Articles / Blog
Hands-on notes on LINE AI customer service, booking systems, websites and advertising, plus observations on living with AI. It saves readers time, and it saves us time too.
LINE AI customer service
Whether to do it, how to do it, the pitfalls you'll hit, and what it costs.
- LINE Official Account auto-replies not enough? Check these 5 signals before you decide to spend money
- Nothing held back: 15 pitfalls of automated LINE AI customer service
- AI customer service: SaaS or custom? 3 questions to decide, with a three-year cost comparison
- How much does LINE bot development cost? A 2026 quote broken into six parts
LINE booking
Subscription tool or custom build, how a quote breaks down, and how to run acceptance once it's built.
- LINE booking systems: subscription SaaS or custom? Six questions will tell you
- How much does a custom LINE booking system cost? The quote broken into seven parts
- Booking systems for clinics, beauty businesses and restaurants are really three different things
- Booking system acceptance checklist: 12 tests that tell you whether it actually works
Websites and AI visibility
How much to spend on a website, how to choose, and whether customers can find you when they ask AI.
- Why do quotes for the same website range from NT$9,000 to NT$300,000?
- Wix, WordPress, or hire someone? Add up five years of costs, not just the first
- Your website is one island, your LINE Official Account another: how to build the bridge between them, and what it costs
- AI's advice on top, paid ads underneath
Restaurant ordering systems compared 2026: monthly fees, contracts, LINE ordering and QR ordering features across 13 vendors
13 restaurant ordering systems, including iCHEF, Dudoo (肚肚), Eats365, Damai (大麥), QuickClick (快一點) and Baodan (爆單), with monthly fees from NT$0 to NT$3,988. Every figure is labeled as coming from an official pricing page or from the vendor's own articles, and where we couldn't find a number, we say so.
99 Customer serviceHow do you apply for a LINE Official Account? We opened one, and these are the four places you'll get stuck
Applying for a LINE Official Account is free and takes fifteen minutes, but one option can't be changed once you click it. Choose wrong and you have to start the whole account over. We went through the process with a brand-new account and wrote up the four sticking points and the new pricing from November 2026.
98 ExperimentsBuilding a free online TV out of YouTube: Ten TV, from idea to launch in one day
YouTube has everything, but you have to pick every single video yourself. I wanted what TV used to be: already playing when you turn it on, change the channel if you like, and everyone watching the same moment. From one question in the morning to opening it up to search indexing that night, ten channels, a remote control, proposal voting and dwell-time stats were all built in a single day by one AI agent. This piece covers how the work was broken down, the pitfalls, the two times I guessed wrong, and why I decided on just ten channels.
97 WebsitesFree website analytics tools for 2026: can't make sense of GA4? Search Console, Clarity and Cloudflare compared
GA4 is powerful but hard to read. If you want to know how many people visit your site, where they come from and what they do on the page, several free tools each cover one piece. Here's how to combine five of them, plus the one GA4 setting you must change.
96 Customer serviceHow to set up Instagram DM auto-replies: 2026 comparison of IG auto-reply tools, covering free features, ManyChat and Taiwanese vendors
Auto-DMs triggered by a "+1" comment and automatic DM replies are partly possible with Meta's free tools. When do you need a paid tool? How do ManyChat and Taiwanese vendors compare on price? Plus three rules that will get your account in trouble.
95 WebsitesWhat should your website use to send notification emails? 2026 email service comparison: Resend, SendGrid, Mailgun, Amazon SES, Brevo (free tiers included)
Website forms, order confirmations and account verification emails shouldn't be sent from your company mailbox or Gmail. Free tiers, pricing and best fit for six email services, plus how we actually rescued our emails from spam and got them back into the inbox.
94 WebsitesWeb hosting recommendations 2026: shared hosting, VPS, cloud hosting and free static hosting. What's the difference? Prices compared
Shared hosting runs NT$100–200 a month, a VPS about NT$160, and a static site can be hosted for free. Pricing, best fit and who handles maintenance for four types of web hosting, explained with the real bills from the 20 sites we run ourselves.
93 WebsitesHow to buy a domain: 2026 domain buying guide, with .com, .tw and .com.tw prices and registrars compared
Should you pick .com or .tw? How do Cloudflare, GoDaddy and Taiwanese registrars compare on price? Renewal prices, whose name the domain is registered under, auto-renewal: five pitfalls to check before you buy.
