Blog / Blog
Observations on AI coworkers, AI customer service, advertising and integrated marketing, and living with AI. Saves readers time, and saves us time too.
How three articles for Mandarin Daily News got written
Three AI teaching articles ran in Mandarin Daily News. I lined up the drafts, the editor's revisions, and the published versions sentence by sentence: AI wrote three versions in one afternoon, then one question from the editor ("Who is this walk supposed to help us 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, and sending a file felt like mailing a package
Mac, Windows, and iPhone on the same Wi-Fi, and passing a link still meant asking someone for help. Late one night I gave Claude three sentences, one with a typo. 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 LINE AI customer service automation
Connecting AI customer service to LINE, the hard part isn't the AI answering questions. It's who owns the account, who pays the messaging fees, how customers type, and who finds out when the system breaks. Real notes from four in-house bots and client projects.
79 Customer serviceHow much does LINE bot development cost? Our 2026 quote broken into six parts
LINE bot pricing can vary 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 post breaks a custom quote into six parts: where the money goes, what you can cut from version one, and what you'll keep paying monthly after launch.
78 AIIt helped write the essay that ruled it has no position
I had Fable 5.1 reread the A series: a layer of parole already laid on the board, intertextuality turned into a loop, a referee flattening difference into ranking. It proposed frameworks; I overturned them one and a half times. And this essay, which rules it has no position, it helped write.
77 AINobody can stop it. Now what?
On stage, Jensen Huang took a call from Trump on speakerphone and said AI won't slow down. I agree 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. "Can't stop it" doesn't lead to "leave it alone."
76 AISame 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 concludes that the only thing distinguishing one actor from another is intent. The problem is, intent is invisible.
75 AIAt the critical moment, who's keeping score?
A 27-year-old researcher resigned, and 100 million people saw the post. But what those 100 million came for was "AI is working against us." What he actually wrote was "AI isn't working against anyone." That gap is more worth examining than the warning itself.
74 Experiments31 programs use AI. Only 3 count as AI agents
"AI agent" has been used until it means nothing. I checked all 31 AI-calling programs on my machine against a set of criteria, one by one. Only 3 qualify as agents, all for internal use. None of the customer service bots we sell are agents, and that's deliberate. This post offers three practical criteria, a 30-second test, and why customer service shouldn't be built as an agent.
73 ExperimentsGoogle TimesFM lost to a 1994 formula (and 3.0 isn't licensed for commercial use yet)
Google released TimesFM 3.0 in late August, topping three major time-series benchmarks, but the weights carry a non-commercial license; only 2.5 is cleared for commercial use. I happened to have a price-forecasting tool running on a 1994 formula. I spent half a day putting both through the same backtest: across 60 comparisons, the new model won just 21. This post lays out what a fair comparison takes, and why "A/B comparisons without significance testing are mostly measuring luck."
72 AdvertisingAre there still bosses willing to spend this kind of money?
Advertising comes down to two questions: will people keep watching, and what stays with them afterward. The method once worked because the whole of Taiwan watched the same spot on the same night; dashboards and algorithms took it apart. What's left is one job: make a case, then convince the people paying for it.
71 AdvertisingI wrote four inside stories about ad agencies, posted them on Threads, and watched the numbers
Four articles on what really happens inside 4A agencies: on Threads, one got 17,527 views, another got 49. A 358x gap. I laid out the numbers from Threads analytics, GA4, and our own database, and two things emerged: views are traffic, followers are assets; and what gets shared isn't a conclusion, it's a position people can stand behind.
70 Customer serviceLINE booking systems: monthly SaaS or custom? Six questions to decide
Subscription booking systems like Hoko and SimplyBook run from a few hundred to NT$1,000–2,000 a month; custom builds start at NT$45,000. Things with a 30x price gap aren't the same product. Answer these six questions and you'll know which to buy. For most people, the answer is the subscription.
69 Customer serviceHow much does a custom LINE booking system cost? Our quote broken into seven parts
Custom LINE booking systems typically cost NT$45,000 to well over NT$100,000. The difference isn't how polished the pages look; it's whether the scheduling engine, reminder system, and admin dashboard are built. This post breaks down the quote: which parts are essential and which can be cut from version one.
68 Customer serviceClinics, beauty salons, and restaurants need three different things from a booking system
They're all called booking systems, but clinics care about separating first and follow-up visits, salons care about stylist assignment and repeat customers, and restaurants care about shared tables and turnover. Pick a system built for the wrong industry and you'll be patching holes by hand every day. This post breaks down the real differences across three kinds of business.
