DemoChen on The Outlier

After Getting Used to Voice Input, I Never Want to Type Again

DemoChen — Sat, 04 Apr 2026 21:06:57 +0800

Let the voice-first era begin. This post covers: why I started using voice input, why I chose Typeless, and the positive feedback it’s brought me.

My first encounter with voice input was in 2024, when WeChat’s keyboard added a voice-to-text function. At that point it was fairly basic — you spoke, it transcribed, word for word. The main requirement back then was accuracy.

But I found the use cases limited. Mostly I’d use it when typing wasn’t convenient, or with people I knew well enough that the informality felt okay. Using it freely felt off — I worried about how it would read on the other end.

Sometime around last year, as AI Agents took off, voice input tools started quietly exploding. By now there are dozens of voice-related products on X alone. For coding and writing, voice input — “speaking your output” — is a remarkably convenient way to input. You don’t type anything. You just express a thought.

I mentioned in my previous post that one of my most significant recent changes is going deep with voice input. There are plenty of products out there; I won’t review them all. Today I want to focus on three things: what voice input actually is, why I use it, and the positive feedback it’s generated for me.

Part One: What Is Voice Input?

(A note: in 2024 I picked up a bad habit of not including images in my posts. Text can describe information fairly precisely, and a quick search finds any product. So I’m keeping this image-light.)

Voice input is the simplest thing in the world: open the app, tap the button, start talking. Most of us have used it — WeChat’s voice-to-text is basically everyone’s first experience. Consider that the earliest prototype. What exists now is far more sophisticated.

Both phones and computers have many such products. I use iOS, and I’ve tried WeChat input, Doubao input, Shandian Shuo, and currently use Typeless. Setting other factors aside, Doubao’s accuracy is noticeably better than WeChat’s — despite multiple requests from our group chat, WeChat’s accuracy is still mediocre.

Voice input is faster than typing. Whether you use Shuangpin, Wubi, or standard Pinyin, voice input is quicker — you’re just talking normally, and it automatically adds punctuation. (Somewhat embarrassingly, I still haven’t gotten used to Shuangpin.)

Voice input also pushes you to clean up speech habits. Even though the mode is spoken, the output is written text — so you start noticing things. Verbal tics like “and then, and then, and then” look terrible in print. So before you start speaking, you pause for a few seconds to think about what you actually want to say.

For long text replies, voice-to-text is actually easier to revise than typed text. That’s counterintuitive — typed text seems quicker to edit — but in practice, especially for journaling, I’ve found it’s faster to revise a voice transcription.

Voice input does have some persistent weaknesses. Homophones are hard: distinguishing 他/她/它 (“he/she/it”) is tricky, and 的/得/地 (different grammatical particles that sound the same) are a constant challenge. Mixed Chinese-English is difficult, too. English words need to be pronounced fairly accurately; Chinese phonetic spellings (pinyin) need manual correction afterward. (That was as of 2024.)

Then I found Typeless. I started using it deeply around January 2026 — set it as my default input method. The core advantage: you don’t set anything up. Open it, start talking. When you finish, it lightly processes the content — removing filler words and errors.

The real magic of voice input, I discovered, is on desktop. Normally on a Mac, voice input is just a mode within the keyboard — one option among many. Typeless is different. It’s a standalone app. Once authorized, you call it in any window with a keyboard shortcut. No switching inputs. No interruption to your current text flow. You can even type and speak simultaneously.

That experience is something else. And the two things Typeless does better than anything else: automatic filler-word removal and real-time error correction and reconstruction. Filler-word removal keeps your output clean from the start. The reconstruction feature is the more interesting one. When you’re speaking, thoughts move fast. If I realize mid-sentence that I said something wrong, I can just tell Typeless: “I said something wrong earlier — please fix it.” It doesn’t transcribe that instruction as text. It reads it as a command and edits the prior content accordingly.

That’s the core reason AI makes voice input genuinely good.

It also has some drawbacks. Network dependency: it’s a foreign app with some latency requirements — typically 1–2 seconds of processing after you stop speaking. And the price is steep. Early on, the free tier was 4,000 characters per week. To stretch my quota, I ran two accounts on separate devices (phone and computer) since they don’t share data, giving me effectively 8,000 characters/week. Recently the limit was raised to 8,000 per single account, so I consolidated. But 8,000 still isn’t enough: by Tuesday of a given week I’d already spoken 6,000+ characters.

So I subscribed. List price is ¥998/year; with a student ID or certain other paths you can get 50% off. The offer I received was valid for 8 years. I paid a little over ¥400 for a year. For what it does to my output capacity, I think it’s extremely good value.

That’s the overview of voice input and Typeless. Since Typeless is my core tool, I’ll leave it at that rather than comparing everything on the market. Different things work for different people.

Part Two: Why I Use Voice Input

The most immediate reason: voice input makes my output dramatically more efficient. I just talk. I don’t have to think about whether my phrasing is correct or whether my idea holds up in real time — the goal is to express the thought, not to produce a polished sentence.

When I type, I can get stuck: a word won’t come, I second-guess an idea, I run slow when I have a lot to say. That friction is hard for me to accept. With voice I don’t stop. My attention doesn’t break. I don’t have to self-edit in the moment, don’t have to make every sentence beautiful. The goal is to get the content down — to make the thought exist somewhere outside my head.

A concrete example: writing a journal. Before, I’d plan the outline, think through the content, second-guess my structure. Now I open the journal app and Typeless, and I just talk from start to finish. When I stop, I often have a thousand words already. I’ve captured the full sweep of the day — memories, reflections, everything — quickly. Typeless handles the light cleanup; the output is maybe 90% what I’d want anyway.

Another example: when I’m working on Coding, describing requirements. I don’t need to organize everything carefully in advance. As long as I know the prompt structure — background, goals, constraints, standards — I can just walk through them verbally. Open the coding tool; it generates the file structure, starts building. I come back to adjust: wrong style, wrong logic. Same voice-based description to specify the fix. The moment I switch to typing, the pace drops. Efficiency is core to why I use this.

Second reason: my Mandarin pronunciation and some English words aren’t perfectly standard, but I want accurate output. Typeless helps with correction. Claude Code is a constant example — it’s easy to transcribe as “Cloud Code.” Some English words are just long: a tool called “Antigravity” is hard to type accurately and even harder to memorize the spelling of. With voice, I speak roughly what I mean; the app corrects it to the right word and right formatting. That’s an experience typed input can’t replicate.

Third reason: voice input has an “input history” feature — everything I say gets logged.

In Typeless, this data is stored locally. I wrote a Skills script to read and parse the local data, then generate content from it. The skill is published now, if anyone wants to try it. The value: excellent corpus accumulation. Traditional typing leaves content scattered across chat logs and specific documents. Voice input is more like a continuous recording tool. Not every app offers this, but the one I use captures everything I say — a word, a sentence, whatever — as a running log.

The exported corpus is extremely valuable to me. In the AI era, I care more and more about accumulating raw personal expression. Feed it to AI and it helps me understand what I used to say. My past experiences, thoughts, and ideas get surfaced again. Very interesting.