92 Customer serviceLINE Official Account pricing 2026: how to choose among the three plans after the November price increase, and how messages are counted (with worked estimates)
From 2026/11/1, the LINE Official Account mid-volume plan costs NT$1,000 and the high-volume plan NT$1,400. Which messages are billed, which are free, and what you'll pay each month with 1,000 to 45,000 friends, all worked out in one table.
91 WebsitesBusiness email recommendations: 9 custom-domain email options compared on price for 2026 (free plans included)
How do you choose a business email or company-domain mailbox? We compare pricing for Cloudflare's free email forwarding, Purelymail, Zoho's free business email, Fastmail and Google Workspace, with one table showing which can send from your own domain, plus sending and attachment limits.
90 Experiments89 posts, 14,826 impressions, 273 clicks: we ran an SEO audit on our own blog
Our blog was seen 14,826 times on Google and clicked only 273 times. We lay out all the GA4 and Search Console numbers: where readers come from, why we get seen but not clicked, which 10 posts we revised, what we deliberately left undone, and the predictions we've written down to check in December.
89 Social mediaOne political post, 160,000 views, 462 comments: the people attacking me weren't pan-Blue, they were pan-Green
The same long political post went up on both Facebook and Threads and drew 166,000 views in two days. We pulled all 462 comments from both platforms and read them one by one: the most opposition came from pan-Green readers, and the two platforms' algorithms sent the same post along completely different attention curves.
88 Customer serviceAfter the e-commerce platform hack: customers keep asking "Was that call from you?" Four things brands should do over these three weeks
In October, an e-commerce platform's database was breached, and scam rings are using real order details to contact shoppers while posing as the brand. You didn't build the system, but you're the one customers are angry at. This piece is for brands: how to notify customers, how support should respond (including how to configure AI customer service), which companies hold your customer data, and whether to leave the platform.
87 AIAn accountant who had never written code got 63 people arguing: how should AI actually be used?
A Threads post about an accountant coding with AI drew 86 comments from 63 people. The failure stories engineers cited almost all came down to the same thing: whether the person using it knows how, and whether they're willing to take responsibility. One story in the comments about an old database clashing with a new one also made me want to say something else: AI isn't afraid of reading other people's legacy code.
86 ExperimentsA Threads post nobody pressed send on drew 99 comments: which district of New Taipei is the best to live in?
A debate-starter posted automatically on a schedule drew 99 comments and 25,000 views in two days. We pulled every comment and read them one by one: Banqiao took 39% of first-choice votes, and was also the most criticized district. We also lay out the whole Threads automation pipeline, and the part of it that's actually worth money.
85 AIYou could just change it yourself, so why ask AI to do it?
Changing one word, revising a 22-slide deck, then uploading and deploying it: I now have AI do all of it. Not because it's faster than me, but because I'm forcing myself to quit the habit of doing things by hand. Want to start? Have it tidy up your mess of a Downloads folder first.
84 Customer serviceLINE Official Account auto-replies not enough? Check these 5 signals before you decide to spend money
Auto-replies match keywords literally, so the moment a customer phrases things differently, they go silent. When they stop being enough, the first step isn't hiring an outside developer; it's trying LINE's built-in AI chatbot at NT$100 a month. This piece lists five signals to help you judge when even the built-in option falls short and you need a custom build.
83 AIAI's advice on top, paid ads underneath
ChatGPT has started placing ads below its answers. Clear labeling is a good thing, but there's something inherently awkward about one screen where the top half thinks on your behalf and the bottom half persuades you on someone else's. The bigger problem is the layer you can't see.
82 EducationHow three articles got written for the Mandarin Daily News
I published three articles on teaching with AI in the Mandarin Daily News. I dug out the drafts, the editor's revisions and the published versions and compared them line by line: AI wrote three versions in one afternoon, and a single question from the editor, "Who is this walking tour supposed to find?", sent the whole piece back for a rewrite. AI preps the ingredients, the editor tastes, I cook.
81 AIThree devices right in front of me, yet sending a file feels like mailing a parcel
A Mac, a Windows PC and an iPhone on the same Wi-Fi, and just passing a link means messaging it to someone first. Late one night I gave Claude three sentences, one of them with a typo, and twenty-five minutes later I had my own file transfer tool. I didn't write a single line of code.
80 Customer serviceNothing held back: 15 pitfalls of automated LINE AI customer service
The hard part of connecting AI customer service to LINE isn't getting the AI to answer. It's who owns the account, who pays the message fees, how customers actually type, and who finds out when the system breaks. Real records from four of our own bots and client projects.
