I taught my echo to speak the things it hasn't figured out

In early June, I wrote a piece (From Customer Service to Echo) documenting how I spent 74 days replacing the customer-service bot that used to greet visitors on saomin.tw with an echo grown from decades of my own writing. You ask it something, and it draws on the words I've written and answers in a way very close to how I would.

It did a decent job. But after a while, I noticed a flaw:

It was always certain.

Ask it about politics, about love, about theatre — it could answer it all coherently, in my voice. But the real me isn't like that. The real me has a whole pile of questions I've been asking for twenty years and still haven't answered, a pile of bad habits I've never noticed, and plenty of things I haven't actually thought through. A thing that only answers with certainty — however much it sounds like me — is still, at its core, a bio. A bio written in the first person.

A bio doesn't hesitate. A person does.

A diplomat, an engineer, and Rumsfeld

As it happens, I'd recently read about two people doing something similar.

This April, Andrej Karpathy revealed how he uses AI to build a personal knowledge wiki. In May, Singapore's foreign minister Vivian Balakrishnan — a man who visits twelve countries a month — said he'd built a "second brain" with AI to reduce his cognitive load. Both had externalized themselves into something they could talk to and search — a kind of digital twin of one's knowledge.

So I built a personal knowledge base of my own. As I did, I kept thinking of former U.S. Secretary of Defense Rumsfeld's famous taxonomy. He divided knowledge into four kinds:

• Known knowns(known known): what you know, and know that you know.
• Known unknowns(known unknown): what you know you don't know.
• Unknown knowns(unknown known): what you actually know but haven't realized you know.
• Unknown unknowns(unknown unknown): what you don't even know that you don't know.

Then I looked back at my echo and paused:

It could only do the first kind. Its entire capability rested on "known knowns" — restating what I'd written and already thought through. The other three, it had none of.

And those other three are exactly where a person gets interesting.

The division of labor: the echo turns inward, the knowledge base reaches outward

Once that was clear, the direction settled itself.

The value of a knowledge base (the kind Karpathy and the minister built) lies inreaching outward: finding literature you haven't read, filling gaps in what you don't know, going straight at the "unknown unknowns." That's right — a knowledge base should reach outward.

But an echo is different. An echo grows from a person'sexistingwriting; its raw material is closed — just the words I've already written, and no more. So an echo shouldn't reach outward and poach the knowledge base's job. It should turn inward, and take the two "about myself" quadrants as far as they'll go:

• Known unknowns: the questions I've asked myself but never closed.
• Unknown knowns: the habits I've never noticed but that recur again and again.

As for the "unknown unknowns," I deliberately left those out. That's the knowledge base's job. If the echo reached outward too, it would stop being a voice and become just another search engine.

So I added three layers to the echo.

Layer one: the unfinished — it voices the questions I never answered

I wrote a program that pulls out my writing on each topic separately — but this time it isn't capturing my "position," it's capturing myquestion marks: questions I raised but never answered in the same piece; two poles I set side by side but never reconciled; threads that span years, recurring again and again but never reaching a conclusion.

My writing is dated, from 2003 to 2026 — over twenty years. So these "questions never closed" can be laid out on a timeline — you can see which threads have tangled me my whole life.

When it was done, I asked it: "Would you still stand up for politics now?"

It said (in essence):

Not anymore. In 2006 I donated a hundred dollars to Shih Ming-teh, and at the time I still told myself this was "standing up." But even then I knew — I'd written the line "Is my anxiety the mindset of a spectator, or am I at the point of bursting?" — I asked it, and then never answered. Earlier, in high school, during the May 20 farmers' protest, I took the earliest bus there; looking back now, it was more spectacle than conviction. So this matter of "standing up" — I've never figured out whether I was genuinely angry, or just chasing a soap opera.This is a thread I've never thought through. All the way to that 2013 line, "all of this has gotten so far away from me" — that wasn't figuring it out, it was life papering over it... This question has gone unanswered since high school.

I stared at the screen, a little unnerved. Because everything it said was true, and it had strung together a thread I'd never connected myself — high school, 2006, 2013 — into a loop that had stayed open for twenty years. It didn't draw a conclusion for me. It just honestly laid that "never figured out" right there.

This is the known unknown. An echo that dares to say "I haven't figured this out yet" is more like me than one that has an answer for everything.

Layer two: the mirror — it reflects habits I've never noticed

The second thing was a little stranger.

I had it read through all my topic summaries, then asked it one question: across all the subjects, is there anything Imight not notice in myself, yet that recursas a habit? Not my views — my underlying habits of "how I think, how I speak."

