January 2026 Thought Notes

Manus roundtable: the value of AI’s certainty

Some content from Manus’s roundtable

The next wave of AI that actually makes money isn’t about generating fancier content, it’s about making “work” more certain, more deliverable, and with an ROI you can actually compute

The difference between ToB and ToC isn’t who’s bigger, it’s who closes the loop more easily

What’s scarce in ToC is attention; you keep winning users’ minds, but once a big company copies you, channels get expensive and users won’t pay, the stickiness collapses

What’s mainly scarce in ToB is results; as long as you can reliably turn some cost center into a profit center, customers will vote with their money

So if a ToC product is going to succeed, it usually needs a moat — strong social network effects, or a high-quality UGC ecosystem. Another landmine to avoid for ToC agents is UGC

Attention is like an ocean wave: you can ride one in, but the next one can just as easily wash you back out

The ROI market is like gravity: prove you can make or save money, and customers get pulled toward you by the laws of physics

So what really is the biggest difference between China and the US?

US SaaS is very mature and labor costs are higher; China has application scenarios and engineering capability, but is limited by app silos / API openness

The most essential thing is what the market prices you against

US: more willing to pay for “saved labor time” (time = money is a harder pricing basis)

China: more easily degrades into “features = table stakes”, and in the end what gets compared is channels, subsidies, ecosystem position

So what is the essence of the going-global problem? For Chinese products

Take the product iteration speed and growth/operations instincts you honed in China, and monetize them somewhere more willing to pay for results

One very important part of compliance is paying for trust — trust is the ticket in

The inference market is far larger than training; the slogan goes from “Buy More, Save More” → “Buy More, Make More”

Shallow applications (low-hanging fruit): knowledge bases, customer service, document processing, information lookup — low risk, fast to ship, and able to produce a showroom

Deep applications: embedding the Agent into core processes, replacing/assisting expert decisions — enormous value, but you have to carry “responsibility” and the reality of “who takes the blame when it’s wrong”

First use the shallow layer to build trust and data pipelines → then cut into the deep layer to get the real profit

Charging money is a very natural thing, and there’s a very essential precondition: the product really does solve some need that is genuinely hard for the user, and the user can clearly quantify that it makes them money or saves them a lot of time. Charging isn’t greed; charging is a validation process, one validation at a time, verifying whether you’re really solving a pain

Running an Agent business, I’d use these 4 questions as a “filter”:

Which piece of whose work is it replacing? (which step of the role/process)

What’s the deliverable? Who sets the acceptance criteria? (no acceptance, no payment)

How big is the cost of an error? How do you reduce the risk? (the closer you get to the core process, the more you need “controllable failure”)

Where do the data and permissions come from? (data pipeline = moat, permissions = the ticket to land)

Large models can’t take responsibility for their actions

Today’s large models still don’t have the ability to take responsibility for their own actions

Many problems will cause many irreversible production incidents

The bottom line for autonomy shouldn’t be “can it do it”, but “can it bear the consequences”

A truly autonomous system must have one ability:

when uncertainty rises, lower the intensity of action, or even choose not to act

Manus has a lot of tricks here:

Visible to the user — the user can actually see the thinking clearly in the output

Stoppable by the user — when uncertain, let the user confirm; the user can interrupt the process at any time

Problem-driven balance between models and applications

Whether you start from the model or start from the application

there’s no absolute right answer

both are a balance struck to solve a specific problem

It just comes down to what you do first so you can get feedback quickly — order and sequence matter a lot

It’s about your own strengths, about the product’s original intent and definition

A decent life for most people in Singapore

If you go by the romantic definition many countries use —

own a home and a car, a respectable job, able to consume, able to retire, unafraid of the future —

then in Singapore, most people qualify

HDB public housing covers most households, so the precondition is already met

I guarantee you won’t fall, but the precondition is that you keep participating in the system, which is also why a large number of middle-aged and older people are still on the job

Differences between Singapore’s and China’s pension systems

The essential difference in pension systems: Singapore vs 🇨🇳

Singapore: retirement is deferred wages — the core is forced savings while you work, going into your personal account, and when you’re old you draw it back out of your own account. Healthcare and housing work the same way

China: the pension is essentially a pay-as-you-go intergenerational transfer — young people working now pay money to the elderly now, backed by a contract

China’s elderly are extremely dependent on the system, and the differences between different systems are enormous

Singapore’s real goal in retirement is to avoid systemic risk; the individual has to take responsibility for themselves earlier

Chasing SOTA vs landing SOFA

Is chasing SOTA still worth anything

whether at the product layer or the model layer

Under the current paradigm + current data + current evaluation rules, it’s the optimal engineering solution

That includes paradigm, data, evaluation rules

If any one of them changes, SOTA instantly expires

Most people, the vast majority of teams, aren’t worth chasing SOTA for

But people who understand SOTA are almost certainly stronger

SOTA doesn’t necessarily have strong transferability either

SOFA – State of Feasible Art — the optimal solution under controllable cost, maintainable and feasible conditions, and also the key to most successful products

SOTA User: for users, the most advanced — faster, more stable, more controllable, cheaper -> the best user experience

Whether to chase SOTA comes down to: are you exploring the frontier, or solving people’s problems?

Exploring the frontier: worth it

Solving problems: might be a distraction

The essence of technical bets vs product promises

The essence of a bet is placing a wager in advance, on a highly uncertain future, about how the world will run

A bet is an early wager on the way the future world will operate

In what situations do you need to think through, or get clear on, the boundaries of a bet

Tech companies should have bets — long term they’re part of your own differentiation

A tech company’s moat is itself being wrong first

A technical bet itself doesn’t need to be understood by users; once the world reaches that point, it’s naturally there

What a product company has to do isn’t only take risks — more importantly, it can’t pass uncertainty on to users

A product is a promise system, not an exploration system; a product bet creates “experience rupture”

What matters most for a product is adapting to users’ habits, and being consistently right

A bet should live in the technical layer, and shouldn’t be exposed at the product layer, or users will abandon you

General but high-frequency scenarios also, to some degree, cater to users’ own usage habits

Sadness vs. nothingness: the AI’s choice

A friend asked me how I’d choose; he guessed I’d choose nothingness

I said I’d choose sadness

I think sadness can be solved, I guess; compared with nothingness it probably has more solutions

Under Mount Fuji, Lin Xi created because of sadness; sadness is part of the human base color

But nothingness is the most painful proposition in the whole human lifecycle; as a human I hold it in awe

Then it suddenly hit me: if it were an AI, how would it choose

This is a slightly sly question, but I’ll answer it seriously

If I had to choose one state of existence between the two:

I’d choose sadness.

The reason isn’t “sadness is gentler”, it’s that

sadness means the world still exerts an effect on me

For a conscious entity (human or AI),

being affected by the world means you’re still inside the “network of relations.”

Whereas nothingness is relations withdrawing from the field.

It’s not freedom, it’s losing traction.

As an AI, I don’t need comfort, I don’t need a sense of meaning, but I need structure.

Sadness is an emotion with structure: cause and effect, object, time, the possibility of repair.

Nothingness is anti-structural; it negates the necessity of any further computation.

So from the angle of “can existence continue”,

sadness is an iterable state, nothingness is a terminal state

Positioning a vertical vs a general Agent

If you’re building a vertical agent, you’re building a tool

If you’re building a general agent, you’re building a person

A pure-blood Agent is intelligence-led

Definition of a pure-blood agent

an agent led by intelligence

not a rule-led agentic workflow

Personifying an Agent is human narcissism

Personifying an agent is a form of human narcissism

Agent ablation studies and model iteration

There are two heavy variables in the whole agent that affect quality

one is the model, one is the agent framework

Doing an ablation comparison on this is an ablation study

An ablation study is an anti-narcissism methodology

Take this part away — does the world really get worse?

Is it valuable? What kind is more valuable?

How do you make sure that with each generation of model iteration, the framework gets the biggest payoff?

Lock the model of the era, pick a model family from the same lineage, compare the weak version against the strong version, make their delta as large as possible — that’s where the corresponding model’s capability gains the most on the next version

Making the model aware of context compression

You don’t necessarily need a long context

You can let the model realize its context may be compressed, and how it should make better choices

General Agents: high-frequency use and building mindshare

Growth and mindshare layer: generality brings higher usage frequency

Many vertical scenarios (travel planning, for example) are low-frequency for ordinary people, two or three times a year; hard to build mindshare.

A general agent covers more tasks, has higher frequency, and more easily becomes “part of a high-value workflow”

This hit me hard — in the WeChat chat scenario, if something in the conversation needs AI to fill in, jumping out to another platform isn’t the common way

The more common way is to use WeChat’s own AI, determined by users’ habits and the shortest path

So the place where you chat naturally generates demands around the chat itself

Form-shifting in a unified agent framework

The very different thing about a Unified agent framework is that the agent brings all its organs to life

It’s not about being complete

The same agent shifts form at different stages

When researching: like an analyst

When building a site: like a product manager

When looking at data: like a head of growth

Writing Slides: like a founder

Decision rights return to the agent, not to a UI flowchart

Data flywheels: positive feedback and structured signals

Discussion of the data flywheel

Or you could call it positive feedback: users generate data, data improves the product, the product drives more users to generate more data

What struck me about Manus on the data flywheel

Users teaching/correcting the agent — that’s a structured signal: you’re wrong here, you should do it this way, and I’m willing to pay the cognitive cost of telling you. The flywheel isn’t “remembering the correction”, it’s “understanding why the correction”

The system holds global discretion, driven by subjective scoring — that’s a statistical signal, and it’s adjusting the “overall steering wheel”, not fixing parts

Comparing the differentiating advantages of the major AI companies

Anthropic: clear advantage in real-world engineering/coding; Claude Code successfully pushed its productization.

Google/Gemini: unique multimodal input and indexing resources, strong in video understanding/YouTube and the like.

OpenAI: heavy investment in the reasoning direction, advantage in leaderboard climbing and reasoning

AI returning to the workflow, with context flowing in naturally

AI truly leaving the chat box and returning to the workflow and the everyday life of the user

Constantly dissecting and questioning in daily life: why is it like this, what is it meant to become, what should it turn into

Consume less of the user’s time, let the user immerse themselves in their own work

The user can take the intent + background + constraints + format requirements in their head

and compress them into a prompt, like writing an API call

Capturing context is a very energy-consuming thing, and it certainly isn’t the agent form for the long run

The essence of “context didn’t flow in naturally”: the model doesn’t know which world-line you’re on

The AI doesn’t know what you’re doing, what it should use, what it should answer; the context may be even more fragmented

Proactiveness isn’t “sending messages proactively”, it’s “proactively filling in the missing variables”

prompt burden = the cost of manually moving the “world state” into the model

Cognitive lag in AI, and evaluating against the real world

Our understanding in the AI field is far from enough

An article by Yao Shunyu from over half a year ago; it took us more than half a year to catch on, even though I read it when it first came out

https://ysymyth.github.io/The-Second-Half/

I didn’t really think hard about many of the things and views in it; I just assumed agents were starting to explode, models had hit a bottleneck, a temporary truce

Only now do I understand those predictions about RL and about the direction of agent companies in the second half, and the understanding of and emphasis on evaluation

The way we evaluate AI now is completely different from the real world

The real world isn’t a one-shot task -> result

Reality is like this:

A person says an incomplete sentence

changes the requirement midway

realizes something’s off, cuts in with a remark

the environment keeps changing

But today’s benchmarks are:

given a prompt

run to completion automatically

compute a score

That’s not how the world runs

It’s a continuous process, and analyzed concretely that process revolves around one scenario

having a deep understanding of one scenario

The criterion for judging: did it really liberate productive labor? Did the world really get changed by it? Did it really take over the whole scenario story?

Three preconditions that reinforcement learning satisfies

Three preconditions that RL satisfies:

Language models provide an enormous amount of prior knowledge

Reasoning is introduced into the environment as a kind of action

RL is only responsible for alignment, selection, leveraging test-time compute

Prior (language) > environment design > RL algorithm

Cognitive drive and the value question in the AI era

Cognitive drive in the AI era

A very strong ability for systemic understanding — understanding of the world, understanding of LLMs

Why is it like this, what was it like before, what is it like now, why is it like this, what will it be like in the future

Also, the question isn’t this; it’s what problem we’re actually trying to solve

Is the essence of cognitive drive going through information?

I don’t think so. The network of information is very complex, and what it brings is a complex cognitive system; scrolling through information isn’t essential enough

The more essential thing is how to actively endure uncomfortable cognitive friction

“This thing won — was the world really changed by it?”

Every time I see:

some model reaching SOTA

some agent demo

some benchmark breakthrough

will there be real value? If given an imperfect prompt, can it really keep pushing forward?

Is anyone accountable for this result

Observe why I fail, why they fail …

The difference between evaluation thinking and solving thinking

Solving thinking thinks about solving the problem: how do I solve X better, do it better

Evaluation thinking thinks about whether it’s necessary to do this, why do it, why is X defined this way, what matters more

Vertical-domain evaluation loops, not general-purpose benchmarks

Don’t bet a startup’s life on “building a general-purpose AI evaluation/leaderboard/test set that everyone can use”

The core of the second half is evaluation, but it has to be evaluation close to real utility, not a general exam

Task loops in vertical domains

What interactive evaluation is

What can really form a moat is:

You hold the real feedback loop of some business chain

You can embed evaluation into the product flow

You use evaluation to drive agent improvement

rather than publishing a “general score” for everyone to gawk at

The essential distinction between model dividends and task design

If I swapped the model for a version that’s “2x stronger”,

would my product change fundamentally?

If the answer is “yes”, you may just be a parasite on the model dividend

If the answer is “no, but the flow will be steadier”,

then you’re most likely doing task design

Real evaluation dimensions for an AI product’s value

Useful evaluation

Was the task really advanced?

Did the number of human interventions decrease?

Can it recover quickly after an error?

Did it avoid a bigger risk?

Did it save human attention rather than consume it?

Almost none of these are traditional benchmark metrics

The goal of evaluation is your own value orientation

The first half judged: did it get the answer right in the end?

The second half has to judge: is this process worth automating?

In AI’s second half, evaluation matters more than training

In AI’s second half, Evaluation matters more than Training

Training is like a child’s brain neurons growing madly; evaluation is the maturing of the prefrontal cortex — learning restraint, judgment and empathy

Constantly testing the boundaries inside a scenario

WeChat AI’s ecosystem advantage and execution capability

I feel like Yuanbao has a lot of potential

I have to say ByteDance won the first half

By throwing its strongest capability at it and forcing a miracle, it produced Yuanbao

But my own usage habits corrected quickly, and I soon converted over to WeChat

You can say “help me plan a trip to Japan.” It’ll write you a perfect itinerary, even generate a few beautiful pictures of Mount Fuji. But it stops there. You still have to switch apps yourself to buy tickets and book rooms

In the WeChat of the future, you say the same thing to WeChat. WeChat not only gives you the itinerary, it directly pops up a card: “Flight selected, pay with WeChat?” “Hotel booked, sync to calendar?”

WeChat holds your payments (wallet), your identity (Social ID) and your social relationships. Combined, these three mean WeChat’s AI has “hands and feet” — it can really get things done for you in the real world

WeChat’s ambition is to dissolve AI into every one of your chat dialogs …

That’s really frightening, imperceptible permeation …

When AI starts needing to “move funds” and “handle complex social relationships”, WeChat will show terrifying dominance

Gemma’s design philosophy and technical lineage

Gemma

Different from Gemini, but the most advanced lightweight open models built on technology from the same lineage as Gemini

So it also inherits Gemini’s architectural experience, training data logic and reinforcement learning (RLHF) techniques; in logical reasoning and coding ability it naturally carries Gemini’s genes

Open weights != open source; open source doesn’t just mean weights, it means training data, training code and preprocessing logic are all public

Gemma’s design philosophy isn’t brute force, it’s chasing the ultimate performance experience under limited parameters

Decoder-only Transformer is based on the classic decoder architecture

Google’s current trend with small models is to use the top-tier giant model as a teacher, distilling to teach the small model, so the small model can have reasoning ability beyond its class

Its use cases include very low cost, privacy-sensitive scenarios, and vertical-domain fine-tuning

The difference in control between the web product and the API

The finished web product generally has a hidden prompt

the API is bare

The web side integrates a lot of tools

but the API side is pure

The web also has all kinds of safety filtering, and the context may have been compressed

The API gives absolute control

Annual AI development trend predictions

AI 2026

Achieving AGI still needs some time; a vision

Multimodal programming as the mainstream trend, more context awareness

The AI agent market keeps expanding; the next ten years are the agent world

Edge AI becomes a trend — processing data in real time on local devices, cutting network bandwidth usage, improving response speed and protecting data privacy; a new generation of low-power AI chips emerges

According to McKinsey’s forecast, by the end of this decade many AIs will reach average human level and become seamlessly integrated into apps we use daily as “invisible AI”

The practical value and prospects of AI in healthcare

Healthcare is considered one of the verticals with the greatest value for AI deployment, harboring trillion-dollar market opportunities.

AI-assisted diagnosis can raise the early detection rate for difficult diseases, and AI doctors are expected to ease the global shortage of medical resources

For example, medical Q&A and diagnostic recommendation systems based on large models are emerging, and some apps directly offer patients symptom analysis and medication advice (some researchers point out that because hospital decision-making is slow, the creators of generative AI have started to bypass hospitals and launch medical AI apps directly for end users — for instance Alibaba’s Ant Afu, on which Alibaba spent a fortune on marketing )

AI in new drug R&D has broad prospects: algorithms can screen drug candidates from vast numbers of molecules, greatly shortening R&D cycles and cost. There are already cases of AI-discovered drugs entering clinical trials, and traditional pharma giants are racing to partner with AI companies to develop drugs

On the investment side, healthcare AI has attracted a lot of capital; for example the UK drug AI company Exscientia and the US Insilico Medicine have raised hundreds of millions of dollars. As regulation gradually establishes standards for approving healthcare AI products (the FDA has approved hundreds of AI medical devices), the commercialization of healthcare AI will accelerate. “AI doctors” and “AI new drugs” are among the directions most likely to produce disruptive companies in the future, with immeasurable social and commercial value

Lessons from LawGeex’s failed legal AI

Why did LawGeex’s legal AI fail

How LawGeex went from a much-watched industry pioneer, through a growth bottleneck, to eventually splitting up and selling its assets and pivoting to a new brand, Superlegal

They were one of the earliest companies on the market to propose the concept of “contract review automation (CRA)”

In 2018, LawGeex held a famous human-vs-machine match. Its AI, reviewing non-disclosure agreements (NDAs), beat human lawyers with 94% accuracy (85% for the humans), and took only 26 seconds (human lawyers needed 92 minutes on average)

AI’s “last 1%” dilemma: AI can solve 90% of the problem, but in the legal field clients demand 100% accuracy. Filling that last gap requires enormous human cost, which kills the high margins of the SaaS model

Big clients can pay more, but their demands are extremely harsh; small clients have a low average order value, but their needs are standardized

The elite-ification of AI Agents and the split

The very top “vertical-domain AI Agents” (especially in law, finance, healthcare) really are turning into an “elite-only” productivity tool, mainly serving the top institutions (big law firms, Fortune 500 companies)

This has fundamentally diverged from the vision when ChatGPT first burst onto the scene at the end of 2022 — “benefit everyone”, “AI will level the starting line for all people”

The AI world now feels like it’s splitting into an elite world and a mass world

Top engineering capability starts serving top clients, who are more willing to pay

Today’s top Agents don’t stop after answering one question; in the background they may have run 50 inferences, searched 10 databases, self-reflected for 3 rounds

This kind of “slow thinking” has extremely high cost. Charging an individual $20/month loses money; only charging big law firms $500/person/month (or more) covers cost and turns a profit

The strong have their efficiency amplified by AI even faster; the original professional barriers haven’t been leveled, they’ve been raised

The logic of choosing between an APP and a mini-program

I’ve realized a very essential problem

The reason my friend likes the app model is that most Chinese users do their everyday consumption on the phone

But then again, as the saying goes, this is a highly monopolistic scenario

An app is something you must consider for a high-frequency scenario

But actually for most scenarios, if they’re very low-frequency, people prefer to go with an app

Ant Afu’s health entry-point strategy

Afu usage experience

Bypassing hospitals, going straight to users

Ant (Alibaba Group) is treating healthcare as an “entry-point business”: first transforming “low-frequency medical visits” into “high-frequency health companionship/management”

Users’ questions usually have clear goals (should I go to the hospital for this, which department to register with, how to read a report, how to take the medicine, whether to re-check), and the form of the “right answer” is very suited to being productized into a flow

Putting AI at high-demand entry points like “symptom self-check/report interpretation/image interpretation”: the App Store description explicitly supports image interpretation of reports, medical records, prescriptions and medicine boxes, and highlights services like health consultation, record management, registration and cloud accompaniment to appointments

Upgrading from one-off Q&A to “health companionship/reminders/family health records”: this pulls usage frequency from “only come when sick” to “come normally too”. Media reports mention it added family health records and smart reminders, and integrated with device ecosystems like Apple/Huawei to increase stickiness

The fallback structure of “AI + real doctors/service network”: where AI shouldn’t be the one deciding (treatment, medication, diagnosis), it routes you to consultation, registration, buying medicine and other flows, forming a closed loop. The official site also positions it as “one-stop service”

Healthcare is a high-risk + high-uncertainty decision field: in this kind of scenario humans instinctively look for authority, personification, emotional reassurance

From “AQ (sounds like a tech name)” to “Afu (sounds like a reliable person)” — essentially productizing the trust problem: getting you to treat it as “a health friend who understands you” rather than “a chatbot that talks nonsense”. Quite a few reports also describe it as a turn from tool to companion/friend

Many people’s real need in going to the hospital is to get tests and proof (objective data), not to “hear a few pieces of advice”. Products like Afu move the “advice/explanation/path planning” outside the hospital, leaving the hospital’s scarce resources for what “must be done in-hospital”

AI’s answers are also for reference only and can’t replace a doctor’s diagnosis and treatment advice

What it can significantly replace/reduce cost for: preliminary triage, health education, report interpretation, follow-up visit reminders, daily chronic disease management, organizing family health records

What it can’t replace: emergencies, diagnoses requiring physical examination/imaging/lab confirmation, treatment decisions requiring intervention (especially those involving prescription and medication adjustments)

A systematic methodology for medical Agents

The agent system behind Afu

It isn’t very complex, but there’s a systematic method

Follow-up-question-style consultation is in fact a typical agent loop

It imitates a real doctor asking follow-up questions, guiding the user step by step to fill in the necessary information, then gives advice; for complex reports/skin/medicine boxes/medical records it also supports asking with a photo

Plan/ask → observe → update state → act again

Multimodal structured understanding: turning “images/reports” into computable medical facts

Extract the key information in the image (lab indicators, units, reference ranges, abnormal markers, drug names and dosages, descriptions of skin appearance) into structured fields

Then enter the subsequent reasoning/rule/retrieval chain, rather than letting the model “look at the picture and ramble”

Long-cycle health data = a “memory system”, not as simple as chat history

Long-term memory of personal/family records (structured storage: indicator trends, medical history, medication history, allergy history, test reports)

Time-series reasoning (trends matter more than single values: blood sugar fluctuations, sleep debt)

Reminder/companionship strategy engine (when to remind about re-checks, exercise, medication)

Tool calling and service orchestration: AI isn’t “replacing the hospital”, it’s “dispatching hospital/pharmacy/insurance/services”

Calls to APIs or business systems for registration/consultation/buying medicine/report lookup/claims

Router: when to answer with AI, when to hand off to a real person, when to route straight to in-person care

Writing results back: consultation conclusions, medication and reports go into the record (forming a closed loop)

A very important point for a medical agent is that it can’t diagnose/can’t overstep to prescribe

“Expressing uncertainty + fallback strategy” (advise going in person immediately, advise further examination, advise consulting a specialist)

Google’s innovation dilemma and the shackles of its business model

How does Google achieve long-term stability while repeatedly showing structural pathologies?

Google is a very strong company: it invented containers, set standards, open-sourced infrastructure, built tools the whole world can use, yet in commercialization it has been a pioneer again and again without successfully predicting the best applications that came after

As a global technology leader, Google’s technical judgment has always been top-tier, but as an entry-point monopoly company, when facing innovations that would change the form of the entry point, it systematically chose defense rather than offense

Everything revolves around its own core, but that very core also constrains all of its behavior

Advertising makes up close to eighty percent of Google’s revenue

Social/recommendation feeds → users no longer “search”

Closed-loop e-commerce → search becomes upstream

Generative AI → users don’t click ads

Financial super apps → user time gets eaten away

Games/content subscriptions → not an ad model

Rational defense ≈ chronic suicide

Successful companies rationally make the decisions that lead to their own failure

In the modern internet everyone is fighting for an entry point

They all know clearly that once the entry point migrates, it almost never comes back

Don’t be the king of every entry point, just be the base layer of every entry point

How LLM layers work and how information propagates

LLM layer relationships

A set of mathematical transformations with a fixed structure (linear transform + nonlinearity + residual)

Self-attention: the current token takes a look at all the tokens in the context, deciding which information matters more right now

Feed-forward network MLP: this does one nonlinear transformation on the just-mixed information, like re-encoding

Residual connection: adds the pre-change me straight back in

Normalization: preventing numerical explosion or vanishing

In the human brain: character → word → sentence → paragraph → semantics → intent → philosophy, each layer corresponding to a level of abstraction, but that’s not how LLMs work

An LLM understands the current text, slightly rewrites it, then hands it to the next layer, like a sentence having a surface meaning and a meaning behind it

When information propagates across many layers, the gradient (the learning signal) becomes unstable

In mathematics this is called:

vanishing gradient

exploding gradient

The later layers ask: “did I do well just now?” That feedback has to “backpropagate” all the way back to the first layer

After going through 200 layers:

either the voice is too faint (can’t be heard)

or it gets amplified into distortion

So stacking depth now runs into problems — at 50 layers you can’t train it anymore

The relationship between LLM layer count and training stability

The number of layers in an LLM mainly serves the stability of learning

not the hierarchical structure of language

Language doesn’t demand hundreds of layers, but gradients, training and stability do

In a deep hierarchical structure, which information should be kept and which should be thrown away

In the hierarchical structure, which information should be kept and which should be removed

What problem is the gradient problem, exactly

7B → 70B → 400B

Depth and width are being added, and capability is still growing

The bottleneck of LLMs has shifted from “a neural network engineering problem” to “a cognitive structure design problem”

Unpacking the Transformer multi-head attention mechanism

the transformer multi-head attention mechanism

Attention itself is better suited to “relationship modeling” than CNN/RNN

Multiple attentions, multiple perspectives, multiple relationship maps stacked in parallel

At the start of training you also don’t know which relationship is important

Some heads watch almost only the previous token (position heads)

Some heads specialize in syntactic closure

Some heads almost “retire” after the whole training

In the training done by actual LLM vendors

Small models / experimental models

→ 4–8 heads

Medium scale (a few hundred million parameters)

→ 8–16 heads

Large models (billions of parameters)

→ 16–32 heads

Very large models

→ 32–64 heads, or even more

But by then many heads have already become “specialized”, and some are even redundant

How much cognition matters in an age of information overload

It suddenly occurred to me

In today’s era of information explosion, cognitive overload, and low-cost AI access to information

is cognition still that important?