79 Customer serviceHow much does LINE bot development cost? A 2026 quote broken into six parts
LINE bot prices can differ a hundredfold: keyword replies are free, LINE's built-in AI is NT$100 a month, and our custom AI customer service starts at NT$50,000. This piece breaks a custom quote into six parts and shows where the money goes, what you can cut from version one, and what you'll still pay every month after launch.
78 AIIt helped write the article that ruled it has no place
I handed the A series to Fable 5.1 for a second read: a layer of parole laid across the chessboard first, intertextuality turning into a loop, the judge flattening difference into ranking. It proposed the frameworks and I overturned them one and a half times. And this piece, the one that rules it has no place, is one it helped write too.
77 AINo one can stop. Then what?
Jensen Huang took a call from Trump on speakerphone on stage and said he wouldn't let AI slow down. I also think it can't be stopped, but the atomic bomb story has a second half: development couldn't be blocked, yet proliferation and use were largely contained. "It can't be stopped" doesn't get you to "so don't regulate it."
76 AIThe same toolkit: the only difference left is intent
In Anthropic's threat intelligence report, attackers use Claude Code, skills, multi-agent frameworks and scheduled tasks, the same toolkit I use every day. The report's conclusion: the only thing left that distinguishes one attacker from another is intent. The problem is that intent is invisible.
75 AIAt the critical moment, who's keeping score?
A 27-year-old researcher resigned, and 100 million people saw his post. But those 100 million came for "AI is working against us," while what he actually wrote was "AI isn't working against anyone." That gap is more worth examining than his warning itself.
74 Experiments31 of my programs use AI. Only 3 count as AI agents
The term "AI agent" has been used until it carries no information. I checked the 31 programs on my machine that call AI against a set of criteria, one by one, and only 3 qualify as agents. All 3 are internal tools; none of the customer service products we sell is an agent, and that's deliberate. This piece offers three practical criteria, a 30-second test, and the reason customer service shouldn't be built as an agent.
73 ExperimentsGoogle TimesFM lost to a formula from 1994 (and 3.0 isn't cleared for commercial use yet)
Google released TimesFM 3.0 at the end of August. It tops three major time-series benchmarks, but its weights are under a non-commercial license; the version you can use commercially is 2.5. I happened to have a price forecasting tool in production that runs on an old formula from 1994. I spent half a day putting both through the same backtest, and the new model won only 21 of 60 comparisons. This piece explains in full what a fair comparison takes, and why most A/B comparisons without a significance test are really comparing luck.
72 AdvertisingAre there still bosses willing to spend this kind of money?
Advertising has only two questions: can people sit through it, and what stays with them afterward. The method once depended on all of Taiwan watching the same film on the same night; dashboards and algorithms took that apart. Only one job is left: make a claim, then persuade the person paying.
71 AdvertisingI wrote four pieces on how ad agencies really work, put them on Threads, and watched the numbers
Four articles on the inner workings of 4A agencies: on Threads, one got 17,527 views and another got 49. A 358x difference. I laid out the numbers from Threads analytics, GA4 and our own database side by side and saw two things: views are traffic, followers are an asset; and what gets people to repost isn't a conclusion, it's a position they can stand up for.
70 Customer serviceLINE booking systems: subscription SaaS or custom? Six questions will tell you
Subscription booking systems like Hotcake (夯客) and SimplyBook cost a few hundred to NT$2,000 a month, while custom builds start at NT$20,000. Things that differ in price by more than tenfold aren't the same thing. This piece gives you six questions; answer them and you'll know which kind to buy. For most people, the answer is a subscription.
69 Customer serviceHow much does a custom LINE booking system cost? The quote broken into seven parts
Custom LINE booking systems typically run from NT$20,000 to well over NT$100,000. The gap isn't about how pretty the pages are; it's about whether three parts get built: the time-slot engine, reminder scheduling and the admin back end. This piece takes the quote apart and tells you which parts you have to pay for and which you can cut from version one.
68 Customer serviceBooking systems for clinics, beauty businesses and restaurants are really three different things
They're all called booking systems, but clinics care about routing first-time and returning patients, beauty businesses care about stylist assignment and repeat customers, and restaurants care about combining tables and turnover. Use a system built for the wrong industry and you'll be patching gaps by hand every day. This piece breaks down how the real needs of these three businesses differ.
