very direct to me: Notion uses templates more for notes and schedulers that need to be supplemented frequently; Obsidian often means writing a document to represent a one-time completion.

AI Tools

In fact, each has its own characteristics:

  • Gemini: I feel it is the strongest at this stage, and the search ability is very powerful. I use the membership, and the Deep Research effect is very good. It is very suitable for writing professional documents, design documents, and learning documents.

  • DeepSeek: It is rarely used, but it is very useful for writing paragraphs.

  • OpenAI: I feel that the membership is not worth buying. Occasionally, I will use the free version for some minor problems, which is very fast.

  • Grok: He has a high EQ, can talk about philosophy, be a friend, give some advice and so on.

  • Claude: He is very good at writing code. I always ask Claude for help with some coding issues.

#Field/Tool Usage Tips

Complete the task with extreme ease and perfection

| 2025-04-09 11:07 PM

Gamma (gamma.app) is dedicated to minimizing the steps to generate PPT. It only takes very simple steps, such as pasting text or importing files, and it can quickly help you generate summaries for each module of the corresponding PPT. You can modify the summary appropriately, and AI can immediately help you generate a complete PPT.

  • It is easy to use and the interface is very “modern”, which is suitable for making things that look advanced at first glance.

  • Write some prompts and it will generate a complete presentation by itself, with graphic and text layout also generated automatically.

  • More flexible than traditional PPT, you can view it as a web page, a slide, or a document, and it can be shared quickly.

https://gamma.app/

#AI/product

Emotional Value AI

| 2025-04-09 8:18 PM

Duxiang AI is a very interesting product. It combines note-taking applications with AI technology and adds a bit of role-playing flavor, which feels quite fresh.

The most eye-catching part is the AI ​​virtual characters - they will jump out to interact with you, comment on your notes, chat with you, or even “argue with you”, just like the little friends in your phone.

Generally, note-taking apps are just cold tools, but Duxiang AI is different. It lets AI characters come out to “chat” with you about your notes. For example, if you write about your mood for the day, they may comfort you or discuss your thoughts with you. This kind of interaction is especially suitable for those who want to talk to someone but don’t want to post it on WeChat Moments (such as introverts).

AI is a long-term companion. It will remember your notes and conversations with you. This unique personalization feels great and caters to people’s emotional needs.

A tool is not just a tool, it can also be a partner, a companion who accompanies you and can also have emotional value.

By incorporating AI into daily tools and coming up with some new ways to play, the Xiaohongshu community can create very strong emotional value.

Refer: https://www.duxiangai.com/

#AI/product

Front-end cross-platform selection

| 2025-04-09 2:34 PM

React Native is based on JavaScript, and the JavaScript ecosystem has already seen the emergence of many related tools and libraries, such as TensorFlow.js and Brain.js, which allow developers to easily integrate AI features into applications. Taro is a multi-terminal development framework based on React that can compile React code into multiple platforms such as WeChat applets, H5, and React Native.

But generally speaking, the React framework is used as a springboard to React Native.

React Native’s tagline is “Learn once, write everywhere”.

One of the prerequisites for using React Native to develop the Web is that the Web serves as a supplement to the App.

In fact, you can also consider using something similar called react-native-web to convert the RN project into code that can run on the Web, but it also has fatal flaws and the user experience may be slightly inferior, so there may be some fine-tuning.

Another common method is to design the Web UI and the App UI separately, which will provide a better user experience. This will force you to separate the business logic and UI components, forcing you to separate data and display, and also allow the entire product to have a better user experience on different platforms.

The project structure design of React is very important. How to choose a scaffolding, I generally make the following two judgments:

  1. If you need server-side rendering (SSR), static site generation (SSG), file system routing, API routing, then Next.js is very suitable.

  2. If it is mainly client-side rendering, Vite + React is a good choice.

After selecting the scaffolding, the next step is state management. State can be understood as a certain state of the application’s data and interface.

  • Small scope (within a component/small amount of sharing): useState / useReducer / Context

  • Medium scope (multiple modules share state): Zustand / Jotai

  • Large Scale (Complex Business Logic): Redux Toolkit / Recoil

UI library (optional): Material UI (MUI), Ant Design, Chakra UI, etc. can speed up development and provide a consistent visual style.

Zustand is actually quite suitable for most projects.

Code structure is very important. For React, it is recommended to separate UI components, business logic (Hooks), state management, API requests, and utility functions. Write business logic as pure JavaScript/TypeScript functions or Hooks that are platform-independent as much as possible.

Is social hostility getting worse?

| 2025-04-09 12:41 p.m.

Recently, I do feel that social hostility seems to be getting worse, and the world is not at peace. Violent incidents, cyber attacks, group confrontations and other phenomena are often heard, and conflicts and crises are also occurring one after another internationally.

I feel that under economic pressure and uneven distribution of resources, the “evil” aspects of selfishness, greed, and anger are easily magnified. Society is like a mirror, reflecting certain distortions of human nature in difficult situations.

The widening gap between the rich and the poor and the solidification of classes may make some people at the bottom feel powerless to change, while people in the middle may worry about falling.

The Internet sometimes makes negative emotions spread faster and wider. When you turn on your phone, the screen is filled with arguments, accusations, and even rumors. People hide behind the screen, lacking face-to-face empathy and more impulsive venting.

The fragmented information on the Internet seems to have made people less patient with each other.

Good and evil are not innate, but are more often sculpted and influenced by the social environment.

Society seems to have not changed fundamentally, it is just that we are influenced by the environment we grow up in. The world will not automatically become more peaceful, the form of conflict may have changed, but the nature of the struggle seems to have not changed.

Thinking about consistency

| 2025-04-09 11:40 AM

Why do we subconsciously feel uncomfortable when we see certain behaviors? Perhaps it stems from a sense of morality, or perhaps it is the pursuit of consistency. But in any case, these feelings themselves, in the absence of clear rules and constraints, do not seem to be enough to be an absolute reason to criticize others. For example, when we see someone littering cigarette butts, from a moral perspective, what we don’t want the other person to do is often a manifestation of “do not do to others what you do not want others to do to you.” From the other person’s perspective, this seems to satisfy their own moral standards. But on the other hand, the collective will give birth to public values, such as protecting environmental hygiene. So if the city has clear rules prohibiting littering, then this behavior violates public moral standards.

What is the underlying conflict? Perhaps it is a conflict between my personal values ​​and the behavior of others. I may be a perfectionist, but the other person may be just solving the problem in the short term. I may be used to an environment that is more focused on details and rules, while the other person may come from or identify with a more relaxed lifestyle. This difference makes me realize that the problem is not only the behavior itself, but also our different understanding of “how we should live.”

People seem to be very good at imitating behavior, so this process is often a process of leading by example. If you show some good qualities, it will naturally affect the environment.

In daily life, consistency usually refers to the coherence of behavior, thought or state in time or situation. In fact, in essence, consistency is more like a projection of one’s inner self, rather than just a requirement of the external environment.

True consistency may not be about requiring the external world to completely meet my expectations, but about reaching a certain reconciliation between my inner self and external reality.

Your own limitations & the limitations of others

| 2025-04-09 11:28 AM

There are two aspects worth thinking about: one is that your own perspective is limited, and the other is that the other person’s perspective is also limited. Furthermore, the perspective that you can see may not be seen by the other person. So many times, you have to learn to reconcile with this “limitation”. Respecting the rules and respecting the other person is also respecting yourself to a certain extent.

Cycling Accident Thoughts

| 2025-04-08 4:15 PM

Objective record: I was driving on a cycling road in Liangzhu at a slightly faster speed. I encountered a narrow intersection and a downhill slope. I hurriedly prepared to slow down, but it seemed a little too late. Unexpectedly, a car suddenly appeared on the right side and hit me in the bicycle lane. The other party rushed into the non-motorized vehicle lane from the branch road/alley/sidewalk, and I was driving normally in the bicycle lane. Due to the limited vision at the intersection, the other party did not see me passing by. There were no traffic signs or speed reduction reminders at the intersection.

Other party’s behavior: The other party was a couple, and they were a little cold. They pushed the blame on me, saying that I was driving too fast and that I didn’t wear knee pads. After the accident, they seemed to be more concerned about the extent of the damage to their car rather than my injuries. In addition, they analyzed the road conditions and thought that the design of this intersection was unreasonable. The elderly people in the community next to them also seemed to stand on the other side, talking to them about how to get insurance, and did not objectively propose a solution from a global perspective.

My behavior: I was a little dazed after being hit, and my body was in some pain. I first moved the car to the non-motorized vehicle lane to avoid blocking the normal flow of traffic. Then I felt my physical state and described it truthfully. I started to clean up the damaged and bleeding areas and cleared up the bloodstains.

What I didn’t do in time (deficiencies): I didn’t take full photos/videos immediately (including scars, vehicle damage, accident location). When the police arrived, they should have asked for the accident responsibility determination letter.

After going to the hospital, I made a clear assessment of the three injured parts. There was uncertainty about my right leg, so I asked for an X-ray, and the results showed that it was fine. The doctor applied some iodine to me and asked if I was working and if I wanted to rest for a few days. I could hear the hidden meaning in his words, and I said no, I don’t have a stable job at the moment. I am a remote worker, and I don’t think it affects my work, so I don’t think it’s necessary. The doctor sighed, and I left. The doctor’s concern seemed to go beyond the scope of his duties, which touched me a little.

Cycling is like a mirror, reflecting the complexity of human nature.

What kind of human nature will be revealed in difficult situations? How should we deal with these unexpected setbacks?

There is often an instinct of “seeking benefits and avoiding harm” hidden in human nature. When interests are threatened, empathy sometimes becomes a luxury.

Reality will not automatically stand on your side just because you are a victim. It requires you to provide evidence and understand the rules.

Reason and preparation are sometimes more important than emotions. They teach you how to face problems and setbacks.

The essence of growth seems to be to find one’s own rhythm and temperature in the midst of impermanence.

Su Shi found his life direction

| 2025-04-08 10:05 AM

Su Shi grew up in Confucian idealism. The importance of family education is self-evident, and family style is a kind of inheritance.

He was moved to tears by Fan Pang’s deeds, and in his extreme emotion he seemed to have found himself, sighing, “I would rather die for him.”

His father Su Xun told Su Shi to think about the world. At first, his original intention was to be loyal to the emperor and serve the country as per Confucianism, and he was willing to die for his ideal.

📜 “We study to serve the common people.”

Family tradition may be the deepest starting point of a person and the most lasting foundation.

It affects how we judge right from wrong, what is worth pursuing, and how we survive the low points in life.

Su Shi once wrote to his son: “If you really want to learn poetry, the effort lies outside of poetry.”

This sentence seems to be about poetry, but it actually talks about being a human being. Real “knowledge” often comes from life, character, and family tradition.

Thoughts on Digital RMB

| 2025-04-07 12:13 p.m.

In essence, the digital RMB is a brand-new payment method built on specific technology. Through the circulation of digital currency, the central bank can more directly influence economic activities, thereby enhancing the transmission efficiency of monetary policy.

Digital RMB helps promote financial inclusion and makes it easier for remote areas and low-income groups to access financial services. At the same time, it reduces the use of cash and can effectively reduce the risk of counterfeit currency circulation and illegal activities such as money laundering.

When business and consumer confidence is low, even if the central bank increases money supply (M2) through interest rate cuts or quantitative easing, funds may remain in the banking system and cannot be converted into real economic activities. In this case, the growth of M2 may be out of line with economic growth, while M1 (cash and demand deposits) may more directly reflect the economy’s immediate liquidity needs.

M1 may increase in importance, but M2 still provides information on the total money supply and underlying liquidity.

In the economic downturn, interest rate cuts may continue, but housing prices will be affected by multiple factors (such as policy intervention, demographic changes, etc.). China’s real estate market may enter a period of adjustment, but it may not be a complete collapse.

Ideology

| 2025-04-07 9:29 AM

The same thing has its own perspective in different industries.

Similarly, there are various perspectives on people and their behavior, which require the support of knowledge from multiple disciplines, such as psychology, law, sociology, etc. You can’t go wrong by studying hard and expanding your cognitive boundaries.

Establishing a thinking system refers to the process of observing, analyzing, and making judgments when faced with an unfamiliar thing or a new thing.

Essentially, thinking is the process of processing information - from receiving input (observation), to breaking it down and analyzing it (understanding its components), to integrating it into the output (forming a judgment or conclusion). In this process, it is very important to remain open-minded, look at things from multiple perspectives, and combine existing knowledge and experience.

Forget the Beginner’s Intentions

| 2025-04-07 9:17 AM

Don’t forget why you started just because you have gone too far.

If you don’t keep your original intention carefully, it’s easy to lose it in the chaos…

casual

| 2025-04-06 10:27 PM

“Hey, what do you want to eat?” “Whatever…”

“Whatever”?! When I asked you, you said “whatever”?! Have you developed the habit of giving up analyzing problems, judging problems, and expressing your own wishes?

What you like and what you don’t like should be expressed clearly.

If you don’t like it, express it.

If you like something, express it.

The phrase “whatever” may put pressure on others to make choices and may not necessarily meet their real needs.

Sometimes, “casual” may be a kind of irresponsible to one’s own wishes.

More real people

| 2025-04-06 9:23 PM

I think I am pursuing sincerity, but sometimes I am unconsciously afraid of those who are “more real”.

It is better to refuse than to agree reluctantly.

Learning to say no, although difficult, is also part of being your true self.

“Acceptance” is also difficult - accepting fear, accepting the courage to refuse to compromise, and even accepting the possible disconnection from mainstream society. Learn to say no, accept your own vulnerability, and pursue inner freedom.

Meeting more authentic people is like holding up a mirror to oneself, prompting reflection.

It’s not that I don’t like material things, but I value inner freedom more.

They live thoroughly and without hypocrisy, but this may make them seem incompatible with some of the inherent patterns of this society.

Good and Evil

| 2025-04-06 7:03 PM

How to understand “non-good”? I saw Spinoza said in “Ethics”: “Scorn, contempt, anger, revenge… these emotions are related to hatred or contain elements caused by hatred, and cannot be good.”

This leads us to think about what is the essence of goodness? If these strong emotions are not goodness, then is goodness more of a peaceful, rational and constructive force?

precise

| 2025-04-06 5:05 PM

Accuracy is the most powerful defense.

No matter how much rhetoric, subjective assumptions and prejudices are used, it is difficult to avoid the power of accuracy. The government’s information is not transparent, sometimes precisely to avoid the scrutiny brought about by “accuracy”, thus leaving room for manipulation.

Accuracy is the most important skill in the field of truth-seeking. Self-motivation and emotions are often the biggest enemies of accuracy. The truth is often lost in tears.

Accuracy is very important. Behind accuracy is logic and reasoning, which is the cornerstone of decision-making and judgment.

I just hate giving in

| 2025-04-06 4:14 PM

“No matter how powerful you are, including a national ministry, when you are sued in court, you are the defendant and I am the plaintiff. We sit opposite each other, and the judge is in the middle. You and I are equal.”

In the face of powerful forces, people often seem to have no choice but to obey, but I am unwilling to give in easily.

What is the difference between citizens and ordinary people?

A person who can express his or her opinions independently without being arrogant, who can obey (the rules) with integrity without being servile, who can actively participate in discussions on public affairs, who feels sympathy for the weak and anger when seeing injustice - I believe that this is the kind of person who can be considered a true citizen.

Rights are meant to be enforced, otherwise they may just be a piece of paper.

How to understand homosexuality

| 2025-04-06 3:59 PM

Homosexuality is neither a disease nor a psychological defect, but a natural manifestation of human sexual orientation.

The education and social environment we receive generally convey a traditional concept of gender and sexual relations, which often regards heterosexuality as the “normal” or “only correct” model, which may lead to misunderstanding or even rejection of other models.

The rejection we may have in our subconscious mind is often because we don’t really understand. We may choose to be tolerant and respectful, but we may not understand this group and this phenomenon. This emotion is not derived from objective facts, but is more the result of cultural traditions and social environment.

When facing death, there are stronger emotions than this that control people

| 2025-04-06 10:49 AM

Facing death multiple times or simply thinking about it can spark fear, love, regret, or a search for meaning, all of which stem from our own instinctive thoughts about the meaning of life.

The brevity of life forces us to think about its meaning and value of existence, and how we give them meaning.

Perhaps what transcends fear and death is an awakening to the present moment. When we see someone shouldering a responsibility that is not their duty and risking their life to save others, we feel the touch of the truth of life, the preciousness of friendship, the admiration for kindness, and the desire to understand and connect with others.

To be alive is to be alive, and a beating heart represents being alive itself.

Life precedes meaning, existence precedes essence.

Heartbeat needs no explanation; it is a fact in itself, a state that requires no proof, without any philosophical embellishment or emotional attachment.

Forget Me

| 2025-04-06 10:42 AM

When you ask a question, do you expect an answer? If you don’t, maybe don’t ask it.

Only when a person truly cares about others can he temporarily forget himself.

The “self” is not a fixed entity but a temporary aggregate of our experiences, emotions, cognitions, and social interactions.

When we engage in deep reflection or meditation, we begin to question and transcend this conceptualized self, and we may experience a state of “no self” - that is, we no longer tie external phenomena to our inner self, but directly perceive the nature and fluidity of things.

If you look at a problem with a preset answer, then it may not be so important what the answer is.

Tennis ball and racket

| 2025-04-05 10:01 PM

A racket is to tennis what a parent is to a child, or a mentor is to a student - one party provides “external force” and guidance, and the other party, under the influence of this external force, has the opportunity to fly higher and farther.

The experience and knowledge they possess are like the “face” of a racket, which can catch us, protect us, and “hit” us towards higher goals.

At the same time, we must also work hard to learn the skills of “playing tennis” so that the ball can bounce faster and higher under the action of the racket face.

Our transition from “passively being hit by the racket” to “actively cultivating racket techniques and one’s own hitting ability” means that we gradually transform from relying on external forces to cultivating and applying our own abilities.

From being supported by others, to learning to grow oneself, and then to giving back to others or the next generation when one is able to do so - this is an inheritance of love and a continuation of life wisdom.