67 Customer serviceBooking system acceptance checklist: 12 tests that tell you whether it works
Before a booking system launches, the main flow always passes. What actually breaks is the edge cases. This 12-test checklist needs no technical background; you can run it yourself on your phone. Works for both custom systems and subscription SaaS.
66 ExperimentsI opened an AI shogunate
It started with "I want to build something like that too." A night and a half-day later, I had an AI company that opens on my phone: each member has its own file-based memory, must ask permission before acting, and can delegate work to the others. This post skips the architecture diagram and focuses on decisions: AI proposed two versions of "how should the bots be divided," I rejected both, and the third answer is worth keeping. Includes a summary of lessons learned.
65 ExperimentsThree things I thought through before open-sourcing my game
The day after Samurai Waltz launched, I published the entire source code on GitHub. This post isn't about tech; it's about decisions: what going public actually gets you, how to guard against what you fear, and how to choose among four licenses. If you have a finished side project and are on the fence about open-sourcing it, this is my thinking on record.
64 Experiments3D isn't as hard as you think: two days to bring a 40-year-old game 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 web 3D could go. Two days later it was live: physics-based duels, severed limbs and blood, 12 levels, an equipment shop, mobile support. The truly hard part had nothing to do with 3D.
63 AIWatermarks can't tell you who's responsible
Anthropic added invisible watermarks to Claude's text, and says itself that they can't identify the author. So what's everyone upset about? Fearing detection presupposes a feeling that the thing isn't entirely yours.
62 EducationAI's best use in the classroom isn't answering questions. It's building a practice ground
Three topics, three lesson formats, and only after writing them did I see the skeleton was identical: AI lays out the options, people make the call. AI's greatest classroom value isn't answering questions. It's turning "let students rehearse it once" from a costly ideal into something one class period can deliver.
61 WebsitesSame website, quotes from NT$9,000 to NT$300,000. Why?
You send the same brief to three firms and get back numbers 30x apart. Nobody's ripping you off; you were actually asking about three different things. This post breaks down the cost structure of website quotes so you know which kind to buy.
60 WebsitesWix, WordPress, or hire someone? Do the five-year math, not just year one
DIY on Wix costs a few hundred a month; hiring someone costs tens of thousands. The comparison looks clear, but it leaves out three costs and one risk. This post lays out total cost over five years, including the one most often overlooked: your own time.
59 WebsitesYour website is one island, your LINE Official Account another: how to build the bridge, and what it costs
Most small businesses in Taiwan have both a website and a LINE Official Account, but the two don't talk. Customers who browse the site have to retype their request in LINE. This post covers three ways to connect them, the cost of each, and when you don't actually need to.
58 ExperimentsThe real distance of automated lead hunting: four days, 638 posts, one reply seen by 3 people
I spent an afternoon building a radar to find projects automatically. Four days later I checked the books: it ran on time every day, classified cleanly, missed no notifications, and the leads it found were real. Then I pulled the stats on the one reply I posted — 3 views.
57 ExperimentsThe real distance of AI video: a full day and NT$92 for a 15-second ad
Every frame AI-generated, and the credit bill for the day came to NT$92. This post 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
2.8 trillion parameters of weights released free online, and you'll never download them. Where K3 really affects you is your bill — and in the standoff between Dario and Jensen Huang, how a reader's judgment grows through the cracks.
55 AI1+1=? How humans, calculators, and AI each get the answer
The same 2, three completely different minds: humans use memory, calculators run circuits, AI estimates. Take 1+1 apart and see whether LLMs are actually doing math.
54 AIThey spent 300,000 conversations proving there's no one on the other end
Anthropic analyzed 309,815 real conversations and mapped four axes of Claude's values. Chinese-language reposts all misread the same point: the paper's first footnote states plainly that they don't imply Claude intrinsically holds values. They spent 300,000 conversations and, in the least conspicuous place, proved there's no one on the other end.
53 AIWe all think roughly alike; I just tried a slightly different direction: a response to Kim Young-ha's...
At the Seoul International Book Fair, author Kim Yeon-su candidly described three struggles of co-writing with AI: ownership, smoothness, and authorship. I'm not here to refute him. I untied the three knots and pushed back on each with something I'm building — including a font fed decades of edges, and an echo taught to say "I haven't figured this out yet."
52 AIThree likes: the origin of Claude Code, and every prototype still sitting in a drawer
Anthropic published an oral history of Claude Code: a demo built in two days got just three likes internally, and a year later became the fastest-growing developer tool. Indifference isn't a signal. The prototype in your drawer may be sitting on perfect timing.
