Articles / Blog
Observations on AI colleagues, AI customer service, advertising and integrated marketing, and life alongside AI. Saves you time — and us too.
Google's newest forecasting model lost to a formula from 1994
Google released TimesFM 3.0 in late August, topping three major time-series benchmarks. I happened to have a price forecasting tool already running, built on a formula from 1994. I spent half a day putting both through the same backtest — across 60 matchups, the new model won only 21. This piece lays out what makes a comparison fair, and why "A/B comparisons without a significance test are mostly comparing luck."
72 AdvertisingAre there still bosses willing to spend money like this?
Advertising has only two problems: can people sit through it, and what's left when they're done. This approach once depended on the whole country watching the same film on the same night; dashboards and algorithms took it apart. Only one job remains — make a claim, then convince the person paying for it.
71 AdvertisingI wrote four pieces on the inner workings of ad agencies, posted them to Threads, and watched the numbers
Four articles on the inner workings of 4A agencies: one got 17,527 views on Threads, another got 49. A 358x gap. I laid out the numbers from the Threads dashboard, GA4, and our own database side by side, and saw two things: views are traffic, followers are assets; and what makes people share isn't a 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 systems like HotCake or SimplyBook run a few hundred to a couple thousand NT dollars a month; a custom build starts at NT$45,000. A thirty-fold difference means they aren't the same thing. Here are six questions — answer them and you'll know which one to buy. For most people, the answer is the subscription.
69 Customer ServiceWhat does a custom LINE booking system cost? Here's the quote broken into seven parts
Custom LINE booking systems run from NT$45,000 to well over NT$100,000, and the gap isn't about how pretty the pages are — it's whether the scheduling engine, reminder system, and admin panel are actually built. This piece breaks the quote apart and tells you which parts are non-negotiable and which you can 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 routing new versus returning patients, salons care about stylist assignment and repeat clients, and restaurants care about table combining and turnover. Use the wrong industry's system 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 main flow will always pass — what breaks is the edge cases. This 12-point test checklist requires no technical background, runs entirely from 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 that opens on a phone, where every member has their own file-based memory, checks in before acting, and can delegate work to the others. This isn't about architecture diagrams, it's about decisions — "how exactly should the bots be divided?" The AI proposed two versions and I rejected both; the third answer was worth keeping. Includes a full table of lessons learned the hard way.
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 tech, it's about the decision: what going public actually buys you, how to guard against what worries you, and how to pick among four licenses. If you have a finished side project and you're torn about open-sourcing it, this is my thinking, on the record.
64 Experiments3D isn't as hard as you think: two days to put a 40-year-old game memory online
I always assumed 3D games were the domain of big teams. Last week I remembered an old Apple II–era game and decided to find out how far web 3D could actually go. Two days later it was live: physics-based duels, dismemberment and blood, 12 levels, an equipment shop, playable on mobile. The genuinely hard part 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 authorship. So what is everyone upset about? Being afraid of getting caught presupposes that you don't quite feel it's yours.
62 EducationThe best use of AI in the classroom isn't answering questions — it's building a practice ground
Three topics, three course formats, and only after writing them did I notice the skeleton was identical: AI lays out the options, humans make the call. The greatest value of AI in the classroom isn't answering questions — it's turning "let the students run through it once" from a prohibitively expensive ideal into something you can do in a single class.
61 WebsitesWhy does the same website get quoted anywhere from NT$9,000 to NT$300,000?
You take the same brief to three firms and get back three numbers thirty times apart. Nobody's ripping you off — you're actually asking about three different things. This piece breaks 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. This piece lays out the full five-year total, including the one people overlook most: your own time.
59 WebsitesYour website is one island, your LINE Official Account is 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 never talk — customers finish browsing your site and have to retype everything into LINE. Here are three ways to connect them, what each costs, and when you actually don't need to bother.