Voice input isn’t just a faster way to input. It amplifies my natural thinking. It’s closer to how I actually am, closer to my real expression. It doesn’t bind me to a structure when I’m recording. It doesn’t make me worry about rhetoric. It captures the scattered thoughts first — turns them into raw material. Then that raw material naturally becomes a record. Later I’ll convert these records into reusable, personal data assets.

Part Three: What Positive Feedback Has Voice Input Given Me?

One of the most important sources of positive feedback is actually AI. It can show me the patterns and problems in how I express myself when speaking or thinking. This depends on the content I’ve logged — I feed it to AI for analysis.

I exported my Typeless records. Only about 300 entries from this account — incomplete, because early on I hadn’t set it to save full history, and this is only the home computer; the more important data is split across my phone and work computer. But I gave the 300-entry export to ChatGPT and asked it to analyze.

I asked several questions: What’s my logic of expression? What are my core recurring themes? What suggestions do you have for improvement? What are the problems with how I express myself?

Reading the report, I found it matched my actual situation at probably 90%+.

ChatGPT said:

1. Your expression has a few stable characteristics. First, you’re naturally goal-oriented — your output is rarely for catharsis. It’s to advance something. Even casual remarks land on how to do it, where to put it, how to use it, how to make it more efficient. Second, you care about conclusion first, key points prominent. You have a native resistance to redundancy and performative structure. You keep saying: reduce ordered lists, use paragraphs to make the point whole. Third, you calibrate as you go — the moment you sense something is off, you correct it immediately. This suggests your judgment isn’t slow; in fact it’s fast, often faster than your words.

2. This also shows your expressive strengths clearly. Strong abstraction: you naturally elevate small life or work problems to principle-level. Strong calibration: you usually know specifically what “wrong” means, not just a vague dissatisfaction. Sensitivity to usefulness: you have a sharp sense of what’s worth keeping long-term and what’s just noise. These capacities are actually more visible in voice input than in writing, because voice preserves original thinking rather than a polished version.

3. There are also things worth improving. The main one: you often say more about your thinking process than about the final instruction. The upside is authenticity; the downside is that collaborators may catch the branches and miss the trunk. A small adjustment would help: lead with the conclusion, then the standard, then the background. Say what you want first. Then what “done” looks like. Then why you care. This won’t lose any of your thinking density — it makes it easier for others to enter your frame.

4. Also: you can start crystallizing your high-frequency value judgments into a stable vocabulary. Things like: accuracy first, conclusion-led, reduce redundancy, real pain point, sustainable, economic impact, social impact. The more stable these become, the more your expression sounds like you — and the more easily it’s inherited by AI and memory systems.

The positive feedback from voice input has gone well beyond “I can talk faster now.” It’s brought me closer to myself. It’s made it easier to see myself. It’s preserved my original judgments and revealed what I actually care about. What comes next is making that rawness converge earlier and become more reusable. Then expression isn’t just output — it gradually becomes my own methodology.

From this angle, voice input really isn’t just an input tool anymore. It improves my efficiency. It helps me accumulate corpus, calibrate expression, and review my thinking. In the AI era, the significance of this keeps growing. Because every raw expression you leave behind may eventually become part of understanding you, assisting you, and inheriting your way of thinking.

A note: Typeless currently doesn’t support exporting history on mobile or fully on desktop without workarounds. On desktop I use the Skill I wrote; mobile still isn’t supported. I submitted feedback; they said they’d consider it. I’m looking forward to seeing that developed — once it’s there, I can merge my data across all devices.

A few habit shifts too.

When you type, you don’t really hear yourself. Voice input means you listen back — is my pronunciation clear? Is my Mandarin standard enough? (It isn’t, but I try.)

To get better accuracy, I’ve learned to make my expression match the context more precisely. The more coherent my spoken input, the less correction I need afterward. That’s something you never feel when typing. So every voice input session is, in a small way, a training session for how you express yourself.

I now genuinely believe: voice input is one of the most powerful tools of the AI era. It makes me more willing to express, and more able to. More importantly, it means what I say no longer just floats past and disappears — it gets captured, organized, understood, and used again.

P.S. I recently saw some discussion in a group about whether AI-assisted content “counts” as AI-generated. This post was produced through Typeless voice input and then lightly adjusted for structure — all content is spoken input. Does that count as AI-generated? I don’t know, and I don’t think it’s worth overthinking. What matters is that my ideas got expressed.

What Actually Scared Me About The Shining

DemoChen — Sat, 07 Feb 2026 11:11:05 +0800

Last night I bought out a late screening of The Shining — walked in with anticipation and a little anxiety. Here are some thoughts.

I don’t remember when I first watched it. What stuck with me then were purely visual jolts: the elevator doors opening and a flood of blood pouring out like it cost nothing; two little girls holding hands at the end of the hallway; Jack pressing his face into the door crack, grinning like a broken toy. I took it for what it was — a classic horror movie. Scary. Impressive. Done.

This time felt different.

I realized the most terrifying thing isn’t anything in the frame. It’s in the silence that slowly thickens outside the frame. The longer you watch, the clearer it becomes: the Overlook Hotel isn’t haunted. It’s more like a container that isolates its occupants — draining time, relationships, and order, leaving each person alone with nothing but themselves. What you thought was creative solitude turns out to be a magnifying glass pressed against human nature.

Something I never paid attention to before suddenly made sense: Jack’s excitement when he takes the job is completely sincere. He doesn’t want the money. He doesn’t even fully want the sense of responsibility. What he wants is a self-contained justification — a closed loop he can present to himself. Five months of quiet. That sounds like the adult fantasy: just give me the right environment and I can restart; just leave me alone and I can fix myself. (He’s also obviously performing, of course — calling himself a teacher and a writer when he’s unemployed.)

The question is whether the environment is actually right.

What makes the Overlook terrifying is how clean it is. Clean to the point of no noise, no variables, no external feedback. Any emotional reaction you have inside it is like shouting into an enormous empty room — all that comes back is your own echo. In a place like that, it’s easy to slip into a particular delusion: it’s not that I’m getting worse; it’s that the world is targeting me. It’s not that I’m losing control; it’s that everyone is forcing me to lose control.

That’s what I think makes the film genuinely brilliant. It takes the supernatural — the go-to device of most horror films — and replaces it with something more realistic and far harder to defend against: entropy.

The entropy in a relationship usually doesn’t come from a sudden betrayal. It comes from a broken communication channel. Jack and Wendy’s channel is broken. Every word of concern from Wendy sounds like a verdict in Jack’s ears. Every effort Jack expresses sounds like an omen of instability to Wendy. Both of them are trying to maintain the stability of the family system — using exactly opposite operations.

Jack seeks stability through control. The more uncertain things feel, the more he reaches for something certain: his writing, his authority, his I decide.

Wendy seeks stability through compliance. She’s not unaware of the problem. She’s just more afraid of the system collapsing than of confronting it — so she swallows it, defers, puts the child behind her.

