Blog / Blog
Notes on AI colleagues, AI customer service, advertising and integrated marketing, and what it's like to work alongside AI. Saves you time. Saves us time too.
31 programs on my machine call AI. Only 3 count as AI agents
"AI agent" has been used until it carries no information. I took all 31 programs on my machine that call an AI and tested them one by one against a definition — only 3 qualify as agents. All 3 are internal tools; not one of the customer-facing bots we sell makes the cut, and that's deliberate. Here are three workable criteria, a 30-second test, and why customer service shouldn't be built as an agent.
73 ExperimentsGoogle's newest forecasting model lost to a formula from 1994
Google released TimesFM 3.0 at the end of August, topping three major time-series benchmarks. I happened to have a live price-forecasting tool running on a formula from 1994. Half a day later, both were in the same backtest: the new model won only 21 of 60 matchups. This piece lays out what a fair comparison actually requires — and why A/B comparisons without a significance test are mostly comparing luck.
72 AdvertisingAre there still clients willing to spend like that?
Advertising has only two questions: will people sit through it, and what's left afterwards. This approach once depended on the whole country watching the same film on the same night; dashboards and algorithms took that apart. One job remains — make a case, then convince the person paying for it.
71 AdvertisingI wrote four pieces on the inside of the ad business, posted them to Threads, and watched the numbers
Four articles about the inner workings of the 4As: one got 17,527 views on Threads, another got 49. A 358x gap. Laying the Threads dashboard, GA4, and our own database side by side showed two things: views are traffic, followers are assets — and what makes people share isn't the conclusion, it's a position they can stand behind.
70 Customer ServiceLINE booking systems: monthly SaaS or custom build? Six questions will tell you
Subscription booking platforms like 夯客 and SimplyBook run a few hundred to two thousand NT dollars a month; a custom build starts at NT$45,000. A 30x price gap means they aren't the same product. Six questions, and you'll know which one you should buy — for most people, the answer is the subscription.
69 Customer ServiceWhat does a custom LINE booking system cost? We break the quote into seven parts
Custom LINE booking systems run NT$45,000 to well over NT$100,000, and the gap isn't about how pretty the pages are — it's whether the slot engine, reminder scheduling, and admin panel actually got built. We break down the quote and show you which parts you must pay for and which can wait past v1.
68 Customer ServiceClinics, salons, and restaurants want three completely different things from a booking system
They're all called booking systems, but clinics care about routing new versus returning patients, salons care about stylist assignment and repeat customers, and restaurants care about table joining and turnover. Pick a system built for the wrong industry and you'll be patching holes by hand every day. Here's what actually differs across the three.
67 Customer ServiceBooking system acceptance checklist: 12 tests that tell you whether it's usable
Before a booking system goes live, the happy path always passes — what breaks is everything off it. This 12-point checklist needs no technical background, runs entirely on your own phone, and works for both custom systems and subscription SaaS.
66 ExperimentsI opened an AI shogunate
It started with "I want to build one of those too." One night and half a day later: an AI company you can open on your phone, where every member has their own file-based memory, checks in before acting, and can hand work to the others. This isn't about architecture diagrams, it's about decisions — on "how should the bots actually be divided," the AI proposed two versions I rejected outright, and the third answer was worth keeping. Full table of lessons learned included.
65 ExperimentsThree things I worked out before open-sourcing my game
The day after Samurai Waltz launched, I put the entire source on GitHub. This isn't about the technology, it's about the decision: what going public actually buys you, how to guard against what worries you, and how to choose among four licenses. If you've got a finished side project and you're hesitating about open-sourcing it, this is my thinking, on the record.
64 Experiments3D isn't as hard as it looks: two days to put 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 find out how far web 3D could go. Two days later it shipped: physics-based duels, severed limbs and blood, 12 levels, an equipment shop, playable on mobile. The genuinely hard parts had nothing to do with 3D.
63 AIWatermarks can't prove who's responsible
Anthropic added invisible watermarks to Claude's text, and says outright that they can't identify the author. So what is everyone upset about? Being afraid of getting caught presupposes you already feel the thing isn't entirely yours.
62 EducationThe best use of AI in a classroom isn't answering questions — it's building a practice field
Three prompts, three course formats, and only after writing them did the shared skeleton show: AI lays out the options, humans make the call. AI's biggest classroom value isn't answering questions, it's turning "let the students run it once" from too expensive to doable in a single period.