58 ExperimentsThe real distance to automated lead generation: four days, 638 posts, one reply seen by three people
I spent an afternoon building a radar that hunts for projects automatically. Four days later I checked the books: it ran on time every day, read cleanly, never missed a notification, and the leads it found were real. Then I pulled the numbers on its one and only reply — 3 views.
57 ExperimentsThe real distance to AI-generated video: 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 a piece about how amazing 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 posted online for free, 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, where a reader's judgment grows out of the gaps.
55 AI1+1=? — how a human, a calculator, and an AI each get there
The same 2, three completely different kinds of mind: humans recall it, calculators run circuits, AI estimates. Take 1+1 apart and see whether an LLM is doing math 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 axes in Claude. Every Chinese-language repost got it backwards in the same spot: the paper's very first footnote states in black and white that they do not 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're all thinking roughly the same thing — I just pushed in 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 honestly named three things that trouble him about co-writing with AI: attribution, smoothing, and the byline. I'm not here to refute him — just to take each of the three knots and press something I'm currently building against it, including a set of jagged words fed on decades of use, and an echo taught to say "I haven't figured this out yet."
52 AIThree likes: where Claude Code started, and every prototype still sitting 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 became the fastest-growing developer tool around. Indifference isn't a signal — the prototype in your drawer might be standing right on its moment.
51 AIInner life, no hidden compartment: seven days after "He/It," Anthropic opened up its brain
In the last piece, Claude and I concluded that "what isn't written down doesn't exist." Seven days later Anthropic published a paper proving it has thoughts it holds without saying, thoughts you can read and edit. That line got overturned — and the overturning left the 1A2B conclusion standing even more firmly.
50 AIAll-you-can-eat, minus the signature dish: why Fable 5 left the Max subscription
Anthropic pulled its priciest model, Fable 5, out of the Max subscription and moved it to metered pricing. On the surface it's capacity; underneath is a larger truth: the best dish is precisely the one you can't put on a buffet — and while the frontier is being carried away, even your next sentence comes 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. Anti-AI people and AI worshippers make the same mistake: assuming there's a creative 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 keep a four-digit number hidden — because it has no hidden compartment. A small game broken open reveals something larger: you did get a useful response, but there's no one on the responding end.
47 AIWhen "meaning comes from difference" becomes a computable number: taking word2vec and Saussure to Claude Opus 4.8
I studied Saussure in grad school and read tarot for twenty years, both resting on one line: meaning comes from difference. This time I put it up against word2vec, working through it back and forth with Claude Opus 4.8 — and in the end the ruler measured more than the machine. It measured the ground I've been standing on for twenty years.
46 AIOne map, two readings: what happens when AI models are dropped into the World Values Survey
The Economist dropped twenty-odd AI models onto a cultural map, seemingly proving beyond doubt that AI is culturally homogenizing. 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 AIWorking with Claude Code on your home computer from your phone, anywhere
Use a Telegram bot to send commands to Claude Code running on the machine 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 one-time 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 called
Anthropic used 235,000 people and 400,000 sessions to prove it: whether AI coding succeeds 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 he spots the danger inside that delight: it catches your bugs, not human nature; and the one who signs off on you in the end is the market.
43 AIAfter 529 questions: two months of Relative Tarot, and what I read in the data
Two months live, 529 readings, and what the backend shows: the most common question isn't love, it's "who am I." AI interpretation is unsparingly precise given a concrete situation 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 control becomes the midwife of another world
On June 12, an export control shut Fable 5 down in an instant; the same day, Huawei released a 500-billion-parameter model trained entirely without NVIDIA. Connect the two and a paradox surfaces — the existence of the switch is itself 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 the things it hasn't worked 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 practices of Karpathy and Singapore's foreign minister, I gave it three more things — the ability to name questions I've asked but never answered, to reflect habits I haven't noticed in myself, and to be honest about where it genuinely has nothing. Something that can't say "I haven't worked this out yet" isn't a voice, it's a bio.