This combination is everywhere in ordinary life. In everyday settings, the friction is cushioned by noise: work, cooking, groceries, television — all buffers. But throw someone into isolation like the Overlook, remove all buffers, and the tension hits directly.

A lot of people read Jack’s descent as the hotel’s seduction. I’d rather read it as a failed man’s self-narrative finally finding a stage. Watch carefully: his madness doesn’t arrive suddenly. It’s permitted, one layer at a time.

First comes complaint. The job isn’t going well. The family doesn’t understand him. He’s being dragged down.

Then comes self-pity. Self-pity is a dangerous emotion because it relocates responsibility — off yourself, onto fate, onto others, onto circumstance.

Then comes rationalization. Rationalization means the violence starts acquiring justification, the control starts acquiring legitimacy. I’m not a bad person; I just need to focus. I don’t want to hurt you; you just don’t understand me.

This arc looks like a lot of real-life breakdowns, just pushed to an extreme. In reality it might not be an axe. It might be the silent treatment, emotional coercion, punishing someone with withdrawal, silencing every conversation with I work so hard and you still don’t appreciate it.

What the Overlook does is one thing: it removes Jack’s external calibration. No friends. No colleagues. No social norms. No daily routine. When a person has no external calibration mechanism, the system runs without monitoring. Errors copy themselves. Eventually they become catastrophic.

That’s why I keep thinking that solitude is not inherently noble. Alone time matters — but it’s a skill, and a kind of luck. It requires boundaries, rhythm, and an exit. Without those, it becomes self-combustion inside a sealed system.

Danny’s Shining is interesting to me, too. On the surface it’s a superpower. But it reads more like a child’s hyper-sensitivity in the face of trauma. Children sense atmospheric pressure shifting in a family before adults know anything is wrong. They smell the gunpowder in their parents’ relationship. Adults think they’re hiding it well. Children have already seen everything — they just don’t have the language to say it, so they absorb it into imaginary friends, nightmares, physical symptoms.

Tony — Danny’s invisible companion — I used to read as horror-movie convention. This time I read it as a child’s self-protective mechanism: creating a buffer between himself and reality. When real life becomes unbearable, send another version of yourself in to give the warning. Don’t go in there. Don’t get close. Run.

Wendy’s transformation is the part of the film I found most heartbreaking and most admirable. A lot of viewers find her annoying — always screaming. But from her perspective, that screaming isn’t weakness. It’s an alarm going off. She’s been living in a state of chronic stress from the beginning. She had no real choice in this. She wanted to hold the family together, to give her son a whole home — while also quietly knowing the home already had cracks.

Then she finds that stack of typed pages: All work and no play makes Jack a dull boy.

That moment is brutal. Not because it’s a grotesque face, or a river of blood, or a sudden apparition. It’s brutal in a more realistic way: you thought he was working; he’s been running in circles. You thought you were both enduring something difficult together; you were never on the same road at all. His so-called creative project has collapsed into one compulsive phrase, repeated forever. A man has locked himself inside a loop — and mistaken the loop for fate.

This is where The Shining most resembles real life. Most disasters don’t happen because someone is irredeemably bad. They happen because an ordinary person keeps repeating their own mistakes while becoming increasingly convinced they have no mistakes to make. Repetition itself is terrifying: it means you’ve lost the capacity to learn, lost the feedback loop, lost the ability to update your model.

The maze is a wonderful image. It’s spatial confinement and psychological confinement at once. Jack isn’t defeated by external force — he’s trapped and exhausted by his own obsession. Danny survives with a very simple strategy: walk backward, erase your footprints. A child doesn’t need a big theory. He just needs to stay alive. He has nothing to prove.

This makes me think of something I see everywhere: adults often don’t lose because they lack ability. They lose because they have to win. Needing to win means needing to save face. Saving face means refusing to back down. Refusing to back down means blocking feedback. Blocking feedback means gradually closing off. And eventually you walk into your own maze, convinced you’ve been thinking deeply.

The final photograph — I used to take it as a creepy easter egg. Now it reads more like a cold, flat verdict: you thought you were chosen; you were absorbed into a loop. The Overlook doesn’t need to create evil. It just needs to provide a stage and let you perform what was already in you. Violence, control, desire, avoidance — compressed by civilization, held down — given an environment that allows it, and they rise naturally to the surface.

So I can no longer watch The Shining as a simple ghost story. It’s more like a mirror — showing what happens to a person who is gradually stripped of feedback under conditions of isolation, failure, and silence. It’s frightening because it’s not distant.

If you ask what I’m left with after watching it, it’s a reminder: don’t turn your life into the Overlook.

Make room for solitude — but also for connection. Connection doesn’t have to mean socializing. It can be a regular sleep schedule, a stable workflow, a weekly walk and talk with a friend, a journal entry made before the next emotional wave arrives. Once a person starts locking all their problems inside their own head, the head begins to self-persuade, then self-indict, then self-destruct.

I’m more and more convinced: what we call rationality isn’t a gift. It’s an external system. Sleep is a system. Exercise is a system. Recording is a system. Relationships are a system. You need them to calibrate yourself, reduce entropy, maintain the closed loop. Without them, you think you’re thinking — you’re just repeating.

The horror of The Shining isn’t the axe, or the flood of blood. It’s that feeling of being in your home — and becoming more and more like someone alone in the wilderness. When that feeling arrives, most people choose to endure it, push through it, hold on a little longer.

But the movie has already shown us where that ends.

Don’t wait for the hotel to choose you. Choose yourself first. Open the door. Let the wind in. Let people in. Let real feedback in.

Most of this piece was written using voice input through Typeless.

How the AI Wave Has Changed Me

DemoChen — Sun, 18 Jan 2026 16:20:13 +0800

Tonight (this piece was written on the evening of January 15th) I had dinner with a friend at a spring pancake restaurant across from the office — haven’t been there in a while. Last time was after a project closed, eating with colleagues. That project’s been gone almost eight months now. One of my teammates said: “In 2025, you and I were in contact for at least eight months.” 2025 brought a significant shift in my work — more project-based, more centered on people.

The work itself doesn’t challenge me. I tend toward introversion, a little socially awkward with strangers — but I open up, especially once I know people. And I’ve found that I genuinely love talking with people: across departments, across roles, across business lines. Swapping notes on how work gets done. Understanding different business models. Talking about ordinary life.

Starry sky painted on a wall by the road

After dinner, I noticed a mural on a nearby wall — a starry sky — and photographed it for our “Four Seasons” group chat. I titled it “Starry Sky by the Road.” What struck me wasn’t that someone had painted the sky on a wall. It was more that someone had hidden it inside the wall — the edges look like concrete being peeled back, torn open, and instead of rebar inside, there’s deep blue cosmos. Behind the hard shell of the city, a wider world. I hope that’s true for all of us.

Probably the photo I’ve liked most recently.

