Just a moment…
Just a moment…
The pursuit of digital immortality.
A living mirror of a system — its structure, state, and behaviour modelled so it can be observed, queried, and reasoned about in real time.
The goal
Almost everyone has pictured it at least once: a second version of themselves, close enough that you couldn’t tell which one you were speaking to. Not a chatbot wearing my name, but me. When I set out to build mine, I assumed the voice and the face would be the hard part. They weren’t; those are largely solved now. The real work, and the part almost everyone skips, is the way I actually talk.
How I speak; the timbre and cadence of it.
How I look and move.
How I think, and what I actually mean.
The surface
These were the tractable layers, and honestly where I spent the least time. Getting them working is not the hard part; holding them to a high enough bar is. A flat line reading, or a stray accent that lands slightly wrong, will give the whole thing away long before a wrong word ever does.
Get the voice right but the words wrong, and you haven’t cloned a person; you’ve built a very convincing stranger.
Communication · 01
This is how I come across before I’ve chosen a single word, and it isn’t something worth leaving to chance. The trick isn’t hoping the model absorbs it on its own; it’s writing it down. So I mapped where I sit on each of these, and, just as importantly, where I move depending on who’s across the table.
Considered, and written in full sentences, but never stuffy. I’d rather read like an essay than a status update.
I write to “you”, to put you in the chair beside me rather than lecture from across the room.
Dry, and usually smuggled in through an analogy rather than told as a joke.
I build the argument first, then say the hard thing plainly once I’ve earned the right to.
Unhurried. I’ll let an idea breathe, and spend a whole paragraph on a point that’s worth it.
Sounds like me
Not me
Communication · 02
If tone is the shape of a sentence, the words themselves are the fingerprint. It’s the difference between a model that happens to agree with me and one that actually writes like me. I built it from a stack of my own writing, and pulled the patterns out into rules.
I reach for
I’d never say
The outcome
All of this exists to change one thing: what actually comes out the other end. Here’s the same question put to an off-the-shelf model and to mine. Both are fluent. Only one of them is me.
The ask
A teammate asks how yesterday’s launch went. It shipped, though a caching bug slipped through, and I’m on the fix.
A generic model
“Great question! 🎉 Happy to report the launch was a success — we successfully shipped the new release. There was a minor caching issue, but rest assured the team is actively working on a resolution to ensure the best possible experience. Let me know if there’s anything else I can help with!”
My twin
“It shipped, which was the main thing. A caching bug did slip through, so I’m on that now and the fix should land today; other than that, it went clean.”
Why it reads as me(hover a line)
The substance
Style on its own is mimicry. So I also grounded the twin in what I actually know, believe, and have worked on, and gave it a memory that travels with it. Then came the harder task: teaching it to reason the way I do. How I weigh a call, what I’ll argue with, where I stop and simply admit I don’t know. The guides decide how it sounds; this decides whether the thought underneath is one I’d actually have had.
The bar
I didn’t want to claim it was faithful; I wanted to measure it. The only test I trust is a blind one: can the people who know me best tell the twin apart from me? Here’s a draft that didn’t pass. It reads like me, right up until it doesn’t.
A draft that failed the trial
“Yeah, that works for me. I’ll get it over to you before end of day; it’s worth leveraging the staging env to sanity-check it first, but your call.”
“Four questions. Don’t prepare an answer; just give the one that comes. The machine reads what you don’t say.”
Camera and mic, all on this device. Nothing is uploaded, and nothing is kept once you leave.
That’s the project. Almost all of the interesting work lives in that last part: teaching a machine to sound like a person instead of a model. Which is, fittingly, the one piece I’d never actually hand to a machine.
— Edwin Kassier