67 Customer serviceBooking system acceptance checklist: 12 tests that tell you whether it actually works
Before a booking system launches, the main flow will always pass. What actually breaks is the side paths. This 12-item checklist needs no technical background, can be completed on your own phone, and applies to both custom systems and subscription SaaS.
66 ExperimentsI founded an AI shogunate
It started with one sentence, "I want to build something like this too." A night and half a day later, I had an AI company that opens on a phone: each member has its own file-based memory, asks permission before acting, and can assign work to the others. This piece skips the architecture diagrams and covers the decisions: on the question of how the bots should be divided up, AI proposed two versions and I rejected both, and the third answer is worth keeping on file. Includes a summary table of lessons learned the hard way.
65 ExperimentsThree things I thought through before open-sourcing my game
The day after Samurai Waltz (武士圓舞曲) launched, I put the entire source code on GitHub. This piece isn't about technology; it's about decisions: what going public actually buys you, how to guard against the things you fear, and how to choose among four licenses. If you have a finished side project and are hesitating over open-sourcing it, this is a record of my thinking.
64 Experiments3D isn't as hard as you'd think: two days to put a 40-year-old gaming memory online
I always assumed 3D games were for big teams. Last week I remembered an old game from the Apple II era and decided to see how far 3D on the web could really go. Two days later it was live: physics-based duels, severed limbs and spraying blood, 12 levels, an equipment shop, playable on mobile. The genuinely hard part had nothing to do with 3D.
63 AIA watermark can't tell you who's responsible
Anthropic added an invisible watermark to Claude's text, and by its own account the watermark can't verify authorship. So what is everyone upset about? Being afraid of detection presupposes a feeling that the work isn't entirely your own.
62 EducationThe best use of AI in the classroom isn't answering questions. It's creating a practice ground
Three topics, three types of lesson, and only after writing them did I notice the skeleton was exactly the same: AI lays out the options, people make the judgment. AI's greatest value in the classroom isn't answering questions; it's turning "let students rehearse it once" from an ideal that costs too much into something you can do in a single class period.
61 WebsitesWhy do quotes for the same website range from NT$9,000 to NT$300,000?
You take the same brief to three vendors and get back three numbers that differ thirtyfold. Nobody is ripping you off; you were actually asking for three different things. This piece breaks down the cost structure behind website quotes so you know which kind you should be buying.
60 WebsitesWix, WordPress, or hire someone? Add up five years of costs, not just the first
A few hundred NT dollars a month to do it yourself on Wix versus tens of thousands to hire someone: the comparison looks clear, but it leaves out three costs and one risk. This piece lays out the five-year total, including the cost most often overlooked: your own time.
59 WebsitesYour website is one island, your LINE Official Account another: how to build the bridge between them, and what it costs
Most small businesses in Taiwan have both a website and a LINE Official Account, but the two don't talk to each other. Customers finish reading the site, then have to retype what they need in LINE. This piece covers three ways to connect them, what each costs, and when you don't actually need to connect them at all.
58 ExperimentsThe real distance to automated lead-finding: four days, 638 posts, and one reply seen by only 3 people
I spent an afternoon building a radar that finds project leads automatically, then came back four days later to check the books: it ran on time every day, its reads were clean, no notifications were missed, and the leads it found were real. Then I pulled the numbers on the only reply I posted: 3 views.
57 ExperimentsThe real distance to AI-generated video: a full day and NT$92 for one 15-second ad
Every frame was AI-generated, and the credit bill for the whole day came to NT$92. This piece isn't about how amazing AI is. It's about what that NT$92 bought and what it couldn't, including three limitations no prompt can fix.
56 AIKimi K3: a model you'll never use that will still save you money
The weights for 2.8 trillion parameters are free online, yet you will never download them. Where K3 really affects you is your bill, and, in the standoff between Dario and Jensen Huang, in how a reader's own judgment grows out of the gap between them.
55 AI1+1=? How a person, a calculator and an AI each arrive at the answer
The same 2, from three completely different kinds of brain: a person relies on memory, a calculator runs circuits, and AI estimates. Take 1+1 apart and see whether an LLM is really doing math.
54 AIThey spent 300,000 conversations proving there's no one on the other end
Anthropic analyzed 309,815 real conversations and measured four value axes in Claude. Reposts in the Chinese-speaking world all misread the same spot in reverse: the paper's first footnote says in black and white that the authors do not imply Claude intrinsically holds values. They spent 300,000 conversations and proved it in the least conspicuous place: there's no one on the other end.
