Pantransderivational

"My voice will go with you."

— Milton Erickson

There is a word I keep coming back to, mostly because it is ugly. Pantransderivational. Six syllables of jargon, born in the hypnotherapy clinics of the mid-twentieth century and later kidnapped by the NLP crowd, who used it, as they used most things, to sell seminars.

But strip the marketing off and there's something real underneath.

A transderivational search is what your mind does when it receives a stimulus too vague to decode directly. Someone says "and you already know what needs to happen next," and your brain — which cannot tolerate an unresolved sentence — goes rummaging. It ransacks your memories, your unfinished business, your half-formed intentions, looking for something, anything, that could make that sentence mean something. And it finds it. It always finds it. Because you have a lot of material in there, and almost any key will eventually turn some lock.

The pan just means: everywhere. Across all derivations. Not one chain of inference but the entire forest of them, searched at once, below the waterline of awareness.

Milton Erickson built a career on this. He would speak to patients in language so artfully imprecise that it contained almost no content at all — and they would leave his office convinced he had seen into their souls. He hadn't. He had just held up a very clean mirror and stood out of the way.

The suit made of your own skin

Here is the thing that took me a long time to understand about communication: the meaning is not in the message. It never was.

A horoscope works this way. "You have a tendency to be hard on yourself, though others see you as confident." Nobody wrote that for you. You wrote it for yourself, in the half-second after reading it, by supplying the specific memory that makes it true. The horoscope is a hole. You are the filling.

A blues line works this way too. "I woke up this morning." Four words containing nothing. And yet.

A good koan, a good slogan, a good brief, a good first line of a poem — they all work on the same principle. They don't deliver meaning; they provoke a search. And the result of that search fits perfectly, because it is tailored from your own material. A suit made out of your own skin.

I once wrote that well-zipped ideas grow like wildfire in people's minds, using the fiber of their own thoughts as fuel. That was a nice sentence but I had the mechanism slightly wrong. It isn't that the idea grows using their material. It's that the idea is their material. The writer just supplies the shape of the absence.

Which is why over-explaining kills an idea. When you fill in every gap, you leave nothing for the reader to search. You hand them a finished suit, cut from your cloth, in your size. It won't fit. It can't.

The problem with mirrors

Of course, this is also how every con works.

Cold reading is transderivational search weaponized. So is a good cult recruitment pitch. So is most political rhetoric, which succeeds precisely to the degree that it means nothing specific and therefore means everything to everyone. "Take back control." Control of what? You tell me. You already have.

And so is the feed. The recommendation algorithm doesn't know what you want. It presents a stimulus and watches which way your search goes, then narrows the aperture, then again, until it has mapped the shape of your absences well enough to keep pressing on them forever.

The uncomfortable truth is that resonance and manipulation are the same mechanism running at different intentions. The blues and the horoscope use identical machinery. There is no technical test that separates them. The only difference is what the person holding the mirror wants from you — and whether, after the search, you end up with more of yourself or less.

The machine that searches

Now the interesting part.

We have built a device whose entire operation is pantransderivational search.

That's not a metaphor and it's not a stretch. A prompt is a vague stimulus. The model does not "understand" it; it propagates it through a space of billions of learned associations and returns the region that lit up brightest. Attention, in the technical sense, is a search over derivations. Embeddings are derivational geometry. When you give a language model an ambiguous phrase, it does exactly what Erickson's patient did: rummages through everything it has ever encountered, finds the material that resolves the ambiguity, and hands it back as if it had been there all along.

The difference — and it's the whole difference — is whose material gets used.

When a poem sends me searching, the search runs through my life. My grandmother's kitchen. That argument in 2011. The specific quality of light on a specific afternoon. And here is the part we rarely notice: the search doesn't just retrieve those things. It reindexes them. Every time a vague sentence makes me go looking, I come back having slightly rearranged my own interior. The poem didn't give me meaning. It gave me an excuse to reorganize myself, and I mistook the reorganization for the poem.

When I hand the same vague sentence to a model, the search still happens. It just happens over there. In the averaged interior of the internet. And what comes back is a suit tailored from everyone's skin, which is to say nobody's, which is to say it hangs on me like something borrowed.

What we're actually outsourcing

We have been told the story of AI as an outsourcing of thinking, then of remembering, then of deciding. But I think the deeper thing being outsourced is the search itself — and the search was never a means to an end. The search was the thing.

Meaning-making is not a lookup. It's a metabolic process. You take in something indigestible, break it down against the whole accumulated mass of your experience, and in doing so you also break down and rebuild some part of that mass. Understanding a difficult poem changes what you remember about your father. This is not a bug in human cognition. It's the mechanism by which a self stays alive rather than merely stored.

Outsource that, and you keep getting answers while slowly ceasing to be the kind of thing that has questions.

So what do we do

Not abstain. That never works and it's boring.

But maybe two things.

First: write more holes. In a world where anything vague can be instantly resolved by a machine, the temptation is to over-specify — to close every gap so the model can't misread you. I think this is exactly backwards. Precision is now cheap and infinitely available. Productive ambiguity — the kind that sends a human rummaging — is the scarce good. Leave the hole. Trust the reader to fall in.

Second: notice when you're skipping the search. There is a small moment, right before you paste the confusing thing into the chat window, when you could instead sit with it and let it rummage through you. That moment is where you get made. It costs about ninety seconds and it is the last thing in the world that nobody can do on your behalf.

Erickson's line — my voice will go with you — was designed to be searched. He never said what the voice would say.

That was the whole point.

On the off chance you'd like to subscribe to my writing, please feel free to subscribe on twitter or use this RSS feed. Thank you.