History
The adjacent possible
At any moment only the ideas one step from what already exists can be built; the rest wait until their parts arrive.
Stuart Kauffman, theoretical biologist; the term appears in his Investigations (2000) and his 2003 Edge essay of that name. Steven Johnson applied it to invention in Where Good Ideas Come From (2010).
Kauffman's adjacent possible is the set of states reachable in one step from the present one: the molecules one reaction away from those that already exist, the moves that are legal from this board position. He observed that biospheres keep expanding into the adjacent possible and, by doing so, increase the diversity of what can happen next. Johnson borrowed the idea for human invention and called it a shadow future hovering at the edge of the present, a map of all the ways the present can reinvent itself. His test case is blunt: there was no way to invent a microwave oven in 1650, however clever you were. The pattern predicts that inventions arrive when their components do, not when the genius does, which is why the same thing tends to be invented in several places at once and why an idea that arrives too early stalls.
It holds because every new part can combine with every existing part, so the frontier of what is buildable grows faster than the stock of parts. In practice the useful question is "what became possible this year that was not possible last year". Deep learning is the clearest recent case. Convolutional networks for reading handwritten digits were published by 1990 (the 2012 paper cites that work), and for twenty years they stayed small. The 2012 ImageNet result needed two things that had only just appeared: a labelled dataset of 1.2 million images and consumer GPUs with 3 GB of memory, and its authors wrote that their results could be improved simply by waiting for faster GPUs and bigger datasets. The idea had been waiting on its parts.
The model explains timing after the fact more easily than it predicts it. Kauffman's own point is that the adjacent possible cannot be listed in advance: nobody could have predicted the Web from the parts that preceded it, so the model tells you where to look, not what you will find. Merton added a caution from the history of science: the ingredients of some discoveries sat in the culture for years before anyone assembled them, so "reachable" is not "inevitable", and someone still has to take the step. The phrase is also easy to misuse as a reason to dismiss an early idea as impossible, when the honest reading is that one named part is missing and the best move may be to build that part.
Example
In 2012 Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained a deep convolutional network on the 1.2 million labelled training images of the ImageNet contest and won with a top-5 error of 15.3 percent, against 26.2 percent for the second-best entry. The network had 60 million parameters and took five to six days to train on two NVIDIA GTX 580 GPUs with 3 GB of memory each. The authors wrote that the network's size was limited mainly by the memory available on current GPUs, and that their results could be improved simply by waiting for faster GPUs and bigger datasets. The ideas were decades old; the parts, a labelled dataset of that size and a cheap GPU, were new.
Where it applies
For an AI roadmap, list what shifted in the last twelve months (context length, GPU memory, a new open dataset, a cheaper model) and ask what is now one step away that was not before. For a product idea that keeps failing, name the missing part; if it does not exist, the idea is waiting, and building the part may be the real project.
Think about it
Which idea on your list is waiting on a part that does not exist yet, and which part arrived this year that nobody around you has used?
Applied it to something real? Log it.
Test yourselfIn adjacent-possible terms, why did the deep learning breakthrough come in 2012 rather than in 1990?
The idea existed but its parts did not: a labelled dataset of 1.2 million images and GPUs with enough memory only arrived around 2012, which brought the result within one step of the present.
Related models
- Multiple discoveryThe same discovery is often made independently by several people at about the same time, because the field was ready for it.
- Network effectsA product becomes more valuable to each user as more people use it, which makes leaders grow faster and makes leaving hard.
- EmergenceSimple local rules among many parts produce patterns at the level of the whole that no part contains or intends.
- Path dependenceEarly choices, including accidental ones, can lock a system onto a track that later choices cannot easily leave.