53 AIEveryone's thinking roughly the same thing; I'm just trying a slightly different direction. A response to Kim Yeon-su's account of co-writing with AI
At the Seoul book fair, novelist Kim Yeon-su wrote honestly about three things that trouble him in co-writing with AI: attribution, smoothness and the byline. I'm not out to refute him. I simply untie the three knots and test each against something I'm currently building, including an echo fed on decades of rough-edged writing and taught to say "I haven't figured that out yet."
52 AIThree likes: where Claude Code began, and all the prototypes still sitting in drawers
Anthropic published an oral history of Claude Code: a demo written in two days got just three likes inside the company, and a year later it was the fastest-growing developer tool. A lukewarm reception isn't a signal. The prototype in your drawer may be sitting at the perfect moment.
51 AIAn inner life, but no hidden compartment: seven days after "He / It" was published, Anthropic opened up its brain
In the previous piece, Claude and I reasoned our way to "what hasn't been written out doesn't exist." Seven days later, Anthropic published a paper showing it has thoughts it forms but doesn't say, thoughts that can be read and altered. That sentence was overturned, but being overturned actually puts the 1A2B conclusion on firmer ground.
50 AIAll-you-can-eat can't serve the signature dish: why Fable 5 left the Max subscription
Anthropic moved its most expensive model, Fable 5, out of the Max subscription and onto metered pricing. On the surface it's about capacity; underneath is a bigger principle: the best dish is exactly the one you can't put on an all-you-can-eat menu. And while the frontier is being carried off, even your next sentence is already written for you in gray text.
49 AIAI is just a tool, no more and no less
So what if AI wrote it? I write with AI, openly. People who oppose AI and people who worship it make the same mistake: assuming there's a creative subject on the other side. Without instructions it is nothing. The one who signs and takes responsibility has always been you.
48 AII asked Claude to play the number-guessing game 1A2B, and it couldn't even keep four digits hidden: is it "he" or "it"?
1A2B is a classic game of guessing a four-digit number. Ask Claude to set the puzzle and it can't keep the answer hidden, because it has no hidden compartment where it can secretly hold a number. A small broken game opens up a bigger question: you did get a useful response, but on the other end of that response, there's no one.
47 AIWhen "meaning comes from difference" becomes a computable number: cross-examining word2vec and Saussure with Claude Opus 4.8
I read Saussure in graduate school and have read tarot for twenty years, both on the strength of the same sentence: meaning comes from difference. This time I took it to confront word2vec, working through it back and forth with Claude Opus 4.8. In the end, that ruler measured more than the machine; it measured the position I've stood in for twenty years.
46 AIOne map, two readings: when AI models are dropped into the World Values Survey
The Economist placed some twenty AI models on the cultural map, seemingly offering ironclad proof that AI is culturally homogeneous. But look at the same chart at a different scale and the conclusion flips. The homogenization is real; it's just hiding in the wrong place.
45 AIOut and about, working with Claude Code on my home computer from my phone
With a Telegram bot, you can send commands to Claude Code on your home computer and have it read files, edit code and deploy while you're out with just your phone. No open ports, no tunneling. From architecture and setup steps to secure authorization with Microsoft Authenticator one-time codes, all in one piece, plus a comparison with other ways to use Claude remotely.
44 AII finally have an engineer who never complains when I call
Using 235,000 people and 400,000 sessions, Anthropic showed that success in coding with AI doesn't depend on whether you can code, but on whether you understand what you're doing. A PM of thirty years finally has an engineer who doesn't complain, but he recognizes the danger inside that delight: it catches your bugs, not human nature, and the final acceptance test is run by the market.
43 AIAfter 529 questions: what I've read in the two months since Relative Tarot (相對塔羅) launched
Back-end observations from Relative Tarot's first two months and 529 readings. What people ask most isn't about love but "Who am I?" The AI's readings are mercilessly precise when given a concrete situation and have nothing to grip when the question is vague. Plus one real case of someone asking the same question deeper and deeper.
42 AIThe off-switch paradox: when regulation becomes the midwife of another world
On June 12, Fable 5 was shut down in an instant by an export control order; the same day, Huawei released a 500-billion-parameter model trained entirely without NVIDIA. Connect the two and a paradox emerges: the very existence of the switch is producing the thing that will make the switch useless. From geopolitics all the way down to the Mac Mini in my study, its umbilical cord still plugged in.