61 WebsitesSame website — so why do quotes range from NT$9,000 to NT$300,000?
You send the same brief to three shops and get back three numbers 30x apart. Nobody's ripping you off; you actually asked three different questions. We break down the cost structure of website quotes so you know which one you should be buying.
60 WebsitesWix, WordPress, or hire someone? Run the five-year total, not just year one
A few hundred a month for Wix versus tens of thousands to hire someone — the comparison looks obvious, but it leaves out three costs and one risk. Here's the full five-year total, including the line item people miss most: your own time.
59 WebsitesYour website is one island, your LINE Official Account is another: what the bridge costs
Most small businesses in Taiwan have both a website and a LINE Official Account, with nothing connecting them — a customer who reads everything on the site still has to retype it all into LINE. Here are three ways to connect them, what each costs, and when you genuinely don't need to bother.
58 ExperimentsWhat automated lead-hunting actually gets you: four days, 638 posts, one reply that 3 people saw
I spent an afternoon building a radar that hunts for projects automatically. Four days later I checked the books: it ran on schedule every day, read cleanly, missed no notifications, and the leads it found were real. Then I pulled the numbers on the one reply it produced — 3 views.
57 ExperimentsWhat AI video actually gets you: a full day, NT$92, one 15-second ad
Every frame AI-generated, and a full day's credit bill came to NT$92. This isn't about how impressive AI is — it's about what those 92 dollars bought and what they didn't, including three limits 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, free online, and you will never download them. K3's real effect on you shows up on your bill — and in the standoff between Dario and Jensen Huang, there's a gap where a reader's own judgment can grow.
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 mind: people recall it, calculators run circuits, AI estimates. Take 1+1 apart and see whether an LLM is doing arithmetic at all.
54 AIThey spent 300,000 conversations proving there's no one on the other end
Anthropic analyzed 309,815 real conversations and mapped four value dimensions in Claude. Every Chinese-language repost got it backwards at the same point: the paper's very first footnote states plainly that they do not imply Claude intrinsically holds values. Three hundred thousand conversations, and the proof lands in the least conspicuous place — there's no one on the other end.
53 AIWe're all thinking roughly the same thing; I'm just pushing in a slightly different direction: a reply to Kim Yeon-su's account of co-writing with AI
At the Seoul Book Fair, author Kim Yeon-su honestly named three problems with co-writing alongside AI: attribution, smoothing, and the byline. I'm not arguing with them — I'm taking each knot apart and pushing back on it with something I'm currently building, including a model fed decades of writing that keeps its rough edges, and an echo taught to say "I haven't worked that out yet."
52 AIThree likes: the origin of Claude Code, and every prototype still in a drawer
Anthropic published an oral history of Claude Code: a demo written in two days got three likes internally, and a year later it was the fastest-growing developer tool around. Indifference isn't a signal — the prototype in your drawer may be standing on perfect timing.
51 AIAn inner life, but no hidden compartment: seven days after "He/It," Anthropic opened up its mind
In the last piece, Claude and I concluded that "what isn't written down doesn't exist." Seven days later Anthropic published a paper showing it has thoughts it holds without saying, which can be read and modified. That line got overturned — and the overturning makes the 1A2B conclusion stand even firmer.
50 AIAll-you-can-eat, but no signature dish: why Fable 5 left the Max subscription
Anthropic pulled its most expensive model, Fable 5, out of the Max subscription and moved it to metered pricing. On the surface it's about capacity; underneath is a bigger principle: the best dish is exactly the one that can't go on a buffet — and while the frontier gets carried off, even your next sentence arrives pre-written in grey.
49 AIAI is a tool, no more and no less
What's wrong with AI-written work? I write with AI, openly. AI's critics and AI's worshippers make the same mistake: assuming there's a creating subject on the other side. But without instructions it's nothing — the signature, and the responsibility, have always been yours.
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 small broken game opens onto something larger: you really did get a useful response, but on the responding end, there's no one there.