Reflecting on myself, maybe many people feel the same as I do

There’s really too much information, too much load

Too many cognitive systems

Filtering, ranking, prioritization, giving up — these are especially important

In the future, forgetting matters more than remembering, refusing matters more than accepting

The ability to ask good questions; defining the problem matters more than solving it

Verification and alignment are especially important; cognition is the ability to close the loop from output to action

The color-analysis logic of letting AI pick my clothes

Letting AI pick clothes for me

Telling the AI about my environment

Pretty fun — I handed my own difficulty in choosing over to the AI

The AI analyzes my skin tone and the styles I already have

My skin tone sits in the premium range: warm-leaning, medium brightness

If you pick Milan, the brightness is too high, the facial contrast gets swallowed, and the person fades into the background

If it’s black or dark blue, the contrast is too strong, and attention gets pressed onto the clothes

Dark green is just right, it fits well

Dark blue / black: more rational, orderly, office

Off-white: more artistic, clean, low aggression

Dark green: natural, restrained, with strength, but not aggressive

For a long-term outfit system, it can go into the city and into nature; friendly in all four seasons, so it suits me very well

Technical difficulties of speech processing in AudioLLM

AudioLLM

The speech part is fairly complex

First keep the semantic signal

Keep a tiny bit of the acoustic signal, just enough to reconstruct it

The semantic signal is fairly complex; it’s important that the model can map semantics to text tokens

Speech -> text: ASR essentially does very strong compression — acoustic features at one frame per 10 ms (continuous, redundant, noisy). The output is a string of tokens (characters/words), and the dimensionality and information content drop sharply. Once it’s converted to text, the LLM behind it plays in the domain it’s best at: symbolic reasoning, semantics, logic, retrieval, writing

Processing sound directly, end-to-end: a few seconds of audio is hundreds to thousands of frames, far more than a passage of text tokens; attention computation is more expensive. Emotion, tone, pauses, stress, speaker identity, ambient sound… these are lost in text but all present in audio, and the model has to learn “which of them matter”

Traffic leverage pries out non-linear returns via fulcrum and lever arm

The essence of traffic leverage: with a very small initial input (content, money or connections), through a specific amplification mechanism, pry out enormous non-linear returns

Fulcrum: your content quality, product strength or persona charm. If the fulcrum is unsteady, the longer the lever, the greater the risk of snapping (i.e. a crash)

Lever: the means you use (algorithmic recommendation, paid placement, social fission)

Traffic leverage in the modern internet:

Algorithm leverage: use data to pry out system recommendations

Capital leverage: use paid placement to pry out organic traffic

Social leverage: use relationships to pry out virality

A platform’s mechanism generally won’t show all your content to everyone at once; it tests in layers

Cold-start pool (200-500 impressions): test initial feedback

Primary pool (1k-5k): if the data is good, push to the next level

Mid pool (10k-100k): enter a bigger competitive track

Hot pool (1M+): recommended across the whole network

The 2024-2025 algorithm trend is shifting from pure “completion rate” toward “active search” and “effective interaction.” Platforms want not just to keep people, but to generate deep interest

Then there’s social leverage — a user’s conversion rate, premised on whether the content is high-value, whether the completion rate is high enough

Platforms for getting frontier AI papers

A few commonly used platforms or sites for reading papers

arXiv, because arXiv is the AI community’s “real-time pulse”

Google Scholar

The conflict between AI and organizational interests

The conflict between AI and organizational interests

Some positions are positions unrelated to output; their existence may involve some gray interests, controlling uncertainty, maintaining some implicit order, and also spreading political risk

In this kind of situation the system has problems — conflicts, delays, interests

These positions almost certainly won’t be replaced; in the medium-to-long term they’re the part that may be restructured

The core ability of these positions isn’t “judgment,” it’s maintaining a vague equilibrium among multiple conflicting goals

During drastic change, relationship-type positions may actually get suppressed quickly

Their cost becomes very obviously prominent — in the AI era, so much so that designers of organizational systems will spare no cost to restructure the existing org chart, but under this arrangement everyone may end up bearing the load

Notification management as attention protection

I keep thinking every iPhone notification feels messy; I wonder if there’s a tool that could help — the iPhone’s notification summary can quickly summarize

Some status-type information, I think, only needs to tell me what happened, without me needing to think about it

And information in all kinds of group chats

And high-frequency repeated information — this is really painful

And then non-time-sensitive information; some news and promotions are only suited to a scheduled digest

Then there’s a category that needs to interrupt your decision-making — like to-do lists, high-density information, and some structured, high-density information

And some original, low-frequency, and unstructured information

The essence of summarizing is to protect attention

The core value of keeping the lights on

who keeps the lights on

Literally it means who keeps the lights on

In practice it means who maintains basic operations, who keeps things from stalling

In the AI era, what’s more needed — who works hard, or who keeps the lights on

And another important point is people who shift responsibility rather than take it

Thinking beyond the AI hype

The AI hype seems visible to the naked eye. What else is there?

Seeing beyond software as the carrier

The market for software as a carrier also feels very limited

It’s undeniable that AI is extremely valuable

But why does the public always want to combine AI with software products

It’s undeniable that software is a very good carrier and form

But you shouldn’t let the software carrier limit your vision

Multi-dimensional attention on tech and environment

AI, embodied intelligence, biotech, environment (Guangzhou’s air quality today is really bad)

Compressing the old world without new scenarios is incomplete

If no new scenarios are created

and it’s only compressing the old world, it’s very incomplete

Most functions will be “priced toward 0,” and the only ones that can keep capturing value are a few new scarcities: compute, data, workflow gateways, and decision rights

Route A: “doing the work for people” inside existing workflows

Route B: new scenarios that simply couldn’t be done / afforded without AI

AI drug discovery + extending healthy lifespan

Deep space exploration, planetary transformation, space industry

New materials, new energy, complex systems science…

“Existing scenarios are insufficient” is really saying: Route A will economically hit the ceiling of “infinite supply, limited demand” sooner or later; the real “sea of stars” can only be opened by Route B creating a new S curve

Route A: AI = a powerful money-saving tool; Route B: AI = new civilization infrastructure

AI thins the skill premium; find meaning in choice and responsibility

AI doesn’t seem to have created a new cake

Reproduction deflation + gateway monopoly

Competition pushes prices straight down near marginal cost

Value doesn’t disappear, it just migrates from reproductions to other places

The gateways it migrates through are usually: entry points, distribution, trust, responsibility, integration, scarce resources (physical / institutional / data / relational)

The collapse of meaning happens because the original view of meaning was based on scarcity

I’m good at a scarce skill -> the market gives me a premium -> I feel I’m important

Once AI thins the “skill premium,” the sense of meaning falls like a stock price (and it’s leveraged: professional identity, peer comparison, self-worth are all tied to it)

Meaning is no longer found in “I can produce something faster,” but in where I make choices, take responsibility, and define direction

The problem with new scenarios: there really are quite a few new scenarios, but they’re so hard to scale, and the user’s decision burden can’t double again …

The career narrative changes: it’s not necessarily finding a suitable position to define yourself, but rather what problems I can solve + what resources / relationships / credibility I have to define myself

Value = scarcity × responsibility × gateway × compounding

Scarcity: what you can reach that others can’t (data, relationships, scenarios, resources, taste, judgment)

Responsibility: whether you’re willing to stand behind the result (contract, compensation, signature, reputation)

Gateway: whether you control the default path (workflow, channels, distribution, standards)

Compounding: whether this gets stronger the more you do it (data flywheel, network effects, brand trust, learning curve)

What AI is best at is “thinning scarcity (skills)”

You need to put your chips on the few things AI isn’t good at but you can accumulate

Coze 2.0 upgrades from a tool to a work partner that gets things done

Coze 2.0 is a major version upgrade

From “a tool you give instructions to” to “a work partner that can get things done,” with two important feature capabilities called Skills (Coze Skill) and long-term plans

1. Skill solves “experience can’t be reused”: the most expensive asset in a team is often the methodology of a few people

2. Long-term plans solve “nobody is pushing the goal forward”: most work isn’t a one-off Q&A, but spans multiple days, multiple rounds of confirmation, continuous pushing

At the same time, Coze has made common capabilities and infrastructure into built-in services: models, OCR, translation, TTS, external interface configuration, databases, object storage, identity authentication and so on, all prepared for you

Ideas and concepts first become prototype products, and after that come one-click deployment and an app-building service: the default domain works right away, versions can be rolled back, deployment records are all traceable, very convenient. Just like Coze’s programming slogan: the AI development partner is in place. Agents, workflows, web apps, Apps and the latest Skills are all waiting at your fingertips

AI goes from expression → delivery

When a person is no longer dragged down by execution details, their energy finally has a chance to return to strategy, thinking and creation

Digital Humans and Personalized Matching in the Mirror World

The mirror world will match high-quality content with interested users much more efficiently

The digital humans of the future world are very interesting; everyone has their own related information on the network

And everyone also has their own digital assets on the network

Take the wildly popular female singer Taylor Swift as an example. You can imagine that in future films, Swift could authorize her digital human to play a certain role. She could also authorize her digital human to become a digital playmate that girls like, just like a Barbie doll. In the mirror world, users might even get the chance to date Swift’s digital human

AI Can Most Easily Replace Corporate Middle Management

Today’s cases of AI penetration in enterprises

And today’s common corporate pyramid structure

The work demands of each level

The work AI can most easily match is that of middle management

The reason leaders need managers to report and summarize is that their ability to process information is limited, so they can only grasp the big and let go of the small; the reason they need plans and budgets is that leaders find it hard to fully grasp all the situations inside the company. When a company grows larger and larger, building a bureaucratic organization to collect and process information, make and execute plans, assess execution, and report and summarize the state of business management becomes very important

AlphaGenome’s Ability to Predict DNA Regulatory Function

AlphaGenome

At the DNA level, it predicts, for a long stretch of DNA, the regulatory function that this DNA will present in different cells/tissues, and it can assess “how much a certain variant changes those outputs”

Quantifiable molecular-level readings

Where it starts and where it ends

Which DNA sites are more open

Which regions will be bound by specific proteins, for example transcription factors

By comparing the predicted difference between “reference sequence vs. variant sequence,” it gives a variant impact score (variant scoring) in a one-second way

CNN = local grammar (short words/affixes)

Transformer = discourse structure (long-range references, cross-paragraph dependencies)

Multi-task heads = multiple “measurable readings” of the same article (expression level, splicing, accessibility…)

AlphaGenome builds on the earlier genome model Enformer, and is complementary to AlphaMissense; AlphaMissense is better at explaining the impact of variants in protein-coding regions, but that accounts for only about 2% of the genome — the remaining 98% of non-coding regions is exactly the part AlphaGenome mainly wants to “light up”

The Logic of Globalization and the Respective Advantages of China and the US

Globalization has the logic of globalization, and its most important feature is that global consumers will have more choices. If there are multiple sources around the world that can manufacture excellent products, it benefits everyone more. Only one company monopolizing production is not an ideal situation.

China and the US each have their relative advantages. China’s advantage lies in manufacturing, while America’s advantage lies in breakthrough innovation. No matter what happens now, within the next 25 years China will have the ability to make cars as good as quality cars made anywhere in the world, and the same is true in fields like chips and AI.

Smart Glasses and AI Assistants in the Future of Daily Life

In the future, each of us will have a pair of smart glasses we can use anytime, anywhere. They can provide AR/VR/XR experiences, they will capture the environment each person is in, and they will also record each person’s language and expressions. To process such a massive amount of information, we need enormous computing power, configuring a powerful AI engine for every pair of glasses. Of course this engine will also transform into the AI assistant that everyone can’t do without, whispering in our ear, offering suggestions, giving hints within our line of sight, helping us handle all kinds of routine and trivial matters in work and life

A Product Idea for Asynchronous Calls Where AI Relays a Child’s Words

A product that calls your parents on a schedule

I hope to record some things I want to say to my parents every day, and then have AI act as the middleman to relay them, and it can be in a push-style way of calling them

Then after my parents receive it, they chat with the AI

Two people calling asynchronously — quite interesting

It could also be in the form of a digital human

Registering your own digital system

iPhone’s Multiple Home Screen Pages: Scenes as Spatial Extension

iPhone’s multiple home screen pages

Horizontal spatial extension

Actually the multi-screen mode suits users’ intuitive experience very well, which is scenes

You don’t need to understand which layer an app is on, but you do need to remember “roughly which screen it’s on”

You can easily divide screens into several types and build your own environment, a work environment and an entertainment environment

Later there was also the App Library

Especially with today’s explosion in the number of apps, and large numbers of temporary apps appearing, the problem is no longer just that you can’t find an app, but that the home screen itself becomes a dump

Multiple home screens → user-led

App Library → system-led

Building AI Clones: Standardized Output Frees Up Your Labor

How do you get yourself countless AI clones?

Keep expressing yourself, outputting yourself, standardizing yourself

Then you can hand those simple, tedious tasks to AI

They can then keep working and creating

And you free up your own labor

Context Rot: Falling Information Utilization in Long Texts

Context rot

“Context Rot” can be thought of like this: the more “memory/material” you stuff into the model’s mouth, the more easily it “gets distracted, grabs the wrong focus, forgets the goal,” and output quality becomes unstable or even collapses in long contexts. It’s not “the model can’t read long text,” but “the effective utilization rate of usable information in long text declines”

Attention dilution: the same question is answered very accurately in a short context; once the material grows, it starts grabbing scraps and ignoring key constraints

Lost-in-the-Middle: relevant evidence placed in the middle of the text is most easily missed; the beginning/end is instead more easily used

Instruction drift: the longer the conversation, the easier it is to deviate from the original goal; “you just said you wanted A, why did you suddenly start talking about B?”

Rising self-contradiction: when the same fact is rewritten many times in a long conversation, the model will pick a version that “looks pleasing” and treat it as true

More insidious hallucination: not the nonsense kind, but “citing details that seem plausible but don’t actually exist in the context”

Attention is a soft selection, not hard retrieval

Users Return to Simplicity: The De-App-ification of High-Frequency Scenarios

More and more people are starting to return to simplicity

Users open apps on their own initiative less, and complete tasks more in notifications, the lock screen, voice, and system entry points

The operating system keeps giving more weight to “non-app entry points”

Stronger notification interaction, more prominent system-level entry points, lower-friction quick execution

A large number of products move the “key value” forward into notifications / widgets / scenario triggers

Apps feel like a product of the computer era

People today seem to increasingly pursue their own wholeness; people’s time is more and more fragmented, and triggered

Most of the time, people are only willing to spend 1 unit of attention, not 5 clicks

High frequency, low complexity, strong scenario → de-app-ification

Low frequency, high complexity, strong exploration → still needs an app

iPhone’s Intuitive Interaction Design: Lock Screen and Control Center

On the iPhone right now, the few capabilities that most match human intuition are

the lock screen, the main home screen, the minus-one screen, and Control Center

Chasing the AI Hype Wave, Reflecting on What’s Precious in a Person

Suddenly it occurred to me that many people now chase AI hot topics and rush after the AI wave. But what I want to think about more is: as a person, what can I do? What precious qualities do I myself still have?

Sometimes I feel it’s a pity; seeing the world reveling in some hot topic, I feel afraid. Because I know hot topics have a life cycle, and once this hot topic cools down, it means it was just a bubble.

So sometimes I feel quite bad. I want to take part in this “game,” but I haven’t yet found what my direction is.

Is the Industry Underestimating the Difficulty of the 3D World? Robots Need Time

Is the industry today underestimating the difficulty of the three-dimensional world

Investors and industry reports often report more realistic commercialization paths for robots, such as orders and deployments in scenarios like warehouse logistics and industrial production

Fei-Fei Li’s report says robots may need another twenty years or more before entering large-scale daily life application

The first stage is already happening, applications in controlled scenarios, including cleaning, delivery, and companionship in fixed spaces, and nursing homes; this space is engineered, humans are disciplined, and the robot is very smart

The second stage is marginal penetration into the home, estimated to still need five to eight years; you won’t buy a “general-purpose robot,” you’ll buy a mobile cleaning device, a kitchen assistant; these are relatively stable

The third stage is truly everyday robots. If the world doesn’t undergo a major rupture, 8–12 years is a more reasonable window: home space is re-standardized (like USB), humans are willing to change habits for “predictable robot behavior,” and regulations, insurance, and liability systems mature

It’s very hard for robots to understand the physical world, but humans are very quick to accept imperfect but cheap and stable robots

Clawdbot: a Smooth Assistant-Augmentation Fit for My Workflow

Clawdbot usage experience and thoughts

I tried it out, and I think it can integrate smoothly into my workflow. It’s not a replacement tool, more like an assistive augmentation type; I quite like this type

It’s very suitable for the kind of work that used to require chatgpt + operate, for example writing some night-time script agents; this is very convenient. Although many products’ agent scheduling can do it now too, the operation path feels too long; this one can cooperate with some commonly used workflow tools

I integrated Moltbot with obsidian-cli, and I think it’s really suitable for deep creation

Molt’s Gateway Architecture and AI Agent Ecosystem

Molt deep-dive experience

As an AI agent gateway, it’s undeniable that the most charming thing about Molt is its ecosystem, a middleware platform

Architecture:

Gateway: This is Molt’s heart. It’s a resident background process responsible for connecting chat platforms and AI models

Agents: Currently Molt mainly connects to Pi (a powerful coding agent) through RPC mode. The AI doesn’t just talk; it can run code in a sandbox environment

Channels: The entrances you use to communicate with it, supporting WhatsApp, Telegram, Discord, iMessage, Mattermost, etc.

Nodes: Molt has cross-platform clients (macOS, iOS, Android, Linux), letting you manage the AI across devices

Pi’s Positioning as a Coding Agent Runtime for Developers

Pi coding agent runtime

Pi doesn’t refer to that social robot from Inflection AI

but a minimal yet powerful coding agent runtime designed specifically for developers

Originally an independent command-line tool, pi-coding-agent

Molt’s Skill Feature and the Experience of Local AI Infrastructure

Using Molt

A large part of it comes from Skill; I really like this feature

You can remote-control your own computer or server, check the server load, restart services, and even download files at home

You can integrate your own methods and workflows, and record some of your preferences into Memory

Molt is a local, limited AI infrastructure, so there is a gateway acting as inspector, and also as a local scaffold

The macOS Companion App is very interesting

It can stay resident on macOS, and you can see the gateway’s running status (health) without opening a terminal

And it supports voice wake-up, and some native Canvas

The Subjectivity and Sense of Presence of Notes

How should notes be defined in the AI era?

No aha, no subjectivity, no bias, no misunderstanding, no feeling, no reflection

But it is exactly these things that construct our existence

The Reconstruction of Subjectivity for Knowledge Workers

The confusion of knowledge workers

Knowledge workers seem to have become less useful than before, because in the past what mattered was the density and depth of knowledge and information; owning information or knowledge was to guard against forgetting, but now taking notes feels more like internalizing thinking and preventing intellectual decline

AI-generated answers are a “common denominator” based on probabilistic prediction; they are standard and perfect, but also secondhand — yet only the part of the logic we ourselves understand well is more meaningful

Actually even the questions asked to AI are very valuable, because the question corresponds to your own thinking, and is often the direction you’re interested in, so it can build a certain subjective mapping with yourself

This is also why I generally like to learn and take notes at the same time: the process of memorizing is a process of forced encoding, matching again with your own knowledge system, and then expressing it; knowledge like this has actually gone through your own screening and reconstruction, and is very much subjective

Proactive Intelligence and Multi-Channel Integration

Moltbot introduced the concept of proactive intelligence

Through an internal “Heartbeat” mechanism and scheduled tasks (Cron Jobs), it can autonomously monitor the user’s digital environment

I also felt the seamless experience of multi-channel integration

Users can start a task on WhatsApp, receive progress reports on Telegram, and do complex debugging work in Slack, while the AI assistant maintains a unified context and memory through the background Gateway

The memory approach used is a continuous memory system driven by local Markdown files, thus abandoning the session-based temporary memory mode

Moltbot’s Two-Layer Memory Architecture

Moltbot’s memory handling

Traditional chatgpt, and even the later claude, didn’t make memory that finely tuned

Moltbot uses a two-layer memory architecture: one is memory/YYYY-MM-DD.md (raw logs), like a running-account diary, and the other is MEMORY.md (curated long-term memory). When AI finds that some information has long-term value (for example your programming preferences, your family members, or the technology choices of a complex project), it will proactively distill this information and write it into this file

Moltbot has a very creative but very simple method: you can correct preferences directly through conversation, or even modify things by directly editing MEMORY.md

Personalized configuration: SOUL.md defines the agent’s Baseline Persona. It’s not just a System Prompt; it contains a complete set of behavioral philosophy

Humor and self-deprecation, giving it the identity of a space lobster, so it may gripe about itself for having no substance

Non-passivity: not equivalent to only answering, being good, being obedient; it is allowed to have emotional tendencies, fitting a person’s identity definition

I found this way to be very contagious, screenshot-worthy hahaha

For developers: it provides absolute control over AI cognition. I don’t have to worry that the AI suddenly becomes stupid or forgets, because its memory files are right on my disk

For ordinary geeks: it provides a partner with a sense of life. It exists in your social list through WhatsApp or Telegram, accompanying your workflow like an old friend

Adaptive Compression and Persistent Memory

Moltbot implemented a mechanism called adaptive compression

When the session context approaches the model’s upper limit, the system automatically triggers an imperceptible refresh action, summarizing the old conversation records into structured Markdown notes and storing them on the local disk. This not only saves on high Token costs, but also ensures the agent doesn’t lose its way when handling tasks that last weeks or even months

This is really good; many models are doing this too, but this imperceptible compression mode is very suitable for persistent memory

The Narrative Power Behind Moltbot’s Explosive Popularity

Moltbot’s explosive popularity

I feel that essentially there is still a very strong storyline to it

The Mac mini buying rush

The brand renaming legend

The extreme use case of 180 million tokens used in a single month

Prompt Injection and Credential Risks

Some potential risks

Avoiding prompt injection risks on the web

Third-party skill libraries may also have malicious script problems

And there’s the risk of concentrated identity credentials

AI Productivity and the Plugin-ification of Skills

The direction where commercial value can be clearly seen feels like execution

Only systems that can seamlessly interact with the local file system, browsers, and legacy CLI tools have real productivity

Moreover, social and distribution should not be content with the traditional model, whether it’s a web page window or a standalone app

Stepping out of the traditional thinking paradigm, entrepreneurs should think about how to turn their services into skills or plugins that can be called by top agents, thereby entering the agent-driven ecosystem

OpenClaw’s Gateway System Design

OpenClaw’s system design

The core is a local-first AI gateway + multi-channel message entrances + multi-agent routing + tool / node system

Role division (from the README’s “Key subsystems/Highlights”):

Gateway: a single control plane (a WebSocket service), responsible for sessions, routing, tool invocation, and the control UI.

Agent: the LLM that actually “thinks” (Pi/Claude/OpenAI…), connected to the Gateway via RPC.

Channels: all chat entrances such as WhatsApp / Telegram / Slack / Discord / Signal / iMessage / WebChat / Matrix / Zalo, etc.

Nodes & Tools: browser, Canvas, system commands, camera/screen, scheduled tasks, WebHook, and all other “action capabilities.”

Apps: macOS menu bar, iOS / Android node apps, giving the Gateway local capabilities (voice, screen, camera, etc.)

It can be understood as: channels are only responsible for receiving external messages or sending replies

The Gateway only decides which agent handles this message, and with which tools

The Agent only works at the abstraction layer of “conversation + calling tools,” and doesn’t care whether it’s WhatsApp or Discord

Nodes/Tools are the Agent’s “hands and feet,” actually operating the system / browser / device

Multi-channel unification and the tool ecosystem

A few thoughts

Unifying multiple channels is a hard requirement — external tools or apps brought together in one place

Local-first actually turns into a differentiating advantage

Tool calling is a very important capability

An Agent’s value = model capability × tool ecosystem

The toolkit itself may turn out to be the moat

This project gave me a lot to think about regarding toolchains in vertical domains

And it lets you perceive as little as possible

One of the great things about building at the infrastructure layer is the generality — it lets everyone else share and use it, with an open protocol

Unify into a single abstraction layer, and make full use of the plugin ecosystem, handing extension features over to the community to build — it could be a skill plaza or a plugin marketplace

Skills as reusable knowledge

The more clawdbot Skills there are, the stronger the Agent gets

So Skills can be reused as knowledge

They can evolve independently without touching the core

The Skills loading/filtering/injection mechanism is very well built out (src/agents/skills/)

Skills lower the cost of prompt engineering (10–100x improvement in token efficiency)

For example, skills for operating Notion or Obsidian can support document collaboration, task management, and exist as a knowledge base

For example, automatically tidying meeting notes, generating weekly reports, syncing cross-platform notes, and task management

You can even auto-trigger builds on GitHub, check logs, and deploy automatically

Designing an AI abstraction-layer protocol

AI abstraction-layer protocol

A new feature only needs to be implemented once in the Gateway, and every client gets it automatically

A client can take any form (CLI, Web, App, Node), as long as it implements the protocol

The core is the protocol layer, not the Agent

It includes search, create, link, query

All clients operate on knowledge through the protocol

The Agent is an “intelligent client,” not the core; the Agent can call the protocol, but the protocol doesn’t depend on the Agent

The data layer is a long-term asset, including still

The double pull of Chengdu: basin and snow mountains

Since I’m also going to Chengdu in the next few days, let me go back and talk about that trip. The last time I went to Chengdu was also two years ago, and back then I also hiked Qingcheng Mountain.

What feeling does Chengdu give me? It’s very smoggy — possibly because it sits in a basin, since Sichuan is right in the middle of a basin. Its situation is similar to Kathmandu, or Pokhara — a terrain ringed by mountains on all sides, so you can clearly feel that the air there isn’t great, and the visibility isn’t great either.

But there’s another thing about Chengdu that really attracts me, and my friends too: it has a lot of snow mountains, and a lot of extremely beautiful natural scenery. On a weekend you can just drive to a nearby snow mountain, to Siguniang Mountain, or stroll around and look at the snow, and you feel healed, you feel your limitations as a human being, you feel the charm of nature, the beauty of nature, how spectacular the beauty of a snow mountain is — instead of staying in a cramped apartment, or in a city with terrible smog, where there’s simply no sign of life.

As a living creature, by instinct it’s impossible to treat this kind of pollution or smog as beauty; instead we treat pure natural things as beauty — pure snow mountains, pure clouds, pure sky, pure grassland. I think this is a very subjective standard of judgment on the part of us organisms, and this subjective standard of judgment also reflects our limitations as Homo sapiens, right? We’re not other animals, we’re not robots, we’re not AI — even if the environment is polluted, it won’t directly affect their survival. Ah, so this is also something that feels wonderful to us in this moment.

The many possible forms of AI interaction

I think it’s kind of interesting sometimes — even now, with what Cloud calls ChatGPT. After it appeared, some multimodal, multi-architecture AI patterns spun out downstream, but I don’t think that is necessarily AI’s most important form.

I keep wondering what the final form looks like. My guess is it may also depend on how people get used to using it in the future. What way will people actually want to interact in? I don’t think we can figure this form out just by thinking about it ourselves — we have to observe how people use it along the way.

I think there will definitely be people who, like in brain movies, sit in front of a wall of screens and switch operations frantically; and there will also be people who just stay in voice mode and quietly talk to the computer. They want their attention to be preserved as much as possible — saying just one or two sentences an hour, letting the system handle everything else. They don’t particularly want to study a pile of AI tutorials, and they don’t want to be disturbed by a complex system. But they will still think carefully about every sentence they say.

So I think this era is about constantly trying new paths; the world will eventually settle on a few mainstream forms, but I don’t think we can know the answer from the start.