That same day, @Qiangu messaged me: “Wow, you’re such a luminous and soulful blogger — don’t let your readers down.” He’d shared @J.sky’s post, which described me that way. Honestly, a surprise. “Luminous and soulful” isn’t a phrase I’d ever heard applied to a blog — and J.sky writes tech posts, which makes it even more unexpected. I went and left a comment: “As a so-called ’luminous and soulful blogger,’ I’m a little embarrassed — 2025 was a busy year, barely updated. This year I’ll try to maintain a better rhythm. Happy New Year!” Got a reply: “You need to post something. It’s been a while.”

Fair. I really had gone a long time — over a year — without sitting down to write properly. I don’t know if time just moves too fast, or if everything feels too urgent. AI is everywhere in our lives. Work pulls at my attention from every direction. There are always family things to handle. Who am I? What am I doing?

On March 28, 2025, I had a conversation with @Dahua and pulled together a small group of people who were still writing regularly — a chat called “Between the Lines.” I’m lucky to have found that group. Even luckier that I get to read their honest shares every day. I’m not particularly good at managing a community, but what I’ve learned from “Between the Lines” is that you don’t have to. Everyone grows at their own pace. When you feel like sharing, you show up. If there’s ever a second group, I’d call it “Letters Like Paintings.”

A lot has changed this past year. Some things haven’t. Let me try to sort it out.

AI Has Become Genuinely Useful — Even Good

I’ve long thought there’s a meaningful gap between a tool that’s “useful” and one that’s “good.” Useful means it can do things. Good means it’s woven into your life, your workflow, maybe even your inner life. Like that wall mural — I don’t need to look at the cosmos every day. I just need to occasionally be reminded that behind the city’s hard shell, there should still be a crack, a little breathing room.

AI is the same way.

I got a ChatGPT account on the third day it launched, in 2022. I was amazed — it could produce content that matched what I had in mind, and fast. No search capability then, no file uploads, no image generation or coding. Still, it felt remarkable. Now, three years on, it’s a completely different thing. Its upgrades have made it genuinely useful — and for me, even good.

I’m not an AI specialist. I can’t fully parse the three pillars — algorithms, compute, and data. The application layer is already more than enough to keep me occupied. As @Yuyi once put it: “You need to top up your technical vocabulary every now and then, just to keep the mind sharp.” My own relationship with AI centers on three things: Chat, Knowledge Bases, and Agents — plus Coding as a kind of infrastructure. The focus has shifted: from Chat, to Knowledge Bases, to Agent. Coding is the foundation (and my current obsession).

To put it simply:

Chat solves “question and answer” — it’s task-based
Knowledge Bases solve “what’s mine” — it’s personalization
Agents solve “do it for me” — it’s project-level
Coding solves “I can string all of this together”

Chat Is Not Giving Commands — It’s Building Context Together

Chat means using a large model’s conversation interface to get work done or find answers — Q&A, document processing, code generation, image generation, deep research, and so on. Two things have made a real difference for me here: understanding context and letting the LLM store memory about me.

Talking to AI requires input before output. I keep marveling at the elegance of tokens — they make it explicit that both input and output have a cost. Even a simple greeting costs something.

I’ve come to accept a fact: input is a form of responsibility. If you want high-quality output, you have to frame the question clearly. “Clearly” doesn’t mean “long” — it means appropriately granular. Too coarse, and the AI can only respond in generalities. Too fine, and you trap yourself in infinite detail, like recursing forever inside your own prompt.

And context matters in AI conversations the way it matters in human ones. Here’s a frame that helps me: don’t treat AI as an employee. Treat it as a collaborator. With an employee, you give instructions — commands. With a collaborator, you need to be on the same wavelength. That’s how you get genuinely useful advice in return. And since AI is the product of pre-training, the more focused your input, the more aligned the output will be with what you actually need.

My most common mistake: I think I’ve explained something clearly, but I haven’t. Especially at work, where my brain is already fragmented by group chats and notifications, it’s hard to tell a complete story in the moment of asking. So I gave myself a small rule: write three lines before you ask. Not to look polished — just to pull out one thread from the tangle. Once I’ve done that, I often find I’ve already half-convinced myself, and the AI fills in the rest.

My general prompt structure:

What’s my current situation (background)?
What outcome do I want (goal)?
What can I provide (constraints/inputs)?
What else might I need to add?

Let AI Remember Me — But Don’t Let It Trap Me In My Own Memory

The second thing: build a consistent voice in your AI conversations so it can remember who you are. Most major models now offer memory features — Gemini, ChatGPT, and others. When used well, this dramatically reduces the time you spend re-explaining yourself at the start of every session. That said, I have a real tension with this.

On one hand, I love being understood without having to repeat myself. The more an AI “gets” me, the more convenient things become. On the other hand — is this just another filter bubble?

My current workaround: use two model windows simultaneously for anything important. One has memory of me; one doesn’t.

The window with memory is like a long-term partner — knows who you are, what you love and hate. High efficiency.
The window without memory is like a stranger consultant — doesn’t accommodate you, more likely to give you advice you don’t want to hear but probably should.

I’ve added a small principle: for any important decision, do at least one “de-memorized” pass. Otherwise, you’ll keep becoming more and more like yourself. That sounds fine — but it can be dangerous. Because you’re not always right. I need the occasional neutral answer.

Pick the Model That Fits the Scene

Even different versions of the same model will produce noticeably different results. Some models are like engineers — rigorous logic, but not particularly attuned to your emotional state. Some are like humanities majors — beautiful expression, but you’ll need to impose structure. Some are like senior product managers — they’ll give you a plan that sounds reasonable, until you push them: What are the boundaries? What’s the cost? What are the risks?

The model doesn’t make the result. The match between model and task does.

Since the second half of 2025, Gemini has become my primary model — both for its capabilities and for the Google ecosystem. I can ask it about specific files in Google Drive or specific notebooks in NotebookLM directly in the conversation. Very efficient.

I still use ChatGPT, Claude, and Doubao. It depends on the situation.

The Secret to Making AI Good: Keep Coding

AI has become good to use, and Coding deserves a lot of the credit — especially Claude Code. Over the past six months, I’ve used AI Coding to build all kinds of small tools. Most of them haven’t been published anywhere; they mostly just optimize my own workflow. In that sense, Coding has become infrastructure for me. A foundation. Down the line, I think everyone’s workspace will be built on top of Coding, and each person’s workflow will look different from everyone else’s.

I like the word “infrastructure.” Infrastructure isn’t showing off. It’s making your future walk smoother.

My experience: don’t think of Coding as a programming tool or something for developers. Think of it as a way of translating your own problems — whatever you’re currently struggling with — into a format the tool can work with. You describe the situation, the constraint, the goal. The tool proposes a solution that can actually run and actually solve the problem. If it doesn’t, you iterate. You keep talking until it’s solved.

Or more bluntly: Coding is how I negotiate with the world. When the world doesn’t give me the tool I need, I make one. Even an ugly little script has value as long as it runs. Programming is writing. Writing is a project. Programming is the same. I recently came across a line about programming I loved: “Programming should be a fluid form of expression, just like writing.”