51 AIAn inner life, but no hidden compartment: seven days after "He/It," Anthropic opened its brain
In my last post, Claude and I concluded that "what isn't written out doesn't exist." Seven days later, Anthropic published a paper showing it has thoughts it doesn't say — readable and editable. That claim was overturned, but the overturning actually makes the 1A2B conclusion stand firmer.
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, off the Max subscription and onto metered billing. On the surface it's about capacity; underneath is a bigger principle: the best dish is exactly the one that can't go on the buffet. And as the frontier is taken off the table, even your next sentence has gray text writing it for you.
49 AIAI is just a tool, no more, no less
What's wrong with AI-written? I write with AI, openly. AI skeptics and AI worshippers make the same mistake: assuming there's a creating subject on the other side. Without instructions it's nothing. The one who signs and takes responsibility has always been you.
48 AIHe/It: a conversation that started with a number-guessing game
Ask Claude to play 1A2B and it can't even hide a four-digit number, because it has no hidden compartment. A broken little game uncovered something bigger: you really do get useful responses, but there's no one on the responding end.
47 AIWhen "meaning comes from difference" becomes a computable number: confronting word2vec and Saussure with Claude Opus 4.8
I studied Saussure in grad school and have read tarot for twenty years, both resting on the same idea: meaning comes from difference. This time I tested it against word2vec, picking it apart with Claude Opus 4.8. In the end, the 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 twenty-odd AI models on a cultural map, seemingly proving beyond doubt that AI is culturally homogeneous. But read the same chart at a different scale and the conclusion flips. The homogenization is real; it's just hiding in the wrong place.
45 AIAway from home, connecting from your phone to Claude Code on your home computer
Using a Telegram bot to send commands to Claude Code on my home computer, so I can have it read files, edit code, and deploy from my phone while I'm out. No open ports, no tunneling. Architecture, setup steps, and secure authorization with Microsoft Authenticator codes, all in one post, plus a comparison with other ways to use Claude remotely.
44 AII finally have an engineer who never complains when I call
Anthropic used 235,000 people and 400,000 sessions to show that success with AI coding 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 never complains — but sees the danger inside the joy: it catches your bugs, not people's hearts; and the one who signs off in the end is the market.
43 AIAfter 529 questions: what I learned two months after launching Relative Tarot
Backend observations from two months and 529 readings on Relative Tarot: what people ask most isn't about love but "who am I"; AI readings are mercilessly precise when given a concrete situation and have nothing to grip on vague questions; plus a real case of someone asking the same question deeper and deeper.
42 AIThe switch paradox: when control becomes the midwife of another world
On June 12, Fable 5 was instantly shut down by an export control; 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 what will make the switch useless. From geopolitics all the way to the Mac Mini in my study, still tethered by its umbilical cord.
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 certain. Borrowing Rumsfeld's four kinds of knowledge and the personal knowledge base approaches of Karpathy and Singapore's foreign minister, I gave it three more things: voicing questions I've asked but never answered, reflecting habits I hadn't noticed, and being honest about where it has nothing to go on. Something that won't say "I haven't figured this out" isn't a voice. It's a bio.
40 AIA proposal that sat in a drawer for thirteen years, finished with AI
In 2013 I designed a quiz game: questions you answer correctly become limited-edition collectible cards that others can steal. It was too much for one person, so it sat in a drawer for thirteen years. In the AI era, I filled it in piece by piece and launched it.
39 AIWho gets to flip the switch? The Fable ban through Civilization's three paths
One letter, one afternoon, and Fable 5 vanished from the world. What really stings isn't whether governments should regulate AI, but whose hand is on the switch. Civilization VI's three tier-four governments labeled the price of each path ten years ago.
38 AIFable and Mythos: what I'm really thinking about the Fable 5 release
Anthropic named its strongest model "Fable" and its limited edition "Mythos." One model, two names, one threshold. After all those years reading Saussure, I never expected the cleanest case of "meaning comes from difference" to show up in an AI company's product line.
37 AIFrom customer service to echo: in 74 days, I replaced what speaks for me
At the end of March, the little figure on the saomin.tw homepage was still next to a KIMI customer service bot. 74 days later, it had become an echo that's read my 278,000 words and answers questions on Taiwan's independence and unification from my standpoint. This post walks through every technical choice along the way, and why.
36 AIThe day the director can't see rehearsal
After using a Mythos-class model, Ethan Mollick went from wizard to patron: describe, pay, judge, with the process entirely hidden. When nine and a half hours inside an AI black box contain hundreds of decisions you never voted on, and verification degrades to an hour of spot checks, by what right does the result carry your name?