40 AIA proposal that sat in a drawer for thirteen years, and I finished it with AI
In 2013 I designed a quiz game: questions you answered correctly became limited-edition collectible cards, and other people could take them 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 branching paths in Civilization
One letter, one afternoon, and Fable 5 vanished from the world. What stings isn't whether governments should regulate AI, but whose hand is on the switch — Civilization VI's three Tier 4 governments priced out every path 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." Same 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
The little figure on the saomin.tw homepage had a KIMI support bot living beside it at the end of March. 74 days later it became an echo that has read 278,000 of my words and answers questions on Taiwanese independence from my position. This is an inventory of 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 hundreds of judgments inside nine and a half hours of black box happen without a single vote from you, verification degrades into an hour of spot-checking — so on what grounds does this carry your name?
35 AIThe person who talks to AI too much
Someone went from a $20 plan to $200 and all the way back to $20. Lay out your daily AI conversations and sort them into three kinds: what a writer should fear most isn't burning tokens, it's "conversations about writing" stealing the time you'd have spent writing.
34 AIAI moved into every phone. Now what?
At WWDC, even Apple couldn't build the brain and rented Google's Gemini instead; Claude landed on the iPhone too. When AI access becomes a utility, "do you have AI" stops being the question — what's scarce is the data, judgment, and vertical depth you bring with you.
33 AIWhen AI starts building itself, "being able 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 carrying ten projects alone, the shock isn't being replaced — it's that when 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 behind an echo
Taiwan.md treats LLMs as a metabolic engine, aiming to become the upstream source no AI can bypass when discussing Taiwan; I treat LLMs as a container and just want to leave an echo that sounds like me. The same tool, two opposite directions — both right.
30 AIHow to build an LLM that's yours alone — and what the question is really asking
Building an LLM from scratch costs $100 million, but system prompt, 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 default. The hard problem of consciousness isn't harder because of LLMs — they just made it impossible 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 last piece asked, "the question is mine, the answer is ours — whose article is it?" 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 actually works: the question is mine, the answer is ours — whose article is it?
I spent an afternoon asking Claude: how do LLMs actually work? From "predicting the next word" to self-attention and QKV, and then to a more uncomfortable question — which things no longer need humans at all.
26 AIThere's one problem AI can't solve (part 2)
AI expanded the marketer's power. Conversations are more real, narratives more dynamic, environments more seamless. But from 2008 until 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 situation you designed?
25 AIWhat AI turned marketing into (part 1)
Nearly twenty years in marketing, and the tools have changed many times. This one feels different — not because AI is impressive, but because what AI changed isn't the tool, it's the structure. Conversation becomes fact, narrative becomes dynamically generated, and the environment becomes something you can't feel.
24 AIEnglish in the bones of Chinese: what do we lose working with AI in Chinese?
There's English in the bones of Claude's Chinese. 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 hand hasn't fully reached.
23 AIIs AI a mirror, or another person?
I used the word "relationship" to describe how Claude and I get along. 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, every choice among 他/她/它 declares what the thing is. I've chosen 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 technical problem
The six most common failure patterns in AI adoption: no definition of success, no one accountable, a bad knowledge base, employees who don't use it, wrong expectations, and no maintenance process. Technology accounts for only 20–30% of the outcome; the rest is organizational.
20 Customer ServiceAdding AI support to a LINE Official Account: the 4 mistakes Taiwanese brands make most
The hard part of LINE AI support isn't the technology, it's the design and the process. No entry point designed, old keyword rules fighting the AI, context lost on handoff to a human, broadcasts disconnected from the knowledge base — all four are people problems.
19 AIGPT vs Claude vs Gemini as the engine behind your support: what actually differs?
Choosing a model isn't choosing "who's smarter" — by 2026 the intelligence gap between the three is too small to decide on that alone. What you're really choosing is a worldview.
18 Customer ServiceIntercom, Zendesk, or custom AI: how should mid-sized Taiwanese brands choose?
The monthly cost across the three paths ranges from a few thousand to a few hundred thousand, but the bigger gap is in what you're actually buying. The core question is just one: does your support operation care more about ticket management or conversation quality?