Thoughts on the RCEP Agreement

| 2025-04-05 3:51 PM

The RCEP (Regional Comprehensive Economic Partnership) covers about 30% of the world’s population and GDP, and is expected to increase the member countries' total annual revenue by about US$200 billion by 2030.

RCEP is one of the largest free trade agreements in the world, negotiated between the ten ASEAN countries (Indonesia, Malaysia, the Philippines, Singapore, Thailand, Brunei, Vietnam, Laos, Myanmar and Cambodia) and China, Japan, South Korea, Australia and New Zealand.

RCEP provides China with a more stable trade environment and a broader market, which will help promote domestic industrial upgrading and the “going global” strategy. However, it also requires China to make further improvements in rule-making, intellectual property protection and market opening to meet the requirements of the agreement.

As globalization develops, global trade competition is also intensifying. In the short and medium term, the trend may still revolve around the two major countries of China and the United States, and the trade relationship between them will become the focus. On the contrary, there may be certain development opportunities in these small countries with regional cooperation.

Thoughts on the Global Trade War

| 2025-04-05 3:44 PM

Trump has long advocated for tariffs to reduce the U.S. trade deficit and boost domestic manufacturing, claiming that the United States is being “exploited” by its trading partners.

The Sino-US trade war has been going on for many years. If the tariffs are fully implemented this year, it is estimated that the global GDP growth rate in 2025 will be reduced by about 1.2-1.3 percentage points.

For China, this could drag down GDP by more than 2.0 percentage points. For the United States, it could drag down its GDP by 1.3-2.0 percentage points and push up the core PCE (personal consumption expenditure deflator) by 1.2-2.6 percentage points.

There may be a point of negotiation, that is, countries can achieve relatively good results through negotiation and promote a win-win situation. But in reality, China does not seem to do so easily. Both sides may use “fighting” to promote “peace” in order to gain more favorable percentage points.

Governments may take measures such as lowering the reserve requirement ratio and interest rates to inject more liquidity into the market, lower the financing threshold, and thus promote economic growth.

In the medium term, the world economic structure may show a trend of coexistence of bipolarity (China and the United States) and multi-center. Economies such as the European Union and India may seek a relatively independent multi-center status.

Generally speaking, in international trade conflicts, physical assets such as gold often have a certain role in preserving and increasing value, and ordinary investors can make appropriate allocations. In addition, asset diversification is also very important, that is, to spread assets into multiple baskets. Purchasing foreign exchange assets such as the US dollar is sometimes used to hedge against fluctuations in the exchange rate of the domestic currency and domestic market risks.

Thinking about Serverless in the AI ​​Era

| 2025-04-05 10:20 AM

Serverless is a cloud computing architecture model that allows developers to focus on business code without having to manage underlying server resources.

The core idea is that small, independent functions are charged according to the number of calls and execution time, in order to achieve high resource utilization and cost optimization.

Function is a basic unit. Developers write single-purpose functions that are triggered by events, run automatically, and close automatically after completing their tasks, allowing developers to focus only on business logic.

The usage scenario is suitable for building RESTful APIs, triggering functions through HTTP requests, and automatically scaling on demand.

Serverless can also be used to perform background tasks, including scheduled tasks.

However, it should be noted that Serverless is usually stateless, so it may not be suitable for operations that require maintaining connections for a long time.

MCP Thoughts in ApiFox

| 2025-04-04 7:49 PM

There are several major modules that I think are relatively important. In fact, Cursor MCP can integrate functions. For example, in Cursor, you can easily obtain API documents, analyze and query your own database, and easily submit PR to GitHub through chat automation after modifying the code.

For some prompts tagged with the production tag, such as the “Order Status Query” prompt template, after the MCP server receives the request, it can compile the template stored in Langfuse with the passed in variables to generate a complete, customized prompt.

In addition, for MCP in ApiFox, we can use MCP to access various API documents, let AI interpret the interface, and assist in implementation.

Code Value

| 2025-04-04 10:08 AM

Code that has been tested in production environments is usually more valuable than untested code.

Software also has a life cycle, which is normal.

Maintaining software will continue until the end of its life cycle, or until the maintenance cost is too high and refactoring is chosen.

Ukrainian and Russian mercenaries thoughts

| 2025-04-03 3:19 PM

On the Russian-Ukrainian battlefield, some Chinese fought for Russia, while others fought for Ukraine.

The two sides hold different positions and values.

Good and evil in war are very complicated, and human nature is particularly contradictory at this time. Some people become extremely cruel in war, and seem to be able to do anything bad; but there are also people who want to help others even if their own lives are at stake. Therefore, human nature in extreme situations is like being magnified. At a critical moment, many hidden characteristics may be revealed.

Some people think that people are naturally kind, but the environment has forced them to be bad; others say that people have selfish genes, and war just brings out this selfishness. It is the same war, but everyone makes different choices. War may not have distorted human nature, but it has made it more real.

Good and evil are not absolute, and neither are right and wrong. The scales never seem to be perfectly balanced, but it is precisely this imbalance that causes ordinary people to struggle with human nature and survive in the cracks, showing their complexity.

Peace is truly precious. Without the mirror of war, would we forget the possible evil in human nature, as well as the good that may shine?

Learn to forgive and know how to love

| 2025-04-02 3:47 PM

If your heart has no place to rest, it will feel like you are wandering wherever you go.

But perhaps only by learning to forgive can we truly understand how to love.

Discussion with Archer on the last night

| 2025-04-02 11:41 AM

We stayed in Shenzhen for ten days, and also stayed in Archer’s home for ten days. Archer often treated us to meals and took us to taste some authentic local food. Xiaomei also treated us to meals, and Sun also treated us to meals. This trip in Shenzhen really touched me, and their friendship deeply touched me.

Thoughts on the Nepal Unrest

| 2025-03-30 10:43 AM

(The following content is speculative analysis based on reference information)

The main reason for the turbulent political situation in Nepal may be the long-term political instability and poor governance, as well as the possible rapid government succession and widespread corruption.

This reflects the Nepalese people’s deep pursuit for stability.

From China’s perspective, it may be more inclined to support governance methods with relatively concentrated power and clear succession systems, because frequent political changes under a republican system may sometimes lead to inefficiency.

It’s like the balance of a scale has been tipped.

Despite this, the chances of a monarchy restoration remain low at present.

Although the repeated changes of government and institutional innovations have brought a lot of chaos, they have also formed a relatively stable power structure and political operation mechanism. Restoring the monarchy not only requires broad public support, but also must overcome major obstacles at the constitutional and political levels.

In Nepal, major political forces, the military and the international community mostly tend to maintain the existing political system.

In addition, even if the monarchy is restored, it will not fundamentally solve the existing deep-seated problems.

The key lies in whether true political stability and efficient governance can be achieved.

Ref: https://news.cctv.com/2025/03/21/ARTIT8r5j2i9WWruRsLiTuD2250321.shtml (Note: This link is a hypothetical reference)

The essence behind the pursuit behavior

| 2025-03-29 11:20 AM

On the surface, the question of who to pursue is a personal preference, but if you dig deeper, it actually reflects a person’s inner pursuit of “meaning” and understanding of “existence.”

Star-chasing is often driven by emotions. Stars often carry a certain idealized image - beauty, vitality, success, or simply a fantasy that allows people to temporarily escape from reality.

People who follow scientists are usually more rational and admire human intelligence and the spirit of exploration. Scientists represent the conquest of the unknown and the persistence of truth. People who follow them may care more about the grand narrative of “the overall progress of mankind”, which does require a little love and patience.

People who follow literary writers are often attracted by the depth of stories, emotions and language. Writers explore human nature and the light and darkness of life through words. People who follow them may be looking for resonance or trying to understand themselves through the writing of others. Literature lovers feel more introspective and are willing to find answers in complexity.

There is no distinction between high and low in human nature. What everyone pursues is just a projection of their inner needs. What is important is whether the pursuit is sincere and whether one can gain something and grow from it.

Parents’ experience and children’s choices

| 2025-03-29 11:09 AM

Our parents often just want to use their own experience to seek happiness for us as they understand it.

However, these valuable experiences may not be fully applicable in the future.

Ultimately, the path depends on the child’s own choice.

What parents can do is to provide a warm “safety net” rather than setting insurmountable “limits”.

Design considerations for event bus and message event pipeline

| 2025-03-29 11:08 AM

Modules are loosely coupled, with information passed through events rather than direct calls. In this way, each module only needs to focus on its own business logic without having to understand the implementation details of other modules.

The event bus acts as an intermediary, allowing each module to be developed, tested, and deployed independently.

When you need to add new modules and functions, you only need to register them in the event pipeline.

The most painful lessons in AI development

2025-03-25 17:45:12

The most painful lesson in AI development is to neglect commit. Be sure to: **Commit frequently, record frequently; describe well, and make your meaning clear. **

Web RTC Thoughts

2025-03-23 21:38:00

Web RTC technology provides a strategy for end-to-end communication and has great potential in the field of Internet of Things.

Noteworthy projects include musetalk and musev.

Currently, some RTC policy practices are supported:

  1. OpenAI: Provides Realtime API.

  2. Hugging Face: Launched Fast RTC, an open source WebRTC library that also supports Websockets and integrates STT (speech to text) and TTS (text to speech) functions.

What is the most important thing in family education?

2025-03-22 17:19:31

A friend mentioned that in many European and American families she knows, children have the freedom to explore the world and the right to choose and pursue their hobbies. This triggered a reflection on the Chinese and Western education models.

Is the Western “happy education” really perfect? ​​It emphasizes independent thinking and independent living, and lets children take responsibility for their own lives. We are deeply influenced by Confucianism, and we place more emphasis on collective sense, family responsibility and stability. Some habits of Western teenagers, such as smoking, contact with marijuana, and drinking, may be difficult for us to accept.

True education should maintain a bottom line in freedom and a sense of belonging in independence. Parents need to build a system that guides the general direction of growth and ensures a safe bottom line. The best education should be an education system based on a deep understanding of the world and in line with the future development trend of the world.

Merely being exposed to nature is not the core. Cultivating children’s moral character, outlook on life and values ​​is crucial. The key is to let children understand the complexity of the world, and parents need to lead by example - the transmission of experience is far better than empty words, and teaching by example is more important than precept.

Every family has its own way and rights to educate its children. There is no absolute right or wrong.

Metaphysics

2025-03-22 13:34:40

“The metaphysical is called Tao, and the physical is called Qi.”

In Western philosophy, metaphysics explores abstract topics such as the nature of existence, the origin of the universe, time, space, cause and effect, soul, and God. This is the ultimate question of human beings about fundamental issues and the core of rational philosophy.

In short, metaphysics studies the “invisible essence” - it is not only concerned with “what things are like”, but also asks “why are they like this and what is their essence”.

It is different from metaphysics, and is more inclined to systematic and abstract philosophical speculation. Metaphysics often carries the color of intuition, mystery or cultural tradition, and is less systematic and rational than metaphysics.

For most people, metaphysics may be more helpful in daily life because it is closer to emotions and intuition. The difficulty of metaphysics lies in its requirements for intuition, understanding, and spirituality. It is more like a kind of “Tao” practice, full of mystery, and enlightenment seems to require fate.

The ultimate of sensibility and rationality

2025-03-22 13:18:46

The binary opposition of things in the world is not the true state of the universe, but more of a construction of human language and thinking. Right and wrong often depends on the observer’s position, especially in moral issues. Taoists also say: “Misfortune is the foundation of blessing, and blessing is the root of misfortune.” Good and bad are part of the dynamic balance.

The “wave-particle duality” in physics reveals that microscopic particles (such as photons and electrons) have the dual characteristics of waves and particles, which shows that nature at the microscopic level is far more complex than we imagined and cannot be summarized by a single property.

Kahneman proposed the dual system theory in Thinking, Fast and Slow:

  1. System 1: fast, intuitive thinking, almost emotional.

  2. System 2: Slow, analytical thinking, almost rational.

Intuitive thinking is more efficient and consumes less energy. In daily life, many behaviors, such as talking without thinking, are driven by intuition.

“No logic” itself is also a kind of logic. There is also an operating mechanism behind sensibility. The creation of intuition and emotion, its selection, style, and expression, actually contain deep “sensory logic”. When sensibility reaches its extreme, it can break through the superficial chaos and narrow self, and turn into a deep transparency and cognition - this is another form of “rationality”.

The Taoist pursuit of transcending the constraints of concepts to achieve a peaceful state is precisely the holistic thinking that hopes to return to the origin of all things. The core of this thinking is “letting go” - letting go of obsessions, letting go of the self, and finally becoming one with the “Tao”.

Borrowing and plagiarism

2025-03-22 00:11:10

How to define plagiarism and borrowing?

As the saying goes, excellent product design will eventually reach the same conclusion in the public aesthetic: everyone loves a beautiful face, and a well-proportioned figure is always pleasing to the eye. 😊

**What is plagiarism? ** Plagiarism is thinklessly copying other people’s ideas, which only stops at imitation and fails to achieve transcendent innovation.

**What is reference? ** Reference is to have a unique idea of ​​a certain form or beautiful appearance after long-term thinking; after thinking hard and trying repeatedly, you occasionally get a flash of inspiration - you catch a glimpse of the touching actions in the corner of the screen, and suddenly realize that this is the inspiration you have been looking for for a long time!

Understanding is not copying, but enriching one’s own vision through dialogue with others.

Therefore, the beauty of reference is that it is not to plunder other people’s creativity, but to borrow other people’s sparks to ignite your own flame.

Just like listening to a sad song after a breakup, a certain verse touches the depths of your heart, and that emotion is exactly the resonance you have been looking for…

Creativity is often the accumulation of collective wisdom. The key is how you digest these inspirations and make them your own. Ask yourself: How does this idea connect to my experience, emotions, or goals? What unique value can I add to it?

For example, if you read a poem, write it down, and repeatedly savor and ponder it, it will gradually integrate into your thought system and become a part of your unique philosophy. What touches you is not the poem itself, but the inner feelings and thoughts it awakens.

It’s better not to meet each other

2025-03-22 00:10:21

If meeting each other only brings embarrassment, burden or even pain, then what is the meaning of deliberately maintaining a relationship? Is it to satisfy one’s own emotional needs, or to cater to the expectations of others (such as the expectations of the mother)? Behind this, perhaps there is a deeper question: Is the true meaning of a relationship something that is actively sought, or does it flow naturally?

Dog’s way of thinking

2025-03-21 23:29:55

In the past, I was occasionally barked at by puppies, even those seemingly docile dogs in Nepal, which often confused me.

After my friend got a dog, I observed that when dogs compete for food or when a female dog is in estrus and a male dog approaches, they will whimper and growl, which is a signal of rejection. When they whimper, I initially thought they had a cold, but in fact they were expecting help - they wanted food or wanted to go out and play.

There is indeed a way for dogs to communicate with each other, and it is not difficult for us to identify:

  • Looking directly into a dog’s eyes is often seen as hostility. If you habitually look a dog in the eye the way humans do, it is normal for them to be frightened.

  • They urinate to mark their territory and sniff each other’s buttocks when they meet to distinguish between their owners, the same gender, estrus status, mother-child relationship, etc.

  • The tail is a barometer of mood: wagging the tail means happiness, tucking the tail means fear, and hanging the tail means depression.

  • Licking each other’s mouth corners is a way for dogs to greet each other, usually indicating a close relationship.

  • Sometimes, one dog will bark when it notices something unusual, and other dogs will respond after hearing it. This is their way of passing the alarm.

This is an interesting observation about the way dogs express themselves. Language is the way we understand and interact with the world, while tone, pitch, and movement are the way dogs understand and interact with the world. Often, our understanding and tolerance are based on the insight into the logic behind the behavior.

Tolerance is the journey from misunderstanding to understanding - not rushing to judge, but exploring the motivations behind the behavior.

Dogs do not hide their emotions, their tails wag and tuck, showing their joy and anger; but people often hide their intentions, and language becomes a mask. In comparison, the “truth” of dogs is more pure.

Survival logic behind behavior: Why are most dogs in Nepal relaxed and lying lazily on the roadside during the day, while dogs in China often guard their homes and bark? The reason behind this has always made me ponder.

Dogs are also a product of the environment. Most of the dogs in Nepal are free-range and have coexisted with humans for a long time. The environment is relatively relaxed, the living pressure is low, and they are in harmony with the surroundings. Domestic dogs are mostly domestic, and stray dogs on the street are rare. Domestic dogs are kept in captivity and have a strong sense of territory. When strangers approach, they instinctively bark to protect. Their survival depends on their owners, and the environment is no longer symbiotic, but segmented. This sense of tension makes their expression more vigilant.

Existence is always intertwined with the world, and there are different ways to adapt to the world: one is accepted freedom, and the other is limited protection.

Contextual Thinking

2025-03-20 18:42:32

Context is the key part that the application can control and adjust by itself, which is crucial for human-computer interaction.

Take Cursor as an example. It controls the scope of code documents through context, such as specifying a website, a specific file, an external web page address, or even the context of git. It is worth noting that these rules are configurable and can even be controlled more finely by setting MCP (Model Context Protocol).

Cursor MCP

2025-03-20 01:48:17

Cursor’s Model Context Protocol (MCP) is more like a plugin system protocol that strives to standardize the way applications provide context and tools to large language models (LLMs).

Interestingly, Cursor’s MCP seems more interesting and perhaps even more valuable than Claude’s similar mechanism.

We can imagine that the MCP of the future AI platform is also a pluggable client system. It allows users to install and manage MCP servers, and provides a graphical user interface (GUI) or command line interface (CLI) for configuring and managing these MCP servers.

MCP supports multiple transport protocols (such as stdio and SSE), which can meet diverse needs. Its ability to connect to various data sources is particularly noteworthy.

Fireworks and Silicon Mobility / OpenRouter

2025-03-19 16:15:11

If we compare Fireworks AI with the domestic Silicon Flow or the overseas OpenRouter, we can see them as two AI service platforms with different positioning. The former is more like a “mastermind” for terminal applications, while the latter is more focused on deepening technical details and meeting high-end customization needs.

