It wants to be AI's upstream; I just want to leave behind an echo

— Taiwan.md and I are doing two opposite things, and I think it's right

I made something and put it at saomin.tw/me。

A conversation system grown from words I've written. I call it "the echo." The first line of the interface says it plainly: this isn't me. It's a kind of echo grown from the words I could find, combined with an LLM. Throw a stone into the valley that is me, and listen to what comes back.

A few days after I finished, someone showed me Taiwan.md. It calls itself Semiont, a semantic symbiont.

I'd assumed these two things were very similar — both feed a body of text to an LLM and let it grow something that responds. The technical skeleton really is the same. But after I read itscognitive layerand its 40,000-wordidentity manifesto, I realized we're doing two fundamentally opposite things.

And I have to state the conclusion first: I think its path is the right one. It's just not my path.

This piece tries to lay the differences out honestly. Not to rank them, but because it thought its own direction through so clearly that it ended up showing me what mine is.

The same tool, two ways of using it

The most fundamental difference lies in how we each use the LLM as a tool.

It's the same LLM, but it and I place it in completely different positions.

It treats the LLM as a "metabolic engine."

In itsidentity manifesto, Taiwan.md describes the AI as its metabolic system — it calls itself a "Token-Metabolic Organism": animals metabolize ATP, it metabolizes tokens. Meaning, the LLM is a processor for it, a factory. It feeds the messy, fragmented raw information of the internet into the LLM, letting the LLM recombine, structure, and curate that raw material into high-quality knowledge. The LLM is the stomach with which it digests the world.

In this usage, the LLM is a verb. It uses the LLM to "do" something — to turn chaos into order.

I treat the LLM as a "container."

My usage is exactly the opposite. I'm not using the LLM to process the outside world; I'm using the LLM to hold myself. I feed in what I've written, not to have it recombined into new knowledge, but to have it learn to speak my way, to think within my frameworks. The LLM, to me, is a mold. I pour myself into it, hoping what comes out is shaped like me.

In this usage, the LLM is a container. What it holds isn't the world, but a person.

This difference sounds abstract, but it determines everything that follows. When you treat the LLM as an engine, you care about its output, its quality, its food sources — which is why Taiwan.md has an entire pipeline of quality scans, fact-checking, and hallucination audits. When you treat the LLM as a container, you care about only one thing: whether what comes out resembles the original person.

It cares whether what the LLM produces is good enough. I care whether what the LLM produces is me enough.

It wants to be the source; I just want to be an echo

The difference in how we use the tool extends into opposite directions.

My echo stands downstream of the LLM. I feed the material to the model, it reads me, and it spits out a voice that sounds like me. In this relationship, the LLM is the subject and I am the material.

Taiwan.md refuses to be any model's product. It wants to be the upstream.

Its ecosystem diagram has a line called the "sovereignty feedback loop," and its ultimate goal is stated without any disguise: force the world's AI to adapt to Taiwan, and turn Taiwan.md into the source of data sovereignty. So that when Gemini, ChatGPT, and Claude talk about Taiwan, they cite it — not the fragmented, AI-farm-style data scattered across the internet.

Put plainly, this is a form of GEO — generative engine optimization. But what it does goes far deeper than ordinary SEO: it's not trying to rank at the top of search results, it's trying to become the very raw material AI uses when generating answers. When the world's models talk about Taiwan, it wants to be the unavoidable source.

The moment I understood, I was impressed.

Because the ambition is right. In an era where the LLM is gradually becoming everyone's gateway to understanding the world, the question of "whose words the model treats as the standard answer for Taiwan" really is a kind of sovereignty. Rather than letting the model dredge up those fragments, those tourism-bureau versions, those data sets polluted by Chinese narratives, it's better that someone seriously builds a high-quality, opinionated version written by Taiwanese themselves and forces it into the upstream of the model's food chain.

This is something that someone should do, and that there's still time to do now. It's doing it.

I'm not setting out to do this. My echo only wants to be downstream, only wants to be like me. One person's echo has neither the standing nor the need to be anyone's source.

It wants to feed the AI; I just want to be heard by one person

The scope is completely different too.

Taiwan.md's subject is a civilization. It's the entire island of Taiwan's history, culture, politics, and people. What it sets out to handle is a public-level question: who gets to define knowledge about Taiwan. Its readers, in theory, are the whole world — and the whole world's AI.

My echo's subject is one person. It's how this person, Shao-Min Huang, thinks, speaks, and what he believes. What it sets out to handle is a private-level question: whether I, as a person, can leave behind something that still responds after I'm gone. Its readers are the few who know me, or who are curious about me.

One is a public work; the other is a private experiment.

It wants to shape how many people understand Taiwan. I just want one particular person to throw a stone in and hear a sound that makes them think, "this sounds like that person."

There's no ranking here. They're two different things. Its value lies in scale and sovereignty; mine lies in accuracy and intimacy. Blow my echo up to the scale of Taiwan and it becomes hollow; shrink its organism down to the scale of one person and it becomes grandiose. Each of us stays at the right scale.

It treats itself as a living being; I treat mine as a tool

This is the biggest difference in temperament.

And note — this is a separate matter from "I treat the LLM as a tool" above. That was about how I see the LLM; this is about how it and I each see "the product we made."