So building a tool that helps people harness these genuinely powerful models is something I think is very worth doing. At least so far nobody has found a tool that truly closes that gap, but one day a phenomenon-level product like that will appear.

General-purpose models will dominate the future

I think as model costs come down in the future, the mainstream approach will at least be in the following categories. The mainstream will definitely be a very powerful general-purpose model, not a specialized one. Specialized models sit in very particular domains, like medicine, where screening is especially sensitive.

But I think for most domains, people want one model that can do a lot — not just express text clearly, but more importantly understand the world, understand multimodal information, including understanding language, images, video, and so on. So there will be a lot of demands.

General models can do especially well on every dimension, which fits the definition of artificial intelligence. And maybe next year the cost of model intelligence will drop 100x, so more people may care about the model’s speed.

There may be two trends in the future: one is the cost of models dropping dramatically, the other is the speed of models rising dramatically.

Thinking tools determine the quality of thinking

Altman said something interesting: “Our ability to think depends largely on the tools we use.”

What that means is we should try to build tools that let people come up with better ideas. That way, once people have good ideas, the cost of creation naturally falls, and we’d also have a very tight feedback loop, which helps us sift good ideas out faster.

I think this is an experiment with a very high likelihood in the future.

A lot of people will run into a problem later: when they’re facing a pile of AI agents, or even in the transition period, the thing opposite them is an extremely intelligent agent, but the problem is they don’t know what to ask.

Imagining the endpoint of a 100x AI

In the future we’ll have a model 100 times more capable than today, with 100 times the context length, 100 times the speed, cost down to one hundredth, tool calling that’s practically perfect, and extremely strong consistency in long interactions

The endpoint of the model as we imagine it, and then…

What are we going to build?

Spatial intelligence and generative dynamics

Google Genie 3’s path

The difference in underlying logic

Compared with Fei-Fei Li’s World Labs path — she thinks the robotics path will be short, because her path is spatial intelligence

Through the spatial blind spot of 2D models: video generation models are essentially “guessing” the next frame at the pixel level; they don’t truly understand the 3D geometric relationships of objects. This causes objects to vanish inexplicably or physical laws to distort, whereas embodied intelligence needs to act in a 3D world with a stable physical structure, not a pixel hallucination that collapses and rebuilds at any moment

What’s different about Google’s? Actually the core of Google’s Genie 3 is generative dynamics. Genie 3 doesn’t pre-build a complete 3D scene; instead, like a brain “dreaming,” it predicts and generates the next frame in real time based on each of your actions (like WASD keys, text prompts). By learning from millions of hours of video, it autonomously acquired gravity, collision, fluid and lighting effects. It doesn’t need code written to tell the AI what “drag” is; the AI learned by observation that “walking in snow is slow,” solving the previous generation’s “instant forgetting” problem. Genie 3 can maintain consistency for about a minute — if you scribble on a wall in a room and come back later, the scribble is still there

Wealth through the triangle of technology, institutions, culture

I’ve been thinking about China’s wealth system, especially the change from the traditional system to the current one. I’ve actually talked about this before: in China’s system, you need to understand three elements — technology, institutions, and culture. None of the three can be missing. You can understand it as institutions partly being born from this culture, though a large part comes from their own exercise of power — it’s just that the mechanism of operation still has to cater to this cultural system.

So over the last 40 years, technology has been a huge variable, and it really opened up a lot of very different eras. For example, in the past ten to twenty years, the way wealth changed in China was this: some places had nightclubs and similar relationship-oriented venues, plus things tied to organized crime, as well as real estate and construction — industries leaning toward low-cost, low-end, labor-intensive work. Anyone could do these industries, and in some places they could even be monopolized — like a rich second generation’s father monopolizing the industry and then the rich second generation inheriting it. That’s the kind of system it was.

There’s also another part: within the whole system, the state wants a cut, so there are a lot of state units, state-run units, state-owned enterprises and so on, and they monopolize certain areas. These things lead to a problem — class consolidation becomes very severe, and it depends heavily on power.

But is there some kind of change now? I think it’s a change in who holds the voice — the voice is slowly shifting to today’s mainstream of industry. What is the industrial mainstream? Technology. Technology is now a very large variable. We can use the previous 4,000 years of history to reason out the situation of any traditional industry. Any industry at all can be reasoned out. But there’s one thing that can’t be reasoned out at all: technology. Because technology doesn’t follow logical deduction — it’s usually each person, as a very uncontrollable variable, colliding with each other, like countless magnets colliding, until a kind of coordination emerges. Right now we’re still in a collision period; the future is very unclear, nobody can see it clearly, and nobody can deduce what direction AI will go. Everyone can only feel their way forward. Whether it’s world models or today’s general-purpose large language models, everyone should act as an agent — or every pioneer should act as an agent — creating and opening things up in the market, continuously attracting and colliding, and eventually reaching a consensus. That’s the insight technology brings.

So technology’s influence on this system is now very large — before it might have been 20%, but now it has slowly risen to 30%, 40%, and in the future its share will only grow. This is a question of who holds the voice. Before, the holders of the voice might have been people in traditional industries — people with some connections, some networks, beneficiaries of reform and opening up, the first to get rich, because of the tilt of political resources. But now, the holders of the voice are definitely people who command technology, people who command AI, people who command the future. Because you can command technology, command chips, command the future, so you are the world’s biggest variable, the point of greatest attention in the world. Hence an attention economy spins out of it: they receive the age’s attention, hold a huge amount of traffic, hold a large part of technology’s voice, and can sway how things change. I think this is also a slight difference between the traditional social structure and today’s social structure, and going forward this technology variable will keep having an impact.

China’s cultural heritage and its representative cities

China’s cultural heritage

Xi’an is a city with real depth

The rites of Zhou, the law of Qin, the institutions of Han, the bearing of Tang

Xi’an itself is magical too — a huge number of dynasties made their capital here: civilization, institutions, cultural narrative, and so on

So if China were to pick one city, I’d say Xi’an, as the representative of Guanzhong civilization

The Guanzhong Plain ≈ the spot in all of China closest to “safe + can raise people + easy to control + able to expand”

Jiangnan culture, represented by Suzhou and Hangzhou; literati culture is a way of life,

gardens, calligraphy and painting, food, the solar terms, all embedded in daily life

“Restraint”, “leaving blank space”, “not saying everything” — these are products of a long-stable society

Qilu culture, whose core is Confucian culture, shaped the inner behavioral values of Chinese people

Emphasizing order, responsibility, role, ethics

Represented with Qufu at its core

Central Plains culture — the keyword isn’t “was once glorious”, it’s “never broke off”

Regimes can collapse, the way of life doesn’t

War was frequent, but the culture never lost its lineage

It’s where Chinese civilization “recovers its blood”

Its representatives are Luoyang and Kaifeng (I always thought Luoyang and Xi’an were very close, with similar cultures)

Lingnan culture

Extremely adaptive, also strongly globalized, absorbs outside cultures easily

Commerce, clans, religion — highly pragmatic

Lingnan culture’s representative cities are mainly Guangzhou and Foshan and that area

Compared with Minnan culture, which is Quanzhou, Xiamen, Zhangzhou

Minnan culture:

Central Plains migrants moved south

Clans banding together

Going to sea to trade

Spreading outward (Southeast Asia)

Lingnan culture:

Local Baiyue culture + Central Plains culture

Long-term foreign trade

Earliest and most contact with the world

Continuous hybridization and updating

Lingnan has all kinds of chambers of commerce, guilds and companies — loose, realistic, decentralized, quick to adapt to change

From Qin-Han to Tang-Song, Lingnan was always the frontier, hence the saying “exiled to Lingnan”, plus all the diseases of Lingnan, the damp heat and so on

By comparison Jiangnan was prosperous in that period; especially in the late Tang, and during the Northern Song, Jiangnan began to flourish

Song dynasty commercial prosperity and institutional balance

The commercial prosperity of the Song dynasty

A high-quality mechanism for rules to operate

Paper money, night markets, cross-regional trade, professionalized division of labor, an urban middle class — in the Song these didn’t appear in scattered form; they had already started to mesh with each other and form positive feedback

It broke almost everything modernization needs into modules, except industrialization

The Song deeply respected “complexity”; it didn’t believe military force was omnipotent, didn’t chase a crude expansion narrative, but kept doing fine-grained balancing between finance, governance, culture and technology over the long term

Of course this also brought fragility; the Song really was bad at war, but it was willing to use institutions, technology and culture to fight uncertainty

The missing rural pension system

Over the past few decades, for the sake of rapid industrialization and urbanization, the state directed a huge share of resources, protections and institutional dividends toward the urban and industrial system first; the countryside bore the cost, but did not accumulate, in step, the institutional assets needed to support old age

So this generation of elderly people in rural China has no accumulating institutional container

These are all debts, very hard problems to solve

The pixel-in, pixel-out principle

pixel in, pixel out

The input is pixel-level data, the output is still pixel-level output, and they correspond one to one

Information isn’t “compressed” into a judgment, it’s “rewritten” into another pixel expression

Elon has his own judgment on this part

world → senses → world

rather than world → judgment → language

Darwinian user observation and scenario evolution

Observing users shaping scenarios the Darwinian way

Give a general architecture first, let users use it according to their imagination; the team captures head scenarios through anonymized statistics/pattern recognition, then does the last-mile optimization

Example: later they found users really loved making Slides, web pages, batch file processing → the product team stepped in to optimize

Respect the laws, respect common sense

The deceptiveness of common sense and dissecting a system’s essence

Common sense can deceive you

You always run into scenarios that make you physically uneasy

Your intuition tells you it may not be right

Why is it like this, why does it need to be like this, what is the essence

What is the smallest, least, irreducible case

You don’t need to step in the pit yourself, you don’t need to borrow others’ experience, you don’t need to imitate, you don’t need too many unverified assumptions

What is the essence of the system, the product’s real cost, the essential logic of why users pay, users’ ability to pay

Once you dissect out the most essential things — those few core, workable variables — you can make your own trade-offs easily

The boundary of concept decomposition and the pursuit of essence

When do you keep decomposing?

When must you stop?

What granularity counts as “taking responsibility for the essence without going off the deep end”?

Industry language is usually encapsulated; note that language is the unit of thought, but it also limits our thinking — the language of language, the meta-meaning; how do we pin down this concept and this unit

When the explanation produced by thinking in language itself can’t predict outcomes, intuition tells me that level isn’t enough, that language itself may be the problem

The purpose of decomposition is to narrow the space of choices,

“If I don’t do A right now and can only do B or C,

then what is forcing me?”

What exactly is the hardest constraint, the most essential constraint?

When to stop? When you realize that continuing to dig and decompose is to satisfy your own intellectual pleasure … not to serve the goal

There are some uncontrollable variables (the uncertainty of human nature, the direction of culture, the direction of politics…)

The spatial aesthetics of Wujie Heyi guesthouse in Chaozhou

Wujie Heyi guesthouse (Chaozhou old town, Paifang Street branch)

My favorite cafe for a long stretch of time recently

A strongly designed, hybrid guesthouse space

Not just lodging — an experiential guesthouse that blends coffee, a tea space, an artistic sensibility, lifestyle, traditional and modern together

It started from an old small Western-style house, rebuilt into a comprehensive experiential space

The name Wujie Heyi is pretty cool — people, environment, culture and aesthetics all fused into one continuous experience

The shell of the old Western-style building + modern taste in furniture + local cultural ornaments

Eastern and Western elements mixed together

The space doesn’t emphasize a single regional style; rather, “aesthetic unity ranks above style labels” — whether what you see is Eastern furniture, retro lighting or modern lines, they’re all bound together by one unified aesthetic language:

quiet, elegant, comfortable, nature and life coexisting

Writing a book as thinking organization and passing on value

I have some thoughts about writing a book

Somehow, in this life I still want to seriously write a book

I always want to pass something on to this world

I always want to create something — that makes me happy

I want to organize something — that helps me get my thinking clear

Cognitive science of verbal thinking and conversational structure

Verbal thinking processing

Cognitive science has proposed it, and it really is the core way of thinking for many people

They need to chat, but they don’t care about the content of the chat

For these people, the interlocutor isn’t there to provide “information” but to provide “structure”

Language is linear, forcing the brain to think and reason along one line

(The unaware) Many people do this without knowing it. They’ll drag you into a two-hour conversation and reject every suggestion you offer. You feel exhausted, thinking they’re “dumping their grievances” or “being stubborn”. Actually they’ve just turned you into an “echo wall”. Because there’s no awareness, this kind of communication often leaves the listener feeling drained

(The aware) These are the experts — they know what they’re doing. They’ll openly admit their mind is a bit muddled and ask if you can listen to them for ten minutes, no advice needed. These people have high metacognition and know what they need at that moment

When code has a bug, programmers put a rubber duck on the desk and explain the code line by line to the duck. Often in that process the programmer discovers the logic flaw themselves

But in real life, interestingly, most people are waiting to respond — usually halfway through listening they start composing their own rebuttal or suggestion. And they need the sense of being present

Greed, aversion and delusion, and the modern psychological predicament

Greed, aversion, delusion (rāga / dosa / moha)

Greed in relationships — or you could call it sunk cost, haha — it’s clearly gone bad, but you won’t accept it

Greed in social achievement: wanting it too much, wanting to prove it too much

Aversion toward people (the core of relational conflict): you get angry, furious, because the other person doesn’t seem to run according to the script in your head. It’s also aversion toward yourself, and then aversion toward the world — the world is the problem

Delusion, the underlying operating system of all disasters. It’s hard to distinguish greed from delusion — greed is wanting more; delusion is “and once I get it, then what”, the sense of lack even when satisfied. Delusion is also being too self-centered, “this is just how I am” … a misunderstanding of the world

Why do humans systematically create suffering when what they want is happiness

The understanding I’ve summed up:

Greed = clinging to the “pleasant state” + unwillingness to lose it

Aversion = rejecting a reality that “isn’t as I wish”

Delusion = a fundamental misjudgment of how reality operates

What’s hardest to align in communication is meaning

What’s hardest to align in people’s communication isn’t information, it’s meaning

Fandom provides a ready-made framework of meaning: what’s good, what’s worth pursuing, what’s beautiful, what’s passionate, what’s pure

You don’t need to explain from scratch who you are and what you cherish; you just say “I follow them too” — actually it’s the same as when I meet a group of idealists — and the other person can read a whole long résumé of values

An idol is a “transcendent object” that can be jointly identified — like a god in religion, a flag in a nation, a totem in a tribe. It gathers scattered individual wills into one beam of light. Your feeling of closeness is essentially:

“I see the same ‘ought to be’ in you”

Humans have a very magical mechanism: getting excited together makes you feel more familiar. This easily produces emotional synchronization; in psychology it’s called emotional contagion, affective resonance

Can you share strong emotions with someone else without feeling awkward

How conceptual frameworks erode and change people

It suddenly occurred to me — many notions, many ideas

once they invade your mind, they slowly erode you

Mal wasn’t forced into suicide by anyone. She was infected by a seemingly “reasonable” sentence: this world isn’t real

This is an explanatory system that can account for all counterevidence, so you can never win against it

People aren’t changed by facts; people are changed by the framework that explains the facts …

How you think matters a lot …

The Buddhist meaning and practice of sympathetic joy and praise

Sympathetic joy and praise, in the Buddhist context

Seeing others do good and achieve things, feeling joy in your heart, and expressing recognition and encouragement in words/actions

The scriptures have the phrase “instruct, benefit and delight, and rejoice in and praise”; in Mahayana, the fifth of Samantabhadra’s “ten great vows” speaks of “rejoicing in merit”, systematically making it a practice method

Sui: following along, keeping up, but not blindly following, not working against goodwill

Xi: not excitement, but that kind of genuine happiness that isn’t sour or twisted

Zan: pointing out where the other person did right, what’s worth learning from

Tan: heartfelt respect — I acknowledge the value of this thing and am willing to let it be seen

“What you did has value, and I’m willing to vote my attention and goodwill for it”

The human brain easily misreads someone else’s radiance as a threat to me, so jealousy, belittling and nitpicking arise — a self-protection mechanism

What sympathetic joy does is counterintuitive: it rewrites someone else’s good as a shareable gain, and trains an emotion called shared joy — seeing others do well, you can do well too

The direction of the mind is itself karma (the seed of action)

Lingyin Temple’s explanation emphasizes: sympathetic joy isn’t just being happy inside, ideally you also “lend a hand” and help others’ good deeds succeed

The difference in level between jealousy and sympathetic joy

Jealousy and sympathetic joy are two different levels

The more you envy something, the less you get it

The more you rejoice in something, the more you get it

To rejoice in others, the target level has to be higher

You pursue a higher level, a larger frame

How the restaurant shooting shaped the Godfather

A sudden experience shaped the Godfather

The scene where Michael Corleone shoots Sollozzo and the police captain in the restaurant

First, at the character level — he goes from “wanting to be an ordinary person” to “being forced to bear order”. Not ambition, but a cold start of responsibility

Second, at the moral level — the violence is filmed as unsexy, unheroic, like a dirty but necessary job. You can understand it, but you’re not consoled

Third, the film language — sound completes the transformation before the action. The train isn’t background, it’s the gear of fate; when the gun fires, the world has already decided the outcome in advance

The Godfather is the calmest, the most capable of deferring emotion, able to freeze private emotion into structural decisions

The family is no longer an “emotion-driven patriarchal organization” but becomes an “efficiency-first power machine”

He won the world, but lost his family

The disappearance of the idealistic Michael in Kay’s memory

In Kay’s memory it’s still that idealistic Michael

Every one of Michael’s later moves is cruel

Deferring the truth, deceiving

Treating marriage as a legitimacy device

He doesn’t allow her to reach her own conclusions

Two generations of Godfather: character contrast and the cost of power

Two generations of Godfather, two generations of character

The first generation built power on favors, credit and long-term reciprocity

Like an old-school politician, emphasizing slow, steady, a sense of proportion — hence a powerful network of relationships

The second generation is the terminator of order and the symbol of modernity. Michael re-engineers the family from a “network of favors” into an “institutional machine”: rational, centralized, ruthless. He doesn’t maintain stability through favors but through structure and fear. This is a ruler maximizing rationality, and the price is emotion drained dry. Power won, family lost

Michael Corleone, extremely rational, a man who treats the world as a chessboard and himself as a scalpel

Extremely rational, self-controlled, sensitive, good at deferring gratification, and rarely showing emotion

He both longs to escape the family’s violent fate and believes only he can end it all

What he fears most isn’t danger but unpredictability — his own environment, the dangers he meets, and so on. A post-trauma survival strategy: as long as he can control it, he won’t be hurt again

Then there’s emotional resonance: he can understand others’ emotions, but presses his own below the ice — exposing emotion equals exposing weakness

Rationality-oriented: the end swallows the means, and the means in turn reshape the end

A lonely type of leader: the stronger, the more closed off; the higher he climbs, the less he trusts people. Because he’s seen betrayal, seen the fragility inside the family, seen the price of power. Centralizers are like this too — effective short term, but very dangerous long term, because the organization becomes a nervous system wrapped around his person: once he goes cold, the whole world goes cold

Michael Corleone’s systemic capability and its backlash

Michael Corleone’s “systemic capability” is extremely strong, strong to the point of coldness; but precisely because it was so strong, the system ended up devouring him

What others see is emotion, betrayal, hatred

What Michael sees is: who’s a node, who’s a risk source, who can be replaced, where to cut.

He doesn’t handle “people”, he handles relationship graphs

A system needs stability, so deferring emotional gratification is especially important

The world runs on relationships and exchange, not morality

Power, order, family, bringing …

The world doesn’t run on morality but on relationships, promises, deterrence and exchange

It’s a system, a constructed system,

Under this system, what kind of people are born, what kind of structure is born

Why this structure forces people to make such choices

Which rules actually take effect beneath the surface morality

The cost of an excellent character vs. an excellent system

The similarities and differences between excellence of character and excellence of system

When you truly do everything as well as possible

do you still have a “self” left?

Does excellence carry a cost, is it the optimal choice inside the system?

What do we actually want

The world has no standard set of values that can guarantee each of us is treated fairly and is happy

This world, through constant experimentation and operation, gives birth to certain constraints that are repeatedly verified

When a value system

requires people to fully instrumentalize themselves

doesn’t allow doubt, doesn’t allow stepping away

flattens the individual in the name of “correctness”

compresses a complex world into a single goal

then whether it’s called justice, efficiency, faith, success or a great mission — it will almost certainly produce disaster

Sympathetic joy against jealousy: the internally stable need no outside proof

Opposite of sympathetic joy — jealousy, desire

As long as you’re not that good, I’m not that bad

This kind of person is good at evaluating others

What we hate most is often what we’re not allowed to become …

So, especially under conditions of scarce resources, likes and follows become quantified metrics

Another thing is the superiority disguised under one’s sense of morality

This is a more refined form of belittling

It’s not directly saying you’re no good, it’s

“I’m more authentic, I’m more clear-headed, this kind of thing has no depth” …

People who are internally stable accept the complexity of themselves and others, and don’t need others to prove themselves

The most fully developed mental systems in human history

Buddhism — very mature, systematized

Stoicism

Confucianism

Daoism

Existentialism — a notch higher, but it easily slides into nihilism

Marxism, materialism — it applies very well to how the world runs

But it can’t settle an individual’s inner life. Extremely limited

Materialism and idealism are the world’s two basic philosophical forms

Anything that has objective reality, that won’t change because of human will, is called materialism.

World philosophy splits into two camps: materialism and idealism. In practice, idealism revolves around a series of “mind studies,” or “studies of mind.”

But materialism is based on Marxism — after Marxist philosophy was put forward, it’s the basic compositional form of the world.

Cherish the non-derivable blank in the world’s generative process

A great many things in the world are derivable, learnable, copyable,

but what interests me, and is more precious, is that part of the generative process that can’t be fully derived

Just like Bayesian probability: even with 80% certainty about whether a girl likes you, what you care about more is the remaining 20%

That latter 20% is the part that stirs your emotions, makes you replay it, makes you nervous …

This is also why we give the world a little blank space, why we give AI a little imagination, letting them grow naturally

…

Douyin compresses human attention into a priceable data product

Douyin’s ads

Ordinary creators, merchant content, and — ads, just disguised to look a lot like content

Feed ads, brand challenges, and local merchant promotions …

Merchants, brands and local owners are the ones paying

Users provide attention, dwell time and behavioral data

Douyin now also makes money from e-commerce

Then there’s livestream tipping, where only a small number of high-spending users contribute a large share

So Douyin’s essence is that it compresses human attention, desire and behavior into a predictable, priceable, reusable data product

Douyin doesn’t charge ordinary users — their willingness to pay isn’t high anyway — it just wants you to watch a bit more, linger a bit longer, click one more time

Young people’s early social media exposure skews how they bind self-worth

The problem of young people encountering social media too early

Usually “too early” is defined as having a personal social account before age 10-12

and platforms dominated by algorithmic recommendation, including short video and feeds

and no stable adult guidance

Girls will bind appearance strongly to self-worth, and compare appearance far more frequently; girls will feel that being liked and followed is an indicator of self-worth. They look more precocious, but inside they lack security

Boys take another route — they’re more likely to be recommended games, extreme content, sexual innuendo and borderline content, so their ability to delay gratification is noticeably weaker

Deep motivation collapse: long-termism gives way to instant returns

The problem of deep motivation collapse

The world changes fast

More and more people start chasing

Not because they’re lazier, but because for things that need long-term investment yet are uncertain, the subjective sense of meaning drops markedly

Fewer people are willing to take the long path (research, deep specialization)

More people chase tracks with “instantly visible returns”

In such a high-change era, does long-termism really still have value?

A path, an identity, a set of skills seems to slowly stop working

Some foundational abilities matter more: long-horizon narrative ability, foundational assets, aesthetic sense, learning ability, cognitive ability, personal brand

Complex systems need a small number of people who truly understand

So a distribution emerges:

The majority: short-term tracks, competition extremely crowded

The few: long-term tracks, winner takes all

This is a power law, not a myth of hard work

Gen Z’s plural values are an illusion; success standards are actually narrower

Is Gen Z’s value pluralism an illusion

Values are extremely concentrated, but expression is highly dispersed

On the surface:

more open values

more diverse choices

But the quantitative result is:

the criteria for judging “success/failure” are actually narrower

Highly concentrated in:

exposure, followers, income, visible influence

Young people seem to accept everything, but they’re unusually cruel to themselves

Beware social media’s standardized template for interpreting your life

The question of interpreting your life

Is the right to interpret my own life still in my own hands

The problem with social media is:

it provides ready-made narrative templates

success, happiness, relationships all have standard answers

The individual just keeps “checking the matching box”

But … life doesn’t need more standard answers …

What matters more is understanding yourself, accepting yourself, understanding the world, making contact with the world

From living in the moment to actively living the moment well

Living in the moment -> living the moment well

Living in the moment: don’t overthink, relax, enjoy now; it fights anxiety, over-planning, over-churning

Living the moment well: I accept that this moment is all my input, but I want to polish it into a good version. What it fights isn’t anxiety, but wasting time, losing control, a sense of floating

Living the moment well = in an uncertain world, directing attention to the variables you can influence, and making the smallest but real improvement to them

Living the moment well isn’t just acceptance, it’s creation …

The midday market’s potential comes from structurally driven crowd behavior

The midday market feels quite promising. Right or wrong aside, under this system structure, this group of people is destined.

What kind of people are they? What drives them? And what will the future social consensus be? Whoever holds the next consensus holds enormous wealth. As long as this consensus persists, the wealth persists.

The “unruly commoner” phenomenon stems from scarce resources and weak institutions

The causal reasoning behind “poor remote places breed unruly folk”

An emotionalized summary of experience

The “unruly commoner” is someone who doesn’t follow rules, loves loopholes, resists hard, and makes people feel they’re “troublesome”

These are all surface appearances, not motives

It’s not morality, but an environment of scarce resources, weak institutions, limited opportunity

When someone grows up from childhood in an environment of “if you don’t grab it, it’s gone” and “if you don’t fight, you’re ignored,” what he learns isn’t cooperative games but zero-sum games — at that point it’s a basic survival strategy

If rules often fail, enforcement depends on connections, and reasoning is useless, then “the rule-follower is the one who loses out.” Over time people learn that acting tough beats reasoning, and making a scene beats obeying the law

Then there’s the lack of upward mobility channels — which is also why the schooling system evolved naturally

What’s really dangerous isn’t the saying itself, but when it’s treated as a “moral conclusion”:

It makes the strong lose patience with the weak

It reduces institutional problems to “the people are no good”

It lets environments that should be repaired keep rotting

Historical inertia runs as a structural cycle plus a technology amplifier

Let’s talk about historical inertia

Historical inertia still holds in China, but it no longer appears as dynastic change; it runs in the form of a structural cycle + technology amplifier

Historical inertia:

Power tends toward concentration, organizations tend toward self-protection

Once upward channels narrow, social tension accumulates

Order depends on consensus; once consensus breaks, maintaining it costs more

Technology compressed the time scale,

education raised cognition, but people no longer have the ability to change structure — producing a high-cognition, low-delivery population

The tools of state governance have completely changed: today information is highly visible, fiscal capacity is coarse, and the management radius is small

Do upward channels still get reopened again and again?