This made me think about the small things I’ve been building lately — a plugin for syncing Feishu multi-dimensional tables, a batch converter for ePub to PDF or Markdown. The point was never to become a professional developer. It was to express my needs through AI, and solve my problems.

Syntax Is Punctuation; Projects Are the Story

When we learn to write, we don’t fixate on spelling and grammar. Knowing every character doesn’t mean you can write a good essay. Knowing programming syntax doesn’t mean you can build a good product. Writing is about telling stories and communicating ideas. Programming is about expressing logic and creating value.

If you’re grinding through isolated algorithm problems just to learn syntax, you’re memorizing a dictionary to learn how to write — tedious, and almost impossible to reach flow. This is why many people (including past-me) give up on programming early: too focused on whether the punctuation is right, forgotten what story we’re supposed to be telling.

The thing writing and programming have most in common: you have to let yourself produce a bad first version. Get it running, then optimize. Get it written, then polish. Close the loop, then refine.

Deadline / Goal / Value: The Three Elements of a Project

What is a “project”? A project is a unique, temporary undertaking with three essential elements:

A clear deadline. Give yourself a finish line.
A concrete goal. Not “I want to learn Python” but “I want to build a bot that auto-scrapes news.”
Value created. It has to be useful — even if only to yourself.

You can also build in a feedback mechanism. Projects without feedback are hard to sustain. Learning without feedback is hard to grow. Life without feedback is hard to improve. When I build small tools, I deliberately design feedback points:

How many minutes did this save me today?
Did this help me earn anything?

Small feedback, but it accumulates. It creates a sense of agency. And that sense of agency is what keeps my anxiety at bay.

The Painful Part Is Discovering the Need

I shared on Jike a while back: to solve the problem of “turning web pages into e-books,” I explored an Epubkit plugin together with @Dahua. That’s a classic micro-project. We didn’t need to learn e-book encoding standards from scratch. I just needed to understand how the tool worked and how to run the workflow. The problems you encounter while doing a real project — those are the knowledge that truly belongs to you. That’s the project-based approach to building infrastructure.

Now we have Cursor and Gemini — essentially all-knowing programmers who can correct any syntax error or logic gap. This actually makes project-oriented thinking more important, not less. Once AI removes the syntax barrier, “knowing what you actually want to build” and “stringing requirements together with logic” become the real competitive edge.

I came across an article that put it well: “The era of ’everyone is a product manager’ is coming — and once everyone can implement their ideas, they’ll quickly realize that most ideas aren’t that good.” I agree completely. When execution becomes easy, what becomes scarce is: What problem are you actually solving? What are you willing to give up for it? Does this problem matter to you?

AI has flattened the barrier to entry. But it’s also lowered the cost of self-deception. It’s easy to produce a pile of things that look impressive but don’t close any real loop. So I hold myself to a stricter standard: every small project has to land somewhere in my actual life or work.

From Getting Answers on the Internet to Finding Them in Myself

To use knowledge bases well, you first need to understand your data. There are roughly three types: publicly available internet data; data produced through interaction with organizations or communities; and data that’s entirely private, like a personal journal.

When talking to AI, only content that comes from you — or relates to you — will actually help you solve your own problems.

If you feed it only internet data, the output will be generic, unpersonalized. You’re not adding a second layer of processing; you’re not organizing it around your own needs. That’s exactly where NotebookLM shines. It makes working with personal content remarkably simple. It’s the tool I’ve used most deeply in 2025.

What this is really about is something simple: AI isn’t best at knowing things — it’s best at organizing them. But what it organizes depends entirely on what you feed it. Feed it the internet, and you get the internet’s average. Feed it your own material, and it might actually produce your own insights.

For Knowledge Bases, I Choose NotebookLM

I like to think of a knowledge base as a middle platform, not a tool. A tool is point-to-point: I use it, solve a problem, move on. A middle platform is systemic: I deposit information into it, and that information compounds over time, generating new connections and new relationships.

I love that closed-loop feeling: read → organize → output → feed back → read again. When your system can self-circulate, you’re freed from the role of information porter. You can spend energy on the things that actually matter — judgment, decisions, communication, creation.

For example: I’ve collected many YouTube videos on project management, entrepreneurship, operations, and business. I put them all into a single knowledge base and ask it to surface connections between them — patterns across videos that might not be obvious otherwise.

I’ve also imported all of Luo Yonghao’s interviews into a dedicated notebook. Inside, I can analyze the full arc of his interview subjects’ careers, surface similarities and differences in tabular form, and ask cross-cutting questions: What’s common to all these people? How did each of them respond to difficulty? What questions did Luo himself keep returning to? All of it answerable through the knowledge base.

I’m not storing information. I’m storing relationships. Once a relationship appears, the world feels a little less chaotic. Finding relationships in information I care about — that’s a genuine pleasure.

In Organizations, a Knowledge Base Is a Management Tool

I once shared a view with a colleague: in a corporate organization, a knowledge base isn’t primarily a service for frontline staff — it’s a strategic tool for the organization itself. A knowledge base is fundamentally a management tool, not an efficiency tool.

This sounds counterintuitive, but I mean it. For example: a company could use a knowledge base to analyze the characteristics of its top salespeople — pull their call recordings and WeChat chat histories, identify their methods, and compare them against lower-performing tiers.

But if you use a knowledge base to build an online customer service chatbot for the sales team, you’ll get very limited value. It might raise the floor for underperformers slightly. But the real conversion skills — the ones that actually close deals — are non-structural: tone, timing, subtext. Those don’t live in documents.

The real pain point isn’t “stop new employees from making mistakes” (that’s a baseline, addressable by process and policy). It’s “how do we scale what our top salesperson does” (that’s growth).

Why is “helping the bottom 50%” low ROI?

Training a 30-point performer to 60 points is possible. But firing 30-point performers and hiring 60-point ones is often cheaper than building and maintaining a large knowledge system.
Sales teams typically follow a power law: 20% of people drive 80% of results. Serving those 20% (extracting their wisdom) has far more leverage than feeding the 80%.

Looking forward: what if the knowledge base became dynamic? It automatically analyzes recordings and chat logs from top performers, identifies their patterns (always anchor on value before discussing price, for instance), and auto-generates a new SOP pushed to the whole team. That’s actually serving the organization — replicating what works at scale.

This line of thinking leads me to something bigger: in the AI era, organizations will increasingly resemble systems rather than groups of people. The core of a system isn’t any particular role — it’s information flow and decision chains. Whoever smooths the information flow makes the organization faster. Whoever shortens the decision chain makes it more stable.

Coding Can Make a Knowledge Base Come Alive

When Claude Code first launched, I wasn’t particularly excited — mainly because I wasn’t familiar with it, especially compared to Cursor’s interface. When Claude introduced MCP, I still didn’t use it; it felt like something for developers, not for users.

But recently, Claude Skills changed my view entirely — to the point where I started learning Claude Code from scratch. An example I saw on X: using a Skill to automatically upload files to NotebookLM and generate articles from them. Essentially automating NotebookLM as a professional knowledge pipeline.