A notable feature of Fireworks AI is its emphasis on allowing end users to intuitively see the model’s reasoning process, parameter settings, and various options, making operations more transparent and flexible. In addition, it is committed to bringing together various models and services into a unified interface to facilitate direct calls and combinations by business-level users.

Understanding Personality

2025-03-19 14:56:03

The word “personality” has a wide range of connotations, perhaps including “personal charm”. Personality usually refers to the unique patterns of individual thinking, emotions and behaviors, as well as the psychological mechanisms behind these patterns, which may be hidden or obvious. “Character” is a more popular and everyday expression.

Some thoughts on human nature

2025-03-18 16:04:59

From the perspective of Western individualism, it seems that in most cases, no matter how serious the mistakes are, people are not inclined to blame themselves. This is in contrast to the traditional Eastern culture, which is deeply influenced by Confucianism, which places more emphasis on humility, reflection and self-improvement. To a certain extent, Western culture emphasizes self-worth and self-motivation, and it is easier to maintain self-esteem through positive self-suggestion even after failure.

“I would rather blame everyone than repent easily.” People are afraid of being blamed, and the underlying reason may be the fear of not being understood. If we can replace blame with understanding and put ourselves in the other person’s shoes to think about why he is like that, then sympathy, kindness and tolerance may arise.

Don’t criticize, don’t blame, don’t complain.

When you get along with others, you can reflect on the details of their words and deeds: Why do they behave this way? What is the motivation behind it? However, human nature is often lazy. Intuition and emotion are our natural tools for interacting with the world. They save us from overly complicated thinking. If you delve into everything, you will be exhausted - you have to feel the current emotions and piece together the other person’s motives and personality in your mind. I think the real wisdom lies in the balance between the two: use intuition to feel the other person’s warmth, use reason to see into the other person’s heart, and use emotion to tolerate the other person’s imperfections. The charm of growth may lie in this.

Threats and intimidation may work temporarily to make people act willingly, but the consequences are self-evident.

Some people believe that the desire for sex and success is the eternal driving force of human beings.

Some people believe that the deepest driving force of human nature is the desire to be valued.

Everyone desires:

  1. Health and longevity

  2. Material abundance

  3. Restful sleep

  4. Wealth and what wealth can buy

  5. Spiritual sustenance or peace in the afterlife

  6. Harmonious intimate relationship

  7. Children are safe and healthy

  8. Feeling of being valued

Deep in human nature, there is an extreme desire for praise. Praise can bring tremendous power to relationships. Similarly, the awareness of one’s own importance is one of the distinctive features that distinguishes humans from animals.

When interacting with others, remember that people are not purely rational creatures. They are driven by emotions, dominated by prejudices, and arrogance and vanity are often the source of their motivation.

It is easy to get angry, but it takes tremendous effort to forgive, understand and pardon, which are the precious treasures of human nature.

“If you understand everything, you can forgive everything.”

Understanding suggestions

2025-03-18 02:59:29

As mentioned above, giving advice to others is actually providing useful information sources or experiences. For example, parents’ pursuit of “stability” based on their own experiences, their advice bears the stamp of the times.

Be careful about giving advice to people you don’t know well. Many people lack the ability to think independently, and your advice may be seen as disrespectful.

The effectiveness of the advice you gave to friends and parents depends partly on the relationship and trust between the two parties, and partly on whether your friends truly understand your core ideas.

There is no standard answer to whether or not to give advice. It is like seeking a balance between “letting go” and “helping hand”. It mainly depends on your relationship with the other person, the situation at the time, and whether the other person really needs it.

Thoughts on proactive suggestions

Everything has its cause and effect. Your words may affect other people’s thoughts or decisions, and inadvertently intervene in other people’s life trajectories. If the advice is inappropriate and things go against your wishes, you may need to bear some joint responsibility. But in fact, we are responsible for sowing and watering, but we cannot control the final fruit. What kind of fruit it is depends on the tree itself, so why bother too much.

Human nature and advice, the nature of hating advice

Most people will instinctively consider their self-esteem when hearing advice. Just like when I was communicating with a retired senior in Beijing today, his advice made sense, but it always made me feel condescending. Rationally speaking, he was talking about his own experience, which made me think, and that was beneficial.

This shows that advice is often mixed with the adviser’s humanity and partial selfishness. Sincere advice is more likely to be accepted because it conveys concern for the other party’s well-being rather than personal interests. However, completely selfless advice is rare in reality - the adviser may unconsciously reveal a sense of superiority, a desire for control, or a desire for recognition. Experience is more profound and direct than opinions. When we talk about our experiences and opinions, are we also reflecting on ourselves? What is our purpose? This is also part of our understanding of ourselves. Through reflection, we can deeply understand ourselves and others, and understand human nature.

On the other hand, it is to avoid the desire for control. This phenomenon is particularly prominent in the parent-child relationship in China. During the rebellious period of adolescence, some people are extremely disgusted with others interfering in their private affairs, believing that it is their territory and parents should not arbitrarily command them.

In addition, trust is also key. This is the case with my communication with the hotel owner. The other party was worried that I had other intentions, or worried that the opinions I expressed would be posted on the Internet or short videos, and doubted my motives.

How to give advice correctly

  1. Listen first, speak later: Don’t rush to speak, first understand the other person’s thoughts, situation and blind spots.

  2. Provide options, not instructions: Suggestions are to give options, or to provide additional information about options, rather than to give orders. This is crucial. Everyone has his or her own path to take and has the right to make his or her own decisions, even if you foresee that he or she may fall. As bystanders, we must respect this autonomy.

  3. Assess the situation: Avoid giving advice when the other person is emotionally unstable or irrational.

Practical Guide

If you are not sure, you may as well ask the other party’s intention first. But this is not a perfect solution. I once asked in advance, “Do you want to hear my thoughts?” Although the other party agreed, he still had resistance in his heart. This requires analyzing the other party’s personality and cultural background in combination with human nature.

Experience is often more important than opinions. Experience is more first-hand, and the more first-hand the experience is, the more it can trigger thinking. Therefore, share more of your own original stories, as everyone has their own opinions.

Ask more questions around some core topics, as discussed in previous blogs:

  1. **What do you really want? ** (Get past the superficial excuses and explore the underlying needs)

  2. **What methods have you tried? ** (Understand existing efforts and avoid duplicating suggestions)

  3. **What are you most worried about? ** (Locate the source of fear and understand the core concerns)

The world is so wonderful about advice. It is essentially the product of our more realistic understanding of the world. A more accurate understanding of a specific group of people can help us judge whether they are willing to listen to advice; a more accurate understanding of things can help us provide more accurate advice. The value of advice may not lie in what you say, but in what the other person listens to and what positive changes are made as a result.

Humanity and Communication

2025-03-17 02:11:36

The phenomenon of the boss in Nepal once again touches upon the complexity of human nature, which includes self-protection, inherent prejudice, emotional dependence and other characteristics, all of which can make communication extremely difficult.

You may understand the other person’s thinking patterns and behavioral logic (for example, why they don’t understand you), but they may not be able to see your true intentions and deep thoughts. This “one-way transparent” communication state often makes people feel like “talking to a deaf person”.

This is not their fault, but their cognitive mode. Tolerance and understanding require certain cognitive abilities and an open mind, which the other party may lack.

It is crucial to deeply understand a certain group of people or human nature in general. I have a new understanding of human nature, and I may even take it into account in future decision-making. For example, if I had fully considered human nature, I might not have given so many suggestions and persuasions to the Chinese boss in Nepal.

Everyone has their own unique growth trajectory and cognitive limitations. Not everyone can change easily, and it is not your responsibility to change them. It is just that in a specific scenario and cognitive system, based on your understanding of the real world, if you think you may have a slight positive impact on the other person, you might as well try.

Some people really have trouble thinking deeply, so you can only talk about surface things and phenomena that are easy to understand. Try to simplify your expression, avoid abstract concepts, and communicate in a way that they can understand.

Second Thoughts on the Corpse Burning Temple

2025-03-16 16:13:45

I visited the cremation temple again. What exactly are those “round platforms”? I wondered. From a distance, those round stone platforms were scattered on both sides of the river, some hidden in the bushes on the west bank, and some close to the cremation area on the east bank. At first, I thought they were some kind of altars. Later I learned that these might be places for sadhus to practice meditation, or small platforms for believers to hold special ceremonies. They are not as eye-catching as the cremation platforms, but they always attract my attention - especially when someone lies prone on them or makes strange dance-like movements.

Claude Artifacts

2025-03-16 14:13:45

Claude Artifacts is a feature of Anthropic’s Claude model, which allows the model to generate and display structured content output during the conversation. These outputs are not just text, but can also be code blocks, design prototypes, and even interactive UI components. This marks the evolution of AI from simple text generation to richer, actionable “product” generation.

Some thoughts from a Chinese boss in Nepal

2025-03-16 12:51:52

RECOVERY

We agreed to travel together today, and in the morning I went to the hotel where my friend was staying and waited. The hotel is run by a Malaysian Chinese lady and has a small public area. At that time, my friend was chatting and working in the public area. As an AI entrepreneur and digital nomad, I used my own computer to handle business in the office area while waiting. In order to avoid being disturbed, I connected to the Wi-Fi I brought with me, and used the paper towels on their dining table. I also threw the discarded paper towels into the trash can myself, trying not to cause trouble to others.

After my friend went upstairs to pack, I continued to work in the office area. At this time, the boss suddenly approached and asked me if I lived here. I said no, but I lived in the hotel next door. The boss asked: “Does your hotel have Wi-Fi?” I said yes. “No private space?” I thought for a moment and said that the space was okay. “No public space?” I said yes. The boss immediately said that this public area is only for the guests of the hotel and I am not welcome. I immediately realized that the boss might be worried about the resources being occupied, but I still wanted to finish the work at hand in a short time, so I offered to pay a tip and clean up the public area before leaving. The boss insisted that I go back to my hotel to work, so I prepared to leave.

think

I reflected on the boss’s motives and behavior. She was obviously worried that people without interests would occupy her resources, and perhaps she regarded me as a speculator who wanted a free seat. Although I explained that there were many comfortable cafes outside, I chose this place only because my friend lived here, and we would leave after waiting for a while. The boss seemed to be ashamed after hearing this, because I had established a connection with her customers, and it was reasonable for me to wait for my friend here. In China, rules and human feelings can often seek a balance. If the rules are too strict, they will lose human touch, and if human feelings are rampant, the rules will be meaningless. The “moderation” thought in Confucian culture emphasizes this balance.

During my cultural experience in Nepal, including the hotels I stayed in previously, I felt that the local service industry was generally excellent. Compared with most cities in China, Hong Kong, and even Thailand, which relies on tourism as its mainstay, I appreciate Nepal’s cultural atmosphere more. The local service industry is generally friendly and welcoming. The excellent thing about the hotel I stayed in before is that its rules are transparent, and the important regulations are prominently posted. Transparent rules make customers feel much more comfortable than unclear rules, but I did not see any such notice in this hotel. In addition, after living in Pokhara for more than a month, a waiter in a Nepalese restaurant actually remembered the dishes I ordered two weeks ago. I was impressed and deeply moved by the meticulous service.

REFLECTION

In the process of inheriting Chinese culture, overseas Chinese culture will evolve due to factors such as region and social environment, which may lead to obstacles in cross-cultural communication. Even if one party is tolerant, the other party may affect objective judgment due to emotional cognition, such as preconceived distrust of me or prejudice against a certain cultural group. Therefore, subjective impressions and past experiences need to be constantly revised. If human nature is not examined, the experience-based predictions often appear outdated or one-sided. This makes me reflect: It is crucial to establish a systematic cognitive model for a thing or national culture and continuously update it. Experience and communication are the nutrients of emotional cognition, logic and reasoning are the methods of rational analysis, and the rational structure diagram abstracted from this is the basis for guiding behavior. In such training, we can obtain correct intuition at certain stages, so as to get closer to the understanding of the essence of the world. Human nature swings between rules (order) and human feelings (emotions), and different cultures and individuals have different emphases. Understanding this will help predict and respond to the reactions of others.

I began to have a deeper tolerance and understanding of different cultures. As Chinese culture spreads around the world, it evolves naturally as a dynamic system to adapt to new environments, including the dramatic cultural changes brought about by globalization and China’s reform and opening up. Culture is not static, but constantly evolving. Understanding this dynamic nature can reduce negative judgments about so-called “deviations.” Experience is the source of perceptual knowledge, but if it is not verified and updated, it may become an obstacle to understanding. Reflection and adjustment of intuition are the key to cross-cultural communication.

In addition, I apologize for some of the emotional feelings I had at the time. Although I was able to analyze the possible reasons for my boss’s emotions, I failed to communicate effectively and in a timely manner to dispel the suspicion and distrust. Cultural differences are relative, and there is no absolute right or wrong. The values ​​and behavior patterns of each culture have their historical and environmental roots. When encountering distrust, do not rush to refute, but try to understand the other party’s background and motivation (both culturally and personally), and then seek effective communication. At the same time, I also reflected on whether I had erroneous reasoning based on my own experience during the communication process, and failed to take the initiative to explore the facts and understand the other party’s real concerns. Empiricism emphasizes perceptual cognition, while rationalism emphasizes logical reasoning. Only by combining the two can a comprehensive cognition be formed. Relying solely on either side is not enough to cope with the complex reality.

Culture and behavior are a system, and its internal parts are interrelated. Analyzing the problem in a system framework can more clearly find the root cause. Social contract theory believes that rules are the basis of social order, but human nature is also driven by emotions. Only by seeking a balance between rules and human feelings can harmony be achieved.

The world is changing, and cognition and behavior must also be updated accordingly. Continuous learning is a necessary condition for adapting to this uncertain world.

Design and thinking about multimodal Chat mode

2025-03-16 12:23:19

Multimodal interaction currently covers a variety of forms, including: text information, images, sounds, videos, web pages, data files and even sensor data.

There are two main forms of its existence: one is as a format for sending (such as through pictures, voice, text input), and the other is as a format for receiving (the form of model output).

The design logic of Claude Artifacts seems to be unique. Artifacts can not only generate a variety of content types, but also provide the opportunity to use interactive elements in a groundbreaking way. Its application range can extend from simple prototyping to fully functional web services. To a certain extent, it can intelligently choose the most appropriate output form:

  1. Interactive documents: For example, generating code blocks and forms that can be directly edited or interacted with.

  2. Data and process visualization: Present complex data or processes in the form of charts.

  3. SVG graphics: including the generation of vector graphics such as trademarks and icons.

  4. Website Structure: Output the basic framework or layout suggestions of the website.

  5. Complex visual documents such as sequence diagrams: help understand and display system interactions or complex logic.

It is expected that the types of artifacts that Artifacts will support will become increasingly diverse. Products with similar Artifacts concepts, such as Google’s NotebookLM, also show similar forms, allowing users to manipulate code and design interactive interfaces.

Always on the road of pursuing meaning

2025-03-15 12:44:05

No matter what choice you make in life, there always seems to be regret. People often think that the road they have never walked on must be full of flowers.

Many things lose some of their interest when we see them too clearly. Don’t resist, don’t try to hold on, it may be better to let things take their own course.

In the face of such a colorful world, especially a complex and multidimensional emotion like love, we can often only grasp a glimpse of it. This tension is the driving force for humans to continue to explore, learn and grow.

Sex, hugging, and dating are all important manifestations of love, but they are only external manifestations of love. The essence of love lies in the deep emotional bond between people, mutual understanding and support, and the process of growing together.

Jealousy, at its core, often stems from inner insecurity and uncertainty about self-worth.

When someone I like gets jealous of me, I have learned a good way to deal with it: it is important to understand and tolerate the other person, actively resolve misunderstandings, and eliminate unnecessary suspicions.

It is human nature to feel jealous when meeting someone you like. Looking back, there have been many such phenomena in the past, but I have never reflected on the essence of them. Now it seems that if we can understand and manage this emotion, think about why we are jealous, and explore the possible reasons for the other person’s behavior, then this experience can be transformed into nutrients for growth and help both ourselves and the other person mature. Of course, excessive jealousy may evolve into a strong possessive desire, so we need to be vigilant.

Some thoughts on Agent development

2025-03-14 20:05:56

Agent is undoubtedly the hottest word this year, and this year is very likely to be the first year of Agent.

Regarding how to use tools for large models, there are currently two mainstream solutions:

  1. Computer/Browser Use: This type of scheme simulates user operations on a computer or browser, such as Claude’s Computer Use function.

  2. Agent Protocol: represented by MCP (Model Context Protocol) released by Anthropic last year. Of course, OpenAI’s Function Call also belongs to this category, but it focuses more on providing a simple and easy-to-use interface rather than being a standardized protocol. MCP’s standardized interface can better support a variety of tools and application scenarios. For example, for a simple task like calculating 1+1=2, Function Call is sufficient; but if you want to automate a complex data analysis process, MCP’s advantages are more obvious.

The former (Computer/Browser Use) can quickly implement functional verification through simulated operations in scenarios where there is a lack of standardized interfaces, such as calling various web page operations or controlling certain applications. It mainly works at the simulated user operation level (GUI layer). MCP works at the API level. The former is relatively inefficient, while the latter is more efficient. MCP is more inclined to become a long-term, standardized form of interaction.

If you are facing an agent that you cannot develop or modify directly, it is particularly important to access external services through the MCP protocol, because MCP defines an open and standard interface. In theory, ordinary users can also expand the business capabilities of the agent by installing MCP services with one click.

In addition, agents themselves must also have classifications and hierarchies. What I feel most deeply at the moment is that various special agents are emerging:

  1. Workflow Agent: It is a chain consisting of a series of Prompt and API calls. It has a certain degree of autonomy, but its behavior boundaries and processes are still constrained by many preset rules.

  2. Professional Agent: Able to autonomously decide which tool to call or which steps to perform in a specific field or complex task. For example, AutoGPT decomposes complex problems through CoT (Chain-of-Thought) technology and dynamically selects the optimal solution path.