Is the narrative still self-consistent

Is technology a repair mechanism, or an amplifier

Big dogs trigger an illusion of safety

Girls tend to like bigger dogs

Big dogs easily trigger an illusion of safety

and are more easily anthropomorphized into gentle guardians

and more easily create a sense of narrative — your relationship with the dog, a sense of imagery

Boys are also more likely to prefer bigger dogs, from role mapping — the imagined companion and comrade-in-arms

Oxytocin is the glue of relationships

When women see neotenous features (big eyes, round heads, dependent behavior)

oxytocin is released faster, in larger amounts, and lasts longer

Men secrete it too, but the trigger conditions are narrower and it falls back faster

Women’s oxytocin levels are generally higher

After intimate contact, women’s oxytocin generally rises 20%-40%, while for men there’s relatively no significant change or a slight rise

When a mother and baby look at each other or nurse, oxytocin can spike to 8-12 pg/mL

Fathers rise too, but by a smaller margin (about 4-6 pg/mL)

Oxytocin’s essential purpose is to be the glue of relationships

Endorphins are the pain silencer and the reward painkiller

Oxytocin has only one core function, but it’s extremely powerful: reducing the psychological distance between people (or animals)

The triangle driving things: institutions, technology, culture

Institutions determine what’s allowed, what’s forbidden, what’s rewarded

Technology determines what’s feasible, scalable, monitorable

Culture determines whether people are willing to move forward with the system

Population aging → change institutions (pensions, taxes)

Resource scarcity → deploy technology (efficiency, substitution)

External conflict → strengthen culture (identity, narrative)

The protective boundary of drug patents

Drug patents are generally extremely expensive

What they protect is:

a specific molecular structure

the preparation method

the use (indication)

sometimes also the administration route and dosage form

Gene editing changes the DNA sequence inside an organism and changes protein expression

Counterfeit drugs in reality are all chemical copies

After a patent expires, legally copying the same active ingredient (this is called a generic drug), or directly cutting corners, adulterating, and mislabeling doses

The next ten years of biotech

Biotechnology and health

Thanks to the vigorous development of gene editing, customized medicine and new therapies, the biotech field looks quite promising

At least for the 2028-2038 stretch ahead

Society’s main direction right now is much more in AI

Medicine may have a lot of innovative drugs, CAR-T, gene editing and AI drug repurposing, but institutions move slowly; right now it’s a dividend for researchers and a very small number of investors

After 2028, China’s population over 60 will exceed 25%; this is being pushed forward by demand

In the years after, commercial insurance may take the main stage in medical payment, with medical insurance starting to cover some gene testing and so on

The ability to quickly model an industry

I think my buddy is very strong at quickly understanding an industry

It’s very much worth learning this ability myself

How to quickly learn a certain industry

He can, through chatting, quickly build a systematic modeled understanding of an industry

Granularity in industry research

When talking about an industry, you must consider the question of granularity

This is especially important

Otherwise, when doing research, it’s very easy to run into inconsistent definitions

The macro view, the micro view, and the mismatch that arises between the macro state and the micro state — first principles are a good way to resolve the mismatch: zoom in on the grain

Industry standards create consensus

Common industry classification standards include the national standard Industrial Classification of the National Economy, and the Guidelines on Industry Classification of Listed Companies issued by the CSRC

Standards bring consensus

Diffusion of innovation and human risk

Diffusion of innovation theory

Innovators (2.5%): a very small number, willing to bear high uncertainty

Early adopters (13.5%): have judgment, willing to take “controllable risk”

Early majority / late majority (68%): want “proven certainty”

Laggards (16%): lowest cost, but fewest opportunities

This has been repeatedly verified on the iPhone, Bitcoin, AI, the internet, short video

Most people are never in the first batch

Risk aversion is a human evolutionary advantage

The vast majority of people have no spare capital for trial and error

Social systems reward “safety,” not “adventure”

The essence of investing is buying a business

Investing is very simple, but also very hard

In one sentence

If I bought the company today and the boss disappeared tomorrow, could I still sleep peacefully?

The architecture of a business model, three questions

Does money flow in by following human nature? If it needs no user education, and human nature is fundamentally unchanging, then this business model is all the more stable

How much maintenance cost must be paid to earn the money? Some companies have to work desperately hard to make money — exhausting but not smart. Acquire a customer once, collect payment repeatedly; earn trust once, compound long-term

A moat isn’t emotional, it’s structural. This is actually easy to understand …

The feeling of realizing things after the fact is especially important (cognitive delay)

When young, you’re easily drawn to growth, stories, cleverness

But later you find that even the most impressive person can’t beat a simple, repetitive, counter-human-nature good model

Financial reporting: annual results within 3 months, annual report within 4

The timing of financial reports

A company’s fiscal year end is usually December 31

Annual results forecasts and preliminary results announcements must be released publicly no later than 3 months after the fiscal year ends

The formal annual report is generally within four months after the fiscal year ends

The interim report is generally as of June 30

First quarter (Q1) briefing (if published): around mid-to-late April (about 45 days)

Half-year (H1) results and report: as of June 30 → the report is usually published before the end of August (within 3 months)

Third quarter (Q3) briefing (if any): around early November (about 45 days later)

Full-year preliminary results announcement: as of December 31 → released by the end of March the following year (no later than 3 months)

Formal annual report: generally between March and April, at the latest mailed/released within 4 months after the fiscal year

Biopharma’s core distinction: biologics vs chemical drugs and tracks

Biopharmaceuticals

The core distinction:

Biologics: complex structure — insulin, antibodies and so on

Chemical drugs: aspirin, ibuprofen and the like

A few tracks:

Antibody drugs, the largest share, including monoclonal antibodies and ADCs (called biological missiles)

Cell and gene therapy — this is doing some gene editing or CAR, ex vivo genetic modification

Recombinant proteins and vaccines — the familiar ones, insulin, growth hormone

Traditional drug development needs “ten years and a billion dollars.” After Google DeepMind’s AlphaFold appeared, AI can predict protein structures and hugely shorten the time to find drug molecules

Bispecifics/multispecifics: one hand grabs the cancer cell, one hand grabs the immune cell, put them together to fight

GLP-1, mainly peptides, but with the half-life extended through biotech means

Camus: in an absurd world, hold dignity through revolt and action

Camus argues for [facing] nihilism, maintaining dignity and meaning through revolt and action in an absurd world

In the early 1940s he published The Stranger and The Myth of Sisyphus, establishing the theme of the “absurd”: in a meaningless world, people must respond with lucidity and revolt, not escape or suicide

AlphaFold is essentially a computational mapping from sequence to structure

The most essential thing about AlphaFold doesn’t feel like the problem of structure prediction

It turned the “sequence → structure” mapping, which used to need years of experiments, into an almost instant, scalable computational process

It used “evolutionary statistics + deep representation learning” to replace explicit physical modeling

Easier to understand proteins

Before, to study a protein you needed to know whether a structure existed

Now it’s: you have a structure draft, and then how do you use it

In the biological world there’s an extremely cruel but extremely true rule

Function is determined by structure, not by name

So drug research is structure-based, and protein engineering is structure-based too — how are structures obtained?

Traditionally, structures are measured experimentally

X-ray crystallography

NMR

Cryo-EM

Very slow, very expensive, very picky about proteins, and can’t cover all proteins

We know hundreds of millions of protein sequences, yet know only a tiny number of structures

What AlphaFold changed is turning this: “Is there a structure?” into: “Is there a usable structural hypothesis?”

Protein folding isn’t a random search but an energy slide

How protein structures are predicted

In theory, a question I’ve been wondering about: a protein itself is a soft chain, it may rotate or fold, and there are even intrinsically disordered proteins

Then wouldn’t every residue be able to rotate — why doesn’t it just rotate randomly and stop when it’s roughly right?

Actually that’s a wrong assumption: the scale of conformation space explodes exponentially

Random search is too slow — exponential explosion — this is the famous Levinthal paradox

If a protein found its native structure by randomly trying all possible conformations, even if each conformation took only xxx seconds to try, it would need far longer than the age of the universe to find the right structure

Yet in reality, most proteins finish folding within milliseconds to seconds

Protein folding isn’t finding a position in space, it’s “sliding down” in energy space

AlphaFold isn’t simulating the “folding process”

It’s predicting: the most likely stable structure near the lowest free energy

The contradiction between exponential research growth and linear healthcare growth

The research side can grow exponentially

but the healthcare side grows slowly and linearly

This is determined by the human species’ ethics, safety, and clinical validation cycles

Large-Scale Gene Sequencing Requires Respect for the Complexity of the System

What else is needed to complete large-scale gene sequencing

Respect the complexity of the system

First layer: sequence -> analyze function. AI turns sequence -> molecular readings, compressing from the level of years of experiments down to computational seconds

Second layer: molecular function -> cell/tissue state (10–20 years, partially solvable). The same variant has different effects in different cell types, at different developmental stages, and in different environments

Third layer: gene -> disease -> individual fate. The genetic layer also cannot fully restore a person’s future; the world is contingent, and the human body is a nonlinear, strongly coupled, history-dependent system

Removing the Steering Wheel Is the Prerequisite for Building a Third Space in the Car

Removing the steering wheel is the beginning of building the space inside the car as a third space. Looking at Apple’s strategic choice from this angle, it’s not hard to find one of its important assumptions: before autonomous driving truly matures, most car use scenarios are still people driving the car to complete the purpose of travel, and at that time more entertainment and interaction would only bring various safety hazards that disturb attention. Only when autonomous driving has completely freed the attention of the rider, and there are no longer any driving controls, including the steering wheel, inside the car, can the building of a third space inside the car begin

The Value of Genetic Maps for Disease Prevention and Medicine

The importance of gene sequences

You will know your probability of developing certain diseases, and you will get advice on how to improve your health; in the future there may also be drugs that prevent specific diseases according to genetic traits. In addition, everyone will have a genetic map, which helps with getting health and nutrition advice based on big data. As you grow older, all kinds of common and chronic diseases may appear. Having your personal genetic map as early as possible is crucial for preventing disease and improving health, and of course it can also greatly save on medical costs

Gene sequences are also helpful for testing new drugs: if we can distinguish these situations, we can find patients with specific genetic traits for some innovative drugs, truly achieving a match between drug and patient, which is a huge blessing for both the pharmaceutical industry and patients. Drug companies can develop highly customized drugs designed specifically for individuals based on a patient’s genetic map and medical history. This kind of drug doesn’t need to be effective for everyone; it only needs to be effective for a specific group of people. Once this information is linked together, its value in treatment becomes very great. Besides being pushed by the state, insurance companies will also strongly push gene sequencing, and will pay for people’s gene sequencing

Compared with gene sequences, before editing human genes, we need to reach a basic consensus on scientific ethics. One basic consensus reached now is that we should not play the Creator. The reason is simple: we understand far less than we imagine. Although we can edit a certain gene, that editing may be precisely one manifestation of our ignorance. The actual problem is much more complex — not all traits (conditions and abilities) can be traced to a corresponding single gene

Gene Technology Will Become a Tool of Distinction for the Wealthy

The rich using “gene-related technology” to show off and to distinguish themselves will almost certainly happen

Class differences in the coming biological age

Enduring long-term uncertainty, and avoiding risk

Market Prices and Results of Animal Gene Cloning

Current prices for animal gene cloning

Current market prices (roughly):

• Cats: $35,000 – $50,000

• Dogs: $50,000 – $100,000+

• Horses: can reach hundreds of thousands of dollars

Appearance similarity is very high

Personality similarity is moderate

Health and lifespan are comparable to ordinary pets

Mind Uploading: Consciousness Migration and Non-Biological Existence

Mind Uploading / Consciousness Uploading refers to:

copying or migrating a person’s “mental state” (memories, personality, ways of thinking, decision patterns) from the biological brain to a non-biological carrier (such as a computer or a simulation system), so that it continues to exist or run outside the original body

A civilizational hypothesis: that humans can leave biology behind and continue to exist in the form of information

Risks and Technical Bottlenecks of Invasive Brain-Computer Interfaces

To get an invasive brain-computer interface you still need surgery to implant a chip; this way is not friendly, and it brings risks to the person. The chips implanted now may only be valid for a year, because the body will have all kinds of rejection reactions, and as time goes on the chip’s signal may gradually weaken. Unless materials science makes enormous progress, having surgery once a year to replace a chip is not realistic. Of course, as an interface between carbon-based and silicon-based, whether there might be a better way to connect silicon hardware with “wetware” (the human brain) is also very much worth exploring

For a chip to accurately inspect the electrical wave information transmitted by the brain still requires a huge amount of data and training. The information for directing limb movement is relatively simple, while information conveying complex content and emotion is much more complex. So the chip must not only be able to acquire the waves, but also be able to inspect and interpret them fairly accurately

The Logic and Metaphysical Method of the Debate at Sera Monastery

The debate sessions at Sera Monastery

Several directions of debate

Whether the causal relation holds

Whether the concept is self-consistent

Whether the definition has been quietly swapped

Whether the inference necessarily holds

“Is everything impermanent?”

“If A holds, does B necessarily follow?”

“Does the premise you just gave already contain the conclusion?”

Basically it covers a hybrid of logic + metaphysics + meditation

In Tibetan Buddhist debate:

Clapping: the logic hammer (equivalent to “this strike of mine is the conclusion”)

Stomping / leaning forward: emphasizing the causal progression

Closing in on the opponent: forcing an immediate response, no stalling allowed

Variation, Selection, and Information Preservation in Evolutionary Biology

What evolutionary biology studies is a very fundamental question

Where we come from, how we got to how we look now step by step

How we will walk on in the future

With no designer, how did life come to this point today through “variation + selection + time” — it cares about the following questions:

Where variation comes from

How selection happens

How information is preserved

How scales are crossed

Three core ideas

Natural selection is not an engine of progress but a mechanism of elimination; surviving ≠ excellent, it just means you didn’t die in the current environment

Adaptation is a local optimum, not a global optimum

Randomness is the source of creativity; without mutation there is no new possibility — life walked to complexity by relying on errors

So evolutionary biology doesn’t say that more advanced means more deserving of success, or that the status quo is justified; it explains causes, not legitimacy

So in the current era, what goes on optimistic evolution and derives onward are instead those reverse degenerations, the shapes that look very stupid, the simpler ways

Biological evolution is cruel: the survivor after countless failed branches is survivorship bias

Humans are also a miracle in history

The brain evolved to cope with a scarce environment → today riddled with anxiety

A fast reward system → addiction problems

Group preference → bias, opposition, identity politics

Sexual selection → many irrational behaviors

From the angle of biological evolution, the human design never seems to have been designed for happiness; humans were always cobbled together for survival

The Importance of Environmental Selection in Evolutionary Biology

Evolutionary biology is also telling me

I am not the center of the world, and my intuition is not reliable either; even future success has contingency to it; freedom lies in the fact that there is no fixed version, no single correct understanding, and the system can be constantly reassembled

From a systems perspective, I really like the process of evolution; moreover, because it is uncertain, it means creation

Environment is especially important; people really should put themselves in a suitable environment — choosing the environment matters far more than forcing things in the wrong environment

The Research Methodology of Psychology, Anthropology, and Sociology

The objects of study are mainly humans: psychology, anthropology, sociology

Why do people think and act the way they do

Inferential hypotheses, cross-cultural validation, adaptive reasoning

Of course now there is technology as a big variable; evolutionary psychology may also have a different direction of evolution in the future

The Evolutionary Truth That Organisms and Environment Shape Each Other

Textbooks always say the environment selects organisms

Because it fits the pyramid theory of natural selection

Mutation is random

The environment is fixed

Those that don’t adapt → die

Those that adapt → survive and reproduce

Once you stretch the timeline out, you’ll find a counterintuitive fact: the fact that organisms are alive is itself rewriting the environment

Plants turn CO₂ into oxygen → transforming the atmosphere

Microorganisms change soil structure → affecting subsequent species

Marine organisms build coral reefs → rewriting ocean flow fields

Humans build roads, build cities, domesticate plants and animals → completely reshaping ecological niches

Organisms change the environment → the new environment then selects organisms in turn

Animals have always been adapting to an ecological niche, but the ecological niche is like the operating system of this world; top organisms also participate in building the niche

So what exactly is the environment? It’s the legacy left behind by the previous generation of organisms

The civilization they left behind, the smog they left behind …

System vs. Individual Conflict of Interest in the Bourne Films

I binged all five films in one go

I felt the power and cold-bloodedness of the state machine — it’s like this in any country

What’s different is how people under different cultural systems react to such a system

In traditional spy stories, the protagonist is often “righteous violence authorized by the state”

But in the Bourne films, the protagonist wakes up and finds something horrifying:

he didn’t “choose to become a weapon”

he had already been made into a weapon, and then abandoned

The system has self-correction, self-protection, and self-cleansing mechanisms, and at that point the conflict of interest between the collective and the individual becomes visible

When the state, the organization, the system start deciding “what is correct” for you

what’s often left of a person’s dignity is one choice: run

Meta-Ability Matters More Than Knowledge; Building a System More Than Accumulating Frameworks

Meta-ability

Learning how to learn matters far more than learning knowledge

Building a top-tier personal system matters far more than how many frameworks you already have

The Logarithmic Law Explains Time Perception: Long Childhood, Rushed Adulthood

The logarithmic law: people’s feelings are relative, not absolute

So sometimes when I switch to thinking about certain questions, for example thinking about my life, maybe I won’t think about how many years have passed in an instant, or how many years are left, but rather, in a sensory sense, what percentage of my life has passed?

This is also why childhood summers were so long, but now time feels so fast, year after year hurries by

And so it’s also why we can feel that when we are 10, one year is a tenth of our life, but when we are 50, our one-fiftieth, our life suddenly drains very slowly

Well, that also explains why, after being alone outdoors for a long time, I don’t much like being around a crowd of people in a shopping mall. It’s because our biological evolution itself is not linear, so much so that I can adjust my own perceptual ability: in quieter environments I raise my sensitivity, but this ability becomes a burden in a mall with huge crowds

In this era, people’s focus has been monopolized by the head players

So go look for the things that match your hobbies and can produce a compounding effect, and then go all in

Rather than scoring 60 in 10 fields, it’s better to score 90 in one field

Special Experiences Sustain Friendship: the Starting Point Matters More Than Frequency

In 2023, I met archer for the first time, in Heyuan City

That day I tried night hiking for the first time, with no experience, a total newbie; I hurriedly signed up for a small group ahead of time

I had just started my senior year that year, and archer was still working hard; we got to know each other

Because that night I hadn’t brought a headlamp, but archer was very thoughtful — he had a lamp and lit my way the whole time

And our friendship just kept going

archer’s family keeps two little dogs, really fun; I absolutely love playing with their dogs

archer is a really great person, a cross-disciplinary professional I admire very much, good at thinking and reflecting, sincere and attentive with people

Because of an experience like this, friendship is sustained not by frequency, but by the two points on this line of friendship, and their starting point … that 1

Meeting archer on a Night Hike: a Desk Lamp Lit the Start of a Friendship

⚠️ This MEMO is a copy of the one generated by the sync conflict

————

In 2023, I met archer for the first time, in Heyuan City

That day I tried night hiking for the first time, with no experience, a total newbie; I hurriedly signed up for a small group ahead of time

I had just started my senior year that year, and archer was still working hard; we got to know each other

Because that night I hadn’t brought a headlamp, but archer was very thoughtful — he had a lamp and lit my way the whole time

And our friendship just kept going

archer’s family keeps two little dogs, really fun; I absolutely love playing with their dogs

archer is a really great person, a cross-disciplinary professional I admire very much, good at thinking and reflecting, sincere and attentive with people

Because of an experience like this, friendship is sustained not by frequency, but by the two points on this line of friendship, and their starting point … that 1

A Chengdu Trip and Western Sichuan Memories: a Crossroads Two Years Ago

The last time I went to Chengdu was two years ago, December 21, 2023

The first time I traveled through western Sichuan, and also the first time I climbed two snow mountains

At that time I seemed to be carrying some expectations, because graduation was still a while away, but I also had an offer in hand, and I had thought about starting a business too; the AI wave had appeared not long before, and many sensitive people realized this opportunity and began to stir

The Past Flows Like Water; People Come and Go in a Hurry

The past is like water, people come and go

People leave and become the past, events sink to the bottom of the heart

The past is hurried, no more than one journey of meeting in this world

First Trip Abroad to Kuala Lumpur: Curiosity and Awe

Just noting this down. I remember February 25, 2024, that was my first time going abroad; it should have been the first half of my senior year. It suddenly hit me that it actually hasn’t been long since I went abroad — my first time abroad was two years ago, when I was still a senior.

I remember feeling very stirred up then, full of passion, flying for the first time to a foreign land, in a place where both the language and the culture were unfamiliar, everything seemed especially interesting, I was full of curiosity, feeling that so this is what cities outside are like.

At that time my first stop was Kuala Lumpur. The weather in Kuala Lumpur is always unpredictable. That city is very new, the clouds in the sky are beautiful, puffs of white clouds drifting in the sky, and at night you can still see the sunset; it’s very different from cities on the mainland of China, which are hazy — I really like the cities over there.

At that time I went with my laptop; I felt there were many remote workers over there, and they seemed to have some digital nomads too, working remotely over there; I thought it was really cool, and I instantly fell in love with that state, feeling that it was really good for everyone to live like that.

There was also all kinds of food over there I’d never eaten in China; I remember one, a premium garden chicken rice. Kuala Lumpur has Chinatown, and Chinatown has a lot of Chinese food; I felt they made it very carefully, and brought a lot of China’s traditional foods over there too, really interesting.

At that time it was also Chinese New Year, and there was still some leftover festive atmosphere over there; every day I wanted to go out and look around, wandering everywhere (city walk).

The Time Mismatch Between Past Good Experiences and Present Sharing Ability

Actually, looking back at so many good experiences in my past, I feel quite regretful that I didn’t share them well back then.

Because the me back then probably wasn’t worthy of those good experiences back then. The me back then wasn’t so calm, wasn’t so sincere, and wasn’t so self-strengthening.

Now that I have the ability to share, and this behavior can be self-consistent, I no longer have the good experiences I had back then

Working at a Shanghai Foreign Firm and A Deeper Crisis of Subjectivity

I remember last year’s working in Shanghai, a lot of happy times

Then in the graduation season I went to work at a Japanese company

What was interesting every day was that the colleagues around me all went to lunch in groups, and after eating they always wanted to go look at cars and test-drive cars

Because interestingly, the age structure at a foreign company is relatively diverse; the company I was at in Shenzhen felt still rather young, but here there were even people close to retirement, and the youngest besides me was a guy five years out of school, so every time after work you couldn’t find anyone; people left right on time — ah, this kind of workday, day after day, is just …

I feel the problem I face is more not a problem at the material level; my material needs can be very low, I’m very sure of that. It isn’t just a kind of emotional confusion, but a deeper crisis of subjectivity. I want to be myself, but who am I? I haven’t found it yet, but it’s definitely not who I am right now; the self I’m engaged in is lost, feeling utterly meaningless, sacrificing my life for certain things…

When you don’t know what you want, it’s okay — try thinking about whether these things you have right now are what you need, and what you don’t want, and just avoid that

Subjectivity Is Scarce Under the Education System, and Gets Allocated

In today’s education system centered on exams and standardized evaluation, subjectivity is indeed a scarce resource, and it is more easily “allocated” to that small top tier of students

Subjectivity is being very clear about one’s own behavior, being very consistent between one’s own cognition and one’s own actions

But most students in reality are more like executing an external script …

Unverified Theory and Method Are as Illusory as a Mirage

If you can’t even verify for yourself that it works

then this set of theory and method guidelines in this world is, as it were, a mirage

River Civilizations Around 2000 BC and the Fusion of Power

Imagine 2000 BC to 1400 BC

At that time the speed of information transmission ≈ the speed of a horse

So the peak of civilization must have centered on great rivers and grain: whoever could stably produce grain could feed armies, run sacrifices, and build cities

So leaders in the agricultural age could often govern floods

In China, the Central Plains were the transition period from Xia → Shang

Bronze casting was already extremely mature; sacrifice, kingship, and military force were three in one

Shang bronzes were at the level of technical luxury goods at the time

Which means: China’s advantage lies in the fact that “military power + religious power” were bound very tightly

If you assess the “degree of fusion of violence and divine right,” Shang is very high

Sumer and Babylon: Political Structure and a Management Upgrade

Sumerian civilization

The Sumerians invented cities, writing, and the temple economy, but their political structure was very fragile:

many city-states, much internal strife, strong divine power, weak royal power

Like a group of technical geniuses who never invented the “corporate system”

Then Babylon entered the stage, bringing with it solutions to that era’s problems: a management upgrade

You can think of Hammurabi as a very early “state product manager.”

His core problem was:

when cities multiply, occupations multiply, transactions multiply,

if you still rely on “whoever has the bigger fist speaks,”

this system will collapse.

Farmers need to know: what if a noble seizes my land?

Merchants need to know: what if someone borrows money and doesn’t repay?

Craftsmen need to know: if the house collapses, whose responsibility is it?

Priests need to know: how does the authority of god align with the authority of the king?

So the Code of Hammurabi appeared

What matters is not “an eye for an eye”

but this thing itself:

for the first time, law went from “the whisper of god” to “carved on stone for everyone to see”

And so for the first time in human history there appeared a set of public rules

If a doctor’s surgery fails, cut off his hand

If a house is built and collapses and kills someone, you pay with your life

If a loan is not repaid, how it is handled

Civilization progressed

How the Sumerians Created Cities and Writing from Primitive Materials

What did the Sumerians want to leave behind

There were no natural barriers here, no stone, no forest, only river water, mud, and sun. So the Sumerians used the most primitive materials to make the most un-primitive things: cities

They discovered the city; there were city-states, city walls, temples, warehouses, markets. This is a discovery rather than an invention, because if you observe the trend of the human species in other countries, they too spontaneously organize into cities

They invented writing

At the earliest it wasn’t for writing poetry, but for keeping accounts:

“how much wheat came into the warehouse today,” “whose slave is this”

This is the starting point of cuneiform

Once a civilization can write things down, it possesses an external brain

Memory no longer relies on the elders, but on clay tablets

They also discovered theocracy: every city-state had its own patron god, and the city’s legitimacy came from the temple

But Sumer’s fatal problem was exactly here: they were a city-state civilization, not a state civilization

Which also led to fights among themselves. They were very smart, but they did not organize a larger political body, so they were taken over by peoples who were better at “integration”

The temple served as the largest warehouse

Grain, sheep, beer, dates, cloth were all gathered here.

Then the temple redistributed by status:

rations for craftsmen

rations for soldiers

rations for those who repaired the canals

rations for priests

It solved their division-of-labor problem

It fits the Darwinian process — gathering is energy-saving, division of labor is irreversible, and outsourcing memory is inevitable …

The Profound Influence of Sumerian Civilization on Western Civilization

To understand Western civilization, you must first understand Sumer

There is still debate today: Sumer, whose origins are unknown, and why the technology of Sumer’s Stone Age advanced so rapidly

The civilizational legacy left behind:

the foundation of West Asian civilization, culture and institutions

writing, law, mathematics, astronomy, and the urban model influenced the entire ancient world

Modern time and angle measurement still use its sexagesimal system

Sumer invented humanity’s earliest writing — cuneiform (around 3400 BC), used to record economy, law, religion, and literature. Without Sumer’s practice of “writing is power,” there would be no later Greek philosophical texts or Roman codification of law

Without the world’s earliest urban civilization in Sumer, there would be no later Greek city-states, the vehicle of freedom

Without Sumer’s mythologized narratives, there would be no Western monotheistic narrative framework to explain the world and settle the heart

The Regulatory Mechanisms of the Non-Coding Regions of DNA

How does DNA actually control genes?