Another Skill I love is “Superpowers” — a complete development workflow and skill library that, before writing any code, first clarifies project goals, then breaks down product design and features, then produces a granular, executable plan. For Coding newcomers, it dramatically helps clarify requirements — much more aligned with actual software engineering than the “one-sentence requirement” I used to give.

MCP requires developers to provide it. Skills are documents — concrete descriptions of workflows or requirements. You can write one yourself; anything AI produces is something you can actually read. Easy to install, easy to use, and you can write custom ones for your own needs. Playability and practicality both went way up.

After getting NotebookLM running smoothly, I had a new idea: use NotebookLM as a content processing middle platform. My company uses Shimo and Confluence for documentation, and different departments use different ones. I used Claude Code to read project documents from both Shimo and Confluence, upload them to NotebookLM, and save the processed output back to Confluence or Shimo — completing a project closure report automatically.

I love Skills. Like people, assembling different capabilities is what creates comprehensive ability. Claude Code makes those capabilities external and concrete. In an organization, these Skills could be shared — anyone can call them — and efficiency multiplies. Collaboration becomes more fluid.

Learning May Be About to Turn Upside Down

Future learning may undergo a radical transformation. The transformation I’m imagining isn’t just “more online courses” — it’s a fundamental change in how learning works. Right now: videos, lectures, tests, assignments. But what if learning became more game-like, more immersive, integrated with world-model products and multi-modal capabilities? Visual, interactive, adaptive? I think it’s possible. And the efficiency might be genuinely higher.

One small sense I have: reading may become a much richer audio-visual experience. When you’re reading a novel or classical poetry, instead of pure text to decode, you could use multi-modal capabilities to understand what a passage actually looks and sounds like — anchored to environments and concepts you already know. Why not?

I’m excited about that possibility. It could turn “learning” from a form of pressure into a form of experience. But I’m also a little worried: the richer the experience, the easier it is to become addicted; the more addicted you are, the harder it is to sit quietly with a slow, difficult text. So in the end, it comes back to a more fundamental question: What are we actually learning?

Knowledge, or how to think?
Skills, or judgment?
Methods, or ourselves?

On Working Differently

I rarely write publicly about work — on this blog, or anywhere else. Every person’s work situation is so specific that what’s true for me may be useless, or even misleading, for someone else. Different backgrounds, personalities, environments, roles. Even common communication norms vary by company. So I tend to stay quiet about it.

In 2025, my work shifted significantly toward coordinating people. As I mentioned above, I’ve found that the deeper pleasure of working with people is understanding them: their goals, their starting points, their professional needs. Once you understand all of that, you realize how genuinely difficult it is to achieve a shared goal across different perspectives. Product, engineering, QA — they all share the same surface goal (ship the thing), but their actual work, their metrics, their anxieties are all different. In some projects I’ve been involved in, those differences are even more pronounced. As the person coordinating the effort, the challenge is real.

What I’ve found more interesting, compared to before: previously I’d track a project mainly through tasks — move this, close that, push here. Now I think about it through people — who needs to be aligned, who needs support, who needs space. That’s more engaging, not because the friction disappears, but because talking with people reminds me that people are the most essential factor in whether anything gets done. You can’t just optimize the work. You have to respect the human beings doing it.

I increasingly believe: the real bottleneck in any project is never the tools. It’s the people. Tool problems are mostly solvable with time. Human problems require understanding, patience, communication, and sometimes structure.

My biggest change this past year: I’ve become more willing to invest time in alignment rather than just pushing forward. Pushing is moving the car. Alignment is making sure everyone agrees on the destination. Alignment is slow, even frustrating — but it prevents rework. It’s anti-entropic: spend a little more energy upfront, have a little less chaos downstream. (AI context and memory work the same way.)

What’s Stayed the Same

One more thing: the constants. My enjoyment of recording hasn’t changed at all — keeping accounts, keeping a journal. As of now, I’ve written more than 1,500 days of journal entries. These are among my most valuable outputs. I’ve fed them to AI, had it analyze and generate comprehensive reports. It’s a window into myself. In the journal, I don’t have to consider feelings or social dynamics — I can say anything. That raw honesty might be the one place where I’m completely myself.

My approach to information hasn’t changed either. I’ve returned to using an RSS reader (Inoreader), not to fight algorithmic recommendations, but to preserve my ability to follow the people and things I actually care about. The shift from “following topics” to “following people” has made the world feel more interesting.

The curiosity about new things hasn’t changed. The interest in AI, the learning and experimenting with Coding — these are ongoing.

I Started Recording With My Voice

Two weeks ago, I bought a recording card hardware newly released by Dedao Notes — it attaches to your phone and captures audio throughout the day, which is then transcribed. Every conversation, every passing thought, recorded and summarized. These collected sounds become text uploaded to my NotebookLM, where I can find connections between things I said that I wouldn’t have noticed otherwise.

I find the act of collecting deeply valuable. It fights forgetting. I write a journal every day, and now I can review that day through what I actually said — what ideas came up, what I was thinking. It’s not unlike having a camera running 24 hours (but more affordable and actually suited to how I live).

Extending this further: more collection supports better self-understanding, including recognizing patterns and problems in daily life that would otherwise slip by.

More recording and creating helps us know ourselves better. I love the phrase “fight forgetting.” People are fragile in ways we don’t acknowledge. We think we remember; we don’t. We think we understood; we just felt strongly in the moment. When the emotion fades, everything becomes fragments. Recording collects the fragments and gives them to a knowledge base for organizing. Slowly, you start to see your own behavioral patterns, expressive patterns, emotional patterns — even your blind spots.

As a note: most of this article — probably 80% or more — was written through voice input. Using a voice input app (Shandian Shuo), I can just start talking. When I stop, it transcribes everything. That’s a significant convenience the AI era has brought.

I end with an image: a tired evening, opening the voice input app, talking for a few minutes. Maybe incoherent, maybe repetitive, nothing that sounds like a finished piece. But I said it anyway. And that itself matters. Because expression isn’t proof that I’m impressive. It’s proof that I’m still alive, still observing, still willing to pick up the small things and hold them. That’s worth more to me than learning another tool.

Finally

The long gap between posts wasn’t because I wasn’t writing — it’s because I wasn’t making what I wrote public. I write 1,000-word journal entries every day. This blog, as a platform with a higher bar for publication, carries a real cost. And one of the bigger changes in how I record things is that I’ve come to prefer voice.

That said, the blog remains my first choice among all platforms. Which is also why I subscribe to others through RSS. A living person updating and expressing themselves — that deserves respect and patience.

See you next time. 👋

7 Notes From Moving

DemoChen — Tue, 25 Mar 2025 21:49:32 +0800

Rather than “drifting,” let’s call it: always in motion.