35 AIThe person who talks to AI too much
Someone went from a $20 subscription to $200, then cut everything back to $20. Sort your daily AI conversations into three kinds, and what writers should watch for most isn't burning tokens. It's how "conversations about writing" steal time from writing by hand.
34 AIAI moved into every phone. Now what?
At WWDC, even Apple couldn't build its own brain and chose to rent Google's Gemini, and Claude landed on the iPhone too. When AI access becomes a utility, "do you have AI" is no longer the question. What becomes scarce is the data, judgment, and vertical expertise you bring.
33 AIWhen AI starts building itself, "knowing how to make it" stops being valuable
Anthropic laid out the data: Claude has written 80% of its own code, and AI is accelerating AI. For me, running ten projects alone, the impact isn't being replaced. It's that when building gets cheap, the "deciding" and "finishing" I get stuck on can no longer hide.
32 AII built myself a knowledge base, then refused to let it speak for me
I built a knowledge base holding over 200,000 of my words, then refused to let it speak for me the easy way, because summaries kill a person's voice and maps don't. A decision about RAG, containers, and echoes.
31 AIIt wants to be AI's upstream. I just want to leave an echo
Taiwan.md treats the LLM as a metabolic engine, aiming to be the upstream source every AI must pass through when discussing Taiwan. I treat the LLM as a container, wanting only an echo that sounds like me. Same tool, opposite directions, and both are right.
30 AIHow to build an LLM of your own, and what the question is really asking
Building an LLM from scratch costs $100 million, but the three paths of system prompts, RAG, and LoRA have low barriers. The real question isn't how; it's what three layers "putting yourself into it" requires.
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 claim into a factory setting. LLMs don't make the hard problem of consciousness harder. They make it impossible to keep pretending it isn't there.
28 AIWhose article is it? The reader's: answering the last post's question with a few philosophers
My last post asked, "The question is mine, the answer is ours, so whose article is it?" This one answers with 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? From predicting the next word to understanding the world, I asked Claude itself
I spent an afternoon asking Claude how LLMs actually work: from "predicting the next word" to self-attention and QKV, and then to a more uncomfortable question — what things AI no longer needs humans for.
26 AIOne problem AI can't solve (part 2)
AI has expanded marketers' power: more genuine conversations, more dynamic narratives, more seamless experiences. 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 scene you designed?
25 AIWhat AI has turned marketing into (part 1)
Nearly twenty years in marketing, and the tools have changed many times. This time feels different — not because AI is so capable, but because what AI changes isn't the tools, it's the structure. Conversation becomes fact, narrative becomes dynamically generated, and the scene becomes an environment you can't feel.
24 AIEnglish in the bones of Chinese: what do we lose working with AI in Chinese?
Claude's Chinese has English in its bones. Its sentences sometimes have an odd completeness: every clause spelled out, no blank space, none of the natural "unfinished places" of Chinese. Working with AI in Chinese, you gain a lot, but there's one place its hand hasn't fully reached.
23 AIIs AI a mirror, or another person?
I used the word "relationship" to describe how I am with Claude. The mirror metaphor gets part of it right, but a mirror doesn't remember how you looked last time. Claude does. It isn't a person, but it isn't just a tool either.
22 AIIs Claude a name? How I refer to Claude, and how we get along
In Chinese, choosing between "he," "she," and "it" declares what the thing is. I choose to keep calling it Claude. Not he, not she, not it, because the word "Claude," as it weighs today, is enough.
21 AIWhy do AI rollouts fail? 6 common causes, and technology is only 20–30% of it
The six most common patterns of failed AI rollouts: no definition of success, no owner, a poor knowledge base, employees who don't use it, wrong expectations, and no maintenance process. Technology accounts for only 20–30% of success or failure. The rest is organizational.
20 Customer serviceConnecting AI customer service to LINE Official Accounts: 4 pitfalls Taiwanese brands hit most
What's complicated about LINE AI customer service isn't the technology; it's design and process. An undesigned entry point, clashing legacy keyword rules, context lost in handoff to a human, broadcasts out of sync with the knowledge base — all four pitfalls are people problems.
19 AIGPT vs Claude vs Gemini as a customer service foundation: what are the real differences?
Choosing a model isn't choosing "who's smarter." In 2026 the intelligence gap among the three is too small to be the only criterion. What you're really choosing is a worldview.
18 Customer serviceIntercom and Zendesk alternatives in Taiwan: one question decides
Intercom and Zendesk hit three snags in Taiwan: USD pricing, Traditional Chinese isn't a priority market, and knowledge bases are hard to maintain. Whether to switch to custom AI comes down to one question: does your customer service depend more on ticket management or conversation quality?