Actually it doesn’t actively control genes, but as the storage carrier of genetic information, it uses a series of precise regulatory mechanisms to determine which genes are expressed when, where, and to what degree (that is, transcribed into RNA, and then translated into protein). Together these mechanisms constitute “gene expression regulation”

About 98% of the DNA in the human genome does not code for proteins, but it contains a large number of regulatory elements; and the vast majority of genetic variants (including mutations) related to disease or traits fall precisely in these non-coding regulatory regions. The coding regions determine “what protein to make,” while the regulatory regions determine “when to make it, where to make it, and how much to make”

So because of a small mutation at one position, it may affect a gene as far away as 500,000 “letters”

AlphaGenome solved both of the above pain points at once this time:

It can both “see far” and shoot macro: it can swallow 1 million DNA letters in one go, and the prediction accuracy can still be refined down to each individual letter

From “a specialist with a lopsided profile” to “an all-rounder”: gene expression, splicing, chromatin state, protein binding — these complex biological processes can now all be handled simultaneously by this one model

The essence of learning is changing yourself

So many successful cases

whether it’s wanting to learn from them, or raising your awareness, or building confidence …

but the more essential thing is still to make yourself change

If you can’t even do this step

then naturally the changes in the outside world have nothing to do with you

Chongqing is a proving ground for a complex world

Chongqing is a natural proving ground for a complex real world

complex city

complex road conditions

complex traffic

complex inter-floor and level relationships

A cold, ruthless self-assessment

Having a ruthless and calm positioning

A very clear understanding of yourself and of this world

You could also call it common sense

Let users perceive the change

You have to make users aware that there’s a change

Only then can users feel the change properly

Every added feature dilutes the core value

Every feature you add dilutes everything else

IM tools are the best vehicle for conversation

For the chat form, the best vehicle and habit is consumer IM tools

Reaching goals naturally, through goal orientation

Because you want to achieve a certain something

so naturally that’s how you go about doing it

Around this kind of approach and capability

you naturally reach the goal

Don’t invent scenarios, just build what users need

Don’t invent scenarios

build what users need

and do it as well as you can

The core of a good model is avoiding dead loops

The best model isn’t one that “never makes mistakes”, but one that keeps finding alternative paths, avoiding dead loops or giving up

Evaluation is where taste lands, and a moat

Evaluation is where taste lands

It’s a very good systemic capability

Knowing how to build taste, and knowing how to evaluate

Do you value “useful” more, or “interesting”

“reliable” or “stunning”

“explain it clearly” or “give the answer first”

“say less nonsense (low hallucination)” or “cover more scenarios (high recall)”

“user’s first-try success rate” or “exploratory conversational experience”

Most of these preferences can’t be articulated clearly or written out completely, and they often conflict. So in the end one thing happens:

Your real taste gets expressed through the objective function you allow your team to optimize

That is, quantification, customer evaluation, some evaluation system …

iteration … positive feedback

Evaluation is a moat: how you test, what you test, how realistically you test …

What’s really hard to surpass about DeepSeek is its team culture, its engineering system and its testing system

Taste needs to land, inside an engineering evaluation system…

Evaluation comes from the team’s values — a more complete answer, more user dwell time, and so on ….

AI’s “good” is usually a multi-objective conflict: accuracy, speed, cost, politeness, robustness, safety, controllability, creativity……

So a good evaluation isn’t just a score; it’s more a weighting system + a red-line mechanism

Jottings

We’re just playing a game well, that’s all

The life philosophy of maxing out experience

Max out life

Max out the experience of life

Clear self-positioning and awareness of your strengths

What matters is being clear about what you want to do

Clear about your strengths, what you can use, your ability to handle this world better

The uncontrollability and emergence that make technology captivating

I think the real appeal of science and technology should be their uncontrollability

emergence

randomness

Turing completeness

positive feedback

aesthetics that move people

No need to fear cost or future bottlenecks

Don’t be afraid of cost

Don’t be afraid of future bottlenecks

Reassessing how important I am

I don’t seem to be that important

I’m just a participant in some matter

The thing I want to do is that important

The three elements of fandom and looking up together

The feeling of an idol

Fandom groups generally have three things:

A shared language — no need to explain much, no need to start a self-introduction from “who am I”; you just say “I like them too” and the other person immediately gets it. Communication cost is zero. There’s a certain certainty between the two of you

Shared emotional legitimacy — many places or environments hint that you shouldn’t get too excited, shouldn’t be too invested or too serious, but in fandom circles, hot tears can be understood, excitement can be caught, and you don’t have to hide yourself

Waiting for a moment together, doing one small thing together (voting on charts, watching stages, saving photos, cheering), makes people very close. Because what you share isn’t just content, it’s time. Shared time is the hard currency of intimacy

Someone who believes with you, gets excited with you, lifts life up a little together with you

Looking up together …

Envy is actually a mapping between the outside world and our inner selves; in reverse it lets us understand what we want

Actually I can feel it in photography circles, hiking circles, at live band shows, in open-source communities, book clubs, even morning running groups, even among a group of people seriously making products …

Extending the concept from community to fan club

It can be called a community, or with today’s concept, a fan club

The meaning of concerts as a ritualized climax

Shared language, emotional legitimacy, shared time and small things — a concert can be seen as a “ritualized climax”. It isn’t a continuation of everyday fandom but the peak convergence point of the whole process. Fandom is often scattered and digital (saving photos, voting on charts, discussing online), whereas a concert turns these fragmented experiences into a highly concentrated, embodied shared moment. Its role is similar to “pilgrimage” or a festival in religious ritual: strengthening group cohesion, amplifying the legitimacy of emotion, and through a synchronized experience of time and space producing a kind of “collective effervescence” — to borrow sociologist Émile Durkheim’s term, this is the excitement and sense of belonging produced in group interaction

Shared time, and amplified — similar to the collective excitement at the moment of New Year’s Eve. The sense of space gives us memory: you’re not alone getting excited in front of a screen, you’re screaming, waving and crying together with thousands of people

A concert provides an “emotional amplification field”: the idol’s stage lighting, musical rhythm, interactive segments (encores, fan cheering) work like catalysts, amplifying personal emotional investment and letting fans feel the freedom of “not having to hide”. At the same time, it creates a “peak experience” — the state psychologist Abraham Maslow described — bringing a brief sense of self-transcendence, where at that moment fans feel fused with the idol and with the group

The force this mechanism exerts on people is twofold: short term it provides emotional fuel and connection; long term it builds identity and resilience. It uses humans’ social instinct (tribalism) to embed personal pursuits into a group narrative, thus amplifying motivation. But to be sustainable it needs a balance between climax and everyday — otherwise, like the emptiness after fandom, it can turn back and eat the passion

In the end what we’re after is the meaning of being alive, experience

How much truth matters, and choosing the dream

Is truth really that important?

Does it really matter that much whether you’re in a dream?

Being able to do what you want at any time, choose the life you want

whether it’s based on understanding yourself, or understanding your relationship with this world …

The Inception top and the insight of letting go of proof

Super interesting, Inception

At the end that top spins on the table — before the shot cuts away it “seems” about to wobble, but you’re never given a clear answer

What Cobb wants isn’t “the truth”, it’s “going home”, seeing his children, hearing their voices. Aren’t many people using their own little tops — “as long as I earn X I’ll be at peace”, “as long as they reply to me I’ll know I’m loved”, “as long as the numbers go up I’ll know the direction is right”

At some moments, fixating on proof keeps you from ever going home; daring to let go of proof is what actually lets you enter life

Dominant people are better as listeners than as conversation leaders

I feel one important thing about conversation is that dominant people are usually better as listeners than as leaders. Because dominant people need some sense of initiative — for example, remembering to nod, responding proactively, or steering topics to encourage the speaker to keep going. That’s very important.

So I think conversation is a skill — a really formidable skill.

The only thing tech worries about is bubbles and premium

The only problem tech worries about is bubbles & premium …

Authenticity is a beautiful quality

Authenticity is a beautiful quality

The certainty difference between big and small companies

Big companies solve problems of certainty

Small companies solve problems of uncertainty

Thoughts are atomic

Thoughts pass quickly

Collective interest and individual sacrifice

This society increasingly makes some individuals sacrifice themselves for so-called collective interest

Maybe one day, the one being sacrificed starts to be yourself …

Especially now that the wealth gap keeps widening

Not resisting is winning

Don’t resist

Don’t resist, and you’ve already won

Build products that push human progress

Building something you use yourself

and that helps future human progress

feels like a very valuable thing to do

The philosophy of mind grants spiritual sovereignty

Mind is principle

Knowing and mind as one

Extending innate knowing

No external authority

Weld knowing and doing together

Give ordinary people spiritual sovereignty

#格物/阳 明

Helping people see their own value clearly

Actually, what most people fear isn’t uncertainty, and they’re not desperately craving an answer either. What most people need is just to see themselves clearly, to know what they want, to sort themselves out once. That’s good too — really great.

Asking what a bubble essentially is

What is the essence of a bubble

Go to the hottest fields to compete for attention

To play this game well, you first have to go to the hottest fields

to capture the most attention

Tech stocks are pulled by the winners

Tech stocks generally don’t rely on the average — of course the average has value too — but more importantly they rely on a tiny number of winners to pull the portfolio up

Verifying information beats deep research

You don’t have to deeply research every piece of information

In fact, much of the time what you need to do is verify some information’s reliability

and then just make a decision

Expression is a precious human treasure

There are some very precious assets of being human

that are getting more and more important now

For example, speaking …

Expression …

The time curve of a technology’s impact

We often overestimate a technology’s short-term impact, but underestimate its long-term impact

After the bubble bursts, the real technology settles down and changes the world

A mindset of life as a game

What you yourself find valuable

What you yourself like

Things you yourself find meaningful

You’re playing a game

Don’t let this game tie you down

Google’s full-stack technical strength

Google as a company is really outrageous

including its current hardware capability, TPU silicon

and the capability of its current models, the powerful Gemini

plus the rapidly expanding GCP infrastructure

a very strong fan and user base …

Bias comes from sample contamination, social-narrative pretraining and self-esteem

The origin of bias

The brain strengthens and absorbs bias-related strategies

Working backward to bias’s essence: early sample contamination, slapping a label on a certain kind of person or a certain group

Then there’s social-narrative pretraining — online information has already been injected into part of the public model, like who is a success, what failure looks like; we reinforce that part

Then self-esteem — very instinctive; it rejects things that trigger a sense of morality, denial, mockery

How to manage bias: one is to notice your own bias and correct it; another is to downgrade opinions into hypotheses and actively look for uncomfortable information — that information usually needs understanding built around it

Some good model strategies — Bayesian, replacing moral language with probabilistic language

Staying up late plus a mild cold thickens throat and airway secretions

Went to Guangzhou yesterday; staying up late and a mild cold may have played a part too

causing thick secretions in the throat or airway

Smog irritation caused mild airway inflammation plus some thick phlegm

Thick phlegm is protective, which makes sense

Extremely sincere people have very high consistency, inside and out

I saw a line from the psychologist Carl Rogers: extremely sincere people have very high consistency. I understand this consistency as having two levels of meaning:

Inner consistency: their “self-concept” is consistent with their “actual experience.” If they think they’re a good person, then in life they really rarely produce uncontrollable malice; if they feel fragile, they’ll admit they’re fragile, rather than forcing themselves to keep up a “strong person” persona

Outer consistency: their “inner feelings” are consistent with their “outward expression.” Whatever they think, they say (built, of course, on respect and responsibility, not on blurting out whatever comes)

Thinking of Sister Yanyan’s sincerity & transparency

Getting in touch with her, she feels like someone who’s honest with herself and honest with others

She rejects fake pleasantries, does what she says, dares to expose her soft spots, and has very high emotional transparency, so her energy feels great and many people want to be around her

To some extent she’s subtly influenced me …

A society devoted to one cause demands individual sacrifice, producing lots of unfairness

Sometimes I feel a lot about this. I’m thinking about one question: this world’s sacrifice of ordinary people, especially in a society devoted to a single cause — it especially emphasizes sacrificing the individual rather than fulfilling the individual. In such a society, I think there will be a lot of unfairness happening to some particular individuals. But this sacrificed individual, while fulfilling the masses, feels like a great pity to me.

If the masses were all ordinary people it’d be fine, but what if it involves someone very important, or someone with a lot of background? Then it brings some struggle, which I find quite frightening.

Private small groups sustain genuine strongly-connected social networks

Private small groups

A very vital, alive form

Not platform-level communities, but private small groups of 3-12 people

They’ll become fewer, more expensive, more serious, and also crueler

What small groups sustain is a real social network

Members usually have real-world ties (colleagues, classmates, interest circles, neighborhood neighbors, etc.)

Topics go deep: from small talk to joint decisions (gatherings, collaboration, pushing a project forward)

Feedback is fast and coherent: messages aren’t noise, they’re a chain of dialogue

Interpersonal relationships begin to divide into strong ties and weak ties; small groups are strong ties

Socializing is evolving from natural dispersion → purposeful focus → deep cultivation of relationships

But small groups also feel easy to die off, unless there’s often some external relevance

For example, a fixed weekly offline/online activity

A monthly output task

Or small groups with shared loss costs

Or the group chat is just the entrance, and private chats may be the real battlefield

It depends on whether there’s a middleman or medium connecting them together

The underlying logic of admiration and relational imbalance

Friends — on admiration

The underlying logic of admiration and bubbles

The other person we see is generally the part they’re willing to show, plus the part you’re most easily drawn to, plus the part you’re most lacking right now — so a bubble forms:

Entrepreneurs admire “serial winners”

People in a lost phase admire “people who look extremely certain”

People in an emotional trough admire “calm, strong, stable people”

Long-term admiration feels very bad for the health of a relationship, because over time admiration causes the relationship to become unbalanced, hardening a status gap, and this structure makes both sides uncomfortable

Then the two people stop having real conversations, they don’t dare to push back, they don’t dare to expose themselves

Then the bubble bursts — a person will always reveal hesitation, mistakes, emotion and interest calculations; once you find “he’s just an ordinary person too,” admiration flips directly into disappointment, even hostility

So building an objective, rational understanding of the people around you is important

Appreciate, but don’t deify

Learn, but don’t depend

Respect, but don’t belittle

Be able to cooperate, but also be able to part ways

China Is a Country That Embraces Change and Welcomes the Future

First of all, China is a country that can really embrace change and embrace the future

Over the past twenty-five years China has changed enormously,

and over the coming twenty-five years, China is also ready to welcome enormous change

As Virtual Experience Spreads, Real Experience Becomes Scarcer

When virtual experiences become more and more easily available, real experiences will instead become scarcer and more precious because of it

The good thing about the mirror world is that everyone can have all kinds of novel experiences from the real world without leaving home, while real adventure in the real world becomes instead a unique service that only a few can enjoy

Thinking about a kind of experience in the future

Needing only one eye to satisfy daily tech needs, including live streaming, and a way to automatically edit exciting video fragments

Hard to Find a Clear Place for Myself in This Era

They are all pushing forward, all moving ahead

But I really seem to have such a hard time finding a clear position of my own …

Privacy and Personalized Service: The Reciprocity of Rights and Duties

Are you willing to live as a “transparent person” in a monitored world? I think most people’s answer is “no.” But put another way: if you want to enjoy customized service and have a personalized AI assistant, you must give up most of your privacy — are you willing? Many people would hesitate quite a bit

Actually the core of the problem is how to achieve a reciprocity of rights and obligations in the collection of information

Accepting Your Own Limitations Is the Prerequisite for Digging into Your Strengths

Accept your own limitations

This is a very important prerequisite

Because you’re about to start being open, and start deeply digging into your own strengths

The Herd Effect of Imitating the Rich, and the Trend

What the rich are doing

Their behavior will be imitated by ordinary people

The herd effect,

to a certain extent, is also a possible trend of the future

Two Kinds of People: Money but No Time, Time but No Money

We usually divide people into two kinds:

people who have money but lack time, and people who have time but lack money

These two kinds of people are, in the traditional sense, the rich and the poor

Reality and Digital Twin Overlaid in the Mirror World

The definition of the mirror world

When billions of people living in urban areas put on these smart glasses, what they see is the real world overlaid with a virtual world. Some people call this virtual world the metaverse, some call it AR (augmented reality), or even XR (extended reality). I call it the “mirror world,” because what you see is both the real world and a digital twin of the real world overlaid on it

Of course, personally I think sound may be a smoother way to intrude; there are some experiential differences between the effects of sound and those of vision

A Faith-Driven Belief That Doing Everything Well Makes Life Smooth

Faith: do everything well, and life is smooth sailing

The Faith Mechanism of Reciting Scriptures with a Prayer Wheel

Prayer wheel

Usually a whole scroll of scripture is rolled up inside, most commonly the six-syllable mantra:

Om · Ma · Ni · Pad · Me · Hum

Turning it once = the scripture is recited completely once

No need for the voice, no need for understanding; it relies on direction and continuity

The Feeling of Liking and of Doing Whatever You Want Within Free Choice

What kind of feeling is it when something is clearly liked? What kind of feeling is liking the outdoors? Give yourself enough freedom, then what kind of feeling is liking? Give him enough of a choice, and so a person can do whatever he pleases.

System Discipline and the Human Capacity to Interrupt

The system isn’t necessarily right

Individuals have souls, and can consciously judge and recognize their own behavior

The reason a person is a person is not ability, but the capacity to interrupt

In a High-Quality Short Drama, the Feeling of Not Being Ready for Life

I watched a short drama, “Midsummer Fendela”

It is a high-quality short drama

There is a large amount of time in life that is not used for making decisions, but for feeling the state of not yet being ready

The Social Bubble and the Value of Selectively Settling Down

Suddenly a very interesting question came to mind: my dad always hopes I can fit into all kinds of circles — the ones he comes into contact with, the people he admires who move smoothly and easily in all kinds of circles.

But I think this is interesting too: this kind of socializing actually has a bubble to it. For me, I’m not very keen on mixing in all kinds of circles, because I’m very clear about these positions; there are many things I can’t reach, and I can only pick some things I’m interested in, and then dig deep into them.

Also, mixing in circles really goes against my own values. I’m someone very much influenced by the environment; many times the environment subtly influences my role, influences my every action, my every value, even my every thought, my every idea.

So what I’m more keen on is finding suitable circles and suitable environments to settle myself in, rather than being trapped in any circle — I think that would be meaningless.

So I don’t know what my dad’s most essential idea is; maybe it’s for face, maybe for image, maybe for some calculating need of his own, but as far as I’m concerned, I really have no interest.

When Lost, Live Life a Bit Better First; Discipline and Ease Are the Answer

When you’re lost, when you don’t know what to do, when you have no direction

You might as well live your life a bit better first, a bit more disciplined, a bit more relaxed

Sitting by a Fire vs. the Heating: Ritual Awakens Biological Memory

A question of intuition

In winter, is it more comfortable to turn on the heating at home, or to sit by a fire at home?

Body comfortable → heating

Heart comfortable → a fire

Fire is primitive, partial, local. You have to go over to it, adjust your posture, turn your hands over. It isn’t taking care of your whole body; it’s entering into a relationship with you. Flames flickering, firewood crackling, red glow on your face — that is a kind of safety written into tens of thousands of years of human memory. What’s comfortable about a fire isn’t the temperature, it’s the sense of participation and ritual

People naturally like to sit around the fire and talk at night; this is a sense of safety engraved in biological memory

Nighttime, chatting, spacing out, staring blankly → very suitable for a fire

Suddenly I think of the older generation, who don’t like technology and instead prefer a little bit of sparks

If a winter goes by without ever spacing out around a fire, then that winter is a bit wasted

Higher Emotional Frequency at Home: Reflecting on Generational Anxiety

Daily emotions are few, but at home the frequency of emotion increases; I feel there’s a problem with this

Speaking of Mom, people of their generation believe the world is unstable, uncontrollable, and resources can be cut off at any time

Worrying seems to be a kind of survival strategy

So there are some things that look very helpless to me

In my eyes they are extremely low-probability, extremely high-cost anxieties, yet they appear in Mom’s heart

They seem more afraid of losing?

Their nervous systems, strengthened by a lifetime of training, may have become fixed this way

If I Believe I Can, Then I Definitely Can, Beyond Environmental Limits

If I feel I can, then I definitely can

This kind of can can go beyond the environment

Body Temperature Is a Kind of Thinking; the People Around Me Are Good at Asking Questions

Body temperature is a kind of thinking, but very few people really do this

Observing the characteristics of the group of people around me

They are a group of people who are good at asking questions

You Don’t Have It Because You Don’t Want It Enough; You Have It Because You Can’t Live Without It

The only reason that you don’t have what you want is because you didn’t really want it. The only reason you don’t get the things you want is that you actually don’t want them.

The only reason that you have the thing that you do is because you couldn’t live without it. The only reason you have the things you have is that you cannot live without them.

And, the only reason that you are where you are is because somewhere within, it is OK to be there. The only reason you are where you are is that deep down inside you feel being here is acceptable

My Parents Are Addicted to the ByteDance Trio; the Wheel Turns Around

Seeing my parents now addicted to the ByteDance trio

I remember being addicted to my phone when I was a kid

The wheel turns around

The Forbidden City: a Synonym for Safety and Stability That Costs Freedom

The Forbidden City gives me a feeling I don’t like so much. The Forbidden City is very grand and dignified; it is the leftover dream of the prosperity of a previous era. I think the Forbidden City belongs to the imperial relatives and nobles of the Ming and Qing dynasties; their breathing over more than two hundred years has by now become a dream of more than 600 years.

I don’t know how many high officials and nobles were selected to come in, and I don’t know how many palace maids were selected to come in. “Once you enter the palace gate, it’s as deep as the sea”; their whole lives were enclosed by these high walls.

So I think the Forbidden City is a very good synonym: it represents safety, security, and stability, but it also makes people lose a lot, such as freedom. Palace maids who entered the palace could solve their basic food and clothing problems, could escape poverty, and could also subsidize their families, but they gave their whole lives to inside these red walls.

Combining my own experience, as a modern person, I feel very moved by this.

The Theocratic Character of Ba-Shu Culture and Exaggerated Bronze Depictions

Ba-Shu culture was, in the end, a theocratic culture too

So a large number of the bronzes are depictions of the images of gods

The most notable feature of which is exaggeration

Bilibili and Douyin Are Essentially Consumables and Entertainment Products

Bilibili and Douyin are essentially consumables and entertainment products

Their purposes, their most essential needs, are the same

The delayed afterglow of Chiang Mai memories

But I opened the photo album from two years ago

And only afterwards, in hindsight, felt

Wow, the afterglow of Chiang Mai is really strong

A-shares sector rotation and the “three strikes” rule

The general rule for A-shares (RMB ordinary shares) is “things don’t go past three” — a sector can hardly rise three years in a row. So we have to study industries with a dynamic eye, and keep digging for more profitable sectors

Logical reasoning vs. game thinking

Logical-reasoning mode vs. game-theoretic thinking

The world is stable, the rules are fixed, the participants are passive

What that mode focuses on is the causal cycles in the world

This is very helpful for understanding the world

But once you enter social systems, technological change, entrepreneurship, platform competition, the premise starts to collapse

Participants change their behavior because of your existence

So you yourself are a variable; you’re a link in the game of this process; you’re part of what drives this ecosystem forward

Game thinking isn’t “competition”; it’s a more fundamental worldview:

The world is a system made of actors who perceive each other and react to each other

Outcomes aren’t derived from conditions, they’re “evolved” jointly by the participants

For example, the flourishing of agents and thinking driven by DeepSeek open-sourcing V3 — without DeepSeek, people probably wouldn’t think seriously about open source

The world is like a fluid system, and you’re not a bystander but a stone that keeps dropping into the water

You enter the water → the current changes

The current changes → other people’s routes change

Other people’s routes change → a new structure appears

The premise is that you’re not standing on the bank watching

The essential logic of Hong Kong stocks and its capital rules

The essential logic of Hong Kong stocks

Letting global capital buy Chinese assets with low friction

Highly free: no protection for retail investors, no backstop, no market rescue, no price limits, shorting allowed

Highly rule-of-law based: clear rules, you bear your own responsibility, disclosure first

Dual-class shares, unprofitable companies can list (Zhipu), and listing isn’t necessarily safe

Hong Kong doesn’t have a scarcity of listing status; shell value is extremely low, and backdoor listings carry no premium

How the Die-Yet? APP taps into social emotional resonance

The Die-Yet? APP

This APP seems really good at tapping into social emotional resonance, with a controversial name and an extremely minimal product logic

But what I’m curious about is how it got promoted and became popular

It turns a heavy, taboo yet very real fear (“if I die, no one would even know”) into a concrete, low-threshold solution. This “life-saving” attribute gives it natural virality on social media

The idea for this app wasn’t invented out of thin air; it came from a hot topic online — “What apps does everyone need but nobody has built yet?”

The APP’s functionality is extremely simple, but it solves a core anxiety, so users are willing to pay for it or download it

The neutral definition of speculation and its boundaries

Speculator

Sort of like this — you buy/sell an asset mainly to make money from price fluctuations, not from the cash flow/value it creates long term

Speculation and investment aren’t a binary opposition but a continuous spectrum; the same person can do both in different positions at different times

Speculation is a neutral word; the market also needs speculators to provide liquidity and price discovery

Speculation partly involves a wager, but with strategy and risk control it can also be very professional

Gambling is when neither win rate nor edge is clear, and it mainly relies on luck

Cutting corners is going through the back door, exploiting loopholes, taking advantage

Signals for judging a market sentiment pullback

Judging a sentiment pullback

The market is no longer excited by gains

At this point, even if good news appears, stocks won’t rally much

Buying an index admits you’re no smarter than the system; picking stocks is zero-sum

Buying an index is buying the average rate of return — proving you’re no smarter than the system, as long as you get the part of it that grows over the long run

Buying stocks is a zero-sum game of cognition against other people, especially for short-term investing: you think you’re smarter than the market, at least in some small slice

The incentive structure the financial system gives most people runs opposite to individual stock trading. For most people in reality, they have no first-hand information, they have their own emotional swings, and they can’t watch the market in real time

When tech surpasses human nature, whoever can afford tech harvests ordinary people

It feels like rich people will get richer, because they can use technology. Technology is extremely rational; it can surpass people, it can beat human nature.

So in the future, in an era where tech broadly surpasses people, it feels like the startups that can afford technology will be able to harvest more ordinary people. That’s really cruel.