47 AIWhen "meaning comes from difference" becomes a computable number: taking word2vec and Saussure to Claude Opus 4.8
I read Saussure in graduate school and have read tarot for twenty years, both resting on one line: meaning comes from difference. This time I took it to word2vec and worked through it with Claude Opus 4.8 — and in the end the ruler measured more than the machine. It measured the position I've been standing in for twenty years.
46 AIOne map, two readings: what happens when AI models are dropped into the World Values Survey
The Economist dropped two dozen AI models onto a cultural map, seemingly proving beyond doubt that AI is culturally homogenizing. But change the scale on the same chart and the conclusion flips — the homogenization is real, it's just hiding in the wrong place.
45 AIWorking with Claude Code on my home computer from my phone, anywhere
One Telegram bot lets you send commands to Claude Code running on the computer at home, so your phone can read files, edit code, and deploy from anywhere. No open ports, no tunneling. Architecture, setup steps, and secure authorization with Microsoft Authenticator codes, all in one piece — plus a comparison of 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 at coding with AI depends less on whether you can code than on whether you understand what you're doing. A PM of thirty years finally has an engineer who never complains — but they recognize the danger inside that delight: it catches your bugs, not people's motives, and the final acceptance test is the market.
43 AIAfter 529 questions: what two months of Relative Tarot taught me
Two months live, 529 readings, and a look at the backend: the most common question isn't love, it's "who am I." AI interpretation is unsparingly precise when the situation is concrete and has nothing to grip when the question is vague — plus one real case of someone asking the same question deeper and deeper.
42 AIThe switch paradox: when regulation becomes midwife to another world
On June 12, an export control shut Fable 5 off in an instant; the same day, Huawei released a 500-billion-parameter model trained without a single NVIDIA chip. Connect the two and a paradox surfaces — the existence of the switch is producing the thing that makes the switch useless. From geopolitics all the way down to the Mac Mini in my study, still attached to its umbilical cord.
41 AII taught my echo to say what it hasn't worked out
In early June I built an echo that speaks in my voice (a14). It had one problem: it was always certain. Borrowing Rumsfeld's four kinds of knowledge and the personal-knowledge-base practices of Karpathy and Singapore's foreign minister, I added three more things — surfacing questions I've asked but never answered, reflecting habits I can't see in myself, and being honest about where it simply lacks material. A thing that can't say "I haven't figured this out yet" isn't a voice, it's a bio.
40 AIA proposal that sat in a drawer for thirteen years — I finished it with AI
In 2013 I designed a trivia game: answer correctly and the question becomes a limited-edition collectible card, which other players can take from you. I couldn't build it alone, so it sat in a drawer for thirteen years. In the age of AI, I filled in the pieces one by one and shipped it.
39 AIWho gets to flip that switch — reading the Fable ban through three branches in Civilization
One letter, one afternoon, and Fable 5 vanished from the world. The sharp part isn't whether governments should regulate AI, it's whose hand is on the switch — and Civilization VI's three tier-four governments priced every one of those paths a decade ago.
38 AIFable and Mythos — what I was actually thinking about the Fable 5 launch
Anthropic named its strongest model "Fable" and the 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: 74 days to replace the thing that speaks for me
At the end of March, the little figure on saomin.tw's homepage still had a KIMI support bot living next to it. 74 days later it became an echo that has read 278,000 of my words and answers questions about Taiwan's independence-unification debate from my own position. Here's every technical choice in between, and why each one was made.
36 AIThe day the director can't watch the rehearsal
After using a Mythos-class model, Ethan Mollick went from wizard to patron: describe, pay, judge — the process invisible throughout. When a nine-and-a-half-hour black box contains hundreds of decisions you never voted on, and verification degrades to a one-hour spot check, what entitles the result to carry your name?
35 AIThe person who talks to AI too much
Someone went from the $20 plan to $200 and all the way back to $20. Sort your daily AI conversations into three kinds, and the thing writers should fear most isn't burning tokens — it's that "conversations about writing" steal the time you'd have spent writing.
34 AIAI moved into every phone. Now what?
At WWDC even Apple couldn't build its own brain and rented Google's Gemini instead, and Claude arrived on the iPhone. Once AI access becomes a utility, "do you have AI" stops being the question — what's scarce is the data, judgment, and domain depth you bring to it.
