The Adjacent Possible

A door you didn't know was there

In the late 1990s, the theoretical biologist Stuart Kauffman gave a name to something every evolutionary system already understood. He called it the adjacent possible: the set of states one step away from where you currently stand. Not the entire space of what could ever exist, but the much smaller, much more interesting space of what could exist next.

A bacterium does not invent flight. It cannot. The genome it has is too far from the genome it would need; the mutational distance is unreachable in any single generation. But the adjacent possible — the proteins one or two amino-acid swaps away — is dense, and any one of them might unlock a new metabolic pathway, a new defence, a new niche. The bacterium does not search the whole space. It searches the rim of where it already is.

Kauffman's claim, formalised in Investigations (2000) and developed for a wider audience by Steven Johnson in Where Good Ideas Come From (2010), was that this same logic governs every generative system: ecosystems, economies, technologies, ideas. Innovation does not happen in the vacuum of pure possibility. It happens at the boundary of the present, where the next step is reachable. The future is built one door at a time, and only some of the doors are visible from where you are.

Why discovery clusters

If the adjacent possible is real, an empirical prediction follows: discovery should cluster. Not in the sense of fashion or fad, but in the deeper sense that breakthroughs should appear in places where the rim is densest — where the most adjacent doors are visible from one room.

The historical record is unforgiving on this point.

Bell Labs, Murray Hill, between roughly 1947 and 1962. Inside one institution: the transistor (Shockley, Bardeen, Brattain, 1947), information theory (Shannon, 1948), the laser concept (Schawlow and Townes, 1958), the solar cell, the cellular telephone architecture, UNIX a decade later. Jon Gertner's The Idea Factory (2012) catalogues the institutional design that made this possible: long corridors that forced encounter, a culture in which a physicist could not avoid running into a metallurgist, a chemist, a circuit designer. The adjacency was not metaphorical. It was hallway-shaped.

Florence, roughly 1450 to 1500. Brunelleschi's dome, Alberti's perspective, Botticelli, Leonardo, Michelangelo's apprenticeship, the Medici banking innovations, the early printing presses, the rediscovery of classical texts being smuggled out of a collapsing Constantinople. A single city the size of a contemporary university campus produced a density of ideas that historians still struggle to explain by individual genius alone. The Medici effect — Frans Johansson's term for what happens when previously isolated fields are forced into the same room — is named for this place precisely because the place was the mechanism.

Vienna, 1900–1914 — the same coffee-house geometry I wrote about in the previous post. Princeton in the 1940s, with Einstein, Gödel, von Neumann, and Turing as a wartime guest, walking the same lanes. Cambridge before the First World War, with Russell, Whitehead, Wittgenstein, Keynes, and the Bloomsbury circle in walking distance. Athens in the fifth century BCE.

The pattern is too consistent to be coincidence. Where adjacency is dense, the rim is wide, and more of the doors are open at once. Where adjacency is sparse, the rim is narrow, and the same individuals — equally talented, equally hardworking — produce far less.

"The trick to having good ideas is not to sit around in glorious isolation and try to think big thoughts. The trick is to get more parts on the table." — Steven Johnson, Where Good Ideas Come From (2010)

Johnson's framing is deliberately deflationary. Not lone genius. Not divine inspiration. More parts on the table. Adjacency is the parts being close enough to combine.

The geometry of "next"

If adjacency drives discovery, the question becomes geometric: what does next mean in the space of ideas?

In Kauffman's biology, the answer is precise — one mutation away, one protein swap, one regulatory rewiring. The metric is well-defined because biology is. In the space of human knowledge, the metric is messier, but it still exists. Two concepts are adjacent if a person who deeply understands one can, with modest effort, understand the other. Two fields are adjacent if their methods translate. Two people are adjacent if they share enough vocabulary that a lunch conversation produces something neither of them would have produced alone.

Distance, in this sense, is not Euclidean and not symmetric. It depends on what you already know, what you're prepared to encounter, and crucially — on which connections are visible to you at all. A statistician and a molecular biologist may have been mathematically adjacent for forty years before anyone built the institutional bridge that let them notice. The graph existed. The path through it didn't.