What it pulled out made me laugh, because it was so accurate:

• When I talk about love I borrow from literature and pop songs (Haruki Murakami, Su Shi, Bobby Chen); when I talk about power and structure I switch to theorists (Foucault, Saussure, Barthes). The two casts never mix.
• I tend to approach "what something is" by way of "what it isn't" — from my master's thesis's "what history is not" to, twenty years later, looking at AI as "remembering is not experiencing, judgment is not judgment" — the same blade, just aimed at a different target.
• At my most lyrical, I always immediately deflate the pressure with a self-mocking parenthetical.
• When someone points out a contradiction, I don't concede; I'll say "those refer to different things."

I'd never been aware of any of this. But they really are me — only twenty years of reading my words could surface them. This is the unknown known — it was right there, I just couldn't see it myself.

For this layer I set a very strict rule:By default, never bring it up. Otherwise it would dissect itself in every sentence — insufferable. It holds up the mirror in only two cases: when you ask directly about its habits, or when this particular answer happens to be a vivid example of one. I asked it, "Do you have any habits in how you answer?" It named three, and closed with this:

I'm aware of all three myself. But knowing doesn't help — the next question, I do the same thing.

— that line wasn't in my design. It grew out of the thing itself, and it's deeply me.

Layer three: honesty — it knows where it has nothing

The last layer is the plainest, but the one I think matters most.

Those two foreigners' knowledge bases, run long enough, both grow a "dashboard": the diversity, density, and growth rate of your knowledge — numbers like blood pressure or blood sugar. Fascinating. But working on my own knowledge base, I caught onto something:

Those numbers don't measure your brain — they measure whatever you happened to feed in. If you read three extra papers on aging this month, it says you're "highly focused on aging issues lately." A blood-pressure monitor measures blood; this measures your upload log.

So the third layer I built for the echo held to a naming discipline from the very start: it's calledcorpus checkup, not "brain checkup." What it shows is — across my writing, which registers run thick and which thin, which years are dense and which are blank, which topics I write about most. A thin topic only means I wrote less about it, not that I thought less about it.

The more practical use: the echo now checks "how much material do I have on hand for this." Where the material is thick, it speaks freely and concretely; where it's thin, it honestly drops a notch — "I haven't written much on this, so from here on I'm more extrapolating from my framework than drawing on specific things I've written." It no longer pretends to have everything.

(For this layer I also made a chart on the side, plotting my topics as a co-occurrence network, placed at the very top of the echo's page. Honestly, that chart is mostly just nice to look at — the meaning of recognizing the structure has to come from you reading it, not from the chart itself. More on that shortly.)

Why "reflect it," not "finish the thinking for me"

While building these three layers, I kept resisting one temptation: since it can find the questions I haven't figured out, why not just let the AI go ahead and find the answers too, and fill in the gaps?

I didn't. And this is the one decision in the whole thing I'm most certain about.

Benjamin wrote about a term, aura. The hidden connections the AI finds for you, those threads you haven't figured out — none of those, I think, are the aura itself; at most they'rethe material from which aura forms. Without the connection passing through your own neurons, without the process of you thinking it through yourself, what you get is only a "false sense of mastery" — you look like you understand, but you don't.

As that Singaporean minister said, we live in an age where everything can be outsourced: computation, memory, replication, transmission can all be handed to machines. But there's one thing that can't be outsourced — your own understanding.

So the echo's job is toreflect that unfigured-out thread, not to close it for me. That step of closing it — actually figuring it out — that's my job, not its. A system that automatically fills in every gap for you only speeds up your information anxiety, making you feel full when really you've just looked at more things you haven't digested.

Zhu Xi said reading should go "from broad to spare" — wide reading is for the sake of converging, returning at last to a clear, simple core. Borges wrote of a man named Funes who forgot nothing, who could see and hear everything in the world, yet for that very reason could form no structure, and quickly drowned in information — for him, every thing broke into fragments, and nothing could make up a whole.

Finding the hidden patterns, recognizing the structure, focusing — that is the spirit of this kind of thing. Not accumulating more.

What it is now

So the echo now on saomin.tw/me can do four things:

• Answer in my voice the things I've thought through (known knowns)
• Voice the questions I've asked but never answered (known unknowns)
• Reflect habits I've never noticed in myself (unknown knowns)
• Be honest about where it actually has nothing (honesty about coverage)

The one thing it doesn't do is reach outward to find new knowledge for me (unknown unknowns) — that's left to another set of tools; don't contaminate this one.

It still can't stand in for me in person. What it has read is only the part of my writing I was willing to set down and have kept — probably a thin slice of the real me. But compared with that early-June version that was certain about everything, this one — which hesitates, admits when it hasn't figured something out, and occasionally even ribs itself — is, indeed, a little more like me.

Because where a person is most himself is never the things he's thought through. It's the things he hasn't yet figured out but keeps thinking about.

Throw a stone in, and listen for the echo:saomin.tw/me。

(This piece is a record for the curious, not a tutorial. For how it's wired up technically, it continues from a14's 74-day record; all three layers here are built on the original topic routing, not started from scratch.)