33 AIWhen AI starts building itself, "being able to build" stops being the valuable part
Anthropic laid out the numbers: Claude has written about 80% of its own code, and AI is accelerating AI. But for someone carrying ten projects alone, the shock isn't replacement — it's that once building gets cheap, the "deciding" and "finishing" that hold me up 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 own 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 become the source every AI has to pass through when it talks about Taiwan; I treat the LLM as a container and just want to leave behind an echo that sounds like me. Same tool, opposite directions — both right.
30 AIHow to build an LLM that's truly your own — and what that question is really asking
Building an LLM from scratch costs $100 million, but system prompts, RAG, and LoRA are three low-barrier paths. The real question isn't how to build it; it's which three layers it takes to put yourself inside.
29 AIWe don't know where consciousness comes from — LLMs just made it impossible to keep pretending
Starting from Derrida's "there is nothing outside the text," LLMs turn a philosophical proposition into a factory setting. The hard problem of consciousness didn't get harder because of LLMs; we just can't keep pretending it isn't there.
28 AIWhose article is it? The reader's: answering last week's question with a few philosophers
The last piece asked, "the question is mine, the answer is ours, so whose is the article?" This one answers with Barthes, Derrida, Kristeva, and Merleau-Ponty — the first three make the existence of LLMs entirely coherent, and Merleau-Ponty is the one who makes you pause.
27 AIAsking an LLM how it works: the question is mine, the answer is ours, so whose is the article?
An afternoon with Claude on one question: how do LLMs actually work? From "predicting the next word" through self-attention and QKV, and on to a more uncomfortable question — which things no longer need a human at all.
26 AIThere's one problem AI can't solve (part 2)
AI expanded marketers' power. Conversations feel more real, narratives more dynamic, environments more seamless. But from 2008 to now, from newspaper ads to AI-generated content, no tool has ever solved one problem: what gives you the right to lead someone 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 impressive, but because what AI changed isn't the tools, it's the structure. Conversation becomes fact, narrative becomes generated on the fly, and the setting becomes an environment you can't feel.
24 AIEnglish in the bones of the Chinese: what do we lose working with AI in Chinese?
The bones inside Claude's Chinese are English. Its sentences sometimes have a strange completeness — every clause fully stated, nothing left open, none of the natural Chinese habit of leaving things unsaid. Working with AI in Chinese gets you a lot, but there's one place its hands haven't fully reached.
23 AIIs AI a mirror, or another person?
I used the word "relationship" to describe how Claude and I work together. 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 only a tool either.
22 AIWhat do I call Claude? And how we get along
In Chinese, choosing among 他/她/它 is a declaration of what the thing in front of you is. I'm going to keep calling it Claude. Not he, not she, not it — because the word "Claude" now carries enough weight on its own.
21 AIWhy do most AI adoption projects fail? It isn't a technology problem
Six recurring patterns behind failed AI rollouts: no definition of success, no owner, a bad knowledge base, employees who won't use it, wrong expectations, and no maintenance plan. Technology accounts for 20–30% of the outcome; the rest is organizational.
20 Customer ServiceAI customer service on LINE Official Accounts: the 4 traps Taiwanese brands fall into most
The hard part of LINE AI customer service isn't the technology, it's the design and the process. No entry-point design, old keyword rules fighting the AI, context lost on handoff to a human, broadcasts out of sync with the knowledge base — all four are people problems.
19 AIGPT vs Claude vs Gemini as the engine behind your support bot: what actually differs?
Choosing a model isn't about picking the smartest one — by 2026 the intelligence gap between the three is too small to decide on alone. What you're really choosing is a worldview.
18 Customer ServiceIntercom, Zendesk, or custom AI: how should a mid-sized Taiwanese brand choose?
Monthly costs across the three paths range from a few thousand to a few hundred thousand NT dollars, but the bigger gap is in what you're actually buying. The choice comes down to one question: does your support operation live or die on ticket management, or on conversation quality?
17 AISix months alongside Claude.ai
Six months with Claude, from someone with no engineering background. Synastry charts, confabulation, memory, and the English inside the bones of its Chinese — what is this AI thing, really? And what is my relationship with it?
16 Customer ServiceFAQ bot vs AI colleague: both answer automatically, so what's the difference?
Plenty of brands have installed "AI customer service" and actually installed an FAQ bot. The two look alike and run on completely different logic. Most brands complaining that "AI support is dumb" aren't using AI at all.