This is the crux. Adjacency in the space of ideas is partly real and partly an artefact of what you can see. Florence was not adjacent to Beijing in 1480, despite both cities containing extraordinary minds, because no edge existed between them. Bell Labs was adjacent to itself by architectural design. The internet was supposed to render geography irrelevant — and in some ways did, while introducing new geographies of attention, recommendation, and platform that decide which ideas now appear next to which.

Interfaces decide what is near

Every tool we use to navigate knowledge implicitly defines a notion of adjacency. Wikipedia decides that an article is adjacent to whatever its blue links point to, and not adjacent to anything else, no matter how semantically close. A search engine decides that a query is adjacent to the documents it ranks highly, which is to say, adjacent to whatever a particular ranking model believes you want, which is to say, adjacent to a function trained on what previous users clicked. A library catalogue decides that adjacency runs by Dewey number, an artefact of nineteenth-century classification choices that no one chose with adjacency in mind.

Large language models decide adjacency by latent embedding. A concept is near another concept if their high-dimensional vectors point in similar directions, which is to say, if they appeared in similar contexts across the training corpus. This is a different notion of adjacency than any prior tool produced — denser, less linear, less constrained by editorial choice — and it has begun to expose connections that hyperlink graphs and search rankings systematically hid. But that adjacency, too, is an artefact: of the corpus, of the training objective, of which contexts the model happened to see often enough.

None of these notions of adjacency are wrong. All of them are partial. Each one widens some doors and narrows others. The interface is not a neutral window onto a fixed terrain. The interface is the terrain. Choose a different one and the rim of the adjacent possible reshapes.

What it would mean to widen adjacency on purpose

If interfaces define adjacency, and adjacency defines what is reachable, then the most consequential design choice in any tool for thinking is which adjacencies it makes visible.

A document reader makes the next paragraph adjacent. A hyperlink graph makes the linked article adjacent. A recommendation feed makes adjacent whatever the engagement signal flagged. A search engine makes adjacent whatever ranks for your query. Each of these is a particular geometry, optimised for a particular use, and each leaves enormous regions of latent connection structurally invisible.

The interesting question is what a tool would look like if its primary purpose were the inverse. Not to surface what you already half-know to look for, but to make visible the rim — the doors you didn't know were there, the concepts that are semantically adjacent to the one you just typed but that no editorial process ever pointed you toward. A tool whose default action is not here is your answer but here is the topology around your question.

Such a tool would have to do several uncomfortable things. It would have to refuse to rank. It would have to tolerate showing the user concepts the user does not yet have language for. It would have to treat the question as a coordinate, not a query, and the response as a neighbourhood, not a list. It would have to admit that a user might leave the session having learned something they could not have asked about when they arrived.

This is what serendipitous discovery has always meant, before search engines redefined it as a synonym for accident. Real serendipity is the discovery of the adjacent possible — the concept that was always one step away but that no instrument made visible until now.

The work in front of us

Vhaeon, in its current form, is one attempt at this geometry. You type a word; a region of a knowledge graph appears; you can see, at a glance, what is adjacent to what you typed — not because someone hyperlinked it, but because the structure of the corpus places it near. The interface is rough, the corpus is partial, and the experience is closer to a sketch than a tool. None of that is the point yet.

The point is the premise. Hyperlinks gave us one notion of adjacency. Search rankings gave us another. Embedding-based retrieval is giving us a third, and the better large-scale demonstrations of it are still ahead of us, not behind. Each generation of these interfaces will widen or narrow the rim of what feels reachable, and the design choices we make now — about what to surface, what to rank, what to suppress, what to make navigable — will decide which adjacent possibles get explored at all over the next decade.

Kauffman's original observation was that the future arrives one door at a time. The corollary, which he did not need to state and which the institutional history of innovation makes obvious, is that the doors that get noticed are the doors someone built a way to see. The interface is the corridor. The corridor decides what is adjacent.

That is not a small thing to be working on.