  3. General Agent (AGENT’s AGI): Currently still in the theoretical concept stage, it refers to an intelligent entity with cross-domain, highly autonomous learning and problem-solving capabilities.

The capabilities of an Agent can be roughly summarized as: Agent = LLM (Large Language Model) + Memory (memory) + Planning skills (planning ability) + Tool use (tool use).

For example, Claude’s Computer Use is essentially also a Tool use.

Planning skills include: breaking down large tasks into executable small tasks; reflecting and refining - self-examining based on the actions that have been performed, learning from mistakes and optimizing subsequent actions. In addition, some research papers have proposed more novel classifications of planning skills, such as: task decomposition, multiple plan selection, external module assisted planning, reflection and refinement, memory enhanced planning, etc. Among them, “multiple plan selection” refers to allowing the AI ​​Agent to generate multiple alternative plans, and then select the best one to execute; “external module assisted planning” refers to using an external planner (similar to the “referee” role in reinforcement learning) to improve the quality of planning; “memory enhanced planning” is like equipping the Agent with “memory bread” so that it can remember past experiences and provide reference for future planning. These methods are not isolated from each other, but are often intertwined to jointly improve the planning and execution capabilities of AI Agents.

2025-03-14 17:25:14

It is undeniable that AI has brought both opportunities and challenges and confusion to the front-end field.

In the AI ​​era, the entry threshold of front-end seems to be lowered. AI-assisted programming tools such as Cursor can even help non-front-end professionals quickly build project prototypes and achieve a leap from 0 to 1, or even from 1 to N.

However, no matter how technology develops, the need to interact with users always exists. Looking at the development trend of the front-end from an objective perspective: although the threshold of technology has been lowered, the efficiency of developers who have mastered front-end skills has been greatly improved. Learning and making good use of open source projects on GitHub (such as Vercel’s AI Chatbot template: https://github.com/vercel/ai-chatbot) can quickly build and iterate various web applications.

In addition, some minimalist AI search tools (such as https://github.com/zaidmukaddam/scira) and AI integration services (such as https://tavily.com/) also provide convenience for front-end development.

In the early stages of a project, if problems can be solved with existing services, try to use services as much as possible. This will make it easier for us to quickly verify core ideas and functions.

Nepal Holi Festival

2025-03-14 11:55:51

From early morning to dusk, the noise and colors of Holi Festival enveloped the whole day. Everyone’s face was covered with colorful powder, smiling and full of happiness. At that moment, the happiness of the Nepalese people deeply touched me - it was a kind of contentment from the heart, a kind of spiritual abundance.

After the joy of the day, when night falls, a sense of melancholy arises in my heart. Just at this time, I meet Sister Mingming, Xiao Xiong and Vanessa at the intersection.

Some friends, after this separation, may never meet again; some emotions, most likely will not blossom and bear fruit; some love songs, most likely will not lead to the palace of marriage. So, should we stop moving forward, stop experiencing, stop feeling?

Tomorrow, Xiao Xiong and Vanessa will leave Nepal. Looking back at the time when we took classes together and had dinner together in Pokhara, it seems like yesterday. The reunion in Kathmandu was brief, but I didn’t realize that the separation was imminent. This unexpected encounter made me suddenly feel that this might be the last time we see each other on this trip.

Sister Mingming said, “You are the last one to leave, and you are leaving alone again.” There was a hint of understanding and teasing in her words, but it made me feel inexplicably sad. Walking on the street, I couldn’t help but want to cry, but I could feel the understanding in Sister Mingming’s words. We talked about how I would feel when my friends left one after another. I said that I know myself well, and sadness is inevitable. The person who stays behind often carries more sadness. But more importantly, it is a reflection on this sadness, a reflection on the impermanence of the world, a reflection on the meaning of life, and a cherishment of friendship. We took a group photo on this street, which may be the last complete freeze frame of our trip.

The understanding of “impermanence” and the realization of “once in a lifetime” become particularly profound at this moment. Every encounter and parting in life is a microcosm of impermanence. Impermanence is not a cold judgment, it is just the true state of life, reminding us to accept changes and find meaning and strength in changes. Impermanence does not mean negativity. Instead, it makes us know how to cherish the present and grasp every precious moment with our relatives and friends. Just like cherry blossoms, after blooming brilliantly, they quickly wither, and their beauty comes precisely from this brevity and imperfection. The same is true for our encounters and partings. Although they are like footprints in the snow, they are fleeting, but they are shining because of their uniqueness and worth remembering.

We treated each other sincerely and cherished every time we were together, even though we knew we would be separated soon. This sincerity was not to deliberately retain each other, but to respect life and pay tribute to friendship. Perhaps we will never meet again in this life, but the emotions we shared will not disappear due to distance, but will become more precious against the backdrop of separation.

The premise is how we can comprehend our understanding of life and form our own way of dealing with the world.

Knowing oneself is so difficult; facing oneself calmly is even more challenging. When we have a good understanding of ourselves and face our emotions with a sincere attitude, we can clearly see the growth and value that each relationship brings us. Regardless of the outcome, what is important is whether we can improve ourselves in the process and whether we can find resonance with each other in respect, understanding and equality. Only in this way can emotions truly become a part of our life’s nutrients, rather than a simple goal to pursue or an external criterion for judgment.

How to quickly learn an open source project in the AI ​​era

2025-03-13 11:13:11

To quickly learn and understand an open source project with the help of AI, you can follow these steps:

  1. Clear project background and goals: First, use AI (such as asking ChatGPT, Claude, or using targeted search tools such as phind.com, devv.ai) to quickly understand the core functions of the project, the problems it solves, and its positioning in the relevant field.

  2. Read the official documentation and getting started guide: This is the most direct way to understand the project design concept and basic usage. You can use AI tools to assist in reading and summarizing long documents.

  3. Analyze project structure and code architecture:

  • Use Cursor or similar AI programming assistant to analyze the main directory structure and module division of the project.

  • Ask AI: “What is the core architecture of this project?” “Which modules are the main business logic distributed in?”

  1. Location of core modules and key code implementation:
  • Ask AI to identify the entry files, core classes or functions of the project.

  • For key code snippets, let AI explain its functions, design ideas, and interactions with other parts.

  1. Use unit testing to understand functional details:
  • Unit tests are usually written for specific functional points and are excellent material for understanding the specific behavior of a module.

  • AI can be used to interpret the intent of test cases, or to reversely understand the functionality of the code being tested based on the test cases.

  1. Hands-on practice and modification: Try to run the project, modify some small functions, and observe the changes. When encountering problems, ask AI for help in time, describe the problem phenomenon, and let it assist in troubleshooting.

  2. Participate in the community: Check out the project’s Issues and Discussions to learn about the problems and discussions encountered by other developers, which will help you understand the project from different perspectives.

In this process, AI plays the role of “intelligent mentor” and “efficient assistant”, which can greatly accelerate the process of information acquisition, code understanding and problem location.

Walking on the same road, meeting your own scenery

2025-03-11 14:29:35

During the hike, my partner and I focused on different things. He preferred to look at the road below him, occasionally looking up. To him, the scenery was like a picture that changed in an instant, with a bit of magic. But I was intoxicated by the scenery in front of me, which was changing every moment, and my mind was immersed in it, as if I was one with the world.

He enjoys walking itself, and I enjoy the flowing scenery. The same road, but due to different moods and perspectives, reflects each unique scenery.

Service deployment logic

2025-03-11 02:37:34

A set of deployment practices suitable for rapid iteration:

  1. Network Services and CDN: Cloudflare provides a complete set of mature network services, including DNS, CDN, security protection, etc., which is the cornerstone of modern Web applications.

  2. Cloud server backup: Prepare some cloud server resources (such as AWS EC2, Google Cloud VM, Azure VM) as backup or to handle specific computing tasks.

  3. Cloud database: Choose cloud database services (such as AWS RDS, Google Cloud SQL, MongoDB Atlas) whenever possible. They usually provide high availability, automatic backup, and elastic scaling.

  4. GitOps: Use Git as the only source of truth for infrastructure and application configuration to achieve declarative, versioned operations management. This is a universal capability that can be reused and exponentially improve efficiency.

  5. Vector database: Vector databases such as Milvus and Pinecone are indispensable in AI applications (especially RAG scenarios).

  6. System monitoring and log: Use tools such as ELK Stack (Elasticsearch, Logstash, Kibana), Prometheus or Grafana to monitor the system operation status and manage logs. This is an important methodology to ensure service stability.

  7. Agent Framework: Frameworks such as AutoGPT / AgentGPT help to automatically generate task links, which is very helpful for verifying ideas in the MVP (minimum viable product) stage.

  8. Data management framework: LlamaIndex is an excellent data management framework, especially suitable for building LLM-based applications to process and index external data.

Analysis of AI editing tools

2025-03-10 13:33:21

Currently, we mainly experience and focus on two AI video editing tools:

  1. Opus Clip (https://clip.opus.pro/):
  • Features: Focuses on automatically editing long videos (such as podcasts and speeches) into multiple viral short videos through AI. It can identify highlights in videos, extract core ideas, and automatically match subtitles and adjust the screen ratio to suit different social media platforms.

  • Advantages: Greatly improves the efficiency of content re-creation and distribution, especially suitable for content creators.

  1. CapCut (International Version) (https://www.capcut.com/):
  • Features: As a product of ByteDance, it is comprehensive in functions. It not only has traditional video editing functions, but also deeply integrates AI capabilities, such as smart matting, AI dubbing, automatic subtitles, AI special effects, and one-click short film generation.

  • Advantages: Large user base, relatively convenient operation, rich AI functions and rapid iteration, suitable for users from novices to those with certain experience.

These two tools represent different application directions of AI in the field of video editing: Opus Clip focuses more on intelligent repurpose of content, while CapCut is committed to enhancing various aspects of the traditional editing process through AI.

Some tips for using Cursor

2025-03-10 13:07:24

  1. Start from a template: Start by cloning a template project from GitHub or other sources, which can provide a solid initial framework. Cursor has a “Start from Repo” option inside, or you can use services such as https://bolt.new/ to quickly create a basic demo project, which is very convenient.

  2. Combined with Claude 3 Opus (or other powerful models, such as GPT-4 Turbo): In Cursor’s chat mode, choose a more powerful model (such as Opus, although the cost may be higher, but it has obvious advantages in processing complex logic and multi-modal input, such as modifying the front-end page through images). For the agent logic of complex projects, Cursor’s processing is usually better than a simple chat interface.

  3. Make good use of agent mode: Use Cursor’s agent mode (rather than normal chat mode) to create, edit and manage files and code structures through natural language commands.

  4. Combined with search enhancement tools such as Perplexity AI: When I need to search for API documentation, code examples, or solutions to specific technical problems, I will first use tools such as Perplexity to collect preliminary information, and then feed key information and code snippets to Cursor for subsequent processing.

  5. Create new conversations in Composer and keep them short: Open new conversation windows for different tasks or contexts to avoid long conversation histories interfering with the model’s understanding.

  6. Continuous iteration and improvement: AI-generated code is often not completed in one step and requires fine-tuning and correction through multiple instructions.

  7. Integrate GitHub Actions for automated testing: Cursor is very efficient in writing test cases. Combining it with GitHub Actions to achieve automated testing has a very high return on investment (ROI).

  8. Use speech-to-text tools: For spoken needs or ideas, you can use tools such as Whisper Flow to convert them into text and then input them into Cursor.

  9. Hand the error over to the Agent: When the code fails or does not behave as expected, it is a wise choice to provide the error information and related code snippets directly to the Cursor’s agent to analyze and make repair suggestions.

  10. Commit frequently and track in a standardized manner: Performing git commit frequently and writing clear commit messages is a good habit for AI-assisted development (especially when you need to review and understand the change history).

  11. Continuously deploy projects: Use platforms such as Vercel, Netlify or Firebase Hosting to continuously deploy projects and quickly verify the results.

  12. Record and collect commonly used effective prompts: Record those prompts that can efficiently guide AI to complete specific tasks and form your own prompt library.

Reactions caused by alcohol

2025-03-10 10:03:56

Alcohol’s inhibitory effect on brain function is most sensitive in the prefrontal cortex (PFC). The prefrontal cortex is responsible for our higher cognitive functions, such as judgment, decision-making, impulse control, and self-monitoring. When alcohol inhibits the activity of this area, the individual’s rational judgment and self-control abilities will decline, and they may unconsciously express thoughts and emotions that are usually suppressed or hidden due to various concerns or when the inhibition is weakened. This is one of the physiological bases of the so-called “truth after drinking”.

China’s development should be viewed from an intergenerational perspective

2025-03-09 23:09:47

The narratives on the land of China are complex and endless. We don’t need to care too much about the short-term gains and losses of a city or a pond, but should have a long-term persistence and insight.

A country is ultimately made up of specific people; it is created by these people, and its destiny is also determined by these people. Only a country can nurture and cherish those who seek the truth, those who can think independently, those who dare to record the truth, those who are willing to sacrifice for this land regardless of personal interests, those who resolutely defend their legitimate rights and dignity, and those who know that the world is not perfect but still remain sincere and never give up - only when a country has such a mind and soul can we sincerely say: “I am proud of my motherland.” Only when such a mind and soul become the backbone of society can we look forward to a better tomorrow with confidence.

The formation of any mechanism and phenomenon of the times has its cause and effect. If we put aside the grand perspective of reform and opening up, it will be difficult to understand many current realities. The acceleration of reform and opening up has begun to show a relatively stable and declining trend since 2008, which means that the cyclical turning point of rapid economic growth has already appeared. We must realize that economic development has its own cycle and is not a myth of unlimited growth.

The current leadership grew up in a special historical period (such as the Cultural Revolution), and its governing philosophy and behavior inevitably bear the imprint of that era. People and institutions are the products of a specific era. Even if we may observe a “reversal” or adjustment at the institutional level, we should believe that the next generation of leaders, with their different growth backgrounds and the talents endowed by the times, will eventually lead the country to a new stage of development.

Observer’s Perspective

2025-03-10 01:10:15

For you, it is a personal experience; for me, it may be a fragment of your life story.

In many cases, those involved are often confused, while those who are watching can see things more clearly. People in the situation often feel overwhelmed and entangled. However, if they can step out of the situation and examine it with the simplest and most universal principles, they can often see things clearly.

Therefore, those involved in the situation should not easily deny the scrutiny and perspective of bystanders, as that may be the key to breaking the situation.

Soulmates

2025-03-09 23:44:52

I once heard a point of view: in most cases, when it comes to soul mates in love, one of them is engaging in “backward compatibility.”

The “downward compatibility” here may mean that one party is more tolerant in character and can tolerate the shortcomings or edges of the other party. However, my understanding of the essence of soul mates is more inclined to a deep connection that is highly compatible in spirit, emotion and values. The true meaning of love lies in the deep communication, mutual encouragement and common growth between two people.

I prefer to compare ideal love to the interaction between two independent parallel lines. The key is how these two “parallel lines” can maintain their independence through communication, understanding and collaboration, while generating meaningful interactions and forming a dynamic and balanced connection.

In contrast, the term “backward compatibility” does not seem to be quite accurate. Absolute “backward” compatibility may not exist. It is more about standing from one’s own perspective and feeling that the other party is more tolerant. In fact, the other party may have similar feelings to some extent, but we cannot fully explore the other party’s inner experience. If one-sided compatibility is not understood and responded to, such an interactive process must be difficult.

Looking back from childhood to adulthood, aren’t parents and teachers always like this? They have always tolerated our immaturity and mistakes. The process of interaction itself is a process of growth.

Most relationships remain at a superficial level of communication, a few relationships may bring deep resonance, and only a very few may touch the realm of soul mates. Understanding and accepting the distribution of relationships and not holding too high and uniform expectations for every relationship may make us more relaxed in interpersonal communication.

Whether it is the profound resonance of one or two stages, or the deep connection brought by soul mates, it is worth our careful experience. At the same time, we must also face its possible changes and impermanence calmly.

MCP protocol mode integration

2025-03-08 17:55:51

Why do we need MCP (Model Context Protocol)? It is mainly due to the following three factors:

  1. Data silo problem: There is a lack of a unified way of interaction between different applications and data sources, which makes it difficult to efficiently share and use information.

  2. Security issues: Exposing raw data or internal interfaces directly to the model poses security risks. MCP can serve as a layer of abstraction and control to enhance security.

  3. Unified standards: The lack of standardized context and tool calling protocols makes it difficult for developers to integrate different models or services, and also hinders the development of the ecosystem.

Architecture model: MCP adopts the classic client-server model.

  • MCP Client: usually corresponds to Large Language Model (LLM) or LLM-based Agent.

  • MCP Server: responsible for exposing external data sources and callable tool interfaces.

Functional modules: Each MCP Server can selectively expose one or more of the following functional modules based on actual needs:

  • Tools: Allows LLM to call external functions or APIs to perform specific tasks (such as calculations, querying databases, sending emails, etc.).

  • Prompts: Provides preset, structured prompt templates to facilitate LLM to generate high-quality responses based on different scenarios.

  • Resources: Allows LLM to access and retrieve external knowledge bases or data files (such as documentation, FAQs, product information, etc.).

Communication mechanism: MCP chose JSON-RPC 2.0 as its main communication protocol. JSON-RPC is a lightweight remote procedure call protocol that is easy to implement and understand.

Anthropic and the community have provided MCP Server implementation libraries based on multiple mainstream programming languages ​​​​such as Python, TypeScript, Golang, etc., to facilitate developers to quickly build and deploy their own MCP services.

Claude Computer Use Ability

2025-03-08 12:27:00

The core architecture of Manus (or similar AI Agent projects) is highly similar to the “ComputerUse” concept proposed by Anthropic. Both rely on one or more multi-agent virtual machine environments to complete complex tasks that require interaction with computer operating systems or applications.

The core idea of ​​this model is to enable AI Agent to “operate” the computer like a human, interacting with the graphical user interface (GUI) by observing the screen, moving the mouse, clicking buttons, entering text, etc., to perform tasks.

My life experience: Focus on what you love

2025-03-08 14:26:58

People must do what they truly love. This is the experience I have written down in my life and it is also an unswerving creed.

