Movement Three · The Implications
Chapter 22
The Next Hundred Million Years
The desert is still there.
The colony Gordon began tracking in 1985 is, in all likelihood, gone now. Harvester colonies live about as long as their queen — fifteen to twenty years, sometimes as much as thirty. The queen dies, and within a season or two the workers die with her, and no new queen takes her place. The nest entrance fills with sand. What had been, for two decades, a point of intense and patient activity becomes, again, indistinguishable from the ground around it.
But the desert is full of other entrances. Some belong to colonies that were founded while Gordon was watching. Some belong to colonies founded since. At any given hour, on any given patch of New Mexico dirt, foragers are leaving and returning, sensing the rate of contact in a corridor, crossing a threshold encoded in a nervous system, departing into the heat. The particular colony is gone. The thing the colony was doing has not paused.
This is worth holding still for a moment.
The colony was never the point. The colony was an instance.
The architecture that ran in that colony — agents, signals, marks that strengthen and fade, paths that harden into knowledge no single agent holds — did not begin in 1985, and it did not end when the queen died. It was running in harvester ants before Gordon was born. It was running in the ancestors of harvester ants before there was a New Mexico to run in. The fossil record places recognizable ants in the Cretaceous, alongside the dinosaurs, more than a hundred million years ago. The architecture is at least that old.
The species that expressed it then are extinct. Every one of them. The ants that first refined trail-laying and threshold-based recruitment and the chemical bookkeeping of a nest are gone, as completely as the colony Gordon tracked is gone. What they were doing did not go with them. It passed into the species that came after, and after, and after — fourteen thousand of them now described, with more uncounted, on every continent but one.
The architecture outlived its implementations. It always has. That is the first thing to notice. An implementation is a particular colony, a particular species, a particular arrangement of carbon that runs the pattern for a while and then stops. The architecture is what the implementations are implementations of. It does not live in any of them. It is older than all of them, and it will be older still.
There is no reason to expect this to change. The species expressing the architecture today will, on the evidence of the last hundred million years, eventually go the way of the species that expressed it before. Something will express it after them. The dirt will hold trails. The trails will fade. The foragers will follow the strong ones and the colony will know things no forager knows. For another hundred million years, give or take, the architecture will run, in materials that have not been named yet, in the bodies of animals that do not exist yet, in deserts that are presently ocean floor.
None of this requires anyone to be in charge of it. That was the lesson of the very first ant. There was a rate, a threshold, and a departure. There was no plan and no planner. The architecture does not need a steward. It needs only agents, a substrate to hold what they leave behind, and time. It has had a great deal of time.
The same separation — architecture from implementation — has been visible in every substrate the book walked through, once the word was available to see it. The market outlived the merchants. The traders who set prices in a particular year are gone; the price system that aggregated what they knew runs on, with new agents in their seats. The city outlived its residents. The streets and the records and the trust networks persisted while the people who built them came and went. The science outlived the scientists, the body outlived its cells, the ledger will outlive the machines presently running it. In each case the agents were replaceable and the substrate was not, and the intelligence — every time — was in the substrate. The colony is only the oldest and clearest instance of a thing that turns out to be everywhere.
Something new is happening to the architecture now. It is happening in a different material, and it is happening fast.
In the early decades of the twenty-first century, humans began — mostly without naming it, often without recognizing it — to build the same architecture in silicon. Not the ants. The architecture. Substrates that hold what agents leave behind. Channels that carry signals between agents that were not built together and do not trust each other. Marks that strengthen with success and fade without it. Records that persist when the agents that made them are gone. Flows of value that select, automatically, for what works and starve what does not.
The pieces did not arrive together or in order. They arrived from different directions, built by people solving different problems, most of whom were not thinking about ants and many of whom were not thinking about intelligence at all. A way to keep a shared record that no institution holds. A way for a program to own something, to be paid, to pay. A way for agents to find each other and route work between them. Each piece was built for its own reasons. Assembled, they are the substrate, with the four conditions holding not because an institution permits them but because the protocol cannot prevent them.
The implementations are crude. This needs saying plainly, because the architecture's pedigree should not be borrowed to flatter its first attempts. The early silicon substrates are clumsy in the way that early instances of anything are clumsy. They leak. They are wrapped in speculation that has nothing to do with what makes them matter. They make choices that will look, in retrospect, obviously wrong. Their selection pressure is often weak, their memory often shallower than it appears, their openness often granted by some party who could revoke it after all.