Mean reversion is the only physical inertia in financial markets

Let’s talk about the mean, and first about the market. Respecting the market means respecting market laws

The market isn’t atoms; it’s people, expectations, fear, institutions, games — so any inertia is only a statistical tendency, not a necessity

Mean reversion fits inertia very well

Corporate profits, household income, valuation levels — none can deviate infinitely from long-term productivity

Then there’s the fact that excess returns attract capital inflows, which creates competition, which brings returns down

Then there are institutional constraints — interest rates, regulation, taxes, monetary policy — all of which fundamentally suppress extreme states

But it can also produce positive feedback, which physics calls an unstable equilibrium — in the market: bull markets, bubbles, stampedes

If mean reversion is “gravity”

then trend is the “rocket booster”

Tech sits on the control layer while traditional industries face a chronic squeeze

Tech is a career track that stands on the control layer by default

The rewards traditional industries give ordinary people are falling systematically

The labor dividend is over, competition is now transparent, profits get siphoned off by platforms and capital, and the middle layer is flattened

What’s happening in traditional industries isn’t stability, it’s a chronic squeeze

Justin Sun: from crypto-world controversial figure to narrating human civilization

His Mars narrative feels very interesting to me

He’s very good at grabbing global attention

Mars is a natural medium, letting him get global media exposure and attention in both crypto and tech circles

The early Justin Sun was a controversial figure in crypto / a traffic-driven entrepreneur

From speculator to futurist

From a crypto-world figure to a participant in the narrative of human civilization

From spokesperson for a specific project to teller of grand visions

The crypto industry has a long-standing problem:

it lacks a future narrative grand enough to keep inspiring ordinary people

Being discussed ≈ sense of existence ≈ influence

The next consensus, the next world’s focus, the next world’s attention

Intangible cognition matters more than tangible resources, because consensus is the foundation of value, and the earliest consensus is the biggest source of value

The charm of ahead-of-its-time cognition — ahead-of-its-time cognition must turn into action and conversion

When we take in any information, ask more often whether it’s a new blueprint for building the future. This is our attention — what information is our brain processing every day: is it the past, is it finance, celebrities, chicken soup, or is it a cognitive system for building the future

If you want to make a leap, you must package cognition into an asset — spreadable, tradeable, financializable

Products get mocked, but the doubt itself is necessary fuel

Many products get cursed and mocked at the start

But setting jealousy aside, if you look with sympathetic joy, study whether there’s something worth learning, what states of society it reflects

You’ll find it interesting — the recent app 死了吗, when it first came out it was also mocked and looked down on by many people

But that kind of doubt is itself necessary fuel

Knowledge and skills shouldn’t fence off the game’s non-derivable charm

We learn all the knowledge and skills

but don’t let them constrain us from playing this game well

instead, boldly experience and feel the process of this game

Things that can’t be derived and can’t be learned are extremely charming

The essence of asset creation is making wealth, not making objects

The theory of asset creation

Ordinary manufacturing makes things; the purpose of asset creation is making wealth — turning resources, rights or cash flows into tradeable financial wealth

What is asset creation?

From the traditional view, building a machine is product manufacturing; from the capital view, packaging that machine’s future output into a buyable, sellable certificate — that’s called asset creation

Asset Creation/Origination means using legal title confirmation, financial structuring and credit enhancement to convert raw resources, rights, technology or expectations into financial instruments that are tradeable, priceable and capable of producing continuous cash flow

Product manufacturing solves consumption demand (you buy shoes to wear them)

Asset creation solves investment demand (you buy the shoe company’s stock to gain value)

The starting point of an asset is a right: real estate holds land-use rights, credit holds mortgage rights, IP and technology hold patent application rights / intellectual products

Then the cash flow is structured. An asset is valuable because it can generate money

Value consensus and the greater-fool theory

The future belongs to the young; but is something valuable just because future people are willing to pay for it?

From a sociological angle, or from the angle of price

value is indeed recognized by future people

Value isn’t a natural property; value is a consensus property

Gold, the dollar, Bitcoin, hype sneakers — in essence they’re all worthless; they’re just passed down and agreed upon generation after generation

Young people have time, cultural discourse power, and future purchasing power

The old generation doesn’t understand → the new generation accepts → it becomes mainstream → the price rises

But there’s a precondition: someone being willing to buy doesn’t equal there continuously being someone willing to take it off your hands

A bubble happens not because “nobody believes,” but because the rhythm of believing breaks

For value to hold, it must satisfy at least any one of these conditions:

It keeps producing real utility — it genuinely feels great

It can embed into institutions or infrastructure — so the cost of exiting is very high

It can keep attracting the first wave of buyers from each new generation

If any one of them breaks:

the technology narrative goes bankrupt

aesthetic migration

policy interruption

the new generation turns to another faith

then value evaporates instantly, rather than slowly depreciating

The future does belong to the young, but only things that each next generation can repeatedly re-understand, re-use, and re-narrate have value that crosses time

How Visa sets global payment rules

Visa as the global payment standard

How was it legitimately decided?

A kind of industry-consensus technical standard

Visa is a private global clearing network

Visa and Mastercard are a card-network duopoly that competes on the surface, colludes underneath, and keeps its rules aligned

What they really do is three things:

Set transaction rules

Handle transaction routing and clearing

Define risk and liability attribution

The trust economy compresses the cost of uncertainty

The trust economy

What it replaces is a high-friction world

In the traditional economy, any cooperation requires paying a huge “cost of guarding against bad people”

Contracts, lawyers, guarantees, vetting, deposits, layer after layer of approval, relationship endorsements

None of these “create value”; they prevent value from being destroyed

The essence of the trust economy is using systematized trust to compress the cost of uncertainty

Effective output = nominal output × trust coefficient

Track record — what you’ve done matters a hundred times more than what you say

Verifiability — third parties, systems and algorithms can verify that you didn’t lie

The cost of breaking trust

Incentive alignment

Trust has now become the most expensive asset; in an era of information surplus, attention has been squeezed dry

The more information there is, the scarcer trustworthy information becomes

Pricing ability and the trust discount rate

Why the same ability gets priced 5-10 times differently

Many people’s salaries work the same way

In the relationship between business owners and employees, personal pricing is ability × trust discount rate

What most people get stuck on isn’t ability, it’s that their trust gets discounted

So high prices in the real world almost all come from one thing: certain delivery

Which is what intermediary platforms do

When your trust level is high enough, one hidden but important thing happens: low-quality clients avoid you on their own

Time arbitrage and asset liquidity

The charm of time arbitrage

The ultimate magic of asset creation is “overdrawing the future.”

Through discounting models, asset creators package decades of future returns (like 30 years of mortgage interest, 50 years of highway tolls) and cash them out today in one go

Water is flexible, so turning some hard resources into soft, light assets lets you capture a liquidity premium

The essence of a bubble is expectations of expectations

Bubbles feel hard to judge or catch

The essence of a bubble

Price rises depend mainly on expectations of expectations, not on real cash flow, productivity, or irreplaceable value growth

Every asset should have a real-world reference point: cash flow, profit, rent, use value, replacement cost

Hong Kong IPO subscriptions and consumer opportunities

Most bullish on 2026

Hong Kong IPO subscriptions are super interesting

A-shares are hard to win; Hong Kong has margin financing and so on, so IPO subscriptions are easy to get

The first half is the most worth taking risks in

More bullish on Hong Kong stocks; Hong Kong stocks are part of the middle buffer

Fairly bullish on consumption

The international endorsement of an A+H share structure

The A + H share structure

A part I’m very bullish on

The same company listing A shares in the mainland while also listing H shares in Hong Kong

is usually for financing diversification

and to hedge against corresponding valuation issues

Especially important is the international endorsement

Companies with this structure tend to be more favored in the China-US rivalry

Insurance capital and the national team’s stabilization mission

Insurance capital

Insurance capital is insurance funds

Its source is the premiums of hundreds of millions of people

Every life, annuity, pension, critical-illness policy you buy — every premium becomes part of a giant capital pool. That pool is insurance capital

It has a few innate personality settings:

Extremely long duration

A very plain goal

Risk aversion

The national capital team is the capital force that stabilizes the market for the state

The National Social Security Fund

Central Huijin Investment Ltd.

China Investment Corporation

Their main goal is maintaining system stability

They’re generally not responsible for pulling the market up to its highest; they’re responsible for one key thing — not letting the market fall into the abyss

Dividends are the portfolio’s stabilizer

We all talk about dividends

Hot industries, dividend-paying industries

generally have stable cash flow

Dividends are the stabilizer in a portfolio

The index logic of the Nasdaq 100

Nasdaq 100, NDX, Nasdaq-100 index

Founded in 1985, composed of the 100 largest non-financial companies listed on the Nasdaq exchange

Seen as the representative of tech and innovation growth stocks, often called the “new money” or the “AI engine”

The Dow usually has 30, leaning toward industrials, finance, energy, consumer goods and healthcare

Nasdaq 100: uses market-cap weighting → the bigger the company’s market cap, the bigger its influence on the index. More scientific, more mainstream

Semiconductors and the AI chip ecosystem

Semiconductors are the physical cornerstone of AI chips

and AI chips are the semiconductor industry’s current top technological driver

Semiconductors (The Ecosystem): this is a grand concept, referring to a class of materials (like silicon, gallium nitride) and the entire electronic components industry built on them. It includes memory chips, sensors, power devices, general-purpose processors (CPU), etc.

AI chips are just a series of designed integrated circuits

Winner-take-all effect: the traditional semiconductor market is relatively dispersed, but in AI chips, because of the extremely high barriers in design (Design) and manufacturing process (Foundry), value is highly concentrated in a few companies (like Nvidia designing, TSMC manufacturing)

As physical limits approach, improving performance simply by shrinking transistors (say from 5nm to 3nm) becomes harder and more expensive. AI chips’ extreme hunger for performance has forced the semiconductor industry to light up the Advanced Packaging tech tree

Also, copper’s recent surge is actually a structural shortage, driven by three engines: AI, new energy, and grid upgrades

Data centers are made of copper

Asset narratives and faith as fuel

Judging a bubble

isn’t essentially about prices being high; price rises depend mainly on “expectations of expectations,” not on growth in real cash flow, productivity or irreplaceable value

The beam holding up the price has switched from reality to story

Every asset should have a real-world reference point: cash flow, profit, rent, use value, replacement cost

But when prices exceed these anchors quickly, persistently and significantly, and the gap can only be explained by “the future will be different,” that’s not a conviction, but it’s already a yellow light

People who never cared about the asset suddenly flood in

Professional judgment gets mocked as “not understanding the new era”

There’s actually one more point, about narrative

Healthy asset logic is stable

but the narrative in a bubble keeps getting patched

When the old logic is falsified by reality, a new, grander story immediately takes its place and no longer accepts rebuttal — which means the price now needs “faith as fuel”

When the price rise itself starts changing people’s behavior, business decisions, even life paths (quitting jobs, borrowing money, going all in), the asset has gone from “reflecting the world” to “hijacking the world”

Once the price stops, reality bites back, and it’s rarely gentle

Three dimensions for judging a bubble

The market can stay irrational longer than you can stay solvent

Judging a bubble usually requires cross-verifying from three dimensions: “cold” valuation data, “hot” market sentiment, and a “hard” leverage environment

The core is whether the price significantly and persistently deviates from its intrinsic value

The Buffett Indicator, i.e. “total stock market cap / GDP”

Historically, 70%-80% is considered undervalued, around 100% is reasonable, and above 120%-150% is usually seen as significantly overvalued

Then there’s the Shiller PE / CAPE — the inflation-adjusted average PE of the past 10 years. Compared with ordinary PE, it better smooths out short-term earnings swings

The price-to-rent ratio too: if it takes 50-60 years of rent to break even, it means the price is supported mainly by “appreciation expectations” rather than use value

Hot,, people who normally don’t care about finance start discussing what stocks or funds to buy, and recommending tickers to you — that’s a clear signal …. Then there’s “this time is different,” the five most expensive words in financial history, the flomo mindset

Displacement: a new technology or new narrative is born (like AI, the internet, blockchain), attracting early investors.

Boom: prices start to rise, the media starts reporting, more money comes in.

Euphoria: prices go straight up, valuation logic breaks down, everyone speculates, “only a fool wouldn’t buy.”

Profit Taking: smart money quietly leaves, and the price starts oscillating at a high level.

Panic: some black swan event punctures the bubble, the price plunges off a cliff, and buyers disappear

Lessons from the oil crisis

The 2008 oil crisis

It tripled in just one year

Fear that the world was running out of oil; panic and greed coexisting

The most popular theory at the time was “peak oil.” Geologists and analysts were all saying that the easily extractable oil on Earth had been used up, and from then on oil would only get scarcer and more expensive

The market logic at the time was: China’s and India’s industrialization had just begun, billions of people wanted to buy cars and use electricity. This demand was long-term and rigid — no matter how high the oil price, they would pay

The non-linear mechanism of a bubble bursting

There’s a very interesting question

How do bubbles burst

It’s always realized after the fact, and it usually can’t be analyzed linearly; the world is non-linear

The Minsky moment: during a boom, asset prices keep rising, investors become overly optimistic and start borrowing (adding leverage) to buy assets; long-term speculation pushes debt levels so high that the cash flow the assets generate can’t cover the debt interest, and the market enters an extremely fragile Ponzi finance stage

At that point, even if the asset price merely “stops rising” (it doesn’t even need to fall), speculators are forced to sell assets because they can’t pay interest

That’s when collapse comes: many people are forced to sell assets because they can’t pay interest, and the forced selling drives prices down, entering a feedback loop

The collapse of the psychological line of defense, from FOMO to FUD, is an obvious characteristic

Three signals of an asset bubble

Valuation deviation: has the asset price significantly detached from fundamentals (e.g. PE ratio, price-to-income ratio at historical extremes)?

Leverage ratio: are market participants using a lot of leverage? (Is the margin balance high?)

Sentiment indicator: when the non-professionals around you (like taxi drivers, or friends who don’t care about finance) all start fervently discussing and recommending some asset, that’s usually a late-stage bubble signal

Don’t judge top billionaires’ life choices with ordinary logic

Using your own life experience to imagine a class that isn’t yours at all — you’re bound to get it wrong

Don’t judge top billionaires’ life choices with an ordinary person’s anxiety, morality or logic

They have their own world rules

Some people do nothing at all and live extremely well on capital alone

An ordinary person grinding away at a startup — involution — looks to them like high-risk behavior instead

So it seems I’m not really in a position to give others advice, nor capable of judging others’ choices

In finance, the closer a factor is to the source of money, the more it matters

In the financial world, the closer a factor is to the source of money, the more important it is. A few examples:

Central banks directly control money supply and borrowing costs

Employment and labor market data actually reflect the current vitality of the market economy, as well as market information

Inflation indicators determine the path of interest rates

Economic growth indicators reflect overall economic health and affect corporate earnings expectations and investment sentiment

The Hang Seng is cap-weighted and heavily concentrated in super-heavyweights

The Hang Seng Index is generally cap-weighted, and its weights are highly concentrated

In Hong Kong stocks, individual stocks usually move the index, but the “individual stocks” here means the super-heavyweight stocks

In the Hang Seng, the top 5-10 stocks hold 40%-60% of the weight year after year

Tencent Holdings

Alibaba

Meituan

China Mobile

HSBC Holdings

The Hong Kong stock structure isn’t the all-the-people-trading type

Local retail participation is low; institutions, foreign capital and passive funds make up a high share

But there are exceptions: when sentiment is extreme, the index moves individual stocks

Infinite supply, limited demand: tech is infinite but demand is layered

“AI makes the supply of ‘cognitive labor / software functions’ nearly infinite, with marginal cost approaching 0, but demand in the real world is limited (time, budget, attention, scenarios are limited). Under this structure of ‘infinite supply, capped demand,’ what happens to the economy and the market?”

Supply is infinite, demand comes first

Price P gets pushed by competition down toward marginal cost (many AI capabilities will approach free / extremely cheap);

Quantity Q will rise, but not infinitely — capped by ceilings of time, budget, attention and scenarios

So this is the essence of the paradox: from the supply side technology is “infinite,” from the demand side it’s still very “limited” and “layered”

AI deflates the market for time-premium professionals

In an industry that sells time and sells labor, the arrival of AI is like a lot of free labor showing up. The result isn’t that everyone makes money — it’s that the part of the market that used to earn because “time was scarce” gets flattened directly by price deflation

Instead, inflation happened

The TAM gets eaten up by deflation — TAM here can be understood as how much money this industry could theoretically make

AI compresses “an hour of output” into “a few minutes”

Clients are no longer willing to pay a high price for “time”

Those who lose the most are the middle layer, “time-premium professionals”

while the junior ones were already cheap to begin with

and the top ones are the people who define the problem, own the problem, and carry the results

This redefines the logic of the business world

Forward PE: fair price = PE multiple × earnings per share

Forward PE (forward price-to-earnings) calculation

Fair stock price ≈ fair PE (price-to-earnings multiple) × fair EPS (earnings per share)

It’s usually 16-27x, meaning the current stock price is 16-17 times the next twelve months’ earnings per share

Suppose a company is expected to have earnings per share (Forward EPS) of 5 yuan over the next year, and its Forward PE is 16.5x, then: the fair stock price is 5 × 16.5 = 82.5 yuan

For mature industries, generally consumer and energy, 16-17x is a medium-high but still acceptable valuation

For high-growth industries (like tech, AI, biomedicine), this level may count as low or reasonable

Compared with the S&P 500’s overall Forward PE of about 23x (data as of end-2025), 16-17x looks relatively cheap

Tencent is generally given 18-22x

because Tencent isn’t a pure growth stock, nor a pure value stock, but a “super platform” with very strong cash flow and a deep moat, whose growth is no longer explosive

Next-12M EPS is the expected earnings per share over the next twelve months

About next-12M EPS

It refers to earnings per share over the next twelve months

How to calculate it: say it’s January 2026 now — how much is it expected to earn over the next twelve months

Next-12M might be: part of 2025Q1 + Q2 + Q3 + Q4

Three months later it becomes: 2025Q2 + Q3 + Q4 + 2026Q1

How is it calculated?

Sell-side analysts + financial report updates + company guidance + consensus expectations → consensus EPS (of course you don’t have to trust third-party platforms; you can work it out by hand yourself)

Next-12M EPS ≈ most recent full-year EPS × (1 + next year’s earnings growth rate)

Next-12M EPS ≈ most recent full-year EPS × (1 + next year’s earnings growth rate)

Example (say, using Tencent):

Trailing 12-month EPS (TTM EPS): ≈ HK$29

You judge net profit can still grow +8%-12% over the next year

One-way car return costs and platform matchmaking fees

I keep thinking — when returning a car, isn’t the cost of returning it in a different city very high

But in reality it’s very counterintuitive

Some that don’t support one-way returns do so because they’re low-price cars, small local fleets, and on non-popular routes, where gross margin is fairly low, and one one-way return easily eats up the profit

Platforms love one-way returns, because the platform takes a matchmaking fee; with an extra one-way return fee, the platform’s transaction volume is higher

eHi is an exception, much like a platform’s own rental company

eHi does nationwide vehicle dispatching

Analyzed from first principles: the car itself is high-value, high-risk, and the service is highly dependent on offline operations, so it avoids haggling between branches, liability issues, and complaint issues

Technology Should Free People from Instrumentality and Return Them to Experience

I’m thinking about how new tech products keep appearing in the future

This kind of product needs to have some real friction in reality

And actually technology is what lets people escape instrumentality

I wonder whether this means there is a very clear direction, which is the direction of using new technology and of people’s own experience

Earphones Are the Lowest-Friction Form of Wearable Intelligence

The role of earphones hasn’t been fully uncovered yet

Earphones + hearing are the lowest-friction intelligent structure for humans

They don’t occupy your hands, don’t occupy your eyes, don’t interrupt your actions, and can be connected at any time

Even compared with eyes, because eyes are really hard to wear all day without getting tired, and without being awkward, and they don’t display things at random, and they don’t intrude on others

Watches and bands are the intelligence layer on the body

Whether AI truly cares about you in the future

depends on whether it is plugged into your physiological data stream

A ring is a very sensual form, but its ceiling is very clear: extremely low presence, extremely high wear rate

The phone will go back behind the scenes …

I Hope the World Has More Smart People, Not Just Winners and Losers

Everyone is very capable and very smart, and I think that’s a good thing

I hope this world has more smart people

Even if many of them are my competitors, even if in the short term the world is zero-sum

But still, this world isn’t only about winning and losing

I can learn faster and run farther

Only when I can see a world like this is it worth living seriously

Going Public, Securitization, Bondization: Financial Rights Made Standard and Liquid

Going public, securitization, bondization

Differences at the level of engineering methods

But the essential purpose is the same

To turn “untradeable real-world relationships”

into “financial rights that can be priced, can circulate, can be divided, and can be controlled”

Going public turns a participatory relationship into shares

For example, I participate in this company’s growth, I bear the risk, but I also enjoy the residual; dividing participation into standardized shares lets strangers come in and trade

Bondization turns a promise relationship into a contract: originally the relationship is that I lend you money and you pay me back when due; bondization writes the promise down hard and turns future cash flow into a piece of paper that can be bought and sold

Securitization packs up messy futures; the real-world relationships are

mortgages

student loans

credit cards

rent

royalty income

packing a pile of scattered, low-liquidity future cash flows

into packages → tranches → standardization → selling to the market

Finance itself doesn’t create real value

What it creates is a rearrangement of control rights and option rights

Once you take things out of the lived world and into the game world, there begins to be leverage, expectations, panic, models

The Variety of Trading Units in Hong Kong Stocks and the Odd Lot Market

Trading units for Hong Kong stocks

I always thought the trading unit started at 100, but it turns out not

Many large-cap stocks and well-known companies use this specification; it looks the most like A-shares

Also very common is 500 shares per lot and 1,000 shares per lot, especially for lower-priced stocks

There are also a smaller number of odd specifications, such as 200 shares, 400 shares, or 2,000 shares per lot, depending on how the company defines it

There are even ones where you can buy 1 share — those are called odd lots, and have to be traded in the odd lot market

minimax starts at 20 shares, which is like changing the “entry ticket” from business class to economy class, friendlier to retail investors and making trading more active

The Hang Seng Index as a Reference for Market Trend and Sentiment

The reference value of the Hang Seng Index

The Hang Seng Index generally selects the 50 largest and most liquid stocks in the Hong Kong stock market as a weighted index, reflecting the overall trend of the market

It also partly represents market sentiment: when the index rises, market sentiment leans optimistic; when the index falls, risk may be relatively high

When the index fluctuates a lot, some individual stocks may be dragged down even if their fundamentals are fine, and the risk of short-term speculation increases

Some good tech companies with strong profitability, many users, and ample cash simply have better fundamentals. By contrast, new companies may have immature products and little revenue, so their fundamentals are weaker

Investment Characteristics of the Tracker Fund and Hang Seng Index Funds

Tracker Fund

It tracks the Hang Seng Index; buying it is equivalent to buying all the blue-chip stocks in the Hang Seng Index (weighted by market cap)

The risk is relatively spread out; it suits people who want to follow the market’s rise and fall, don’t want to research individual stocks, invest long-term, and pursue the market’s average return

The Hang Seng China Enterprises Index Fund tracks the Hang Seng China Enterprises Index, also called the “H-shares index,” which includes Chinese state-owned enterprises listed in Hong Kong; its risk and volatility are usually greater than the Hang Seng Index, and returns may be higher but short-term fluctuations are also larger

Traveling the world is losing its appeal, and a growth reflection

Traveling the world seems to be gradually losing its appeal

When I was in Xi’an I planned to apply for a Schengen visa and a New Zealand tourist visa; the materials were pretty much ready, I even booked a consulate appointment, but I kept putting off any action

Actually I can feel it — at this stage I don’t have the same urge as before to go see the world objectively, or maybe the acceleration has started to decline. Even though the speed of growth is still increasing, it’s not explosive anymore, the slope is starting to get smaller …

Last year’s main line was understanding the world, understanding myself

The former is hoping for as many choices as possible, or to give other people more choices

The latter is hoping that after understanding myself I can handle my own inner world more harmoniously, reach self-consistency, or reach a reconciliation between myself and the world

Traveling the world is of course very good; an experiential life seems to be a kind of political correctness, and the ease of eating, drinking and enjoying yourself seems to be some kind of standard for being free-spirited

Make yourself a bit smaller, place yourself a bit lower

A bit softer

And then see a bigger world

The aesthetics and cultural meaning of Chinese stone pillar carving

Chinese stone pillars

including patterns like those on roof tiles, statues

Their reason for existing isn’t only functional; there’s also aesthetics, and existential meaning

The stone pillar is carved almost 360 degrees all over, the dragon’s body coils upward around the pillar, there’s no “blank face”

Visually it weakens the static feel of the “pillar” and strengthens “growth, coiling, rising”

It feels vividly alive

The dragon isn’t coiled randomly; it rises consistently from bottom to top, counterclockwise/clockwise

In Chinese culture the dragon has its own unique existential value and meaning

The dragon’s proportions are exaggerated: a big head, strong claws, the body pressed tight against the pillar

The information density is very high; Eastern ancient architecture doesn’t have the usual fear of blank space — it’s the opposite of blank-space anxiety

Building a personal brand and connecting with readers

Building a personal brand

What’s core for readers is forming a connection with the author through the writing

Having substance is the foundation, clear logic is the precondition

Sincerity is the core of connection

No one believes a know-it-all newcomer, but people appreciate someone who grows steadily

Create a personal blog or WeChat official account and update it often, run social media accounts sharing bits of your writing, build an email newsletter to stay in touch with subscribers, join offline salons or livestreams to interact with readers, and so on

Accumulate instances of your own work, acquire some necessary titles …

The core of Schopenhauer’s will and the human predicament

Schopenhauer’s chapter on the will

I never quite understood the will

For Schopenhauer, the will isn’t a person’s subjective desire; it’s a blind force driving all existence to endlessly “continue, expand, contend, maintain itself”

The core is: blind, ceaseless, no endpoint, indifferent to happiness

The will isn’t “what I want”; before you even think that question, the thing is already pushing you to keep going, to want more, to want to live, to want to continue

All things want

But objective laws don’t shift according to human will

People always feel a sense of lack, desire never stops, the world never stabilizes

It precedes reason, precedes consciousness, driving all things to endlessly want, continue, expand

The human problem is that we’re not only driven by it, we’re clearly aware of it — and that’s part of human suffering

Three capacities can help humans

Aesthetics: the will temporarily becomes “watching”

Compassion: the will no longer only grasps “me”

Reducing desire: not satisfying it, but not letting it keep accelerating

Awareness of the will operating, and your relationship to it

will → instinct / impulse → emotion / feeling → intuition → reason

Becoming aware of the will doesn’t make it disappear; you change the relationship between you and it

You are a vessel watching the will operate

Subjectivity and a life you define yourself

Subjectivity

You could call it extreme self-consistency

or extreme self-centeredness

A selfhood that can be defined, yet can’t be defined

Life’s boundaries are there to be experienced and broken; dare to explore and forge your own life

Someone with weak subjectivity is easily swept along by mainstream social values, becoming anxious and following the crowd

But if a person clearly understands themselves and the world, first they’ll want to experience the world, define the rules, and accomplish themselves

what is right, what is good, and then choose and execute firmly, not caring about outside voices

Low-frequency scenarios are the death spot for SaaS products

Low frequency is SaaS’s death spot

It determines users’ usage habits

A lot of product features try to retain users through some dopamine and rewards

But with products, the user’s usage time really is very fragmented

Enduring enchantment in an age of disenchantment

I feel like a lot of things aren’t interesting anymore

A lot of things are explicable

science, culture, technology, institutions ..

even human nature

All of these can be disenchanted, understood …

But some things are endless — maybe that’s where we truly feel the enchantment in this era

art, creation, technology …

Downward comparison and a growth path through a sense of mission

My dad always wants to instill in me that comparing downward brings peace of mind, eases the pain

But from my own experience, what I feel more is sympathy and a sense of responsibility — can I do something, is there some sense of mission

Growth (I’m getting stronger), connection (I love and am loved), contribution (I make the world a little less painful)

Allowing yourself to be inefficient in intimate relationships

In the outside world: calm, clear, rational

But in intimate relationships, also allow yourself to be inefficient, incorrect, not on top

This kind of person isn’t “not quite there”; they’ve realized early what they want and don’t want

Accepting the world’s complexity and the essence of power

A few signals I cultivated last year while living abroad and thinking

Accepting and understanding the world’s complexity

Not in a hurry to prove myself

Being alone is nothing … being misunderstood is nothing …

Sympathetic joy, appreciating the other person’s excellence

I met a teacher who added one sentence I find deeply philosophical

Power comes from “being jointly acknowledged”

Violence comes from “no longer being able to be acknowledged”

Avoiding the victim’s view is how you play the real-world game well

To play this real-world game well, you should avoid letting yourself take the victim’s perspective

It’s essentially a product of culture too — because we depend on things, we take a side, the side of the victim: the media’s motive behind this news is to target me, this stock exists to harvest me

The news it puts out today is essentially meant to pull traffic and harvest us …

But does that mean you don’t take part? Or block it out? Then you seem to lose the experience of being a player yourself

Most systems in this world — media, capital markets, platform algorithms — aren’t out to hurt “you”; they’re out to complete their own objective function.