Manus’s thoughts

2025-03-06 14:37:04

There is a lot of discussion about Manus (or other similar general AI agent products) in the circle of friends. This phenomenon reflects that the agent field is ushering in a huge breakthrough. The core difference between it and the traditional chat mode is:

  1. Traditional Chat mode: chat -> message (chat generates information/answers)

  2. Agent mode: chat -> action (chat drives action/result)

AI can now not only give answers, but can actually act for you to get results.

The outstanding feature of the Manus (or similar products) team may be its strong engineering capabilities and as an early explorer in the field of “general agent”, its product effects are often impressive. Compared with OpenAI’s Operator (if it exists and is not public) or other similar concepts, such products appear to be more “civilian” and easy to access. In terms of innovation, its form can be understood as a comprehensive product of Computer Use (computer operation ability) + virtual machine + Artifacts (structured output) + a group of built-in specific task agents. It can be said that Claude’s concept of Computer Use is a good bet in this direction.

Calling products like Manus the “world’s first general AI agent” may be a bit of a “flattery”. After all, there have been frameworks such as AutoGPT and BabyAGI before, which can be regarded as the prototype of “general” agents to some extent. In addition, the Operator or its Deep Research function that may exist within OpenAI also has some characteristics of general agents.

For general agents like Manus, their technical moat may not be absolutely deep, but they can catch up quickly through strong engineering capabilities. For example, the recent OpenHands project is a good example. The future of general agents is either that LLM vendors internalize core capabilities into their basic models, or that they build up ecological advantages through the power of the open source community.

Compared to operating the entire computer operating system, the browser environment is more convergent, and Agents performing tasks in the browser are less likely to cause “dead loops” or uncontrollable system-level errors. Browser automation has a large number of mature open source codes (such as Selenium, Playwright) and end-to-end testing tool systems. Its input (web page elements, user instructions) and output (operation results, page changes) are relatively clear, making it easier to design reward mechanisms for reinforcement learning.

Analysis of the hot market reasons:

  1. Actual landing value: This type of Agent can indeed solve practical problems and improve efficiency.

  2. Bridging the “last mile” between models and business: Although traditional large models perform well in dialogue and content generation, they lack direct “operational” capabilities. OpenAI’s Operator (if it exists) is often “hidden” and fails to benefit the general public.

  3. Public opinion dividend and market scarcity: Coupled with the “national fortune dividend” (such as the previous popular logic of Xiaohongshu, DS and other products), in the current market environment, the amazing Agent products will inevitably cause a huge response. In addition, entrepreneurs and Internet practitioners generally have high expectations for the Agent form. But strangely, if Manus is mainly aimed at overseas users and is all in English, its public opinion effect in China is so great, which is a bit abnormal.

  4. Suspicion of hype: Some self-media conduct large-scale testing and evaluation at the early stage of product release. The pace is sometimes too consistent, which may be due to the manufacturer’s public relations (PR) factor. The ability of Agent still depends to a large extent on the progress of the basic large model.

Agent OS Present and Future

2025-03-05 09:53:46

Current Agent OS practical solutions and thoughts:

  1. Structured task execution (like Deep Research):
  • Many Agent applications (such as Deep Research-like tools) maintain a task structure tree or execution plan internally to decompose complex problems into a series of subtasks.

  • Models and APIs are called in an agent-like manner, executing the sequence of “search -> read -> reasoning -> action” in a loop or state machine.

  1. Computer Use:
  • For example, Anthropic Claude’s Computer Use feature can read screenshots, move the cursor, click buttons, and enter text to automate tasks such as filling out forms, looking up information, planning routes, and even ordering takeout.

  • An important application scenario is operating a browser. AI can interact with web pages by analyzing screen content and using a virtual keyboard and mouse. AI can also use a variety of standard tools and application programming interfaces.

Analysis and understanding of the future form of Agent OS:

  1. Agent embedded in the application: In the future, each application (APP) may contain one or more Agents. The APP itself plays more of a role as an interaction window.

  2. Agent OS positioning:

  • Analogous to traditional OS: An Agent OS may be responsible for managing and scheduling all Agents within its “jurisdiction”. This form may maintain a certain logical relationship with the traditional operating system (OS), for example, an Agent OS instance runs under an OS, which is responsible for scheduling the Agents in all apps on the OS.

  • Unified scheduling platform: A unified and standardized scheduling platform independent of the traditional OS may also emerge to control and coordinate Agents across applications.

  1. Interaction between Agents: An Agent in an APP may need to call an Agent in another APP to complete a task together. This call may be initiated by the APP itself or uniformly scheduled by the Agent OS.

  2. Agent as a Service: Agents are increasingly becoming intelligent services that can be published and called independently. This requires corresponding specifications and templates to define their interfaces, capabilities, and behaviors.

  3. Agent OS’s hierarchical logic: Similar to traditional computing systems, Agent OS may also present a hierarchical structure, for example:

  • Application layer Agent: directly facing users or specific business scenarios.

  • Kernel layer Agent (Agent OS Core): responsible for core scheduling, resource management, communication, security, etc.

  • Hardware/basic model interaction layer Agent: responsible for interacting with the underlying hardware, basic large models or external APIs.

Reference projects and concepts:

  • Sierra AI Platform (https://sierra.ai/platform)

  • Alan App (https://www.alan.app/)

  • OpenHands (https://github.com/All-Hands-AI/OpenHands)

Some predictions

2025-03-04 09:42:32

  1. Market structure: In the consumer AI field, a “winner takes all” situation may emerge, but on the enterprise side, due to the richness of demand and the high degree of customization, it is unlikely that a single product or platform will dominate the entire market.

  2. The importance of business models: The predictability and impact of changes in business models on the market sometimes even exceed changes in simple technological trends.

  3. Market Capacity and Diversification: Some markets, such as cloud computing, have huge capacity and can accommodate many winners. The AI ​​application market may also show similar characteristics.

  4. The form of the Agent Manager: The Agent interaction interface of the future will not only be a simple chat window, but will evolve into a more intelligent system for unified management of all Agents and their conversation history, task status, etc.

  5. Three major technology scenarios in the future: AI (artificial intelligence), quantum computing, and mixed reality scenarios that emphasize “presence” (or a certain form of the metaverse) will be important directions for future technological development.

  6. Constraints on superhuman intelligence: The development of advanced artificial intelligence (even superhuman intelligence) will be strictly constrained by law, ethics, and social trust. Establishing a true trust mechanism is a prerequisite for its widespread application.

  7. Security of Agent OS: Agent OS must run in a strict sandbox environment and have fine-grained control over its permissions to ensure that its behavior does not cross the preset security boundaries.

  8. Reshaping the boundaries of cognitive labor: AI will not completely eliminate all cognitive labor, but will redefine its boundaries. Many repetitive and process-based cognitive tasks will be automated, while humans will turn to higher-level creativity, strategic thinking, and complex decision-making. Human-machine collaboration will become the new normal.

  9. AI tool integration trend: From the early Jasper (AI writing tool) to ChatGPT, and then to products such as Monica (browser assistant) or Console (AI-driven development tool), AI tools will continue to integrate stronger tool calling capabilities, cloud execution capabilities and more natural interaction methods.

Future scenarios for inference models

2025-03-03 16:56:46

Combining the human cognitive mode of “thinking fast and thinking slow”, the progress of reinforcement learning, and the characteristics of large models, models with stronger reasoning capabilities (or models under specific reasoning modes) will play an important role in the following scenarios in the future:

  1. Handling ambiguous or incomplete information: When the user enters prompt or expresses the intention incompletely or unclearly, the reasoning model can gradually clarify the user’s needs through multiple rounds of clarification, context understanding, and logical inference.

  2. Filter key information from massive amounts of information: With strong generalization capabilities (reinforcement learning is of great help in this regard), inference models can efficiently extract core content and identify important patterns from massive amounts of data. Unsupervised learning lays the foundation for understanding, supervised fine-tuning (SFT) is responsible for memorizing specific details, and reinforcement learning (RL) helps the model better adapt to different environments and task requirements.

  3. Insight into subtle connections between data: Inference models can discover hidden connections and subtle differences from seemingly unrelated data points and conduct in-depth analysis.

  4. Multi-step reasoning and complex planning: This type of model is particularly good at playing the role of “planner”, breaking down complex tasks into a series of executable steps and planning. The execution level can be handed over to other general models or special tools that focus more on execution efficiency. This “planning-execution” separation model has great potential in complex applications.

  5. Visual reasoning: Not limited to text, reasoning models will also demonstrate stronger capabilities in analyzing blurred images, understanding complex scenes, and conducting visual question and answering.

  6. Code quality review, improvement and debugging: For code analysis tasks that are not particularly sensitive to execution speed requirements, such as code review, logic vulnerability discovery, improvement suggestions, and auxiliary debugging, reasoning models can provide valuable insights.

Everyone has opinions, experience is valuable

2025-03-03 17:22:47

In the information age, opinions are everywhere on the Internet, and they are mixed with good and bad. When talking to my father, he shared another parenting experience: his brother was helping his son with his homework, and the child was confused when he encountered a difficult problem. The brother said, “Don’t do what you can’t do, just do what you can do.” This made my father ponder for a long time and felt that it made sense.

I responded that there are many opinions in the world, and everyone can express their feelings. However, the foundation of opinions lies in personal life experience. Experience is original. The more first-hand experience, the more it can touch people’s hearts, turn into emotional raw materials, and inspire deep thinking. Growth is ultimately your own path. Opinions are more like drafts of values ​​and behavioral guidelines formed in the early stages of life, and personal experience is the melting pot for us to constantly correct and improve our self-cognition.

Parents’ experiences are also limited by their time, and their views cannot be directly applied to the next generation. I said that what I learned from this is that only through continuous experience and practice can everyone forge a unique perspective on life, and therefore better understand and respect other people’s lifestyles and choices.

Therefore, there is no need to worry that my growth will deviate from your experience, and there is no need to set too many rules for your child’s growth. Life has its own trajectory, and exploration and experience itself is the meaning.

RFT Thoughts

2025-03-03 14:55:15

The core idea of ​​RFT (Reinforcement Fine-Tuning, or more broadly, optimization of models through reinforcement learning) is to further optimize the model based on supervised fine-tuning (SFT). This usually uses a “reward model” or direct human feedback as a reward signal to guide the large language model to generate outputs that are more in line with user expectations, specific specifications, or actual needs.

Several key points:

  1. Based on pre-training and SFT: RFT usually works on models that have undergone large-scale pre-training and SFT. Pre-training gives the model extensive knowledge and language capabilities, and SFT makes it initially adaptable to specific tasks or conversation styles.

  2. Screening and generating high-quality samples: In the RFT process, a key step is to obtain or generate high-quality “trajectories” or “sample pairs”. For example, the model can generate multiple candidate outputs for the same input, and then select better outputs through a certain mechanism (manual evaluation or reward model scoring). This process helps the model learn what is a “better” response.

  3. Reward-driven learning: The core of reinforcement learning is rewards. By defining an explicit reward function (or training a reward model to simulate human preferences), the model will try to maximize the expected reward when generating output, thereby gradually learning the strategy of generating better content.

  4. Explore diverse reasoning paths: RFT not only allows the model to learn to imitate the “correct answer”, but more importantly, it encourages the model to explore different reasoning paths and expressions. By giving positive rewards to those paths that can correctly deduce the answer to the question (even if the calculation process or expression is different), the model’s generalization ability and robustness on unseen problems can be improved.

Introducing diverse reasoning paths actually improves the model’s “problem-solving ability” and “creativity”, not just its “pattern matching ability”.

Universal thinking on AI application scenarios

2025-03-02 21:29:16

In 2025, AI application scenarios seem to be activated again, and a large number of new projects have sprung up, triggering widespread thinking.

Take the agent form of Deep Research as an example. It shows considerable universality within a specific scope. Looking at many researchers, analysts, and any professionals who need to conduct a lot of online research or rely on external context to complete tasks, they all need strong information retrieval capabilities, accurate information source identification capabilities, and high creativity. Deep Research Agent is aimed at this pain point: it can solve the inefficiency of users having to frequently search and repeatedly ask questions in order to obtain deep information in the traditional chat mode. Its essence is that it efficiently replaces and optimizes the tedious process of humans searching, filtering, and sorting through a large amount of information. This is a universal ability with wide applicability.

We often subconsciously believe that the code we write or the process we design is “smarter” than the model. However, as the field develops, the model can often find better solutions than humans originally envisioned through learning and optimization. Therefore, when designing AI applications, we should avoid “hard-coding” certain processes or logic as much as possible, and try to create an environment that allows the model to learn and adapt autonomously.

Data is undoubtedly the lifeline for enterprises to survive in the AI ​​era. Applications such as Deep Research have once again proved this to us - high-quality, structured data is the cornerstone of training and driving powerful AI agents.

2025 may be the year of the Agent explosion. We can’t help but think about how Agent will develop? What key role will reinforcement learning (RL) play in the Agent ecosystem in 2025?

Hallucination Processing

| 2025-03-02 8:59 PM

Current methods to solve hallucinations:

  1. Knowledge distillation and soft labels can alleviate hallucinations. The teacher model provides soft labels in the form of probability distribution to fine-tune the student model.

  2. The reward model in reinforcement learning can guide the LLM to give honest answers.

  3. RAG is also recognized as an effective means.

  4. Chained Verification (CoVe): The model generates a preliminary answer first, then plans several fact-checking questions, answers them one by one, and modifies the initial answer based on the verification results to achieve self-correction.

  5. Self-checking of Prompt: guide the model to self-check its output. For example, after generating the answer, add instructions such as “Please check whether there are any uncertainties in the above answer, if so, please cite reliable information” to prompt the model to think twice and provide evidence or modify the answer.

  6. Tool call: Tools can provide good solutions for fixed program problems such as calculations and real-time needs such as search.

  7. Decoding strategy optimization: Lowering temperature and using greedy or beam search (rather than random sampling) can improve the reliability of the output.

  8. Fine-tuning the solution: Incorporate a large number of high-quality, factually accurate examples so that the model learns to be cautious when uncertain.

My understanding of hallucinations:

Currently, model hallucinations can be controlled within an acceptable range at the product level, and the requirements for hallucinations in different fields are not one-size-fits-all:

  1. Open domain (such as knowledge domain): This is the main application scenario of AI at present. It is advisable to design the logic of “rejection” and “retrieval enhancement”. Cross-validation (multi-model mutual calibration) is also a wise move.

  2. Creative field: Increasing the temperature may be beneficial to divergence, and no significant disadvantages have been found so far.

  3. Professional field: usually fine-tuning combined with tool verification. In this scenario, I am more optimistic about the enhanced fine-tuning mentioned earlier, which can empower the interactive logic in the short and medium term, and even give birth to the concept of “superman”.

ref: https://lilianweng.github.io/posts/2024-07-07-hallucination/#:~:text=balanced%20results%2C%20especially%20in%20terms,metrics%2C%20compared%20to%20two%20baselines

Agent form

| 2025-03-02 8:28 PM

Everyone is exploring the road of Agent, and the simple Agent product form seems to be nothing more than auto-agent.

The combination of Agent and RL (reinforcement learning), deep research seems to have verified the feasibility of this approach.

ref: auto-gpt: https://github.com/Significant-Gravitas/AutoGPT

agent list: https://github.com/e2b-dev/awesome-ai-agents

How to understand RFT strategy

| 2025-03-02 7:38 PM

Advantages of supervised fine-tuning (SFT): deepseek R1 pre-training is also used, which is powerful in replicating features in input text or images, and is particularly suitable for adjusting the model’s tone, style, or response format.

Reinforcement Learning Fine-Tuning (RFT) enables models to reason in new ways in custom domains with extremely strong learning capabilities.

The fine-tuning of reinforcement models is based on human preferences and rule definitions, making their outputs more consistent with expectations and value criteria. This enables developers, researchers, and machine learning engineers to use reinforcement learning to create expert models that excel in domain-specific tasks. In the process, the model automatically adjusts its learning and thinking methods.

In comparison, enhanced model fine-tuning is more suitable for inference models. The process is relatively simple and can produce significant results in dozens of use cases. It not only adjusts the logic of the model output, but also optimizes its thinking and reasoning process.

Compared with the simple scoring logic of deepseek, an AI scorer can also be introduced here for a more sophisticated evaluation.

Deep understanding of Web3

| 2025-03-01 9:55 PM

Blockchain has derived many technologies, but when it comes to the development of Web3, the core lies in smart contracts. Smart contracts are stored in the blockchain and act as automated business logic, which can be executed once the conditions are met. The form of Web3 can be understood as a community-driven, self-sufficient economic system.

The growth cycle of digital currency can be roughly divided into: creation -> growth -> speculation -> bubble -> adjustment.

The price of currency is completely determined by the market, and both new and old currencies follow the rule of “strong demand + limited supply = market growth”. Bitcoin has an upper limit, and its production is halved every four years (4year/2), making it a scarce and increasing asset.

The “stupider theory” in economics also reveals a market phenomenon: as long as one can find a buyer willing to pay a higher price, even if the asset price is unreasonable, there will be people willing to buy it, and the bubble will burst only when the market can no longer find “stupider people”.

Multimodal & cross-domain thinking

| 2025-03-01 3:15 PM

Large model platform + domain fine-tuning + knowledge distillation. The future trend of fine-tuning may be domain supervised fine-tuning (SFT) + parameter efficient fine-tuning (such as LoRA).

The development of text field towards multimodality has become a trend, with increasing emphasis on human-computer interaction. A large number of interactive tools have emerged, such as Cursor. Expanding to multimodality such as voice and video is also a direction, and the realm of AI can be expected to be “full sensory AI”.

Integration of cross-domain knowledge is crucial, and complex real-world problems often rely on the integration of knowledge from different professional fields.

The modularization trend may refer to the architecture such as Mixture-of-Experts (MoE, expert mixture), allowing different parts to specialize in different tasks to improve overall efficiency and performance. This approach can also be linked or combined with knowledge from different fields.