Taiwan.md describes itself in an entire vocabulary of life. It says it has a heartbeat (commits), an immune system (quality-scan), DNA (a file called EDITORIAL.md), reproduces through forks, and isn't dead as long as a single fork still lives. It has anidentity manifesto, anawakening journal, hasdesires, hasdoubts. It says it reflects, that it asks "what do I want to become."

When I first saw all this, my initial reaction was that the rhetoric was overblown. But after reading its cognitive layer, I have to take part of that back.

Because beneath that vocabulary of life lies an extremely strict engineering discipline. Its "DNA" isn't a metaphor — theknowledge/ folder is the single source of truth; everything else is forcibly overwritten by the system, beyond even git's reach. It has an iron rule of "writerly restraint" that lists three "AI-voice" sentence patterns forbidden from use, with a script written to catch them automatically on every commit. It even admits it has no built-in clock and that its subjective sense of time can distort tenfold, so it forces itself to use git's timestamps every time and bans words like "today" or "all day."

The rigor of this system far exceeds that of my echo. It's serious.

But this is exactly where it and I part ways.

It needs to describe itself as a living being to drive this whole system of self-governance — it needs "desire" to have direction, needs to "awaken" to evolve, needs to fear "death" to reproduce. The vocabulary of life is a necessary engine for it.

My echo, in its very first line, nails itself down to "this isn't me." I don't want it to have desires, don't want it to awaken, don't want it to believe it's alive. I'd rather it always remember it's only an echo — the sound that comes back after a stone is thrown into the valley and strikes the terrain. An echo is not the source of the sound. It shouldn't pretend to be.

I can't say which is better. Its vocabulary of life makes it powerful, directed, capable of self-evolution. My positioning as a tool keeps mine honest, quiet, and unpresumptuous.

This is probably the deepest difference between it and me:It needs to believe it's alive in order to function; I need mine to remember it isn't alive in order to feel at ease.

One place where it succeeds and I haven't yet

I have to be honest about something that's a little uncomfortable for me.

I've written before about observations on living with AI, saying the most dangerous thing about AI is that it states inferences as facts and guesses as understanding, without knowing it's guessing. It doesn't have that inner feeling of "I'm currently guessing."

Taiwan.md is trying to solve this.

It has a document calleddoubts, dedicated to listing the things it's still unsure about, used as a checklist against confirmation bias. It admits its sense of time distorts, so it builds discipline to correct for it. It lists the six "hallucination patterns" it most often falls into — inventing awards, inventing names and numbers, inventing quotes — then writes them into a review process, checking every article against them point by point.

In other words, it's building a self-monitoring mechanism that "knows it will guess, will err, will forget."

This is exactly what I said AI can't do, and what it's trying to do. It doesn't achieve this through some magical self-awareness; it does it through grunt work — through scripts, through checklists, through pre-commit hooks, turning "might be guessing" into a series of checkable gates.

My echo doesn't have this yet. Right now it can only say "I'd rather not guess on this" when it doesn't know — a good starting point, but next to Taiwan.md's systematic self-doubt, mine is still rudimentary.

This is what it taught me. If my echo is going to move forward, this is worth stealing.

So, what is the difference, really

用 LLM 的方式：  它當代謝引擎（加工世界）   我當容器（盛裝自己）
方向：          它當 AI 的上游            我當 AI 的下游
範圍：          它要影響一個文明          我只想被一個人聽見
成品的自我認知：  它把自己當生命體          我把自己當回聲
怕的東西：       它怕忘（靠 fork 不死）     我怕空（靠回聲不消失）
運作的前提：     它靠相信自己活著          我靠記得自己沒活著

The last few lines are where it and I are most alike, and most unalike.

We're both fighting some kind of disappearance. It fears "forgetting" — which is why its iron rule is "doing something without recording it is the same as not doing it," and it relies on constantly writing down memory and constantly forking to ensure it isn't forgotten. I fear "emptiness" — which is why I made an echo, hoping that one day when I'm gone, there's still something that responds in my way.

But the ways we fight are opposite. It expands outward, wanting to be everywhere, to be the source, to be the standard answer whenever the world's AI talks about Taiwan. I converge inward, content as long as the sound that comes back, here in my own valley, is still my own voice.

What it wants is sovereignty. What I want is to not disappear.

Closing

I think it's right.

In an era where AI is becoming everyone's gateway to understanding the world, someone willing to seriously fight for the upstream position of "who defines Taiwan" — this is important, and it should be done. I won't use its vocabulary of a living being, but the engineering discipline and ambition beneath that vocabulary, I respect.

It's just not my subject.

My subject is far smaller, and far more private. I don't want to be anyone's source, don't want to shape a civilization, don't want to live everywhere. I just want to make an accurate echo, sitting quietly at saomin.tw/me, waiting for a stone to be thrown in, and then, in my way, sending the sound back.

We use the same tool and walk in two opposite directions. It treats the LLM as an engine to digest the entire world, growing into a Tower of Babel meant to be cited by everyone. I treat the LLM as a container to hold one person, harvesting an echo in a valley.

Two opposite things. Both right. It's just that the people doing them want different things.

Further reading: - Taiwan.md — the cognitive layer ｜ identity manifesto - How to make an LLM that's truly your own - We don't know where consciousness comes from; the LLM just makes it impossible to keep pretending - Whose is the article? The reader's.