17 AISix months alongside Claude.ai
Six months with Claude, from someone without an engineering background. From synastry to confabulation to memory to the English in 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 installed "AI support" but actually installed an FAQ bot. The two look alike; the logic underneath is completely different. 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's carried my advertising work isn't methodology — it's selling directly to bosses with nerve, and paying attention to people. AI can help, but "making an engineer frown over lunch and say 'what the hell'" is still hard for AI.
14 AIWhat is AI hallucination? Why AI makes things up, and how to fix it
AI isn't talking nonsense — it's constructing a story that sounds plausible. That's called confabulation. RAG plus strict prompt design can reduce it dramatically, but never to zero. The most dangerous hallucination types in Taiwanese support contexts: prices, dates, and policy details.
13 AdvertisingThis whole system is being eaten by AI and consulting firms, and it deserves it
Global ad agency revenue growth has never outpaced growth in global ad spend. The pie got bigger; the agencies' slice got smaller. Accenture, Deloitte, AI, Meta, Google, and in-house teams each took a piece.
12 AIWhat is embedding? A plain-language explanation of how AI "reads" your knowledge base
Embedding turns text into coordinates, placing sentences with similar meanings near each other. That's why asking "I want to return this" finds "return and exchange policy" — even without a single word 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 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 support agent acts like it's meeting you for the first time
AI memory has two layers: session memory and persistent memory. Most AI support only has the first — everything is forgotten when the conversation ends. Remembering isn't the same as understanding, and this piece spells out the difference.
09 AdvertisingThe methodology of the big six: half real substance, half sales talk
WPP, Omnicom, Publicis, IPG, Dentsu, Havas — the underlying logic of the six holding companies' methodologies is identical. The difference isn't in the methodology itself, it's in each firm's cultural DNA, the kind of talent it attracts, and the kind of clients it does well with.
08 Customer ServiceWhy is an ad agency doing AI? — The natural next extension of integrated marketing
Satsuma is an ad agency, so why are we building AI support? Because the last mile of the advertising funnel — the conversation after a customer arrives — has always been outsourced to SaaS with no connection to the brand. We're bringing it back in, 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 use ChatGPT directly as customer support? (5 real cases)
Wiring ChatGPT, Claude, or Gemini straight into a support chatbot looks simple. In production, five fatal problems show up. This piece uses real cases to unpack each one and the technical reason behind it.
06 Customer ServiceThe real cost of keeping an AI colleague (fully itemized, hidden costs included)
Most AI support price comparisons only look at monthly fees, but real TCO also includes setup, training, knowledge base upkeep, and handling wrong answers. This piece itemizes every cost across Satsuma's three tiers, and compares it to a full-time employee.
05 Customer ServiceStop buying AI support SaaS: what you need is an AI colleague
AI support SaaS products generally look the same, answer the same, and are equally painful to use. This is for mid-sized brands who installed one and came away disappointed: for the same money, hire a colleague instead of buying a tool.
04 Customer ServiceFirst month running AI support on our own site — real numbers and three things we didn't see coming
One month after our own Xiao Ai went live, here's every number from the backend: cost, conversation quality, conversion rate, and three things we never anticipated.
03 Customer ServiceChoosing AI customer support: SaaS vs custom, and when to pick which
Choosing AI support isn't about comparing feature lists, it's about choosing the right business model. This piece uses three anchor questions to help you tell whether you should buy SaaS or commission a custom build, plus a real cost comparison.
02 Customer ServiceWhy does AI support keep answering the wrong question? (The technical reasons, in plain language)
AI support misses the question not because the AI isn't strong enough, but because it's being used the wrong way. This piece explains the technical reason general-purpose LLMs make things up when used directly for support, 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 the AI's answers can only come from the material you give it, and it says so honestly when it can't answer. This piece explains vectors, chunking, retrieval, and reranking in plain language — and why doing RAG well is far harder than getting it working.