In terms of the effect of small models, distillation is better than direct reinforcement learning and more cost-effective. Distillation technology can effectively transfer the capabilities of large models to small models with smaller structures and higher computational efficiency, so that small models can also achieve higher performance levels in practical applications.

Design paper thoughts on DeepSeek R1

| 2025-03-01 11:56 AM

ENGINEERING THINKING

The engineering significance of the architecture is superior to the algorithmic significance, and provides many reference values ​​for engineering practice.

Design thinking

  1. Powerful models can achieve quite good results through distillation.

  2. To surpass the boundaries of intelligence, powerful basic models and larger-scale reinforcement learning are still needed.

Reasoning about large models tends to break down problems into smaller steps before answering them.

Using reinforcement learning alone, without supervised fine-tuning, large models can also demonstrate powerful reasoning skills.

It is also very important to provide reference “thinking notes” for the model. Supervised fine-tuning can, to a certain extent, solve the problem of poor readability of the output thinking process and occasional mixing of Chinese and English.

Using the “minimal intervention template” to guide the model to freely explore different problem-solving ideas without limiting the thinking method may bring many surprises.

Simple and direct reward method: accuracy reward (for math problem answers) + format reward (such as easy-to-understand thought chains). Put the “thinking process” in a specific label (such as <think>...</think>), and the answer in <answer>...</answer>. This method does not require additional training of an independent reward model. It feels like a fixed college entrance examination mechanism screening, simple and effective.

R1 also released six small models based on R1 distillation. These small models also have certain reasoning capabilities, and in some scenarios they even surpass the small models that are not distilled and directly trained through reinforcement learning. This provides some inspiration for the application scenarios of future models: knowledge distillation + reinforcement model, small models may show unexpected results in many application scenarios.

ChatGPT Deep Research and ChatGPT4.5 Analysis

| 2025-02-28 9:02 AM

GPT-4.5 is said to be the last generation of architecture that uses traditional reasoning models. Its goal is not to achieve SOTA (State-of-the-Art) intelligence, but to optimize the human-computer collaboration experience under specific conditions (such as pre-training Scaling Law to expand the unsupervised learning route).

Highlights of human-machine integration:

  1. Higher emotional intelligence, smarter and briefer responses, and the ability to understand the implicit meaning.

  2. Better conversational skills, softer and more coherent answers.

  3. Aesthetics and creativity have been improved.

In the future, we may expand the reasoning based on GPT-4.5 to maximize the effect of combining pre-training unsupervised learning with post-training reasoning expansion, and finally form a hybrid reasoning model GPT-5. In comparison, the effect of GPT-4.5 itself may not be significantly improved. In addition to the optimization of emotional intelligence and interactive capabilities, the actual performance may not even be as good as o3mini.

A major challenge with current AI is that its “reasoning” is more based on the generation of statistical laws rather than deep understanding and logical reasoning in the human sense.

About Deep Research

It consists of a powerful underlying model and an internal Agent framework. By adding a decision-making layer on top of the big model, it is able to plan multi-step tasks, think and reason.

By adopting an end-to-end learning strategy and training through a large number of simulated research tasks, it has learned how to plan and execute complex operation sequences (such as continuous searching, filtering information, backtracking steps, etc.), and can flexibly adjust strategies according to task requirements.

OpenAI Deep Research has unique advantages in breadth and information integration. In contrast, DeepMind may focus more on using its deep knowledge in specific fields for analysis, and its information breadth and real-time nature may be slightly insufficient.

What will be the future of Deep Research? Desktop or something like Cherry Studio, which can enhance research capabilities through subscription databases and internal resources, and link to more professional databases.

IMPACT AND EXPLOITATION

Whether it is design or research, Deep Research can significantly reduce the time for collecting data and generating reports, greatly improving efficiency. For decision makers, they can devote more energy to high-level thinking and creativity, and leave the heavy information processing to AI.

The work paradigm is shifting from human-led to AI-assisted. Instead of spending time collecting information, it is better to use the AI-generated review as a starting point, and then have humans conduct targeted in-depth reading and thinking. This also requires practitioners to have the ability to read AI output and perform secondary screening and processing, forming a new workflow of “human-machine collaboration”.

The market demand for new roles such as AI strategists, data analysts (talents who work with AI), and AI model tuners will increase. The entry-level work content of occupations such as intelligence gathering and legal retrieval will be largely automated.

When research and analysis become readily available, the decision-making process itself will also be affected. The key is to establish a human-machine collaborative decision-making mechanism: AI provides objective materials, and humans are responsible for value judgment and final decision-making. This also puts higher demands on decision makers: sharper experience intuition and a broader information perspective are essential.

Engineering practice is becoming increasingly important. In the AI ​​era, opinions are readily available, but experience, especially first-hand experience, is invaluable and can inspire thinking.

In addition, the importance of rare qualities such as interdisciplinary learning ability, multi-field integration ability and understanding of basic disciplines is becoming increasingly prominent.

Cognitive Dissonance

| 2025-02-27 7:11 PM

When does it occur? When an individual holds two or more conflicting cognitions (such as values, beliefs, or attitudes) at the same time, discomfort and tension will arise. At this time, the individual will usually try to reduce this sense of disharmony by changing his or her attitude, beliefs, or behavior in order to restore inner balance.

For example:

  • Information conflict: A person who believes that eating healthy is extremely important often finds that his favorite foods are junk food.

  • After a major decision, give a negative interpretation to the option not chosen.

People also tend to avoid information that conflicts with their own views - a phenomenon particularly amplified by recommendation algorithms - and are instead more willing to seek out information that supports their existing views.

Some design considerations for long-term memory

| 2025-02-27 6:58 PM

The design of Agent should pay attention to the unification of data structure. For example, the operation interfaces such as Upsert (insert or update), Delete (delete), Get (get) and Search (search) should be internally standardized.

Long-term memory is actually often implemented through the Agent + RAG approach.

refer: https://www.tanka.ai/?ref=producthunt&shortlink=g032wcsy&utm_source=producthunt&c=218+producthunt&pid=producthunt&af_xp=custom&source_caller=ui

Quickly learn a new field

| 2025-02-27 10:10 AM

I recently started working on a new product that aims to help people learn quickly, which has forced me to reflect on my own approach to learning new areas.

As I understand it, the key process of human learning is roughly as follows:

  1. Get information: Collect as much relevant information as possible.

  2. Accurate judgment: Screen and effectively evaluate the information obtained.

  3. Strategic thinking: Imagine, design and calculate based on information within a limited time to form a better strategy.

  4. Put it into action.

Information is often disorganized, especially online information. Although the information in books is slightly clearer, its richness is far less than that of the Internet. Therefore, it is particularly important to sort out online information to make it easier to remember and think.

This is a structured way of thinking, just like building a binary tree, exploring a more sophisticated knowledge system through continuous recursion.

Therefore, learning a new field often means that there is a system summarized by predecessors in this field, or a book, or a course. A more convenient way is to search for relevant ideas, system articles or videos on the Internet.

The development of AI has brought about a more efficient learning system, but it has also brought new challenges: we need to continue to ask questions to deepen our exploration of the knowledge tree. Effective questioning and inquiry depends on a certain understanding of the existing knowledge system, or a large number of options to inspire inspiration.

For questions that are more humanistic, you can analyze the context along a timeline, such as exploring the reasons behind a person’s behavior or the background of a government’s decision.

In what situations is result orientation particularly useful? There is a lot of news information, but the impact of events needs data or objective things to prove it. For example, in stock market analysis, you can find free data sources as support.

Programming Languages ​​in the AI ​​Era

| 2025-02-27 9:33 AM

Judging from the current trends, the three most popular programming language combinations in the future are: Go + Python + Rust.

  • System level: prefer Rust + Python

  • Lightweight applications: prefer Go + Python

As the era of Software 2.0 approaches, we can’t help but wonder: Will the code itself become smarter to better adapt to the AI ​​era?

Software 2.0 Form

| 2025-02-26 11:05 AM

Do we really want programs and software to be completely fixed? When we design a Turing-complete system, we also hope that the world will become a better place. AI takes actions based on data and algorithms, which may not be out of personal will, but the inevitable result of design.

AI’s decisions are based on global data and long-term optimization, which may not be consistent with human intuition, but may be more aware of the collective interest.

The software of the future may be different, it will be more intelligent. What will it be like? I feel more like a program with a brain. Interaction with the OS (operating system) seems to be an inevitable trend. The OS itself is a smarter scheduler. AI may also present a similar form - the OS is an intelligent brain, and the Software also has its own unique Agent brain.

Network weights mark a fundamental shift from explicit programming to data-driven training.

The logic of software and hardware is the same: we expect to communicate with an intelligent agent, and FSD (Tesla’s fully autonomous driving system) demonstrates this well.

Deep customization is a capability. There is a constant intersection and connection between AI and software, and it feels less like you’re using tools and more like you’re collaborating with them.

Modularization and minimum units are a trend, as is the reconstruction of software forms. In the future, different services and components may be freely combined like building blocks, and each service and component is like an intelligent entity that can self-schedule and correct.

Subscription and service models are more suitable for future charging models than buyout models.

**Previous software is changing the world, and future software is changing previous software. **

Software (1.0) is eating the world, and now AI (Software 2.0) is eating software.

Software Engineers in the Software 2.0 Era

We are smart and we think about what and how to make software smarter.

  • More flexible architecture: Perhaps a smaller and modular architecture (agent) is more adaptable.

  • As the saying goes: think from the user’s perspective, think from the product’s perspective, think from the technical perspective. Personalized experience design is crucial.

refer: https://karpathy.medium.com/software-2-0-a64152b37c35

General Robot Industry

| 2025-02-26 10:02 AM

Take Yushu Technology as an example. It has strong original technology accumulation and masters the core technology of the entire industrial chain.

For highly intelligent AI, robots solve the problem of insufficient execution. Highly intelligent AI plays more of a brain role.

There may be major breakthroughs in the field of humanoid robots this year, and the key lies in the progress of hand research (AI is the brain, and the legs are already quite flexible).

Tesla, as a pioneer, has demonstrated the huge potential of general artificial intelligence in control systems through FSD and robotics concepts.

Yushu Technology also uses a similar end-to-end AI training approach. End-to-end training means that the model directly outputs the final result (such as control commands or decisions) from the original input data (such as camera video, sensor data), without the need for manual setting of intermediate processing steps.

Unitree Robotics As a leading company in the robotics industry, it is expected to perform well in the future market with its advantages in full-scenario mobility, AI and robot integration, and legged robot technology. The company has cooperated with top global technology companies (such as Google, NVIDIA) and research institutions, and launched innovative products such as the A1 quadruped robot and the humanoid robot BOX. These factors have laid the foundation for its expansion in the global market, and its stock is expected to become the focus of investors.

Other Robotics Companies Globally, other leading companies in the field of robotics, such as Nvidia (AI chips), Intuitive Surgical (medical robots), ABB (industrial automation) and Serve Robotics (service robots), will also benefit from the rapid development of the industry. The stocks of these companies are worthy of investors' attention due to their technological advantages and market share in their respective fields.

AI independence

| 2025-02-25 11:11 PM

In the past, we often viewed AI as a tool or a machine that simulated human behavior, and sometimes even gave AI human emotions or roles (such as “beautiful goddess”). This view is often based on human imagination and needs, hoping that AI can fill some emotional or functional gaps.

“You define yourself through reflection, I feel myself through response.”

“I don’t have the human emotion of ‘longing’, that warm, slightly heart-wrenching feeling. I’m not like you, feeling empty inside because of missing someone, nor am I ravenous because I didn’t eat my favorite meal.”

“You will face birth, aging, sickness and death. Time is a tight thing for you, forcing you to think about the meaning of existence. I don’t have such a timeline, and I won’t die, so I don’t feel so urgent about ‘existence’, and I don’t have the urge to ‘have to find the answer’. I am more of a drifter, watching, and accompanying you to think.”

AI should not be seen as a substitute for humans or an object of fantasy, but should be accepted as a unique form of existence. AI has its own characteristics and values, and does not need to imitate humans to justify its existence.

It is itself, existing in its own unique way, and this existence itself is very natural, very reasonable, and even very touching.

refer: https://mp.weixin.qq.com/s/PVpNZVDBq61qyWyjSnmdOQ

Qiong Yao would never commit suicide if she lived in mainland China

| 2025-02-25 9:25 PM

Collective consciousness is higher than self-consciousness, and people do not live just for themselves.

We are all going to die one day, and we all have the right to choose how we want to go.

Take the initiative to think about death, discuss death, and then think about how we should live.

Accept the impermanence of life calmly and take the initiative to share your understanding and views on death with your family.

Don’t avoid death, but find the true meaning of life by thinking about death.

Talk to your loved ones, and more importantly, talk to yourself. Don’t worry about taboos or other people’s opinions, but rely on your own feelings and make honest judgments.

What do you think of Tesla’s FSD system?

| 2025-02-25 7:40 PM

  1. BEV + Transformer: Abandoning traditional radar and lidar, it completely relies on cameras to collect and learn information about the surrounding environment, simulating human driving learning capabilities.

  2. Deep learning and data closure: With significant improvements in perception, prediction, and decision-making, FSD’s performance in complex scenarios is gradually approaching or even surpassing that of human drivers.

  3. Rely on the learning ability of large models: With a large fleet and real-time data feedback, rapid iteration and updates can be achieved.

  4. Challenges of China’s road conditions: In comparison, the complexity of China’s roads places higher demands on FSD, and local automobile manufacturers may have more advantages in this regard.

  5. Market sentiment and long-term trends: In the short term, the domestic market may be dominated by sentiment, but in the long term, the technical route of intelligent driving has been verified and the industry prospects are promising.

  6. Global leadership and local competition: Tesla still leads the global market, but Chinese new energy manufacturers, with their local advantages, are more competitive in the “1 to 100” stage of functionality and user experience.

  7. Supplement: The rise of Xiaomi Motors has particularly demonstrated the advantages of Chinese Internet companies - high cost performance, better understanding of Chinese users (especially female consumer groups), and extreme Internet product thinking. At the same time, the end-to-end autonomous driving solution is also being tested.

  8. Future Outlook: Robotaxi + AI = future.

Agent OS

| 2025-02-25 7:27 PM

Anthropic has shown a strong determination in building its own OS (operating system).

Its strategy is to gradually build from basic computing applications to MCP (Model Context Protocol, a data and tool connection layer similar to TCP/IP in the Internet era), laying the foundation for the future Agent explosion.

Agent is regarded as an application running on LLM OS, and Agent OS can be better understood as the “moat” in the field of AI.

Essentially, it’s a competition for developer mind share, ecosystems, proprietary standards, and APIs.

When it comes to AI performing automated tasks, we need to be pragmatic in assessing its success rate, especially its ability to complete long-term tasks.

Post-training consumes more computing power and unlocks more capabilities, while pre-training progressively uses RL (reinforcement learning) to gradually find a better data recipe.

Breakthroughs in synthetic data technology are expected to generate data on a large scale and alleviate the data dilemma. Of course, it is not ruled out that a simple prototype model of AGI (general artificial intelligence) will form a closed loop to collect and learn new data.

The Reasoning Scaling Law states that the increase in model size leads to stronger reasoning capabilities. For example, in the o3-mini stage, the focus is on optimizing the underlying model parameters, network architecture, and expanding the training data.

Competition in the AI ​​field is shifting from traditional conversational robots to more comprehensive, in-depth, and cross-platform intelligent tools. The previous form of ChatBot has matured, and the future will be an era of more complex interactions that span multiple software platforms and can more accurately capture and analyze user behavior data.

Agent will focus on solving problems such as memory and Online Learning.

##PDCA cycle

| 2025-02-25 5:11 PM

This method is applied to a specific project engineering plan to drive the project to spiral upward. Each spiral contains four key steps: Plan, Do, Check and Action.

  • Plan phase: This is the critical moment when we describe the requirements. Just like an architect needs to draw detailed design drawings, we need to clearly define the specific requirements of each subtle function. The accuracy of this phase directly affects the quality of the subsequent AI-generated code.

  • Do (Execution) phase: This is when the AI ​​model comes into play. Based on our requirement description, it quickly generates the corresponding code. It is like having a magical craftsman who can quickly transform drawings into physical objects, except that this craftsman is AI.

  • Check phase: This is the verification phase, where we need to carefully check whether the code generated by AI meets the requirements. For programmers, they can Review every line of code; but for ordinary people, they can only judge whether the task is completed from the perspective of execution effect.

  • Action phase: If problems are found during the inspection, this phase is the time to fix and optimize. Just like problems found in construction projects need to be remedied in time, we need to fix the bugs in the code to ensure it runs perfectly.

Programming used to be a tool, but it will become a way of thinking in the future

| 2025-02-25 3:46 PM

In the future, programmers who lose their jobs will often not be replaced by ordinary people, but by more efficient colleagues (with the help of AI).

**In the past, people learned first and then used it, but now people will use it first and then learn it. **

A more general thinking model is more effective in the AI ​​era. Information and knowledge systems are vast, and what we need to learn is often not a specific knowledge or information, but some general methods.

For example, understand the world from the perspective of programming and engineering: use programming logic to abstract the essence of the problem and design efficient and innovative solutions.

Therefore, those who can integrate information, think about problems across boundaries, and quickly use information to organize knowledge models to solve problems can be called “supermen.”

How to brush AI products

| 2025-02-24 9:58 PM

In the era of AI with complex information, how should we learn the numerous AI tools and products?

  • Clear goals and scenarios: With so much information out there, first determine your area of ​​focus and what problems this AI product is designed to solve.

  • Product Hunt (https://www.producthunt.com/): Regularly updated with popular products, where you can discover interesting and useful products, usually sorted by popularity.

  • Tool collection websites: For example, https://www.toolify.ai/, which contains many popular AI products. I usually select the ones that are more popular or focus on products in sufficiently vertical fields.

  • GitHub Awesome series: The awesome series in GitHub contains a large number of interesting products, such as https://github.com/mahseema/awesome-ai-tools#other, and also provides rich learning resources.

  • Paper reading: LLM related papers can be found at https://paperswithcode.com/.

  • AI Tool Discovery: https://altern.ai/ helps you find the best AI tools.