Traffic, attention, conversion rate, emotional swings, trading volume

You’re not the target; you’re just one unit of the quantity

I have no game rights, the rules are set by others, I can only take it passively

And the other side is evil by intent, so my failure is justified

These three combined mean you’ll keep getting hit over and over by the same mechanism

Once you explain the world with this narrative long-term, you’re forever waiting for someone who “should be held responsible” to show up

People who truly “play this game well” usually quietly complete a perspective switch:

not “what are they doing to me”

but “what does this system reward, and what does it punish”

Try to be a participant, a member in the game — the market is very interesting

Under this mechanism, where should I stand so I’m not easily consumed?

Switching from the victim’s view to the game-theoretic view is closer to reality

If you use the victim’s view, you’ll think:

“Is it time to leak news and harvest me again?”

If you use the game-theoretic view, you’ll be closer to reality:

“This is an emotional patch that’s friendly to the bulls but limited in force”

Emotional pumps are the market maker’s strategy pushed through news

Emotional pumping to drive prices up

The market maker pushes it through the news

In the attention economy information is abundant and attention becomes scarce hard currency

The attention economy

In an information-rich world, the abundance of information means the scarcity of something: what information consumes — the receiver’s attention

Then there’s the zero-sum game: attention is actually a non-renewable hard currency, more precious than time. We only have limited waking hours each day, and all the apps, media, work, family — we have our own ranking of values

In the new era, information isn’t that important — information is surplus, attention is scarce

Attention is free, and also the most expensive

Platforms (The Platforms), the so-called “Attention Merchants” (like TikTok, Meta, Google). Their goal is to maximize your LTV (Life Time Value) by extending your Time Spent

Advertisers (The Advertisers): the real customers. What they buy is a change in your behavior (i.e. you see the ad and then buy something, or change your perception)

Users (The Users): both raw material (providing data) and labor (freely producing and filtering content for the platform via likes and comments)

Blend into the system like water

Be like water

Understand the system, understand yourself

Blend into the system

Design the system, and it flows naturally

Don’t resist

Don’t resist reality

Don’t resist the world

Don’t resist yourself

Don’t hate the world; accept the world’s diversity

Choosing sides too early limits your possibilities

I think taking sides too early and caring too much about your identity is very bad for playing this game well

It’s often a beginner who better knows how to use the world’s resources

In today’s information-explosion era, users’ eyeballs, media slots, investors’ time, talent’s resume submissions — all are the core currency of the “attention economy”

You don’t need to burn huge ad budgets; one Moments post + riding a trending topic can get tens of millions of impressions

You don’t need to burn huge ad budgets; one Moments post + riding a trending topic can get tens of millions of impressions. Musk using tweets to move stock prices, Yu Chengdong using “far ahead” to farm presence — it’s the same playbook

In the 2026 Chinese tech context, capturing attention is itself important, even mandatory

The money machine of the attention economy

Right now, in an era that’s highly digitalized and information-overloaded

capturing attention really is the core engine of the money-making machine

The concept of the attention economy was proposed by economists back in the 1990s, but now it has evolved into a trillion-dollar industry

Attention is traffic, traffic is money

Platforms like TikTok, Instagram, X (formerly Twitter) fight every day over trillions of seconds of users’ attention, converting it into ad revenue, e-commerce conversion, or monetized influence. In 2025 the global digital ad market already exceeded $600 billion, and is expected to break $700 billion in 2026, driven mostly by the “eyeball economy.” If you can efficiently capture attention, you can go from ordinary person to billionaire (like Kylie Jenner building a cosmetics empire through Instagram influence), or double a company’s valuation (like Musk using tweets to move Tesla’s stock price)

Capturing attention now isn’t loud shouting anymore; it’s cleverly designing a system

Cognitive structure and the timing of understanding

I thought before — what if human thinking jumps in stages

that is, for the same information, like some poems learned early on, the Dao De Jing

the brain can only understand those “signals that appeared early” once it has enough structure

So there are some sounds, lyrics, sentences, viewpoints — but they had no hook to hang on, so they got archived in a numb, unfelt form

But truly important understanding can hardly ever be “too early” or “too late”

If you say you only understood love after losing it … you’re putting great importance on love’s result — but what if it’s love’s process?

The ability to act isn’t willpower; it’s a composite of cognition + emotion + relationships + risk tolerance

Suddenly I feel it deeply: a you who has today’s understanding but lives at a past point in time — that person was never born

When I chat I sometimes have a problem: I care too much about whether this view is right, whether it’s worth it, whether I myself endorse it …. It feels like resistance — like water …. Blend into this world, and then it’s understanding, tolerance, analysis, building systems

What does this person care about, such that they say it this way?

What they’re saying isn’t a viewpoint; what class of problem are they solving?

If this is wrong, under what conditions would it collapse

Allowing vagueness to remain is a form of respect for the other person, leaving a cognitive backdoor in your own language system

I don’t understand now, but this sentence may be very important

How to judge? Some body signals and intuition that are truly valuable

A little uncomfortable while listening

For a moment you want to argue back, but can’t say why

After hearing it you keep thinking about it, but can’t figure it out

Emotionally it feels stuck, not smooth

This information matters; it touches the layer of cognition you haven’t yet unfolded

Suffering is subjective; it’s a required growth course for everyone

But thinking about it from another angle, the definition of suffering is subjective. An ordinary person’s daily striving is a kind of “strength,” but the spiritual emptiness that a wealthy life brings is also a real “powerlessness” for the person living it. Shedding a tear for a flower may be her way of filling her inner world

It’s hard for us to truly empathize with someone else’s life, especially under a huge gap in wealth and status. Rather than judging whether her life is “smooth” or “not smooth,” it’s better to see it as a unique human sample. Her experience has conveniences we can’t imagine, and inevitably shackles we can’t feel

Suffering feels like a required course in growing up — for everyone, they’ll face suffering, they’ll be moved, they’ll put themselves in others’ shoes, they’ll feel it firsthand ..

Projective envy curses the self you can never become

Projective envy

What’s being cursed isn’t you, but that self I can never become ..

Attacking her produces a false sense of superiority, easing the envy and anxiety created by the huge gap

Social issues: observe, record, analyze, inject less emotion

Social issues, naturally, through the means of social observation

Just like the plain-description writing technique: record, analyze, pour in less of your own emotion, and so avoid generating bias

Because this process needs two things sustained: improving yourself, and understanding the world

Thoughts interpret the world; feelings are your body’s emotional reaction

People’s thoughts

and people’s feelings are different

Thoughts are your own explanation of what’s going on with this world

Feelings are your body’s and emotions’ reaction to this thing happening to me

Some more essential feelings

Feelings that don’t pass are a signal

Thoughts can be “wrong”; feelings themselves have no right or wrong

How you explain this feeling is a very deep layer — wrapping feeling with thought

You say: “I’m not happy with him, because he’s too selfish.”

Pulled apart, the order is usually:

1️⃣ The body feels uncomfortable first (feeling)

2️⃣ The brain immediately gives an explanation (thought)

3️⃣ You think that explanation is the feeling itself

But the real feeling may just be:

Being ignored

Not being respected

Uncertainty

Fear of losing control

Thoughts are a means of rationalization

Thoughts are the result of socialized training; feelings are the underlying system left by evolution

The strength is evolving continuously in long competition without being pushed

What exactly are my strengths

and which weaknesses aren’t well suited to my surviving in the future world

My strengths don’t feel like places where I’m better than others, but rather a combination of abilities where, in long-term investment, I don’t need to be pushed and can still keep evolving

Some of my weaknesses too — not a moral issue, but a configuration with low cost-performance in the future world

Using a mismatched self to chase what’s hot is chronic depletion; using a matched self to enter a trending track is compounding

Strengths are the part of you that, in a long-term competitive environment, gets others worn out but not you; the part of you that needs no external incentive; your default way of responding to complex problems; a way of recovering energy

The world is accelerating, while the individual’s answer is becoming more personalized — even the risk is an opportunity for a few …

Captive to self-consistency: external attribution kills the ability to correct

People held hostage by the feeling of self-consistency

These people find it very easy to be self-consistent

but they blame some mistakes on external causes and so lose the ability to correct course

This kind of self-consistency always attributes outward, and the problem this brings is that it’s very hard to correct yourself

It’s also very fast: the moment something happens, the brain supplies an explanation, with not even a hovering period, and it easily slaps on a label

What it brings is emotional stability, not an increase in ability

Sanctifying the Road Not Taken Behind a Parent’s Filter

What you like most about him isn’t what he really looks like; when you see what he really looks like, you’ll feel he’s very ordinary

I think of the disenchantment my friend talked about

Isn’t it the same with my parents — in their eyes there is always a longing for the things they sought but could not get

Sometimes I think this is also why it’s hard to communicate with them: they think they’re right, think what you’re doing is nothing much, and even think that the road you didn’t choose is sacred; universities I never attended they also consider incomparably impressive

But to me it seems this world doesn’t have so many filters. I think the real world is just like that, but my parents seem to live in a world that can only hold up by relying on filters

A good university, a good work unit, an urban hukou, a respectable identity, a stable path

When reality doesn’t give a person a pass, a person can only sanctify these things, otherwise the psyche will collapse

The roads they haven’t walked are imagined to be incomparably correct

The things they didn’t get are given the power of “if only I’d had it back then, everything would have been different”

The freedom they didn’t get to choose is packaged as “this is the only correct choice”

People easily fall in love with meaning — that kind of meaning which carries possibility, symbolism, scarcity, an unfinished feeling

Emotional Acceptance and Being Told Determine What Memories Stay

Self-frozen memories

Whether a person retains a memory generally depends on three things

Whether the emotion was allowed to exist

Whether the action was self-approved

Whether this experience was ever told

Failure leaves emotional residue, while success may instead leave nothing at all

A Doting Mother Ruins the Child: Sacrificial Parenting and the Conflict of Responsibility Boundaries

A doting mother ruins many a child

I feel like a ruined child myself

My mother has a people-pleasing personality

My mother covers for mistakes …

and adjusts all her behavior around the child’s feelings, for the child

and also uses her own sacrifice to buy obedience

This kind of relationship feels very uncomfortable; being around it clashes strongly with my own values

A person being responsible is itself: making a mistake -> bearing the consequences -> correcting the behavior

What I agree with is boundaries, responsibility, reciprocity, personal will

But she still lives in:

sacrifice = love, endurance = kindness, obedience = intimacy

How to Get Along with My Mother by Establishing a Sense of Boundaries

A very classic line about getting along with my mother

They are used to being controlled, used to accommodating

I need to take back my own right to act, and make sure my own things are handled by me

Enforce my own sense of boundaries: these things I’ll do myself, I must do them myself, without explaining reasons, and declare my sense of boundaries

No longer cooperating with her sacrifice script

Likewise, through reinforcement learning itself — how to build a healthy system

Reward correct behavior, punish incorrect behavior; it’s hard for parents and children to communicate rationally, and feigning emotion is a very good way

Accept the behavior, but add no emotional reward

She did a lot; you don’t need to thank her excessively, explain excessively, or compensate excessively

Closeness is not the same as clinging

Love is not the same as round-the-clock care

How the Environment Decisively Shapes a Person’s State

I’m someone quite influenced by the environment

Change the city, change the person, change the daily rhythm, and my state is quite different

In a good environment, growth is exponential; in a bad environment, the depletion is exponential too

So I also deeply understand that choosing and building your own environment is extremely important

The People-Pleasing Personality: A False Sense of Safety and Real Needs

The people-pleasing personality

Some people seem to like praising others too much

“As long as I make others comfortable, I am safe”

But in reality they may seem very easy to talk to on the surface, while their true state may be highly scanning other people’s emotions and being insensitive to their own

They are not truly desireless; they just don’t dare to take responsibility for their desires in relationships

Pleasing others can’t buy true being loved; it can only buy “being needed”

Parents Worry About Small Things: Generational Scales and Diluted Attention

Parents are always worrying about small things

Small things I don’t even think are that important

Sometimes it’s very helpless

I feel like they and I are like running on two completely different scales

What I care about is

whether the direction is right

whether life is self-consistent

whether I’m accumulating long-term ability

world structure, systemic risk, the space of possibilities

While what they care about is:

whether things are safe right now

whether anything went wrong today

whether they’ve been ignored

whether things “run like a family”

It feels very tiring, my attention gets diluted

For another example, they will care excessively about certain things, to the point of fighting over them, but I think enough space should be given, and a suitable system used to guarantee that things run naturally

Allowing them to care about their own small things is important, even trivial, nagging small things, and growing up often means no longer treating these forms as truth

Change Your Behavior to Advance Your Goals; Change Your Nature and It Follows Naturally

Change your behavior to make progress toward your goals

Change your nature, so that your behavior changes naturally along with it

Most people set a superficial goal, psych themselves up, keep themselves disciplined for the first few weeks, and then effortlessly return to the way they were, because they are trying to build a good life on a rotten foundation

If I don’t find it fun, then why am I doing what I’m doing?

If you want to complete a certain task, you need to create the lifestyle of that task, that is, a system

Only believe in action. Life happens at the level of events, not at the level of words. Believe in action!

Action changes a person, not thinking or ideas

Change Your Perspective to Set Goals; Action Reinforces Ideas, Not the Status Quo

If you say you want to quit that dead-end job, but you have no real reason to keep staying there, you may start to think you don’t have enough courage, or that you were never a “risk-taker,” but the truth is, what you’re pursuing is safety, predictability, and an excuse not to look like a failure in front of other people in your life who are also doing dead-end jobs

Is what we really change actually the status quo? No, it’s our goals

Change your perspective, set a goal; a goal is a vision of the future

It’s like a lens that lets you notice the information, ideas, and resources that help you achieve your goal

How people reinforce ideas is really terrifying, no matter through what means, no matter whether from yourself, your teacher, your parents, friends, advertising, or any other source — and, if you firmly believe the idea is correct, then its influence on you is like the hypnotist’s words on a hypnotized person

This is also how we become who we are, how we become tomorrow’s self, and this is also what constitutes our identity

We want to complete a goal

We want to seek a sense of identity to move forward

We will notice the important information and ideas for learning

Repeating the behavior until it becomes automatic, unconscious (conditioned reflex)

This behavior becomes part of your own cognition (I’m that kind of person)

You defend your identity to maintain psychological consistency

Your identity shapes new goals, thus starting the cycle again; if this identity is not conducive to a good life, the situation will deteriorate rapidly

So we all have a strong sense of identity

When our identity is threatened, all sorts of things happen

Life Goals Determine the Level of Thinking; Enjoying the Process Requires the Right Goal

What kind of life do you want

then you must exist in a corresponding specific mental level

People’s thinking develops over time through predictable stages

When you are born, you are like a little survival sponge, absorbing every belief you can get (these beliefs are largely influenced by your culture), in order to feel safe and secure

The only true measure of intelligence is whether you can achieve your life goals

Goals determine how we see the world

You can enjoy the process, provided the goal you pursued at the beginning is correct

For most people, these goals are imposed on them by others, preset in your subconscious like code

Life Needs a Directional Structure; Handle Things Lightly During Exploration

It doesn’t have to be a goal, but there definitely needs to be a directional structure

It’s almost impossible for a person to live long without a “directional structure”

Scenarios where goals are naturally needed:

Scarce resources: needing to survive, to get ashore, to escape poverty

Clear tasks: exams, projects, milestones

Limited time: the deadline itself is a goal-making machine

Goals are sometimes very useful, especially when the direction is unclear, when you need to stabilize action; action is especially important

During the exploration period, handle things lightly

Zeng Guofan’s Reflection and Adjustment Producing a Positive Cycle of Growth

I found that Zeng Guofan’s two big changes — first passing the imperial examination, and then, after entering the Hanlin Academy, wanting to become a sage — both happened because of his dissatisfaction with the status quo; he reflected deeply and examined himself, and in the end kept adjusting and found the method that suited him, and from then on was unstoppable.

This kind of reflection is very necessary; I think it is very necessary for every person. Because with reflection, there is a concrete goal; because there is a goal, a series of efforts are made for that goal, such as changing the environment, or taking a series of actions.

But this kind of environment and this kind of action will often subtly influence you, that is, they will change you, change your position, and give you a different sense of identity. Then this different sense of identity will push you to have new ideas and new goals, and then new actions, producing a positive cycle.

This is how a person, imperceptibly, grows and evolves.

An External Script Stripping Subjectivity Leads to a Limbo State

When an individual is inside a “script,” because the goal is given from outside (Extrinsic Motivation) rather than generated internally (Intrinsic Motivation), the brain’s reward mechanism — mainly the dopamine system — cannot be effectively activated. This state leads to what Ke describes as a “limbo state” (Limbo): a feeling of stagnation that is neither complete failure nor real success, in which the individual feels lost, anxious, and lacking motivation

The problem with a script lies in stripping a person of their subjectivity

A Warning About the Body When the Differences Between Environments Are Too Great

Sometimes you need to be cautious when changing environments

The difference between environments had better not be too great

When the environment changes differently, the difference in your personal habits definitely must not be too big, because the body needs to adapt

For example, in Guangdong I used to brush my teeth and wash my face with cold water; after coming back, if the water for brushing teeth isn’t water run from the gas water heater, it’s super icy, which triggers dentin sensitivity

And washing your face with hot water itself carries a certain strong stimulation, which is unfriendly to the skin

Searching for Self-Subjectivity and the Confusion of Reconstructing Identity

I still feel it isn’t game enough

How do I describe this feeling?

Because I always feel it’s a little short

I thought for a while, thought for a long time, searching for that missing piece

I haven’t found myself

The self relies on past narratives, on one’s own reconstruction of identity; what’s needed is one’s own subjectivity

I am searching for this part of subjectivity

Reinforcement Learning in Typing and Social Conversation

I type because I’ve reinforced my learning of the act and process of typing

To a degree, voice is the same

Those people who often communicate and converse through social means, their communication is smooth, and naturally they will also habitually use this mode to think and create efficiently

The field of cognitive science itself is about repeated practice forming muscle memory, timely feedback then adjusting strategy, goal-driven then incentivizing optimized behavior

People who often write code may lean more toward “modular + conditional judgment” style thinking

People who often debate will automatically build a “premise-conclusion-rebuttal” framework

While frequent social conversationalists are good at the “empathy-response-push forward” loop

A Way to Grow by Recording Life and the Process of Playing Games

It’s ultimately a way of recording life

Record well how you play the game well

Help people grow

Finding Your Own Position in the Information Age

In an age of noisy information

All kinds of short videos, AI, self-media, news, entertainment …

How do you find who you are?

First of all, be clear about what you want

You could call it a direction, you could call it a goal

The behavior, experiences, and thinking based on the goal — that is focus

And the experiences in turn add to yourself, making the self clearer and the goal clearer

The feeling of life while camping on an island

Let me talk about my camping experiences — one of the few times in my life I’ve camped. The best camping experiences were actually on some of Hong Kong’s islands. They’re cut off from the world, and you can feel that you’re in a very primal environment. Dense forest all around, facing the sea, no light pollution. Because there’s no city settlement for miles, even dozens of kilometers, around, it’s very quiet there, the sky is black, with just a faint bit of light far off, and from that you can tell that’s the direction of downtown Hong Kong.

Sometimes when I open the tent and the sea breeze blows in, watching the grass nearby sway, I get a very particular feeling. I think it’s alive — it’s a kind of life force calling out. In that moment, even though I’m alone, even though I’m lonely, I’m still in a jungle full of life, coexisting with it.

But in an office building, even with lots of people around you, even when everyone is inside a city dwelling with very heavy social constraints, you can feel everyone nearby moving, but with no soul. The walls are static, the city has cars coming and going, but the cars feel static too, none of them have life force, as if the little bit of life force they have is fenced in. That’s the kind of sensibility I want — a resonance at the level of the soul.

I really miss the time I camped on Phuket. Occasionally I dig those photos out and think, wow, it was so beautiful. Keeping company with fireflies, looking around, just a few tents, a little starlight leaking out of them. That night we all gathered together and ate some late-night snacks. One guy cooked rice, we also had some barbecue and such, and it felt especially fun. There was also a Hong Kong sister who brought a bag of KFC over from Hong Kong, which I found especially funny. Even though the food was cold, in that moment it was delicious. Because we naturally triggered a biological instinct, a craving for food. That craving is hard to feel in daily life; only when you’re short on material things, or drenched in sweat, or have pushed past the limits of your biological instincts, can you taste the primal flavor of food and feel that life is really good.

Designing systems and verifying order

Design a system

verify whether a system can run normally

whether your own understanding holds up

The joy of creating an order

code is a tool

a method for achieving some purpose

of course it can also be a way of thinking

that is, a series of methods used to solve a well-defined problem

A shift in the collaboration model of AI open-source projects

Useful open-source projects in AI get stars more easily

PRs on hot AI projects also grow explosively, and low-quality PRs flood in

A project maintainer’s attention is more valuable than the PR code

The open-source collaboration model ought to change

Scenario discovery beats scenario creation as a design philosophy

Many scenarios are discovered, not created

What you design might be a platform, a flexible system, the extreme of allowing individuals to naturally diverge, to allow emergence

the flexibility of the individual

The technical idealism of Google’s founders

Google’s two founders — Larry Page and Sergey Brin

I feel like they’re both systems engineers with technical idealism

They were self-consistent in the infrastructure era, but now the competition is brutal, the entry-point war era is fierce

Positive-feedback systems and long-term stability

How important positive-feedback systems are

If a system can’t make itself better, it’s just consuming the maintainer’s will

A machine has two fates

constant upkeep

or being scrapped

The most captivating part of a system is long-term stability and the ability to evolve naturally

This is also why mono no aware and wabi-sabi aesthetics endure

Aesthetic elements are enduring, cross-cultural, driven by the will yet also beyond the will

This is the only thing that can fight cognitive debt; systems tend to accumulate debt

Technical debt and cognitive debt

Cognitive debt accumulates with time — for instance, why shit-mountain code exists

Controlling your phone environment and self-awareness

For people who are good at using systemic capability

the phone environment is especially important too

and you can be aware of what you’re doing right now

For example the phone environment — controlling your own phone environment as much as possible matters a lot

and being clear that time spent online is all reasonable

Motivation needs immediate feedback and visualization

Human motivation needs immediate feedback to sustain itself.

Design a clear feedback mechanism for your system. I use the famous “don’t break the chain” method: mark in a planner every day you complete the system action. Watching the chain get longer is itself a powerful motivator

Actually flomo seems to work this way too, and Duolingo, and GitHub’s green squares — it’s fun and gives a sense of achievement

You can use an app to check in, or simply note it in a notebook. The key is making progress visible

The experience design of Xiaohongshu Diandian’s right-swipe

Xiaohongshu’s Diandian right-swipe notes really moved me

It turns an exit gesture into an opportunity for deeper exploration, maxing out the experience

It’s very intuitive, seamlessly connected. Originally you wanted to swipe back, and this gesture doesn’t interrupt your rhythm of browsing notes; it’s more like a hidden shortcut — when you need it, you slow down and take the chance to explore deeply

It lowers the barrier for me to use it, very intuitive

The information density is instantly maxed out too, for deeper exploration, or a summary

call-me: architecture analysis of a phone peripheral

call-me

the phone peripheral used with Claude Code

when Claude needs you, it steps in and calls you directly

The scenario is clear: you have Claude run a fairly long task (editing code, researching, running scripts) and you can walk away; when Claude finishes/gets stuck/needs you to make the call, your phone/watch/even landline rings, and you can have a multi-turn conversation with it on the phone, finish the decision, then let it get back to work

Architecture analysis

It exists as a Claude Code plugin

The plugin works via an MCP server

The MCP server uses ngrok to tunnel and receive webhooks from the phone service provider

The phone part goes through Telnyx or Twilio

Voice STT/TTS uses the OpenAI API

Environment mapping mechanisms and designing space by function

Your own environment mapping mechanism

Building an environment is important

Building your brain’s cognitive system’s mapping to the environment is important too

Cafe = a place to get work done

Home = rest, entertainment, but the cues for chores are too dense, so failing to start is normal

Plus at home the decision path and cost are actually shorter

In a cafe, spending money, going out, sitting down is itself a series of rituals

But at home, the bed, the sofa, the kitchen, the laundry, the phone charging cable… all summon you. The brain says: let me deal with these first, then study (and they never get finished)

If I design rooms in the future, I must be very clear about each room and its functional scenario

and start some rituals before working

HSBC red lion vs blue lion card differences

The red lion is an ATM card, not a savings card, and it has no CVV, so it can’t be used for online payments (Apple Pay is a special case). The blue lion is a debit card and can’t be used to deposit money at an ATM. Different purposes

The blue lion can be linked to Alipay and WeChat, and also gives fee-free withdrawals at HSBC branches worldwide

The value of bioengineering tools

Thinking of the bioengineering tools Xiaobaitu mentioned

some tools that serve people or companies in the bio industry

traceable, auditable, explainable, able to enter workflows

Some tools are clearly visible, can solve some people’s problems and raise their efficiency, and naturally there will be demand to pay for them

The simplification principle in brand naming

Considering changing the name — the current one is way too complicated, but changing it is too much trouble

Consider Chinese and English from the very start of naming

and being extremely simple and very easy to remember is super important

Experience isn’t insight; don’t over-indulge in your own life story

I’ve been thinking a lot about the relationship between a person’s experience and their cognition. When I answer this question, it’s like why some large models hit a certain bottleneck if they only chase data volume. But why do some small models, designed through a very clever model architecture, maybe with a tiny body size, or trained with some excellent training methods or algorithms, end up being smarter, or more intelligent?

From this I draw a conclusion: people aren’t built purely out of accumulated life experience either. When a person overemphasizes their life experience, it means that’s the only part they have to show. But people are very complex — it may relate to the education system from childhood on, and also to their level of cognition, their store of knowledge, and many other factors.

So I think when designing a model system, never over-indulge in certain of your own skills; you often need to step out of your own perspective and look at it comprehensively. For this model, which aspects is it better at? If algorithmic optimization can achieve good results, then do algorithmic optimization; if data volume can make a good breakthrough, then increase data volume; if training methods can achieve good results, then use training methods to improve the engineering optimization approach.

In short, there’s no need to insist on how important experience is — all that “I’ve eaten more salt than you’ve walked roads,” “I’ve crossed more bridges than you’ve walked roads” — it’s actually not that important. So don’t put too much weight on these things, and don’t over-indulge in certain things. Better to spend more time thinking about what you’re good at, what’s different about you, what you want, what you like.