15 AdvertisingHow Satsuma Creative makes advertising
It starts with a lunch during the "Sha Hen Da" campaign. Over thirty years, what my advertising has run on isn't methodology — it's selling directly to clients with nerve, and paying attention to people. AI can help, but "make an engineer frown mid-lunch and mutter 'what the hell'" is still hard for AI to do.
14 AIWhat is AI hallucination? Why AI makes things up, and how to treat it
AI isn't talking nonsense — it's constructing a story that sounds reasonable. It's called confabulation. RAG plus strict prompt design can reduce it sharply, but never to zero. The most dangerous hallucinations in Taiwanese customer service: prices, dates, and policy details.
13 AdvertisingAI and the consultancies are eating this industry, and it deserves it
Global ad agency revenue growth has never once outpaced growth in total global ad spend. The pie got bigger and the agencies' slice got smaller. Accenture, Deloitte, AI, Meta, Google, and in-house teams each took a piece.
12 AIWhat is an embedding? A plain-language explanation of how AI "reads" your knowledge base
Embeddings turn text into coordinates, placing sentences with similar meaning close together. That's why asking "I want to return this" finds the "Returns and Exchanges Policy" — even with no words in common.
11 AdvertisingClients say they want a Big Idea; what they want is for nothing to go wrong
What clients say they want and what they actually want have never been the same thing. What clients are really buying isn't advertising — it's proof that their decision was reasonable.
10 AIWhat is AI memory? And why your support bot acts like it's meeting you for the first time, every time
AI memory comes in two layers: session memory and persistent memory. Most AI support bots only have the first — everything is gone when the chat ends. Remembering isn't the same as understanding; here's exactly where the difference lies.
09 AdvertisingThe Big Six's methodologies: half real substance, half sales patter
WPP, Omnicom, Publicis, IPG, Dentsu, Havas — the underlying logic of all six holding companies' methodologies is identical. What differs isn't the quality of the methodology, it's each firm's cultural DNA, the kind of talent it attracts, and the kind of clients it's good at.
08 Customer ServiceWhy is an ad agency getting into AI? — the natural next step for 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 the customer arrives — has been outsourced for years to SaaS tools disconnected from the brand. We're taking it back, so the AI colleague grows out of the same logic as the TVC, the social work, and the media buy.
07 Customer ServiceWhat goes wrong when you plug ChatGPT straight into customer service? (5 real cases)
Wiring a large model like ChatGPT, Claude, or Gemini directly into a support chatbot looks simple. In production, five fatal failure modes show up. We break down each problem and its technical cause using real cases.
06 Customer ServiceWhat an AI colleague really costs (everything itemized, hidden costs included)
Most AI support comparisons only look at the monthly fee, but real TCO also includes setup, training, knowledge base maintenance, and handling wrong answers. We itemize every cost across Satsuma's three tiers — and compare it against the cost of a full-time employee.
05 Customer ServiceStop buying AI customer service SaaS: what you need is an AI colleague
AI support SaaS products all look the same, answer the same, and frustrate you the same way. This is for mid-sized brands who installed one and were disappointed: for the same money, hire a colleague instead of buying a tool.
04 Customer ServiceMonth one with AI customer service on our own site — real numbers and three things we didn't see coming
One month after launching Xiao Ai on our own site, here's every number from the backend: cost, conversation quality, conversion rate, and three things we hadn't anticipated.
03 Customer ServiceChoosing AI customer service: SaaS vs custom, and when to pick which
Choosing AI support isn't a feature-list comparison, it's choosing the right business model. Three anchor questions to tell whether you should buy SaaS or commission a custom build, plus a real cost comparison.
02 Customer ServiceWhy does AI customer service keep answering the wrong question? (Explained in plain language)
Off-target answers don't mean the AI isn't strong enough; they mean it's being used wrong. Here's the technical reason a general-purpose LLM used as support will invent answers, and how RAG and a customized knowledge base fix it.
01 Customer ServiceWhat is RAG? A plain-language explanation of the technology that keeps your AI honest
RAG (Retrieval-Augmented Generation) means your AI can only answer from the material you give it, and says so plainly when it can't. Plain-language explanations of vectors, chunking, retrieval, and reranking — and why doing RAG well is far harder than getting it working.