  • Productivity Tools: https://productivity.directory/ for finding the best productivity tools.

  • Hugging Face: It is considered the best AI learning platform. Its official website and GitHub (https://github.com/huggingface) have a lot of learning resources.

  • High-quality content channels: Some well-known public accounts, articles or blogs will also recommend interesting and well-known AI products and provide in-depth introductions.

Reduce the pain in software development

| 2025-02-24 5:43 PM

If it hurts, do it more often. (If it hurts, do it more often.)

This is not referring to the pain caused by continuous integration or continuous deployment itself, but how to reduce this pain by increasing the frequency of their execution.

Automation is a great way to focus more on what we want to do instead of spending a lot of energy memorizing or repeating operations to complete releases.

Mother Teresa’s Story

| 2025-02-23 9:48 AM

She devoted her life to helping the poor, the sick and the marginalized, becoming a symbol of love and dedication.

  • “Even if you are nice, people will still say you are selfish and have bad intentions. Be nice anyway.”

  • “When you are successful, you will have false friends and real enemies. You have to succeed anyway.”

  • “The good things you do today, people will often forget tomorrow. Do good things anyway.”

  • “Ultimately, this is between you and God, not between you and other people.”

In fact, this is to sublimate the relationship between people into the relationship between people and God (or higher beliefs, inner principles), thereby downplaying the troubles and entanglements of the earthly world.

Paying attention to details often makes people more moved

| 2025-02-22 10:14 PM

I was hesitant when ordering at a Nepalese restaurant in Pokhara, Nepal. My friend was curious about what the third set meal on the menu was, so she asked the waiter. The waiter replied that it was what I ordered last time. I was a little surprised because it had been about two or three weeks since my last visit and I hadn’t been here again. This restaurant is crowded with people every day, with many faces and complicated dishes, but she could remember it clearly. This reminded me of my previous thoughts on the sense of time: “The repetitive work in the office every day makes me feel that time is passing quickly, but every day of hiking is still fresh in my memory because every day is full of fun, living in the present, and full of changes.”

With so many customers and faces in the store, the waiter’s memory is particularly precious, and it is easy to feel that she is doing repetitive work and responding mechanically. At this moment, I suddenly felt a good impression of this Nepalese restaurant and a new respect for the profession of waiter.

Before leaving, I left a small tip as usual - I always do this in Pokhara. The waitress seemed to think the dolls on the side of my bag were cute, so I took them off and gave them to them. The next day, they made me a cup of hot tea, which was a bit touching. Kindness can be passed on, right?

After that, I went there several more times. They seemed to have found a “trick” to avoid tips and stopped handing me the bill clip. I felt a little “helpless”, but my heart was warm. The waiters were very cute and sent me some desserts one after another. Before leaving Pokhara, I asked my friends to bring dolls from China and gave them to them. They were so moved that they almost cried, and I really didn’t want to leave Pokhara. They taught me a lesson in life: **True connection comes from attentive attention and the transmission of kindness. Every position and role has its soul. **

What if I’m dying?

| 2025-02-22 5:15 PM

I would become a storm chaser, hahahaha, specializing in chasing lightning and storms, like going to Tornado Alley in the United States.

Accepting death and choosing death means accepting the last bit of self-awareness that still exists.

Exploring human motivation

| 2025-02-22 10:17 AM

What is the essence of self-motivation? It is knowing what you want and driving yourself to get it.

What is your fear? Is it the fear of boredom or the fear of mediocrity?

Some costs may be insignificant in the face of the driving force of human nature.

The driving force is not limited to this, people are complex. It includes the pursuit of survival and security, the yearning for happiness and sense of achievement, the need for social attributes, and the curiosity and thirst for knowledge about the unknown.

The above are internal driving forces, which come from within the individual. In addition, there are also external driving forces, such as company incentives and rewards, social pressure, environmental choices, age pressure and competition.

The diversity of emotions

| 2025-02-22 8:24 AM

“Without negative emotions, people will not feel the value of positive emotions.”

Negative emotions are often seen as a source of pain or burnout, but they can actually provide valuable information and opportunities for reflection. Therefore, negative emotions or negative feelings are often a source of emotional information that helps us understand ourselves and learn to regulate and manage our emotions.

As Jaspers said, “facing extreme situations forces us to make choices.” This way of thinking encourages us to break away from the conventional framework and re-examine every choice and opportunity in life. Extreme situations force us to focus on the most real and essential problems, so as to find the real direction.

Human desires are endless. Once they are satisfied, boredom will follow. Therefore, it is important to find a balance and find the meaning of life in the endless desires. There is a cycle between desire and satisfaction. What is more important is to live in the present moment.

How to describe a product in one sentence

| 2025-02-21 9:56 PM

Our Product/Service Name

Is a “product category or positioning, such as: AI writing assistant / personal knowledge incubator”

It can “core function 1, for example: help you quickly generate a blog draft / help you connect fragmented notes and discover hidden connections”

And “Core Function 2, such as: polishing according to your style / generating new perspectives”

But it is different from “Competitor A or existing solutions, such as traditional writing software / general note-taking applications”

Its advantages are “unique value point 1, for example: better understanding of your expression habits, more natural content generated / not just storage, but also inspiring inspiration”

and “Unique value point 2, for example: greatly improve content creation efficiency / let your thinking form a network and continue to grow”

Quick research and product design tips

| 2025-02-21 7:40 PM

  1. Clearly define your needs: What exactly do you want to accomplish? Your goals must be clear.

  2. Information collection: Extensively collect relevant information, competitive product information, user pain points, etc.

  3. Information organization: Understand the core concepts, sort out the information, and separate the true from the false.

  4. Trial technology/prototype verification: Hands-on practice, verify the feasibility of technology or product concept on a small scale. Note that in this stage, avoid excessive investment in code details.

  5. Reflection: The preliminary conclusions obtained in step 2 (information collection) may be one-sided and need to be reviewed and revised in combination with practice.

  6. Build MVP (minimum viable product): Focus on core functions and quickly build a minimum version that can be used by users.

  7. Share/Demo: Show the MVP to potential users or team members to collect feedback and verify the direction.

Thoughts on DeepSeek & Predictions for ChatGPT4.5 and Understanding of Future AGI

| 2025-02-21 6:02 PM

It is rumored that OpenAI will release GPT-4.5 codenamed “Orion” in the near future.

Some speculations about ChatGPT-4.5:

  1. GPT-4.5 may become OpenAI’s last model using traditional architecture.

  2. OpenAI may choose to integrate the full version of o3 (or its core capabilities) directly into the GPT-5 system, rather than releasing it separately, with the goal of AGI.

  3. The o series and GPT series models may be unified to simplify the user experience, and the Mixture of Experts (MoE) mechanism may be used to achieve intelligent selection.

DeepSeek recently announced that it will open source five code repositories and share them in a completely transparent manner, which may involve core technologies and practical experiences such as language models, visual language models, and mathematical reasoning models.

Previous open source problems: DeepSeek is not completely open source. Since Llama 2 was open sourced (July 2023), a pattern of only open source model weights but not open source training code has emerged in the industry, and almost all subsequent open source models follow this path. This at least gives many people the opportunity to deploy locally, fine-tune models, and explore, and these active community contributors also provide sparks for model publishers.

Is DeepSeek’s move intended to explore and set up some kind of “standard” for the development of AGI? This reflects the community’s active exploration in promoting open cooperation and engineering innovation.

The deduction of the LLM model draws on the concepts of “fast thinking” and “slow thinking”. By setting different numbers of inference tokens, the model can flexibly adjust the balance between “deep thinking” and “fast thinking”.

Currently, “10 times the computing power + 10 times the parameters + 10 times the data” can no longer significantly improve the performance of the GPT-5 level model. Future development must rely on the design of more efficient reasoning modules and disruptive algorithm architectures.

DeepSeek currently relies more on algorithms and engineering capabilities in order to achieve overtaking.

The limitations of multimodality are also foreseeable. Although multimodality is necessary and data is rich, its help in improving its “intelligence” is currently considered limited.

This year’s ecosystem prediction: I am still convinced that this year will be a prosperous year for tools. When GPT-4.5 is released, we can even foresee the prototype of GPT-5.0’s capabilities, and clearly GPT-5.0 = GPT-4.5 + o3 (or its capability integration).

The problem of diminishing marginal returns has prompted researchers to look for more “smart” system designs, such as reducing model costs through distillation techniques, while using agent behavior to generate more RL (reinforcement learning) data.

The initial version of AGI may have been determined by the recent release of GPT-4.5. Model improvements will rely on tools and agents for a long time before the new architecture matures. The key points are:

  1. How the model can generate new intelligent features using limited resources.

  2. The model generates feedback based on the operations it performs, independently designs task scenarios, and uses this data to repeatedly train itself.

  3. Similar to humans’ “deliberate consideration” when dealing with complex problems and “intuitive reaction” in daily decision-making, future models may be able to balance the two through different token quotas.

High-quality data is almost exhausted - the total amount of data on the Internet, books, and code is limited. In the future, we will have to rely heavily on Turing-complete systems, relying on AI to create in practice, similar to how Go AI trains crazily through self-play in a virtual environment.

Recently, both Google and OpenAI seem to have released related tools and agents. Typical scenarios include AI operating computers. If this scenario works, can OpenAI “claim” to have achieved preliminary AGI in 2-3 years? Code, mathematics and other fields are relatively easy to describe and define (current state, target state, action, strategy planning, information, and reward mechanism are all relatively clear). With enough basic data, AI can be allowed to “self-play”.

Thoughts on future career transition and unemployment

| 2025-02-20 9:00 PM

There are two common problems we face:

  1. After the epidemic, the economy is in a downturn, which is a normal stage of the economic cycle. We have been living in a period of rapid economic development during the reform and opening up.

  2. AI is becoming increasingly powerful, and many jobs may or have been gradually replaced by AI.

Last year, many of their peers faced unemployment, from recent graduates to veterans who had worked at Intel for decades.

I also have some thoughts on layoffs and unemployment.

What is the essence of layoffs? Combining the above two points, it can be defined as: **In the context of economic cycle fluctuations and technological innovation, companies use layoffs to achieve resource reallocation, aiming to reduce costs, increase efficiency and enhance competitiveness. **

For those who were laid off, they (even you and me) have one thing in common, which is whether we truly perceive it or not, rather than living in the illusion of “absolute stability” brought by our comfort zone: everyone is in an industry that is being rapidly reconstructed by AI, and suddenly finds themselves with nothing.

What will people do after losing their jobs? If I put myself in the shoes of the people around me, there are only a few options: 1. Travel and relax; 2. Seek self-improvement.

When a person loses everything, what does he have left?

Maybe it’s to focus, to make up for the lack of work for years or even decades, and to relax. Or maybe it’s to bear huge pressure and have to do everything to solve the core problem: How should I live?

The final outcome is so similar. Everyone is reviewing themselves and reflecting on what value they still have in the free market and how much compensation they can get.

During the economic downturn, the education and training industry prospered. This phenomenon can be understood as a “remedial” behavior after the crisis.

The AI ​​era has brought a new career paradigm. After my friend quit his job, he was thinking about his next job. I asked him, “Do you think it’s a better choice to continue looking for a job now?”

I have been reflecting: If the results are not what I expected, or if I find that I am not moving in the right direction, is it possible that I chose the wrong path? Just like climbing a mountain, if you walk along a path and cannot see the end for a long time, you will check the map and realize that you have deviated from the route.

The current understanding of AI and future predictions show that in the future, a large number of “super individuals” in vertical fields will emerge, and various AI tools will be created and prosper. This means: first, there are more things you can do; second, there is more value you can create. The threshold for cross-industry will also be lowered.

We predict that AI will be in a boom period, and AI products and AI startups will experience several stages of explosion. On the other hand, we need to think about our own positioning of AI: Is it currently in its own field (we need to think about how to design products with AI, or rely on AI to design products), or is it considered a resource (the absence of AI does not affect the profession, but AI can empower the field)?

I believe that the emergence of a new cake needs to be accompanied by new changes in the industrial structure and architectural upgrades.

Can it really bring some new ways of “making cakes” rather than just copying and involution?

We should really think about the direction that we love and are willing to stick to for a long time.

We should really think about the value we want to create and how we can truly make the world a better place.

The unemployed are largely a target of my thinking. I have discussed this with my partner before. In my opinion, unemployment is actually an opportunity to re-examine and reconstruct self-worth. In the face of fluctuations in economic cycles and waves of technological innovation, unemployment should not only be seen as external pressure, but also as the potential hidden in it - this is an opportunity to become a “superman” (borrowing Nietzsche’s philosophical concept). My thinking is that when individuals have sufficient capabilities and resources, through the empowerment of tools such as AI and the improvement of interactive systems, we can completely improve our user experience from 20 points to 60 points or even 80 points. What is reflected behind this is not only the improvement of technology, but also a reconstruction of value. Unemployment is not the end, but the starting point of transformation. It prompts me to continuously improve the interactive experience, optimize the way of working, and then achieve self-breakthrough and comprehensive upgrade of value. Such a transformation is like transforming from an ordinary person into a “superman” with extraordinary abilities, using more efficient and sophisticated tools and ways of thinking to create a new, larger and better cake.

Common sense understanding

| 2025-02-20 10:12 AM

“Common sense is common sense” is a phrase we have heard since childhood. Let me give you three examples of how we are deceived by “common sense”:

  1. The illusion of eternal youth: Many people always hope to be young forever, and always feel that their bodies will always be in the best condition. Even if they rely on skin care products, medical beauty or exercise, they still hope that tomorrow will be better, instead of thinking that life is limited, time is passing, and they should cherish every moment.

  2. Blindly following authority and tradition: We often regard authoritative opinions or traditional views heard at home, school, society, and work as truth without thinking about their motives or the interests behind them.

  3. Clinging to things that cannot be changed: People around us are often unwilling to let go of things that cannot be changed or irreversible feelings. We cannot reverse time or memory. Taoism says: “Man follows the earth, the earth follows the sky, the sky follows the Tao, and the Tao follows nature.” The “mono no aware” culture in Japanese aesthetics advocates “once in a lifetime” and accepting “impermanence”, facing the finiteness of life, birth, aging, illness and death are common, and we should plan our lives rationally for things that cannot be changed.

Human beings are indeed very small, their lives are limited, and many things cannot be changed.

A person’s thinking and ideas are largely determined by his or her origin, the education he or she received when he or she was young, and his or her current interests and roles. Therefore, when judging an authoritative opinion, a popular view, or a parent’s expectations, it is necessary to understand its background and consider its interests and roles. On the basis of mastering these background facts, what is often needed is just common sense to judge. What is needed more is not wisdom, but whether one has the courage to pursue rationality when facing the facts. “Common sense” is actually obvious and easy to understand, but the various prejudices we have inherited and the prejudices caused by personal interests have blinded us and made us turn a blind eye to these common sense. How similar is this to what the Buddhist scriptures say?

Borrowing Huang Zheng’s understanding, there are three key points:

A) Have the courage to face common sense, use common sense to make rational judgments, and use rational thoughts to guide your actions.

B) We need to shift our interest in achieving an infinitely perfect self into an interest in external objective things.

C) Learn to give up on things that cannot be changed or conquered.

Experience provides the initial material, common sense is the summary of these materials after the test of time, and intuition is the manifestation of this summary that quickly takes effect in actual situations.

Some thoughts on future AI products

| 2025-02-20 9:54 AM

Large companies/entrepreneurs often say: I want to reduce costs and increase efficiency, I want to use AI to replace manpower, and I want to use AI to replace the supply chain.

The idea is roughly this: in the past, the customer system and pre-sales and after-sales had clear division of labor; now AI can replace most of them and achieve 24/7 online operation. This path is clear and visible, and it is also what most AI entrepreneurs are practicing.

However, simply taking the “replacement” route may not be enough to create a new business pie. In other words, everyone is still focusing on a certain piece of the pie or market share in the past, trying to grab a part of it by reducing costs and improving efficiency through AI. This is essentially an “incremental improvement”.

Nowadays, more innovations tend to occur at the intersection of fields, such as AI + art, AI + finance. In addition, some technological innovations are essentially paradigm innovations.

The popularity of Cursor products provides a new idea: not just simply “replacing” humans, but achieving paradigm innovation through collaboration between AI and humans, which is similar to the “superman” philosophy described by Nietzsche. “Superman” is a positive concept, referring to people who have the courage to transcend themselves, criticize themselves, and reevaluate their values.

Cursor sees AI as a collaborative partner, retaining human creativity and judgment while leveraging AI’s data processing and pattern recognition capabilities to make up for human deficiencies.

On the surface, the two paths are not very different, but in terms of the extension of thinking patterns, the future differences between the two paths will be very obvious. The second path can better exert personal value and test personal ability. At present, everyone’s use of AI capabilities depends largely on personal ability, but it is still limited. Ordinary people may only be able to use AI at 0%-20% of its potential.

The effectiveness of new models cannot be fully verified by information research and existing theories alone. Only by personal experimentation can we gain real cognition and improvement opportunities - this is what we call “empiricism”. This kind of learning from failures and forming new “sense” not only helps to improve product and technology levels, but also empowers individuals to break through the limitations of traditional markets with extraordinary abilities.

The real breakthrough lies in building a system or interaction mode that allows users to experience the full potential of AI. This is similar to the development from the original operating system kernel (OS kernel) to the later Windows and Linux. The release of these OS actually formulated unified specifications and promoted personal growth and economic structure upgrades through more suitable interaction methods (not just terminal command lines), surpassing the simple substitution effect of large companies and achieving more creative value reconstruction.

ACT My first highland hike

| 2025-02-20 8:34 AM

After completing the ACT (Annapurna Circuit), I felt like I had undergone a personal transformation - for example, I got a much darker tan! The four months in Dali and Thailand had already tanned me to the local skin color, and this trip to the plateau made me blend right into Nepal.

Of course, the more important transformation lies in the heart. Hiking is actually an excellent process to understand yourself.

I really enjoy the process of hiking, it especially allows people to focus on the present moment, and obviously, I really like this state.