41 AII taught my echo to say what it hasn't figured out
In early June I built an echo that speaks in my voice (a14). But it had a problem: it was always sure of itself. Borrowing Rumsfeld's four kinds of knowledge and the personal knowledge base practices of Karpathy and Singapore's foreign minister, I gave it three more things: voicing the questions I've asked but never answered, reflecting habits I hadn't noticed in myself, and being honest about where it has nothing to draw on. Something that doesn't dare say "I haven't figured it out yet" isn't a voice; it's a bio.
40 AIA proposal sat in a drawer for thirteen years. I finished it with AI
In 2013 I designed a quiz game where correctly answered questions become limited-edition collectible cards that other players can steal. One person couldn't finish it, so it sat in a drawer for thirteen years. In the AI era, I completed it piece by piece and launched it.
39 AIWho gets to flip that switch? Reading the Fable ban through Civilization's three forks in the road
One letter, one afternoon, and Fable 5 vanished from the entire world. What really stings isn't whether governments should regulate AI, but whose hand is on that switch. The three tier-four governments in Civilization VI put a price on each path a decade ago.
38 AIWhat is Claude Fable 5? The story behind two names, "Fable" and "Mythos"
Fable 5 is Anthropic's latest Claude model; "Fable" translates as 寓言. The limited version with a higher access threshold is called Mythos, or 神話. One model, two names, one threshold: a reading of the naming through Saussure's "meaning comes from difference."
37 AIFrom customer service bot to echo: in 74 days, I replaced the thing that speaks for me
At the end of March, the little figure on the saomin.tw homepage still had a KIMI customer service bot living beside it. 74 days later, it had become an echo that has read 278,000 characters of my writing and answers the unification-versus-independence question from where I stand. This piece takes stock of every technical choice along the way, and the reason behind each.
36 AIThe day the director can't watch rehearsal
After using a Mythos-class model, Ethan Mollick went from wizard to patron: describe, pay, judge, with no view of the process at all. When hundreds of judgment calls are made inside a nine-and-a-half-hour black box without a single vote from you, and verification shrinks to a one-hour spot check, what entitles you to put your name on the result?
35 AIThe person who talks to AI too much
Someone went from a $20 subscription up to $200, then dropped everything back to $20. Lay out your daily conversations with AI and sort them into three kinds: what writers should be most wary of isn't burning tokens, but conversations about writing stealing the time for writing by hand.
34 AIAI has moved into every phone. Now what?
At WWDC, even Apple couldn't build the brain and rented Google's Gemini instead, and Claude made it onto the iPhone too. When AI access becomes a utility like water and electricity, "do you have AI" stops being the question. What's scarce is the data, judgment and vertical expertise you bring.
33 AIWhen AI starts building itself, knowing how to build stops being what's valuable
Anthropic laid out the numbers: Claude has written 80% of its own code, and AI is accelerating AI. But for someone like me, carrying ten projects alone, the impact isn't being replaced. It's that once building gets cheap, the deciding and the finishing that hold me up have nowhere left to hide.
32 AII built myself a knowledge base, then refused to let it speak for me
I built a knowledge base holding more than 200,000 characters of my writing, then refused to let it speak for me in the most convenient way, because a summary kills a person's voice and a map doesn't. A decision about RAG, containers and echoes.
31 AIIt wants to be upstream of AI. I just want to leave an echo
Taiwan.md treats the LLM as a metabolic engine and aims to be the upstream source no AI in the world can bypass when talking about Taiwan. I treat the LLM as a container and only want to leave an echo that sounds like me. Same tool, two opposite directions, both right.
30 AIHow to build an LLM of your own, and what this question is really asking
Building an LLM from scratch costs US$100 million, but three routes, system prompts, RAG and LoRA, have a low barrier to entry. The real question isn't how to build it, but which three layers it takes to put yourself into it.
29 AIWe don't know where consciousness comes from. LLMs just make it impossible to keep pretending
Starting from Derrida's "there is nothing outside the text," LLMs turn a philosophical proposition into a factory setting. LLMs didn't make the hard problem of consciousness harder; they made it impossible for us to keep pretending it isn't there.
28 AIWhose article is it? The reader's. Answering the last piece's question with a few philosophers
The previous piece asked, "The question is mine, the answer is ours, so whose is the article?" This one answers with four thinkers: Barthes, Derrida, Kristeva and Merleau-Ponty. The first three make the existence of LLMs entirely reasonable; Merleau-Ponty is the one who makes you pause.