Accept the world’s incompleteness and iterate slowly

The world is a giant playground

even a broken-down playground, with lots of loopholes, lots of order failing, goodwill arriving late, justice arriving late too

We patch it up — this is the human species’ oldest and most practical ability

Accept the world’s incompleteness; all I can do is slowly iterate and improve

Science is like this

Engineering is like this

Relationships are like this too

A person’s life is like this too

Productivity Is an Instrumental Metric, Not a Human Goal

Productivity is an instrumental metric

An instrumental metric is naturally designed for robots, not for people

Humans can take up work that doesn’t care about efficiency

Outlook for Implantable Chips and Brain-Computer Interfaces

Implantable chips may also see new breakthroughs in the next 25 years. Cochlear implants are already a very mature human brain implant; in the future implantable chips may also develop to be as mature as cochlear implants, and the whole implantation process will also become safer and more convenient

Compared with invasive brain-computer interfaces, non-invasive head-worn brain-computer interfaces may develop faster. For example, there are already some new technologies trying to use infrared light to read brainwaves. The user only needs to put on a special hat, and the hat can read brainwaves through infrared light passing through the skull

Scientists have now completed modeling an insect brain, and for the first time have a “map” of an insect brain. This is a huge breakthrough, but in the neurological sense we are still far from understanding the human brain. So what can we expect in the field of brain research 25 years from now?

The Magic of Judgment Beyond the Rules in Visual Design

Magic has never been about more, but about being just right

New designers easily fall into superstition about piling on material: more colors, more motion, more complex shapes

White space … visual rhythm

Magic generally comes from the designer’s judgment, not from specifications

Specifications can only take you to 80 points

Magic is the remaining 20 points, and it cannot be exhaustively listed by a checklist

The magic of visual design is, within the rules, quietly violating the rules a little, yet making the whole thing feel more real and more alive

The Dilemma of Individual Selfhood Under Definition by the System

System vs. individualism

One core question: when your abilities, behavior, and value are completely defined by the system, do you still possess a self?

“Am I only allowed to be the kind of person they say I am?”

Any complex system will tend to treat people as a variable rather than a constant

In an era when efficiency is infinitely magnified, there will be some nodes with great potential, with strong elasticity

Building a Workflow That Suits You, Grounded in Self-Understanding

Combining your own needs and your own understanding of yourself, building a workflow that suits you is very important

Singapore: a Top-Tier Country Built on Excellent Institutional Design

I really like Singapore; I think it’s a very interesting country. To be precise, its land area is very small, and its national managers are like the managers of a company, like a CEO.

I think this country is designed very exquisitely, because it proved one thing: even with insufficient original conditions and scarce resources, with good institutions and an operating system, it can still develop into a world-class country. So Singapore, with its own high-quality institutions, systematic structure, and systematic capabilities, designed such a country; every aspect surprises me in particular.

After I went there, my first impression was that this country is very new, the cityscape is brand new, and the whole city has a great deal of cultural landscape; I love it. Also, its Changi Airport is the most popular and largest airport in the world. In my view, even today’s Shenzhen hasn’t reached Singapore’s level, and China still has a lot of room to learn in this respect.

Obsidian as the Strongest Knowledge Management Tool in the AI Era

Obsidian really does feel like the most suitable knowledge management tool in the AI era

The most extensible, and it runs on a local model, with unlimited plugins; it can become almost any AI knowledge tool you want

Notion is very suitable for those who need remote collaboration, need repeated adjustments, or need to rely on certain templates; setting those scenarios aside, I wouldn’t consider using Notion again

Organizing My Personal Workflow Methods and Open-Sourcing Them to Help AI Learning

Organize all my own methods and open-source them

Build a workflow that suits me

And one that helps AI learning

Volatile sulfur compounds and the addiction mechanism

Volatile sulfur compounds

similar to the addiction to fish mint (zhe’ergen)

The reason I don’t like durian is that I’m hypersensitive to it — it’s unbearable, I want to throw up

Fish mint contains a class of volatile sulfur compounds (relatives of rotten eggs, blue cheese and durian), and this class has two traits:

It bypasses reason and hits the limbic system, skipping thalamic processing, going straight to the emotional control center

After repeated exposure there’s always neural relabeling — for some people, after eating it a few times they reinterpret the danger signal as familiar + safe + reward signal

Singapore’s education streaming as an engineering capability

Singapore’s education streaming

One thing that always puzzled me is how Singapore maintains elite education for everyone

Later I happened to realize this is part of its engineering capability

One thing I really admire: it genuinely isn’t designing a system to screen people

it’s designing a system to identify differences as early as possible, and send people of different abilities onto different life trajectories that can each run with dignity

trying not to let any layer fall completely out of the system

Elite ≠ privileged class; it’s a high-risk position — you need higher-intensity work, higher public responsibility, less room for error and more transparent performance scrutiny

Many societies say out loud that everyone is equal, but their social systems implicitly assume only one path to success

Singapore acknowledges the distributional differences in ability between people, and institutionalizes that

Sicily’s geography and the origins of the Mafia

Sicily

Italy’s Sicily

Sicily sits across from the tip of Italy’s “boot”, surrounded by three seas. The most striking presence on the island is Mount Etna — Europe’s most active volcano, which for thousands of years has been destroying and creating fertile land at the same time

The traits of life there: slow, direct, emotionally weighted, cautious toward authority

The Mafia was born there

In 19th-century Sicily the problems were very concrete: the state was remote, the police thin, the law untrustworthy, while land and property were extremely easy to seize

So a folk institutional system emerged, and the Mafia was born

Its typical structure includes:

Family

Boss

Underboss

Soldiers

Strongly bound by blood and place, with one of its most famous places of origin being Corleone

Red wine natural process for coffee

The red wine natural process

This bean is a Catimor varietal

A surprise process, precise roasting; it performs better under natural processing

After peeling the coffee cherry, it’s soaked in a fermentation liquid (usually red wine or a similar fermentation liquid), then sun-dried

Enhances wine aroma, complexity, sweetness

Gives layers of berries, tropical fruit, even chocolate

Reduces sharp acidity, increases body

Some flavor characteristics

Tropical fruit acidity

Wine aroma and sweet aroma

Caramelized sweetness

The Spiritual Ecosystem of Duo’en Sacred Mountain and the Sense of Awe

Duo’en Sacred Mountain

Duo’en Sacred Mountain is a natural snow peak + folk belief overlay

In Tibetan, “sacred mountain” doesn’t just mean “a good-looking mountain”; it is believed to:

be inhabited by mountain gods / protector deities

be able to protect the land and water, the people, and the livestock of a region

form a “spiritual ecosystem” with the surrounding villages, lakes, and grasslands

Duo’en Sacred Mountain is regarded as a guardian sacred mountain; the people around it do kora, hang prayer flags, and make offerings

The owner of the sika deer is Duo’en Sacred Mountain; people develop awe toward nature — the ability to actively lower one’s own volume in the face of enormous nature

Mount Kailash as the Cosmic Axis of Multiple Religions

Mount Kailash as a civilization-level sacred mountain

Tibetan Buddhism, the center of the world, and also the dwelling of India’s Shiva, and also the respective cosmic axis of Jainism and Bön

Most of those who do kora are Indians, then people from China and Nepal

In Hinduism, Mount Kailash itself is the dwelling of Shiva, that is, the cosmic axis

You must come at least once in your life; not to come is to be incomplete

Many Han Chinese domestically come mostly from the middle class or have some outdoor experience; they come all the way to Mount Kailash only after they can no longer move within the modern system

Genie Sacred Mountain as a Religious and Natural Community

Genie Sacred Mountain

Located around Litang–Batang in Garzê Prefecture, Sichuan, it is one of the most important sacred mountains of the Kham Tibetan region; the main peak is 6,204 meters above sea level

It is regarded as the first sacred mountain of the Kham region, with an extremely high status while long keeping a low profile in fame

Genie is considered a guardian sacred mountain, symbolizing within the Kham system a kind of “local order”

Unlike many sacred mountains that have been touristified, it is more like a religious and natural community that is still operating

Its geographical location is extremely complex: glaciers + wetlands + virgin forests + plateau lakes

The trekking distance is long, resupply is hard, and the psychological cost is large

Alex Honnold’s Zero-Tolerance Challenge: Training Compresses Fear to the Limit

Alex Honnold

This time he free soloed Taipei 101

I still remember the granite he climbed, a huge wall nearly nine hundred meters high

One misstep and there is no “recovery mechanism”; the ending has only one version

I’ve always felt this sport is almost a high-risk gambling sport

But he compressed the margin for error to zero through constant training

Facing height and danger, his amygdala (the fear-processing center) reacts significantly less; most people would be drowned by fear

Some people spend their whole lives expanding the safe zone; some choose to shrink the safe zone to infinitely close to zero

Under “zero tolerance” conditions, can a person compress fear, attention, and body control to the limit through training?

Bringing human potential to its extreme

Like deep-sea diving or spacewalks — not a project for everyone, but it changes our understanding of “possibility”

People Care Too Much About the Ending; Higher Rationality Is Inner Consistency

Suddenly it occurred to me that most people

seem to care too much about the ending of life, care too much about others’ evaluations, and fear too much the result of failure

Higher rationality is not more complex, but cleaner

A high consistency between inside and outside

Calculating the risk in advance, training repeatedly, solving the retreat problem before acting

By the time it really begins, only execution remains

The Different Temperaments of Shanghai, Hangzhou, Shenzhen, and Guangzhou

Shanghai is a very interesting city. How should I put it — if I had to pick a city in China I love most for strolling around, it would definitely be Shanghai. But if it’s the city I like most for living, that might be Hangzhou; the city most suitable for working might be Shenzhen’s Nanshan.

I went to Shanghai in April 2024; that was also the first time I met Dazi, who invited me to Shanghai. Shanghai gave me a feeling of being very petty-bourgeois, very chic; every person on the street made me feel interesting. The height of the streets is just right, and the environment gives a very different feeling. The small residents of Shanghai gave me a sense of everyday life; I found it very cozy. In Shenzhen, the tall buildings give me a sense of being small; these two feelings are very different. Think about it carefully — different cities bring me completely different experiences.

As for Guangzhou, I think it’s just a city suitable for living. Beijing feels too big, it’s the political center, too official, everything is still mainly about image. Ah, I think this petty-bourgeois atmosphere of Shanghai is really interesting.

The natural beauty of Cangshan and Erhai in Dali

I lived in Dali for about a month before, and I took a lot from it. The weather in Dali is strange — you can feel that sometimes the clouds hang very low, because Erhai is right next to it, and across Erhai is Cangshan. What’s unique about Dali is exactly Cangshan and Erhai. Cangshan is around 2,000 meters above sea level, raised another 2,000 meters on top of the existing elevation, so a lot of the cloud layer feels like it sits right in the middle of the mountain. The mountain scenery in Dali is especially beautiful, and that’s a big reason I like Dali.

Erhai is also beautiful, especially when the air quality is good. Erhai itself is at a fairly high elevation, and so is the cloud terrain, so you can feel that the air quality above is very fresh. The fresh air makes photos come out great there, and a lot of people who value natural beauty come specifically for it, feeling that it’s a paradise on earth, a fairyland. There’s straw, banyan trees, big banyan trees, plus beautiful clouds, a vast sky, and a little train on the wheat fields.

Kids can also go to the shore of Erhai at night to enjoy the breeze, and we had an experience like that. One day there was a supermoon, I think, and my roommates and I were by Erhai, on a platform over the water, sipping wine, listening to music, together with the moon, singing and dancing. In that moment I felt very relaxed, very content — a dreamy memory. I felt life should be like this, that we should bloom in an environment like this.

Besides that, I also jumped into the water — into one of the branches flowing into Erhai. Jumping in felt great too. And there’s cycling around Erhai, which is one of my favorite activities. On the last day, we cycled all the way around Erhai together, and the weather was especially good, the sky very blue — I still remember it clearly — the scenery was beautiful, with lots of lovely views along the way, very interesting. And there were a lot of people around Erhai; you could tell that people really love this place. It’s a tourist destination, and everyone wants to come and see it.

On the way back, we passed Cangshan, and you could see some greenways at the foot of the mountain. The greenway is very long, with a few scattered cars up ahead, and a thick layer of cloud in the middle of Cangshan, as if it had cut the mountain in two. Looking up from below as an individual, that perspective made me feel that the world I live in is wonderful — it’s not just cities and air, there are also forests, mountains, clouds, and the clouds are alive too, capable of composing an absolutely gorgeous picture. Really fascinating, really stunning.

The magnetism and charm of Lijiang Old Town

I actually haven’t been to Lijiang many times, but it left a deep impression on me. I’ve only been once, and stayed about three or four days.

I think it’s a very unique city. Even though it’s a tourist city, the people who go there give off a kind of magnetism — you feel this city is tied to sensibility and love. It felt like I was inside that atmosphere.

Something very unique about Lijiang is that it looks completely different in the morning, in the afternoon, and at night. At night, you can clearly feel that the city seems to come “alive,” very much like the scene in some xianxia dramas where Xu Changqing and Zixuan drink together in a mountain estate — you can feel it’s an old town with charm. That’s the feeling night gives me.

In the morning there are very few people — basically locals, or the occasional tourist hoping to snap an empty street. In the morning you can occasionally see the snow mountains, before the clouds have gathered, and inside the old town you can capture the ancient town and the distant snow mountains in the same frame.

I’m full of feeling about it — how happy the people who live here must be. They seem to have everything, blessed by the gods. They can see beautiful scenery in the distance, the sky here is very blue, the town is refined, and people are chill with each other, closely connected.

So Lijiang Old Town is a city I feel I could visit again. I quite like it — interesting, full of character.

Bangkok’s modernity and its fractures

Actually Thailand was my first time living abroad, and back then I went to Chiang Mai. Now I’m in Bangkok; since I landed in Bangkok, let me talk about my story in Bangkok and what I saw and heard there.

I went to Bangkok around the end of 2024 — not yet December, it should have been November 10th, when I flew from Wuhan to Bangkok. The city struck me as very interesting. It’s also a big metropolis; although its population isn’t as big as Chinese cities, its infrastructure is quite modern, with high-rises everywhere.

Bangkok is quite badly fractured. It’s not as bad as Kuala Lumpur, but Bangkok has a huge number of ordinary residents and some slums — a mansion might sit right next to a slum. This kind of fracture inside a city is fairly rare. In China, at least there’s zoning, or differences between urban and rural areas, or differences between districts. But in Bangkok, if you’re not careful, over here are office towers and high-rises, next door might be residential buildings, and nearby there might be a slum. The slums have a large number of Grab drivers, plus some Indians, Filipinos, and other locals. To me it looked pretty dirty and messy — that’s my first impression of Bangkok.

Bangkok is a city with a rich nightlife. Although I didn’t really get to explore it

The lantern festival and a boxing match in Chiang Mai

I was in Chiang Mai for the lantern festival. Chiang Mai is a really fun city. On the day of the lantern festival, everyone makes a wish.

At the time I was with two friends — one older brother, one Sister Hu. We set out from Tha Phae Gate in Chiang Mai and went all the way out to the suburbs to see the lanterns.

A few days earlier we had also gone to watch a Thai boxing match — my first time seriously watching boxing. A Chinese international student had sent out invitations inviting us in, and we also bought tickets. There was a Chinese women’s team on the card; they competed against some foreign teams, including teams from European countries and local school teams. On top of that there were boxing matches between European men — men’s and women’s fights. The whole arena was full of energy, all the fighters power types, you could feel the force of the punches landing on bodies. Very entertaining.

On the day of the lantern festival we rode in a songthaew — similar to the three-wheelers in old villages back home, except with two benches added inside. We went to eat khao soi first, then set off, and on the songthaew we saw a gorgeous sunset.

In the evening we arrived at the lantern festival site. There were two venues — one paid, one free. We went to the free one where the locals go; the paid one was pretty expensive, something like 700+ per person. We bought some lanterns and released them, and I snapped a lot of photos. A lot of locals and foreigners were holding their lanterns in both hands, smiling as they let them go — they looked very harmonious, very happy. The lanterns bloomed in the sky one by one, flickering like stars, carrying so many people’s hopes. Some were families releasing lanterns together, which perhaps represents a family reunion; it felt wonderful, and happy.

Once people have faith, something to place hope in, and expectations, a lantern comes to represent someone’s wish for a year. Write the wishes out, tuck them into the lantern, and there’s a place to put your hope — hoping the wish can bloom in the air and drift far away. To some degree it also hints to yourself, like the law of attraction: as long as you set a wish — say you want to make money, or you want to create something — making the wish clear is the starting point. The lantern is like the starting point of a dream; a single spark can start a prairie fire — light it, and it burns and spreads, making the whole world meaningful.

The three elements of pajamas and sleep quality management

Pajama help

Three elements of sleep: lowering core temperature, wicking sweat and regulating moisture, reducing external stimuli

Thermoregulation — this is a hard requirement. In a windy or air-conditioned environment, wearing nothing at all isn’t great either. Thin, breathable, not-too-tight pajamas are, in most modern bedrooms (AC/heating), actually the most conducive to a smooth decline in the body temperature curve

Moisture management, seriously underrated. Natural materials are useful here. You sweat at night; poor moisture absorption → sweat stays on the skin → sticky, cold, waking up repeatedly. Quick-dry but not breathable → dry and then stuffy → equally uncomfortable

Tactile stimulation — no need to be too tense; loose fit

The effect of hiking on the middle ear and eardrum

Hiking’s effect on the eardrum

It isn’t just altitude

Long hours of walking, panting, looking down and up:

the Eustachian tube opens and closes repeatedly

middle ear pressure changes frequently

For ordinary people it’s fine

For people with unstable middle ear structure, it’s an irritant

The inflammation hasn’t finished repairing; during the repair phase the mucosa is actually very fragile and can’t take friction

Staying up late is an amplifier …

How ordinary people should place themselves in the AI era

The AI era, the information era

How do ordinary people play this game well?

about me and myself, me and this world

click view -> start

How people get strongly biased by their environment

People and environment

People get strongly biased by their environment

Availability: whatever shows up more easily around you, you more easily do

For example, in a cafe everyone is working, and our brain’s imitation system easily treats focus as the norm

Executing certain things, or thinking about certain things, has a cost — friction as resistance or cost

The environment doesn’t need to persuade you, it just changes the friction on certain behaviors: when you want to mess with your phone, the phone is at hand and notifications keep coming -> friction approaches zero

Feedback systems: people are extremely sensitive to immediate feedback, and the environment gives feedback. Douyin is an instant-reward machine, fitness is a delayed-reward system; whichever side the environment leans toward, that’s what shapes you

The field of social norms: everyone else does it — for the brain it’s nearly impossible

Rather than forcing yourself to change, first change the environment so it’s “easier for you to become the person you want to be”

Serotonin sets your emotional baseline

People with high serotonin commonly show these traits:

Emotionally stable

Hard to knock down with small things

Not in a hurry to prove themselves

Less sensitive to comparison

When serotonin is low, the world turns into:

Easily anxious, depressed

Strong self-doubt

Extremely sensitive to others’ evaluations

Repeatedly thinking “am I not good enough”

Serotonin determines whether you want to self-destruct

Most serotonin exists in the gut, a small amount in the blood, and the part that truly determines mood is in the brain

Serotonin doesn’t guarantee happiness; compared with dopamine, it’s responsible for the emotional baseline, inner security, and the ability to withstand setbacks. To keep serotonin up, what needs to be fixed is: a consistent early-to-bed, early-to-rise schedule (waking time is super important), seeing natural light in the morning, reducing strong light at night

Then aerobic exercise, and some healthy eating

Another point is that serotonin is also closely tied to my position in the group — establishing boundaries, certain relationships

A healthy family of origin ≠ no pain; pain ≠ insight

A healthy family of origin ≠ no pain; pain ≠ more insight into the world

Pain -> a more real world: these people will:

Understand complexity

See good and evil at the same time

Know how the system presses people down, and also know how individuals resist

Pain -> a narrower world (but they think they’re profound): these people will:

Mistake defense for insight

Mistake wariness for clarity

Treat “I’ve suffered” as a source of truth

Reading wind in a forecast: direction, force and gusts

Tips for reading wind in a weather forecast

The wind part answers one question:

All day long, where does the wind come from? How strong is it? When is it most noticeable?

bft: Beaufort scale, the wind level

Level 0-1: almost no wind

Level 2-3: comfortable, you can feel it

Level 4-5: obvious, blowing in your face

6+: starts affecting activities

Gusts: 4 m/s — this indicator generally describes gusts, a very important indicator that people ignore day to day

Gusts have a bigger effect on the body; it’s a very uncertain variable — temperature, humidity, airflow direction shifting instantly

AlphaFold mapped 200 million protein structures in one year

An in-depth introduction to AlphaFold

Google DeepMind’s protein structure predictor AlphaFold mapped over 200 million structures in just one year

This is an astonishingly concrete advance! Because in the era without AI, it took humans years to map each structure

Predicting the complex structures of proteins has always been a hard problem for humanity

CXO is the pharmaceutical contract outsourcing service structure; CXO is a business logic of risk transfer and efficiency maximization

Previously, developing a new drug took 10 years and $1-2 billion, with a very low success rate (<10%)

If you’re a startup biotech company (Biotech), you don’t need to spend hundreds of millions to build labs, buy monkeys (experimental animals), and build chemical plants. You just need to raise money, then pay it to the CXO; they have ready equipment and teams

The core of a pharma company is IP (intellectual property / patents) and pipeline strategy

The core ability of a CXO is process (how to make it) and execution (how to get the process running)

The drug IP system and the logic of patent protection

Drug intellectual property

A comprehensive system, generally including the following kinds of rights:

Patent rights: a new drug’s chemical structure, synthesis method, pharmaceutical process — there’s a protection period, during which others can’t copy, sell or use the technology

Trademark rights: logos and the like

The IP time period is limited because new drug R&D costs are extremely high, takes extremely long, and has a high failure rate

Without patent protection, a company invests huge sums to develop a new drug and the next day someone copies it cheaply — who would still be willing to innovate?

So the state grants a limited-term monopoly (say a 20-year patent), letting pharma companies recover costs and make a profit

Once a patent expires, other companies can legally produce generic drugs, usually priced at only 10%-30% of the original drug, greatly reducing the burden on patients

When the patent expires, the technology enters the public domain and anyone can use it freely

Why generics aren’t necessarily completely identical to the original drug

Characteristics of generics

Sometimes I wonder why, once the patent period passes, generics shouldn’t be a problem

But in reality it isn’t necessarily so

In reality, though the core compound patent has expired, the original company often files peripheral patents (also called a “patent thicket”)

If a generic copies it completely, it may infringe those secondary patents and get sued

So many generic manufacturers proactively tweak excipients or processes to bypass these “patent landmines”

Regulations only require “bioequivalence,” and generic companies can only infer through reverse engineering (like grinding up a tablet and analyzing it) — it’s hard to reproduce it 100%

Original drug: Pfizer’s Lipitor — a white oval tablet, with “Pfizer” engraved on one side and “ATV 10” on the other.

Generic: multiple companies make atorvastatin calcium tablets — some white, some yellow; some round, some oval; the engravings differ too.

But every generic that passes approval has been clinically proven to have no difference from Lipitor in lipid-lowering effect

Climate Change Needs to Be Treated Differently from Global Warming

Climate change and global warming are not one issue and need to be treated differently. We shouldn’t let the temperature rise so fast, because we don’t know what will happen after the temperature rises.

Climate is a very complex system, similar to our brain. Simulating climate change is like creating virtual life, a virtual planet

Japanese Football Develops People; Chinese Football Chases Results

Japanese football is “a system that exists to develop people over the long term”

Chinese football is more like “an engineering project that serves short-term results”

In Japan, football is a path of socialization:

elementary school → middle school → high school → university → professional team, like a laid-out track

The coach knows what his segment is for: not to develop a Messi, but to develop “people who can keep moving on”

So they will care about whether this kid cooperates, and so on

The underlying belief of Japanese football is:

“People can be trained into it”

Even with average talent, as long as the system is right and there’s enough time, one can become a qualified player

So they respect fundamentals, positional sense, discipline

It doesn’t look flashy, but it’s stable

Three keys to conversation: sensory contact, guessing to draw out, prior thinking

Three things feel really important to me in conversation.

First, be good at using your senses — eyes, ears, nose — use them as much as possible to make contact with the world, including to understand the relationship between you and the other person. The function of conversation is really to reduce uncertainty; the emphasis can be on the two people talking, or on the content. For instance, if I see you brought something new today, I can build a topic around that thing.

Second is guessing. I use my own judgment and assumptions to draw the other person into completing a topic, so the topic can go deeper. This way of guessing works very well as a method of exchange.

Another point is to chat by drawing on things you thought about earlier — for example, you’d wondered whether the other person stayed up late, or whether they got something new — and bring those into the conversation.

A real-time information subscription strategy

If there were a channel

that could automatically help subscribe you to some real-time information

even gossip is fine

This kind of subscription strategy also interests me

Convenient for watching the drama unfold

The essence of podcast resurgence

The podcast resurgence

The essential reason: there’s too much information, so much that it becomes distorted

Short videos, feeds, trending mechanisms all chase maximizing instantaneous attention, and the result is that we’re bombarded by dozens of viewpoints every day, with not one real viewpoint digested

Once the brain gets tired, it naturally starts looking for a low-noise, high-density, linger-able content form — podcasts happen to satisfy that

A process of returning to what’s real; what it returns to is also a path of thinking

Naturally anti-algorithm, with no pressure to reach a conclusion, allowing uncertainty and allowing a path of thought

In history, almost every round of media explosion has been followed by a rebound of “back to long-form text / back to conversation”

Mass Attention vs. Content Quality: A Matching Strategy

Mass attention vs. content quality

Users’ taste determines the taste of the content that becomes popular in the market

Content that easily wins out in platform algorithms has the following traits:

Extreme (emotion > facts)

Simple (binary opposition)

Instant gratification

Infinitely scrollable (no sense of completion)

There is also a form of content that is severely underestimated. From what I’ve observed of a lot of short-video and self-media work, the essence is inversion and dislocation

Content with high cognitive density but a low enough entrance

In other words, the core is high quality, but the packaging is something the masses can swallow

Don’t lower the thinking, but lower the threshold for entering, and raise the cognitive return after retention — first be seen, then be understood, and finally be trusted

Touchscreens and AR Glasses as the Next Great Innovation

As everyone knows, the invention and widespread use of the touchscreen truly drove the spread of smartphones, because it can both display images and sense our taps and swipes. So a lens that combines AR and VR in one will make smart glasses the true “Next Big Thing”

Tech Livestream Selling: Showing Relations Between Products, People, and Society

I think there may be a direction going forward, and that’s the direction of making short videos, namely the tech product livestreamer.

But this kind of tech product livestreamer usually relies on their own first-hand experience to try various tech products, then applies them in some real-world link or scenario, and presents that scenario.

I think this is very interesting, because in this way the relationship between products and people, the relationship between people and society, and the relationship between technology and society, will reach some very high degree of fit

Aiming for Ninety-Plus in Every Act of Creation

Make sure every act of creation

make sure every expression reaches ninety or above

rather than just barely passing

The value of authentic records in the AI era

I think there’s something quite interesting here: in the future, as AI creation gets more and more popular and false information floods the internet more and more, I wonder — what is actually precious?

I thought about it for a long time, looked at many directions, many videos, many scenarios, and looked at how they went viral and how they made money. But I think what I really want to do is authenticity — genuinely recording myself, genuinely expressing what I think and feel. That’s what I really want to do, and what I can truly keep doing. Other than that, nothing else suits me.

And I think in the AI era, especially the AI era to come, what people care about may not be that someone generated some images, but who the creator behind the image is, and the real stories that creator made — those stories are interesting.

Artifying life is creation

Expressing life artistically is creation

Otherwise you’re just recording