1. Professional deep cleaning: temporarily overrated. After we moved in, my mom pointed out it wasn’t really clean — grime still on the stove grate, rust the cleaners had insisted couldn’t be removed. My mom had them spotless by that evening. In the end, the professional team wasn’t much better than a thorough DIY scrub. We’d booked a private arrangement (not through a platform) because the platform packages tend to exclude things like sofas and curtains. We have a chaise longue, a four-seat sofa, and curtains across three bedrooms plus the living room — buying those add-ons separately would’ve been expensive. Going private meant everything was included, better value overall. That said, I’m still curious: why can’t a “new home deep clean” package just include everything? It’s all fairly standard stuff.

2. Settling in takes longer than the move itself. We spent about a day packing. But the preparation — cleaning, washing appliances, sorting through what to keep — took nearly a week before moving day. And we’re still not fully settled; there are small things left to buy. The previous tenants had lived in this apartment for over nine years, so some items were beyond use. The washing machine, a Siemens front-loader, was flagged by a JD cleaning technician as something that would definitely leak if disassembled and reassembled. So I took the opportunity to use a government appliance-replacement subsidy (国补) to buy a new one. The old machine got a surface clean and a tablecloth draped over it — now it’s a storage surface. (I’d throw it out if I could.)

3. Moving day meets staircase: a test of stamina and gratitude. In 2020, the movers used a stair-climbing machine to bring things down from the sixth floor — noisy, but effective. This time we had stairs again. Not many floors, but a lot more stuff — a full moving truck. I booked through the Lánxīniú platform; two guys showed up to help. Watching one of them tie four boxes of heavy books to his waist with rope, hold both ends with his hands, and nearly stumble on the first step — something about that hit me. I’d originally planned to go with the Japanese-style moving service, but for under six kilometers, ¥1,500 felt like a lot. I bought them a McDonald’s combo set to say thank you.

4. For an exit permit, try the security guard first. I called the property management office to ask where to get an exit permit for moving. They said to come in after 8:30 a.m. When I got there, they said they needed to verify with the landlord that I’d returned the keys — which I hadn’t yet — and refused to issue one. I went back and forth with the landlord about whether that was actually required. Eventually I walked over to the security booth and asked directly. The guard said yes, no problem, handed me a form. I filled in the movers’ vehicle info and my residence details, and that was it. On my way out I mentioned to him: “Could you help direct the truck to Building 2 when they arrive?” He said: “No worries — anything we let through, we guide in.”

5. A balcony makes a real difference. The previous apartment was 76 sqm but had no balcony, which made it feel smaller than it was. The new place is three bedrooms, two bathrooms, and a balcony — the difference in feel is substantial. Also: moving out of Ziroom and going direct with the landlord was very much worth it. The last two apartments I’ve rented have both been direct-to-landlord, and what I’ve saved on agent fees would more than cover a full professional cleaning and some minor renovation. Our rent here is ¥4,600/month — the landlord originally listed it at ¥5,200, we negotiated to ¥5,000 at the first viewing, and eventually got it down to ¥4,600. Compared to the ¥3,900 two-bedroom we had before, it’s a step up, and I can live with the downsides (no elevator, fewer in-compound amenities) given the upsides (two shopping malls across the street, five-minute walk to the subway).

6. What’s “lost” during a move isn’t always really lost. Something went missing from the bedroom — none of us could remember it, and we started wondering if it had fallen off the moving truck. Then someone looked at photos and found it tucked into a small pocket of another bag. The relief of finding a lost thing is disproportionately satisfying. The less satisfying part: several parcels I’d ordered were still sitting at the pickup station near the old address. No reminder text, no call — and they’d arrived on the 16th. I’ll have to make a trip back on the weekend. Can’t have it all, I suppose.

7. Thank you to everyone who read, liked, shared, tipped, or commented on the last post. Seeing the comments, I realized I’m not the only one who drifts. But life is our own, and how we relate to it is up to us. “Every new place brings new people, and opens a new chapter.”

Rather than “drifting,” let’s call it: always in motion.

The Drifting Feeling That Comes With Moving

DemoChen — Thu, 20 Mar 2025 17:30:36 +0800

Just before our two-year lease was up, the landlord sent a WeChat message: “Hey, Little Chen. Sorry about this — it’s not going to work for me to renew after the lease ends. Wanted to give you a heads-up. If the timing is a problem, staying an extra month or two is totally fine; I’m not in a rush. The place won’t be needed until June or July, so just do what works for you.”

Understood. Time to start looking.

Snow in spring

When I pulled up the “Documents/Home/Housing/Rentals” folder, I counted five different addresses and ten rental contracts. A lot of drifting packed into a folder. But every new place brings new people, new stories — a new chapter, always.

I’ve thought about buying. Owning would mean stability — no more packing boxes every couple of years. But a large mortgage for a small apartment feels like the wrong trade-off, and in Beijing, where every square meter costs a fortune, renting often makes more sense. To be rigorous: if you want mountains and water, generous space, and a view, you’re looking at the sixth ring.

Finding an apartment isn’t so different from being set up on a blind date — or so I’m told. You have to know what you want before you go looking. When you’re actually viewing places, you weigh each against your requirements. In the end you settle for something that “works for you” — because the perfect place, the one that meets every criterion and fits the budget, almost never exists. Neither do perfect matches in people, I suppose.

My own requirements have shifted over the years: from renting a single room in a shared flat, to a one-bedroom, to a two-bedroom, to a three-bedroom. A bookshelf used to be enough. Then I needed a living room and a kitchen. Then a room I could call my own. Some things haven’t changed: I want a space without too much furniture — we don’t watch TV or lounge on heavy sofas, so simpler is better. I’ve always wanted a subway station within walking distance and a wet market across the street. Constant change brings constant discovery. That might be the best thing about moving so often.

“Embrace uncertainty and benefit from it” — that’s Taleb’s antifragility. And looking back, our last two moves have both turned out wonderfully.

The previous move happened because Ziroom raised our rent from ¥4,200 to ¥4,900. We decided to leave. The commute stretched from 40 to 60 minutes, which was still bearable. Once we were in, I discovered the neighborhood had a river, a park, lotus flowers in summer, beautiful sunsets — a place made for running and cycling. Many of the photos on Today Looks Like This were taken there. I photographed that neighborhood through every season, and I have memories of COVID lockdowns there too, including a memorable standoff with the building management committee. The best part: because we were all working from home, a small group of neighbors started playing basketball together, working out together.

Extended reading: A Record of Centralized Quarantine

The current place is a 70–80 minute commute. The trade-offs are worth it: there’s a bus stop at the door, the Grand Canal runs behind the compound, a wet market sits across the street, and best of all, a newly built library, museum, and art gallery are a short walk away. Most weekends I head to the library to read, watch something, daydream, and catch the sunset. I didn’t know any of this before I moved in. That’s the thing about moving — the good surprises only reveal themselves after you arrive.

Of course, unpleasant surprises happen too. In 2020, I moved into what I can only describe as a windowless dungeon — a living room with no natural light and, worse, cockroaches. That was the only move I’ve genuinely regretted. In hindsight, the warning signs were there: I viewed the apartment at night, and got too excited about a one-bedroom that had a subway station literally on the other side of the front door. After the lease ended, I scrubbed every item before it crossed the threshold of the new place; anything that could go, went. No cockroaches since.