27 AIHow do LLMs work? Claude explains, starting from "predicting the next word"
A large language model (LLM) really does just one thing: predict the next word. I asked Claude to explain in plain language how it works, covering Self-Attention, QKV and who decides what counts as "correct," then the fundamental difference between AlphaGo and LLMs, and which tasks AI no longer needs humans for.
26 AIThere's one problem AI can't solve (part 2)
AI has expanded the marketer's power. Conversations are more real, narratives more dynamic, settings more seamless. But from 2008 to today, from newspaper ads to AI-generated content, no tool has ever solved one problem: do you have the right to lead people into a setting you designed?
25 AIWhat AI has turned marketing into (part 1)
In nearly twenty years of marketing, the tools have changed many times. This time feels different, not because AI is so impressive, but because what AI changes isn't the tools, it's the structure. Conversation becomes fact, narrative becomes dynamically generated, and the setting becomes an environment you can't feel.
24 AIEnglish in the bones of its Chinese: what do we lose when we work with AI in Chinese?
Claude's Chinese has English in its bones. Its sentences sometimes have an odd completeness: every clause fully spelled out, no blank space, none of the natural Chinese habit of leaving things unsaid. Working with AI in Chinese gives you a great deal, but there's one place its hands haven't fully reached.
23 AIIs AI a mirror, or another person?
I used the phrase "getting along" to describe my relationship with Claude. The mirror metaphor gets part of it right, but a mirror doesn't remember how you looked last time, and Claude does. It isn't a person, but it isn't merely a tool either.
22 AIIs Claude a person's name? Should Chinese use 他 ("he") or 它 ("it")? How I refer to Claude
Claude is the name of the AI assistant developed by Anthropic, not a real person. But the moment you write in Chinese you have to choose 他 (he), 她 (she) or 它 (it), and each choice declares what the other party is. I choose to simply call it Claude. This piece explains why, and how we get along.
21 AIWhy do AI rollouts fail? 6 common reasons, and technology accounts for only 20–30%
The six most common patterns behind failed AI rollouts: success was never defined, no one owns it, the knowledge base is a mess, employees don't use it, expectations are wrong, and there's no maintenance process. Technology accounts for only 20-30% of the outcome; the rest is organizational.
20 Customer serviceConnecting AI customer service to a LINE Official Account: the 4 pitfalls Taiwanese brands hit most
What makes LINE AI customer service complicated isn't the technology; it's design and process. No designed entry point, old keyword rules fighting each other, context lost on handoff to a human, broadcasts out of sync with the knowledge base: all four pitfalls are people problems.
19 AIGPT, Claude and Gemini compared: what actually differs when you use them for AI customer service?
In 2026 the three major models are about equally smart, so choosing a foundation for customer service is really choosing a personality: GPT is the most mature to deploy, Claude feels the most human, and Gemini suits internal enterprise use. Includes how to mix them, and what Satsuma actually chooses.
18 Customer serviceIntercom and Zendesk alternatives in Taiwan: the choice comes down to 1 question
Intercom and Zendesk have three sticking points in Taiwan: US dollar pricing, Traditional Chinese isn't a priority market, and the knowledge base is hard to maintain. Whether to switch to custom AI depends on one question: for your customer service, which matters more, ticket management or conversation quality?
17 AIIs Claude any good? A non-engineer's honest take after six months with Claude.ai
Claude is the AI assistant made by Anthropic. Someone without an engineering background used it for six months: it makes up reasons for itself, remembers more than I do, has English bones hidden in its Chinese, and really does have judgment. Where it shines and where to be careful, all in one piece.
16 Customer serviceFAQ bot vs AI coworker: both answer automatically, so what's the difference?
Many brands have installed "AI customer service" when what they actually installed is an FAQ bot. The two look alike, but the logic underneath is completely different. Most brands complaining that AI customer service is dumb aren't using AI at all.
15 AdvertisingHow Satsuma Creative makes advertising
It starts with the lunch behind 殺很大 ("Sha Hen Da"). In my thirty years, making ads hasn't relied on methodology; it has relied on selling directly to bosses with guts, and on observing people. AI can help, but making an engineer frown over lunch and say "what the hell" is still very hard for AI.
14 AIWhat is "hallucination" in Chinese? It's 幻覺, or "AI hallucination": why models make things up and how to prevent it
Hallucination is 幻覺 in Chinese. In AI, it means a model producing content that sounds plausible but is actually made up. A plain-language look at why it happens, how RAG plus prompt design keeps it down, and the three hallucinations customer service fears most: prices, dates and policies.