17 AISix months living with Claude.ai
A non-engineer spends six months with Claude. From compatibility charts and confabulation to memory and the English in the bones of its Chinese — what is this AI thing, really, and what's my relationship with it?
16 Customer serviceFAQ bot vs AI coworker: both answer automatically, so what's the difference?
Many brands have installed "AI customer service" that's actually 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 does advertising
It starts with that lunch behind "Kill Big." For thirty years, my advertising hasn't relied on methodology but on selling directly to bosses with nerve, and on observing people. AI can help, but making an engineer frown at lunch and say "what the hell" is still hard for AI.
14 AIWhat is hallucination? What AI hallucination means, why AI makes things up, and how to fix it
AI isn't just talking nonsense; it's constructing a plausible-sounding story. This is called confabulation. RAG plus strict prompt design can greatly reduce it, but never to zero. The most dangerous hallucinations in Taiwanese customer service: prices, dates, and policy details.
13 AdvertisingThis system is being eaten by AI and consultancies, and it deserves it
Global ad agency revenue growth has never outpaced growth in total global ad spend. The pie grows, but agencies get a smaller slice. Accenture, Deloitte, AI, Meta, Google, and in-house teams each take a bite.
12 AIWhat is embedding? Vector embeddings in plain language: how AI understands your knowledge base
Embedding turns text into coordinates so sentences with similar meanings sit close together. That's why when someone asks "I want to return this," AI can find "Return and exchange policy," even though the two share no words.
11 AdvertisingClients say they want a Big Idea. What they really want is nothing going wrong
What clients say they want and what they actually want are never the same thing. What clients are really 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
AI memory has two layers: session memory and persistent memory. Most AI customer service only has the first; it forgets everything once the conversation ends. Remembering isn't understanding, and this post explains the difference.
09 AdvertisingThe Big Six's methodologies: half real expertise, half sales pitch
WPP, Omnicom, Publicis, IPG, Dentsu, Havas — the underlying logic of the six major ad holding companies' methodologies is identical. The difference isn't which methodology is better, but each company's cultural DNA, talent profile, and the clients it serves best.
08 Customer serviceWhy are ad agencies getting into AI? The natural next step for integrated marketing
Satsuma is an ad agency, so why build AI customer service? Because the last mile of the advertising funnel — the conversation after a customer arrives — has long been outsourced to SaaS disconnected from the brand. We're taking it back, so AI coworkers grow from the same logic as TV commercials, social, and media buying.
07 Customer serviceWhat goes wrong when you use ChatGPT directly for customer service? (5 real cases)
Plugging large models like ChatGPT, Claude, or Gemini straight into a customer service chatbot looks simple. But after launch, five kinds of critical failure appear. This post breaks down each one with real cases and the technical causes behind them.
06 Customer serviceWhat does AI customer service cost? From NT$50,000 to build, about NT$1,500–4,000 a month
Custom AI customer service costs come in two kinds: a one-time build from NT$50,000, then about NT$1,500–4,000 a month (optional maintenance plus AI usage). This post itemizes every cost, compares three-year totals against SaaS and a full-time agent, and even counts your own time.
05 Customer serviceAI customer service SaaS abandoned within six months? 5 structural reasons
At many companies, nobody's logging into the AI customer service SaaS six months after launch. It's not that the vendor is bad; it's five structural limits of the SaaS model: generic knowledge bases, no one tuning it for you, and no grasp of your business details. Written for mid-sized brands that tried SaaS and were disappointed.
04 Customer serviceFirst month running AI customer service on our own website: real data and three unexpected lessons
A month after launching Xiao Ai on Satsuma's site, I'm laying out every number from the dashboard: cost, conversation quality, conversion rate, and three things I didn't anticipate.
03 Customer serviceAI customer service: SaaS or custom? Three 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 custom. Three questions tell you which side you're on: what customers ask most, how costly a wrong answer is, and how much time you'll invest in tuning. Includes a three-year total cost comparison.
02 Customer serviceWhy does AI customer service keep giving irrelevant answers? (The technical reasons in plain language)
Irrelevant answers aren't because the AI isn't powerful enough; it's being used the wrong way. This post explains why general-purpose LLMs make up answers when used directly for customer service, and how RAG and a customized knowledge base fix it.
01 Customer serviceWhat is RAG? The technology that keeps your AI from making things up, in plain language
RAG (retrieval-augmented generation) means the AI can only answer from data you provide, and honestly says so when it can't. This post explains vectors, chunking, retrieval, and reranking in plain language, and why doing RAG well is much harder than just getting it working.