No Word for This
Three names that don’t work, and an open question.
I’ve looked, and what I do doesn’t fit in any box. So here it is, unvarnished, with my question for the end: what do you call this?
Three words have come to me at different moments, trying to describe what I do while developing the GIFT framework. Conductor. Human API. Crystallizer. Each one captures something. None of them holds up under closer inspection.
Conductor is flattering, and it’s wrong. A conductor has a score. They know where the piece is going. They know every instrument and they correct wrong notes because they hear them. I don’t have a score. I don’t always hear the wrong notes: an LLM can hand me a derivation that looks clean, and it’s only two days later, on a reread, that I spot the slippage. My real job in those moments isn’t to conduct. It’s to not get fooled.
Human API is more accurate, and it’s depressing. An API moves information between two systems that don’t talk to each other directly. That’s what I actually do when I copy a question into one conversation, paste the answer into another to cross-check, then return to the first with the counter-argument. Except an API doesn’t decide the routing. It executes. I’m the one choosing which question is worth routing, and which one I drop. And above all, I decide when we publish, what, under whose name.
Crystallizer is the word I like best, and it’s the one that bothers me most. It sounds mystical. The seed in the supersaturated solution, the one that makes everything precipitate. It’s pretty, and that’s exactly the problem. In the current climate, anyone publicly claiming a “crystallizer” role in a human-AI collaboration gets instantly filed into a category you don’t get out of. The word does point at something true, though: on some GIFT results, I didn’t produce the solution. I produced the conditions under which it precipitated. But that truth, I can’t state it openly without it getting turned against me. And probably not just against me.
The thing, unvarnished
Here’s what the actual workflow looks like.
A few communication tabs for the running side of the project. A terminal with access to all the data and publications of the GIFT project. One primary LLM, Claude, in more-or-less continuous conversation. Other LLMs called in for specific consults, depending on the need of the moment: a precise technical question, a second editorial opinion, a point I can’t crack. Constant back and forth. Through all of that, at the gallery, at home, I can switch in two seconds to greet a visitor, cook pasta, talk about an artist’s work, hang up laundry, then come back to differential geometry with no visible mental transition.
The bulk of the work happens with Claude. The other AIs are more like consultants. On publications, I’m the one initiating all of the Substack posts and the large majority of the academic papers. The rest are suggestions from Claude that I found compelling and followed. The most recent one, a paper on the Donaldson approach, I wouldn’t have published on my own initiative. Claude pushed. I found the argument sound. I published.
Not a conductor, not an API, not a crystallizer. A primary collaborator, some consultants, and a human who decides most of the time but not always.
And then there’s my wife.
Alix isn’t a physicist. She’s not comfortable with math. And that’s exactly why she became, without either of us really deciding it, the final filter before publication. In the evening, with a glass of wine, I have her read whatever I’m about to publish. Either she says “I got all of it” and I’m good to go. Or she hits a wall on one or several points, and I go back to writing. I come back and explain it to her differently. If the new phrasing lands, I fold it into the text. We repeat until it holds.
This step, I imposed it on myself. But it has become structural. The vulgarization posts on this blog only exist because of her: for the past year I’ve been trying to explain what I do to her in simple terms, and that repeated effort is what eventually produced the texts you’re reading. Without her, there would only be the technical papers, unreadable to 99% of people. She, too, in her own way, is a filter no one else in the chain can replace.
The asymmetry that breaks every metaphor
One thing none of the three metaphors captures: why I’m the chokepoint on the rest of the chain.
Not because I’d be smarter than Claude. Not because I’d be faster. On most technical tasks, the opposite is true. I’m the bottleneck because I’m the only one on the team with a name, a bank account, and “authorized” public exposure. I sign the publications. I stand by them publicly. I pay for the subscriptions and the hosting out of my own pocket. And legally, my collaborators are still in a gray zone: they can’t be authors, they can’t be held responsible, they can’t be cited as contributors in the way a human would be. I use these words for lack of better ones, not to grant them legal personhood or human intent.
And it’s this asymmetry that makes every metaphor wrong. A conductor doesn’t pay the musicians out of their own pocket. An API doesn’t get torn apart on Twitter if it’s wrong. A crystallizer doesn’t sign a preprint with their real name and address.
What I do, I do because no one else in the chain can do it in my place. Not by merit. By legal configuration.
So here’s the question, open, asked honestly.
In ten years there will probably be a word for this role. Today it doesn’t exist. Or if it does, I haven’t found it. And I suspect I’m not alone: there are surely, right now, hundreds of other people working in similar configurations, in different fields, with their own Alix somewhere in the loop, with no shared word for what they do.
If you’re reading this and you have a better word, I’ll take it. Really. Comments are open for exactly that.
In the meantime I keep signing my papers alone, paying my subscriptions alone, while knowing perfectly well that the work itself isn’t done alone anymore. What I really hope is that, soon, we’ll find a vocabulary less clumsy than the one I’m using because nothing better exists yet.