The Rhizome and the Trail
Nine years ago I ended an essay on stigmergy with two words: stay tuned. This is me, belatedly, tuning back in.
The argument then was simple, and I still believe it. Below roughly twenty-five people, groups coordinate by talking to each other, what Mark Elliott called social negotiation. Above it, talking does not scale, and the successful collectives switch to a different mechanism entirely: they coordinate through marks left in a shared environment. A termite lays a pheromone-laced mudball; the next termite is drawn to build atop it; a cathedral rises with no architect and no meeting. A starling minds its nearest neighbours and a sky-filling figure appears that no bird intended. Stigmergy. Coordination as a consequence of simple local acts, never as a decision anyone made.
What I did not say, because I had not yet worked it out, is that stigmergy is only half of an answer. It tells you that large groups coordinate through the environment. It does not tell you what the environment should be made of. A pile of pheromone is not a mound. The mark only does its work if the medium it lands in has the right shape: a shape that lets a mark left here be felt over there, that lets structure accrete without anyone holding the master plan, that survives the loss of any single piece. Get the medium wrong and the marks just pile up and rot, which is, more or less, what a comment section is. So the real question, the one that has taken me the better part of a decade to take seriously, is: what is the shape of an environment that supports stigmergy?
I think the honest answer was written in 1980, by two Frenchmen who had never heard of a database.
The tree and the rhizome
Gilles Deleuze and Félix Guattari opened A Thousand Plateaus by drawing a contrast between two ways the world can be organized. The first they called the tree. A tree has a root, a trunk, and branches; there is exactly one path from the top to any leaf; everything has its place, and its place is singular. Western knowledge, they argued, is relentlessly arboreal: the outline, the taxonomy, the org chart, the file system. One thing, one location, one parent. To find anything you trace it back toward the root.
Against the tree they set the rhizome: the botanical kind, the underground stem of crabgrass and ginger and mint, the mycelial mat under a forest. A rhizome has no root and no trunk. Any point of it can be connected to any other point. It is, in their phrase, all middle; you enter it anywhere. Cut it, and it does not die; it simply grows back from wherever it was severed. It is a different kind of thing from a tree that has merely grown tangled, organized by connection instead of descent.
Strip away the philosophy (and I will, because my collaborators are builders, not theorists) and the rhizome leaves you with three plain commitments:
- Any point connects to any other point. No relationship has to be routed through a parent.
- There is no trunk and no root. Structure is not imposed in advance; it is whatever the connections happen to make.
- It decomposes without erasure. Remove a piece and the rest survives, scarred but legible, regrowing around the gap.
Read those three lines again with stigmergy in mind. That is exactly the medium the marks needed. A rhizome is what a healthy trail-system looks like from above.
The engineer’s word for a rhizome is “graph”
Here is the part I want a non-specialist to actually keep, because everything downstream depends on it, and it is genuinely not hard.
A graph, in the mathematical sense, is two things: a set of nodes (the things: a person, an idea, a moment in a video, a claim) and a set of edges (the relationships between them: cites, contradicts, is-the-same-as, replies-to). That is the entire definition. A node is a dot; an edge is a line between two dots. You already think in graphs every time you say “this reminds me of that.”
A file system is a tree: every document lives in exactly one folder, and to be in two places you must duplicate or alias. A graph is a rhizome: a thing lives nowhere in particular and everywhere it is connected. It does not have a place; it has its relationships, and those are its place.
One more idea and the toolkit is complete: bidirectional linking. An ordinary link, a footnote or a URL, points one way. You can follow it forward, but the thing you arrived at has no idea you exist. A bidirectional link is visible from both ends: the destination knows who points at it. The humble version of this is the “what links here” box on a wiki page. It sounds like a small convenience. It is not. A one-way link is a citation; a two-way link is a conversation. The moment every link can be walked backward, the graph stops being a pile of documents with arrows drawn on it and becomes a single connected fabric you can traverse from any thread to any other. That is the rhizome’s first commitment, made mechanical.
Graph discipline
None of this is free. The cost has a name, and I have started calling it graph discipline: the steady refusal to let knowledge sit as orphaned prose, or to bury it in a single hierarchy because a hierarchy was the easy thing to type. Discipline means that every claim gets tied back to nodes and explicit, named links. It means writing the smallest thing that can stand alone and stating what it connects to, instead of writing a thousand-word document that connects to nothing because it is about everything. It is more work in the moment, the way washing your hands is more work in the moment. You pay it because of what it buys, and what it buys turns out to be enormous, and almost no one talks about why.
What the discipline buys: the graph becomes a matrix
Here is the punchline, and it is the reason this old idea is worth reviving now and was not worth building in 2019.
A graph that has been kept disciplined stops being a picture and becomes mathematics. Write down which nodes connect to which, and you have written a table of numbers: a matrix. And a matrix is the native food of linear algebra and of every machine we have built to think. You can go further: you can hand each node a short list of numbers, an embedding, chosen so that nodes sitting near each other in that numeric space are the ones that are related in meaning. Now “related” is a distance you can compute. Clusters of allied ideas fall out as regions. The rare node that bridges two camps that otherwise never touch, the most valuable thing in any intellectual community, is findable by arithmetic. Duplicates that three different people wrote in three different words land on top of one another in the space, and announce themselves.
(The term I kept groping for, in conversation, is exactly this: a continuous representation of the graph, embeddings over an adjacency matrix. The discrete tangle of dots and lines, re-expressed as something smooth enough for calculus and matrix multiplication to grip. That re-expression is the whole bridge.)
And once the graph is a continuous object, a large language model can read and write it as comfortably as it reads a sentence, because a language model is, under the hood, a machine for moving through exactly this kind of meaning-space. The graph speaks its native tongue.
The melt
Now put the pieces together, because this is the thing I have actually been building toward.
Recall what small groups spend their negotiation on. Not, usually, on the ideas themselves. They spend it on the coordination overhead: Are these two notes saying the same thing? Where does this belong? Whose version wins? Who has the standing to merge them? That overhead is the friction (the ego, the loudest voice, the dread of editing someone else’s words) that makes collaboration exhausting and makes it fail at scale.
Every one of those questions, asked of a disciplined graph, is now a computation rather than an argument. “Are these the same?” is a distance in the embedding space. “Where does this belong?” is answered by its links; it belongs to all of them. “Merge these” (the operation the rhizome would call a collapse, a disambiguation, a pruning that loses no information because nothing is ever destroyed, only tombstoned) is something the machine proposes and a human merely blesses. The trail-following that stigmergy promised but that no human has the patience to do by hand, the machine does tirelessly, legibly, and reversibly.
That is the melt. The socially frictioned parts of working together do not get resolved. They get dissolved, handed to a substrate that has no ego and keeps perfect receipts. Agreement stops being something a group has to negotiate and becomes something the environment detects. Consensus without consensus. The wiki is not written; it precipitates.
There is an irony I have made my peace with. The wiki, the great hope of stigmergic collaboration, the page that writes itself, quietly lost. Not the mechanism; the venue. Where people now actually think together, teach, and pass ideas along is no longer the edited page but the stream of talking faces: the lecture, the recorded conversation, the clip forwarded with a timestamp. Video overtook the wiki as the medium of collaboration and dissemination, and it did so with none of the machinery in this essay.
So the question I will take up next is the obvious one: what happens when you give that medium (conversation, watched and replied to in the open) the precipitating, self-organizing substrate the wiki never quite earned?