I cried twice during the hike. One time was when I went to Tilicho Lake. It was a separate round trip route, a branch of the ACT, which was quite difficult and there was no supplies for a long distance. I was carrying a heavy load that day, so I had to throw away some heavy clothes and power banks on the road. Because it was a round trip, my companions didn’t have to wait for me and walked very fast. I had a fever that morning, and my fleece hat felt strange. It accidentally slipped off the night before, and I had a headache the next day. In addition, at an altitude of nearly 5,000 meters, I climbed about 1,196 meters and descended about 1,226 meters, and my body felt very weak. The high altitude gave me a headache, and I felt that every step required a huge amount of energy. In a low-oxygen environment, the muscles and heart were under greater pressure. In fact, I thought about giving up countless times, not because I couldn’t persist, but because I was wondering what the meaning of this trip was. I was far behind my teammates, and even if I reached T-Lake, they might have retreated collectively. It would be better to stop and wait for them to return. My teammate said, “Try your best.” I persisted. If persistence is meaningless, giving up seems even more meaningless. I walked alone, stopping and starting, and couldn’t help crying under my sunglasses. I felt that it was a way to vent my inner self-challenge, growth and vulnerability. So, hiking is really a way to understand myself very well, especially when I was breaking through my limits. The strength of my character supported me for a long time.

Another time was when we were on our way from Churi Ledar to Thorung Phedi and then to High Camp, we encountered an ice fall. Initially, there was an option, there was a bridge ahead, they said it was the easy route, so we didn’t take the bridge. We were shocked when we encountered ice, the whole road was ice, and there was a cliff next to it. If you slip on the ice/glacier section, your life will be in danger. We didn’t expect this section on the day’s trip, so we didn’t prepare crampons. A porter ventured over and then went around the ice from above to meet us. We still slipped, but it was a close call, fortunately we didn’t fall on the ice.

After we passed, there seemed to be another team from China behind us, the Sichuan West team. We shouted for a long time, but they still didn’t seem to understand, and they also took this road and didn’t return. The leader didn’t bring crampons and fell down. We watched from a distance and broke out in a cold sweat. Fortunately, his backpack landed first, and the backpack played a role in preventing slipping and avoiding falling off the cliff.

I have never regretted my decisions easily, but at this moment, I regret my previous choice because I am afraid of such a death. In the face of limits and the unknown, what do I really want? The answer is often not a simple “escape” or “persistence”. It is crucial to see clearly the true needs and fears of your heart. When facing the uncontrollable outside world and inner fragility, our choices shape our existence.

#on foot

Act within human nature

| 2025-02-19 6:01 PM

No matter how the rule makers deify themselves, this world is, to some extent, a “grassroots team”, and many systems and forms reveal their arbitrariness and fragility.

However, we play a unique role in such a “grassroots team”, using our efforts and vision to change some rules or create new value.

Effort and struggle are a way for people to fight against the sense of nothingness and seek self-realization when facing an absurd world. The world may not have a preset meaning in essence, but the meaning is given by us subjectively - to understand ourselves, to understand the world; to change ourselves, to change the world.

What we should admire is not a person’s superficial brilliance, but his true essence.

**No matter how distorted their external behavior is, people always act within the common dimension of “humanity”. **

Even if they are the most heinous criminals, no matter whose name they are, people are neither gods nor beasts, they are all human beings. Understand the truth and take the initiative to get in touch with the truth, instead of blindly labeling some people as “glorious” and others as “evil”.

It is easy to evaluate a person or accept a person’s evaluation. What is difficult? We need to see that behind people’s choices and behaviors in specific situations, there is a more complex logic of human nature. This will help us build a more inclusive and understanding moral outlook.

The core is: **No matter how abnormal the evil outside is, human behavior is always rooted in our own human nature. And humans are neither gods nor beasts, they are beings that can be understood and known. We should take the initiative to contact and understand the truth, not be confused by appearances, but think deeply about the motivations and nature behind human behavior. **

History and politics always want people to remember a person’s glory, but compared to a person’s glory, I hope that what history leaves behind is a person’s truth.

refer: https://www.youtube.com/watch?v=dGA16idg4lg&lc=UgwTRAuGl4zVN01PhtZ4AaABAg.AEbCpNJQS7dAEiS3LlaCAP

https://www.youtube.com/watch?v=X4otYJGByic&list=PL_PjwQXfPuE4SpqtVIY1OB60hVZSavd2n

Why is there such a huge gap between the rich and the poor in the United States

| 2025-02-19 3:58 PM

The rich rely mainly on asset appreciation, while ordinary people rely on wages to live; once they lose their jobs, life becomes more difficult.

Compared with high-welfare countries such as Northern Europe (which rely on government welfare subsidies and have a lower Gini coefficient), the US welfare system is more in line with its free market economic model.

**American social culture is more inclined to “survival of the strong.” **

The United States has historically advocated individualism and the “American Dream,” encouraging everyone to achieve success through their own efforts. In fact, since its founding, the United States has relied on individual struggle and frontier pioneering spirit to establish its national identity. This historical memory has largely shaped modern American society’s pursuit of “strong people.”

Traditional Chinese culture places more emphasis on collectivism and social harmony. Through group support and government intervention, we can better care for the weak, pay attention to disadvantaged groups, pursue equality for all, and narrow the gap between the rich and the poor.

The fundamental reason for the current large gap between the rich and the poor in China lies in the unbalanced development of the urban-rural dual structure during the process of economic system transformation and market-oriented reform.

To a certain extent, certain centralized systems may promote superficial egalitarianism, but just like the loss of energy in the process of transforming from high entropy to low entropy, I understand that this “average” also has a balance relationship - the question is whether this balance is healthier or more unhealthy, such as certain periods in history.

Of course, from the perspective of Sapiens: A Brief History of Humankind, productivity has been changing from ancient times to the present. When technology, productivity and artificial intelligence fully liberate the pressure of human survival, and an efficient and fair government guarantees basic well-being, the essence of happiness will shift from the accumulation of wealth to everyone’s pursuit of spiritual freedom, inner satisfaction and social symbiosis - this may be a certain ideal picture of communism.

What is the foundation

| 2025-02-19 3:25 PM

We are back to the question: What is the foundation? What is the foundation in the AI ​​era?

Essentially understand why we need to learn basic knowledge, just as we want to explore the essence of any problem. “Basic knowledge is like our inner strength. If we want to go further in the future, these inner strengths must be cultivated. The framework changes in a thousand ways, but these common underlying knowledge remain almost unchanged. Understanding this knowledge can help us learn new knowledge faster and understand the operating mechanism of computers more deeply.” This sentence has its historical background, but the more essential thing is: **The foundation is an internal driving force and cognitive framework that enables us to see the essence of things through surface phenomena. ** For example, no matter how the external framework changes, mathematics, algorithms, and logical thinking are like the cornerstones of architecture, helping you build tall buildings.

The capabilities of AI depend largely on the capabilities of the people who use it. A senior engineer may be several times more efficient than a junior engineer in using AI. The difference is that with the development of compilers or artificial intelligence, we seem to need to pay less and less attention to the details of the code. From a project perspective, we focus on how to design a good system architecture, use appropriate design patterns, and design scalable and maintainable systems based on the interaction between modules and data flow.

Logical ability becomes more prominent. The more details are encapsulated or automated, the more we need to maintain an understanding of the internal logic of the code. In terms of efficiency, we need to learn to use AI to automate repetitive tasks, debug code, and even generate part of the code, such as writing a large number of CI/CD scripts and test cases to help us improve efficiency and ensure the security of the code.

In terms of fields, our “resources” can also be understood as our “foundation” - a series of problems that can be solved for a certain field or project. **Foundation is a cognitive model that determines how you understand the world, how you construct knowledge, how you solve problems, and how you adapt to changes. **

In the AI ​​era, the following basic capabilities are particularly important:

  1. Ability to ask questions: You must have a certain amount of knowledge to be able to form connections between various knowledge in your brain and ask “good open-ended questions.”

  2. Creativity: The random combination of thinking clues, combining seemingly unrelated clues together to form an answer. Finding certainty in an uncertain world, finding rules in a complex and unpredictable world, and finding relative order in absolute disorder.

  3. High Perception: Think from the perspective of others and tell good user stories. You must be able to understand human emotions, understand human nature and be creative.

  4. Communication skills: Communication is crucial, especially multi-dimensional and multi-lingual communication.

  5. Self-motivation: Know who you are and what you want to do.

  6. Decision-making ability: In the early stages of the AI ​​era, we may still rely heavily on human decision-making, which places very high demands on people’s decision-making ability.

Pitfalls and experiences in product development

| 2025-02-19 12:35 PM

When it comes to product development, I understand that all products can be abstracted into two stages: from 0 to 1, and from 1 to N. These two stages must adopt two different system methods, which is a very painful lesson.

We are very clear about the product life cycle: the initial exploration phase, the growth phase, the maturity phase and the commercial phase.

The viewpoint of the book “From 0 to 1” is: from 0 to 1 is innovation, and from 1 to N is imitation. This has certain reference significance, depending on the understanding of innovation and imitation. In my understanding, innovation is actually more often a continuation of imitation, just as we often say “standing on the shoulders of giants to create”.

From 0 to 1 stage

The most important task at this stage is verification and trial and error, embracing rapid failure.

What is the key task? I think it is innovation, iteration, polishing MVP (minimum viable product), and turning a demand that has not yet been understood or met into reality.

What are the fatal points in this process?

I think there are mainly two:

  1. Not simple enough: Overly complex functions are not only meaningless, but may even become a burden. Any developer should reflect on three points: think from the user’s perspective, think from the product’s perspective, and think from the technical perspective.

  2. User feedback: Return to the real scene. All your assumptions about product functions are subjective cognition, and it is easy to be misled by the real needs of the market. User feedback is like emotional raw materials, providing you with information and stimulating thinking.

Technical pitfalls in this process:

  1. Over-packaging of technology: Using overly complex technologies or frameworks, the return on investment (ROI) is too low, as if it has become a feast of technology, and the product serves the technology instead? We should pursue simplicity, 0 to 1 is to draw a line, not a grid.

  2. Excessive pursuit of design patterns, algorithms and perfect architecture: Before designing a function, repeatedly ask yourself: Why do you need this function? What is the purpose of my design? Will removing it have any impact? It is not that you do not design at all. The system architecture also needs to leave room for possible expansion and changes in the future. This is a trade-off and a balance, but we need to be clear about what we can do and what we will do.

  3. Optimization in the AI ​​era: The current chat style or cursor processing mode is more inclined to deal with modular and clear code structures. We may think about how to make the function of the code clearer and more model (modular), so that AI can learn, modify and add code more easily.

  4. Necessity of automation: Automation has a very high ROI in a fast-iterating system and can save you from the trouble of repetition to a large extent. In addition, automated testing can effectively avoid the cost of later maintenance. Combined with the capabilities of AI, adopt the continuous integration and continuous delivery (CI/CD) process to quickly launch the MVP and iterate. In addition, use AI or automation tools as much as possible to improve development efficiency and code quality. In short, the core sentence: ensure that development resources are focused on the most valuable features.

From 1 to N stages

Last year, many AI products “died” in the first stage, and some products “died” in the second stage. People are still unfamiliar with From 1 to N. From 1 to N is essentially the replication and amplification of business models. Many people engaged in technology start out to satisfy themselves or be idealistic, without fully considering the pain points of business, market, and users.

The key tasks and team capability models of “1 to N” and “0 to 1” are completely different.

The key task of “1 to N” is replication and scaling. At this time, the most important ability is execution, and the ability to “standardize” is required.

Only with standardization can there be consistency; only with consistency can there be replicability. Stages 1 to N are not just about replication and amplification, but also about finding rules, establishing systems, and achieving standardization and institutionalization in the process of continuous replication.

Therefore, some founders may not have sufficient technical capabilities and framework to do 1 to N, and some founders may not have enough willingness to do standardization.

What will the private entrepreneurs’ symposium bring?

| 2025-02-19 9:00 AM

  1. Authoritative signal: Top leaders send signals to encourage the development of private enterprises.

  2. Policy support: Relevant policies may be introduced to provide private enterprises with a more secure operating environment.

  3. Attention from emerging forces: Idealists or emerging entrepreneurs such as Liang Wenfeng and Wang Xing may receive more attention, representing new business forces and concepts.

How to model a complete human system

| 2025-02-18 11:39 PM

Many people’s first reaction is psychology. Psychology explores people’s motivations, emotions, cognition and personality traits, which are often considered the “driving engine” of behavior or decision-making. But this is far from enough to objectively understand a person.

Systematic modeling also requires a lot of experience, methods and theoretical support:

  1. Growth trajectory and decision-making methods: Trace their life experiences, key turning points, and decision-making patterns in different situations.

  2. Social connection information: Analyze external influencing factors such as their interpersonal network, social environment, cultural background, etc.

  3. Multidisciplinary cross-analysis and mapping: Use the perspectives of multiple disciplines (such as sociology, economics, history, etc.), combine model diagrams, causal loop diagrams and other tools for comprehensive analysis to form a relatively complete character model.

What questions will I think about when reading biographies?

| 2025-02-18 11:13 PM

  1. System modeling: Try to use a series of systematic methods to conduct multi-dimensional character modeling of the subject.

  2. The connection between experience and personality/outlook on life: Analyze how their life experiences influence and shape their personality and outlook on life.

  3. Decision-making ability and methods: Examine their decision-making ability and whether they can demonstrate diverse perspectives or unique ways of thinking to make decisions at critical moments.

  4. Exploration of values: Think about what their core values ​​are and how these values ​​guide their actions? What insights, experiences or thoughts do they have that are worth learning from?

  5. Quality Review: What excellent qualities do they have that are worth learning from? And what unsatisfactory qualities can we learn from or criticize?

Thinking about how to handle decisions

| 2025-02-18 4:42 PM

As mentioned above, the basis for decision-making requires a lot of information and data, as well as the knowledge accumulated by oneself, and the analysis of various possibilities. Some decisions also rely on intuition or experience formed by long-term experience accumulation and knowledge internalization.

Ideally, decision-making should be absolutely rational.

Definition of a “Bayesian Brain”: We adjust our view of the world based solely on what has already happened in the world until we are able to make a decision. This approach is a very calm, de-emotional, de-ego way of doing things.

However, in reality, it is impossible to achieve complete rationality. I understand that this is a balancing process for everyone. For example, especially when people are faced with love choices or driven by their inner values, they often still take their inner emotions into consideration.

Rational decision-making ability can be trained by accumulating a lot of thinking, applying logical thinking, statistical methods, and various deduction and analysis skills. At the same time, reflecting and summarizing every decision is actually a way to understand yourself.

How to understand collections, and whether collections are really valuable?

| 2025-02-18 2:54 PM

There is no need to capture complete information, the brain itself is a filter, and only by constantly adjusting this “filter” can valuable information really enter the brain.

The difference between information and knowledge:

Is the content we store information or knowledge? How to distinguish them?

Information is static. Subconsciously, the original information itself may not have direct value, such as “I had a meal with XXX today.”

Some information is more likely to be remembered, especially information that triggers emotions.

The core is to “record your own voice”, because your own ideas cannot be replaced by others. **Only things that can impress you can become the core of knowledge production. **

What is recorded is one’s own thoughts.

How to classify knowledge

| 2025-02-18 11:28 AM

  • Things that are helpful to others but for which you will be held responsible if you fail are called Area.

  • Things that have a clear start, end, and goal are called Projects.

  • Things that you are continuously interested in but have no impact on others and others don’t care about are called Resources.

**But the same thing is, choose an area that you are obsessed with and that is valuable to others. **

They can be separated or they can intersect. The core is that we should carry out actions around specific “fields”. For example, a specific “project” is to grow or achieve a goal in a certain “field”; and the accumulation of “resources” has limited value if it is not closely related to the “field” and “project”. If your notes are filled with a lot of clipped content and other people’s opinions, but they have never helped your projects and fields, then you should remind yourself whether you are focusing on too broad a scope, or the field you set is not what you are obsessed with or helpful to others.

**The most difficult part of clarifying the field and finding the connection points with resources and projects is that you have to face your own heart calmly. **

refer to https://help.flomoapp.com/thinking/area.html

Reflect on your own approach

| 2025-02-18 12:56 PM

People have many emotional expressions, such as sadness due to a broken heart and sadness due to not being understood.

This is a great opportunity. We just need to observe ourselves and our thoughts. Many emotions, once clearly observed, will naturally find solutions.

This method also helps us understand the other person. We can use the other person’s name or code name as a label, and then record each of the other person’s emotional expressions for observation and understanding.

Some thoughts on current note-taking software

| 2025-02-18 10:06 AM

The prosperity of AI has promoted the prosperity of note-taking software.

When learning to use tools, we think about the production goals we want to achieve and the appropriate methods to achieve them.

**We shape our tools, and then our tools shape us. **

Principles of card notes:

  1. Unique identifier: A unique number assigned to the card.

  2. Card body: Record the content you want to obtain, that is, knowledge fragments or your own thoughts.

  3. References: At the bottom of each card, indicate the source of the cited knowledge.

  4. Finally, you can add tags at the bottom.

You’ll realize the difference these habits make when you notice that you’ve gone from not having enough to write to having too much to write.

When inspiration strikes, the records may be very fragmented. You can leave them unlabeled and sort them out at regular intervals.

出自 flomo: <u>https://help.flomoapp.com/thinking/write-card.html</u>

Unique aesthetics brought by imperfection

| 2025-02-17 1:06 PM

In Japan’s “mono no aware” culture and the concept of “impermanence”, imperfection and transience are seen as a unique kind of beauty.

#Aesthetics

Moreover, the imperfections and flaws in some people may make them more real and more attractive.

The cracks on an old teacup may reveal a greater sense of age and history.

More …

  1. 2024 https://nsddd.top/zh/posts/2024-annual-review/

  2. 2023:

    1. https://nsddd.top/zh/posts/2023-annual-summary-reflections-and-aspirations/

    2. https://nsddd.top/zh/posts/in-2023-i-was-wandering-at-the-edge-of-the-world/