Does all of this feel like drifting? Honestly, sometimes. Moving is draining. Each new place requires rebuilding familiarity from scratch — learning the layout, earning the trust of the security guards and building managers. But if you never have the chance to move, staying put is also a valid choice. Neither is better or worse. Each is just a new beginning.

Appendix: A friend, @Na Zhang, shared these apartment-viewing tips with me years ago — still useful today.

Test every appliance, the gas stove, and every light while you’re viewing. Anything that needs replacing or cleaning, negotiate into the deal upfront.
Check doors and windows: does the security door lock automatically? Are the window screens intact? Can the windows be locked?
Check kitchen pipes and drains for leaks. Test the toilet flush.
If there’s a mattress, test the firmness and check the dimensions against your bedding.
Ask which internet service provider is in the building — ideally one you already use.
Close all doors and windows: any strange smells? Good soundproofing?
Open every cabinet and drawer in the kitchen and bathroom. Check for cockroaches.
Test the wall outlets.
Opt for a gas water heater over electric; a gas stove over an induction cooktop.

@Dahua also recently shared a booklet: City Rental Guide. I read it and reformatted it into an e-book. The overall framework is solid — especially useful for first-time renters who want a quick orientation. It would be even better with scenario-specific checklists: a kitchen checklist, a bedroom checklist, a contract-signing checklist, and so on.

Today is the Spring Equinox — equal parts day and night, plants greening, flowers opening. A good day to go outside and photograph something. March’s small gift this year: it actually snowed on the 16th, making up for Beijing’s snowless winter in 2024.

Extended reading: Moving Diary, 2020

The Value of AI in Insurance

DemoChen — Fri, 07 Feb 2025 14:08:57 +0800

DeepSeek has been making waves lately — its impact rivals ChatGPT’s debut at the end of 2023.

Today I chatted with my friend @Yuzi, who works at an insurance brokerage, about AI and insurance. Here are some highlights from our conversation, along with a few thoughts of my own — mostly for my own record.

I’ve recently been trying out Keyou Health’s “personalized health checkup” service (¥299 per session). I started wondering if I could build something similar myself, because the intake questionnaire is entirely structured: do you currently have any of the following conditions? Does your family? If yes, when was it diagnosed, what treatment did you receive, how did it go? You also need to upload lab reports, medical records, and so on. I’m curious to see whether something interesting emerges from the handoff to a real human advisor.

Keyou also offers a health report interpretation service. Once my checkup is done, I might give it a try — though honestly, I’m skeptical. The conclusions are right there in the report, and you could just ask an AI. So what justifies the ¥500/session price tag? I genuinely don’t know. Staying curious.

This naturally led us to insurance. I’d started learning about it last year. As I understand it, insurance is fundamentally a contract — and the terms of insurance products are fixed; they can’t be customized. That part is highly structured and standardized. The complexity lives on the customer side: most people buying insurance don’t know what they actually need, what a specific policy covers or excludes, or what to watch out for. And most people (I’d guess) aren’t fully aware of their own health status — they breeze through the underwriting process without a second thought. That gap between policy complexity and buyer awareness is exactly why insurance agents and brokers exist.

A good agent or broker starts by understanding the customer: why do you want insurance? What’s your budget? What’s your household’s financial situation? From there, they match products and build a suitable plan. They also need to understand the customer’s health history — reviewing insurance records, medical records, lab reports, physical exams, medication history — and use all of that to navigate the underwriting process. Underwriting is probably the hardest part, especially for customers with pre-existing conditions (so-called “non-standard” policyholders), where it needs to be done case by case and can be exhausting.

I also learned that insurance products are heavily commoditized — the differences between competing policies are often minor. This is where AI shines, I pointed out. It can sift through a product library and surface the relevant differences for each customer. Whether it’s gathering customer information, organizing health records, or matching someone to the right plan — AI can do all of this accurately and efficiently. Things that used to require an agent’s time and expertise are now within reach of a well-designed product.

First takeaway: the efficiency gains on the sales side will be dramatic. Today, putting together a customized plan for one customer might take 1–2 hours. With AI, you could do it for dozens — or hundreds — of customers simultaneously. And if you watch how DeepSeek reasons through a problem, you’ll notice it outpaces human analysis and catches errors faster. Greater per-agent capacity means more customers served, which means higher conversion rates overall.

So what happens to the salespeople displaced by AI? They move to the back end: service and relationship management. Right now, brokerages mostly hire more salespeople to handle acquisition and conversion. Once an AI system can do that work, the front-end value of a human salesperson declines — and the backend, where human judgment and relationship-building actually matter, becomes more important.

At its core, buying insurance is a trust transaction, not a tactics game. Whatever sales scripts, product knowledge, and industry expertise are currently being trained into salespeople will be absorbed by AI. Trust cannot be absorbed. What does trust look like? It’s eating hot pot together, playing basketball together, sitting in a sauna together. What does a tactic look like? “You’re one step away from claiming ¥300 cash,” “You’ve outperformed 99.3% of users.”

Because insurance is a long-term, recurring product, customers also need to trust that their agent will still be around when they need to file a claim. I’ve noticed that agents with genuine personal relationships have a real edge. A friend’s agent — 19 years in the industry — plays basketball and has meals with my friend regularly. My friend doesn’t even need insurance right now, but he’d recommend this agent to anyone. Because there’s real rapport there, because this person is the first who comes to mind. Online-only brokers face the opposite problem: if you’re not present in someone’s life, they won’t think of you when they need help — or recommend you to their friends.

My friend made a sharp point: if you’re not going to invest in service, you might as well sell simple, transactional products — travel insurance, accident insurance. I’d argue that’s bad for everyone involved: the policyholder gets no real support, the insurance company gets no loyalty and low ticket sizes, and the brokerage makes thin margins.

Second takeaway: in an AI-enabled market, choosing insurance is really choosing service — service that’s tangible, visible, and trusted. Low-level conversion is tricks; Pinduoduo has mastered that. High-level conversion is trust: when you buy an iPhone, you go to the Apple Store or JD.com, because you trust the brand. I’m not saying every agent needs to send gifts on every holiday — just that a deeper, more genuine connection is what ultimately matters.

More broadly: a business that only sells standardized products is hard to sustain. What keeps you alive is something others don’t have — something the customer genuinely needs. That’s your moat. That’s your core advantage.

How should the insurance industry build that kind of service? How can AI be used to genuinely help policyholders, rather than just cut costs? I don’t have the answer. Happy to discuss if you do.

P.S. DeepSeek is everywhere right now — free, open-source, accessible without a VPN, and genuinely capable. I’ve been watching people in my circle monetize it left and right, and I’ll admit I’m a little envious. I keep asking myself: which of these applications are actually solving real problems, and which ones are just riding the hype? Staying tuned. Continuing to learn.