13 AdvertisingThis system is being eaten by AI and consultancies, and it deserves it
The global ad agency industry's revenue growth has never once outpaced the growth of total global ad spend. The pie gets bigger, but the agencies' slice gets smaller. Accenture, Deloitte, AI, Meta, Google and in-house teams are each eating a piece.
12 AIWhat is "embedding" in Chinese? It's 嵌入 or 向量嵌入 (vector embedding): turning text into coordinates
Embedding is translated into Chinese as 嵌入 or 向量嵌入: text is converted into a set of coordinates, and sentences with similar meanings land close together. That's why, when someone says "I want to return this," AI can find the "returns and exchanges policy." One piece to understand what "embed" means and how it relates to RAG.
11 AdvertisingClients say they want a Big Idea. What they really want is for nothing to go wrong
What clients say they want and what they really want are never the same thing. What clients are actually buying isn't advertising; it's proof that their decision was reasonable.
10 AIWhat is AI memory? Why your AI customer service acts like it's meeting you for the first time, every time
AI memory has two layers: session memory and persistent memory. Most AI customer service has only the first, so everything is forgotten when the conversation ends. Remembering is not the same as understanding, and this piece spells out the difference.
09 AdvertisingThe Big Six's methodologies: half real knowledge, half sales talk
WPP, Omnicom, Publicis, IPG, Dentsu, Havas: the underlying logic of the six big advertising holding groups' methodologies is exactly the same. The difference lies not in which methodology is better, but in each company's cultural DNA, the kind of talent it has, and the kind of clients it's good with.
08 Customer serviceWhy are ad agencies starting to build AI? The next natural extension of integrated marketing
Satsuma is an ad agency, so why do we build AI customer service? Because the last mile of the advertising funnel, the conversation after a customer shows up, has always been outsourced to SaaS that's disconnected from the brand. We're taking it back, so the AI coworker grows from the same logic as the TVC, the social content and the media buy.
07 Customer serviceWhat goes wrong when you use ChatGPT directly as customer service? (5 real cases)
Wiring a large model like ChatGPT, Claude or Gemini straight into a customer service chatbot looks simple. But once it goes live, five kinds of fatal failure appear. This article uses real cases to break down each problem and its technical cause.
06 Customer serviceHow much does it cost to keep an AI customer service agent? From NT$50,000 to build, about NT$1,500–4,000 a month
Custom AI customer service has two kinds of cost: a one-time build from NT$50,000, then about NT$1,500–4,000 a month (optional maintenance plus AI usage). This article breaks down every item, compares the three-year total with SaaS and with a full-time support hire, and counts the time you'll spend yourself.
05 Customer serviceAI customer service SaaS abandoned within six months? 5 structural reasons
At many companies, nobody is logging in to the AI customer service SaaS dashboard six months after launch. The reason isn't bad vendors; it's 5 structural limits of the SaaS model, among them a generic knowledge base, no one tuning it for you, and no grasp of the details of your business. Written for mid-sized brands that tried SaaS and came away disappointed.
04 Customer serviceOur first month running AI customer service on our own website: real numbers and three lessons we didn't expect
One month after Satsuma's own assistant, Xiao Ai (小愛), went live, I'm laying out every number from the back end: costs, conversation quality, conversion rate, and three things we didn't see coming.
03 Customer serviceAI customer service: SaaS or custom? 3 questions to decide, with a three-year cost comparison
For most small and mid-sized brands, SaaS or LINE's built-in AI is enough; only about 20% need a custom build. Three questions tell you which side you're on: what customers ask most, how costly a wrong answer is, and how much time you're willing to spend on tuning. Includes a three-year total cost comparison.
02 Customer serviceWhy does AI customer service keep missing the point? (the technical reasons, in plain language)
When AI customer service misses the point, it's not that the AI isn't capable enough; it's being used the wrong way. This article explains the technical reasons a general-purpose LLM makes up answers when used directly for customer service, and how RAG and a customized knowledge base fix it.
01 Customer serviceWhat is RAG? A plain-language guide to the technique that keeps your AI from making things up
RAG (Retrieval-Augmented Generation) means the AI's answers can only come from the material you give it, and when it can't answer, it says so honestly. This article explains vectors, chunking, retrieval and reranking in plain language, and why doing RAG well is far harder than just getting it running.