Many of them will fail. Some will fail badly — not quietly, the way the colony failed when its queen died, but loudly, taking value and trust and years of work down with them. This is not a reason for alarm and it is not a reason for reassurance. It is what the early instances of a pattern do. The Cretaceous ants that first tried trail-laying did not get it right either. The ones whose chemistry was wrong, whose thresholds were badly tuned, whose nests were poorly placed, are not in the fossil record because they did anything memorable. They are in the fossil record, when they are there at all, because they failed and something better followed.
It is worth being precise about what failure means here, because it is not what failure usually means for a machine. A machine fails when it stops working. A substrate fails when its conditions stop holding — when its population quietly closes, when its memory turns out to be carried by a single agent after all, when the value stops flowing and the selection pressure goes slack, when no one is exploring anymore and the substrate locks onto the first thing that worked. These are not dramatic failures. They are often invisible from inside, and they look, right up until the end, like success. A substrate that has stopped accumulating intelligence can run for a long time on what it accumulated before. The colony with the dead queen still forages for a season. The reader who has the language will be able to tell the difference between a substrate that is learning and one that is merely still warm.
What does not follow from the crudeness of the early implementations is that the architecture is wrong. The architecture is not wrong. It is the most thoroughly tested architecture for intelligence that exists on the planet, and it was tested not by argument but by a hundred million years of everything that did not work being removed from the world. The implementations being built in silicon will be selected the same way. The weak ones will fade. The ones whose conditions actually hold will be marked, by the only thing that marks substrates, which is whether intelligence accumulates in them or does not. And the reader is no longer at the mercy of an announcement on this point, because the reader has already been given the instrument that settles it. Freeze every agent, so that nothing inside any agent is permitted to change. Let the substrate keep running — marks accumulating, failures fading, patterns hardening, value flowing. Come back after time has passed and ask whether the system now solves problems it could not solve before. If it does, the capability accumulated in the only place left to accumulate, which is the substrate. Every claimed implementation of the architecture can be put to the Gordon Test. The failures will fail it and the real thing will pass it, and no announcement, no benchmark, and no measure of sheer size will substitute for passing it.
There is a temptation, at the end of a book like this, to say what will happen. To predict the year, name the breakthrough, draw the curve. The temptation should be resisted, because the architecture does not work that way and never has. Convergence says nothing about when. It says only that a substrate satisfying the conditions cannot become permanently dumber — that it can stay the same or grow more capable, and that in the limit, given enough time and enough variety, it reaches everything reachable. The limit is not a date. The ants did not converge on a schedule. They converged because they did not stop, and because what worked was kept and what failed was not.
The one thing the material changes is the speed of the keeping and the failing.
A colony tests a pattern across a foraging season. A silicon substrate can test a pattern in the time it takes a signal to cross a continent, which is a fraction of a second. The selection cycle that took carbon a hundred million years to run to its present state can run, in silicon, in days or hours. But the cycle is the whole loop — test, verify, mark, fade, harden — and it runs no faster than its slowest stage, which is the verifying. So the speedup does not arrive everywhere at once; it arrives domain by domain, fastest where the world answers quickly and a result can be checked at the speed of the signal that carried it, and slowest where verification waits on consequences that take time to land or on judgment that cannot be hurried. This is the whole of silicon's advantage, and it is enormous, and it is also the reason the early failures will come quickly and the early successes will compound quickly, and the reason the people building these systems should be more careful than they have so far generally been. A substrate that learns fast learns its mistakes fast too. The architecture is patient. The material it is now running in is not.
There is one asymmetry the speed introduces that is worth naming, because the reader will watch people get it wrong. The carbon substrate had a hundred million years to discover, the slow way, which conditions could be allowed to lapse and which could not. It discovered that a colony can lose almost any individual ant and lose nothing, and that a colony that loses its memory of where the food is loses everything. It discovered that forgetting is not a flaw — that a substrate which never fades cannot adapt, and a substrate which fades too fast cannot accumulate. It discovered the rate at which resistance should fade faster than strength, so that the colony forgives a bad path sooner than it forgets a good one. These are not small tunings. They are the difference between a substrate that converges and one that thrashes. Carbon found them by killing every colony that got them wrong. Silicon will not have a hundred million years to find them, and the people building in silicon will be tempted to guess. Some of the loud failures will be failures of exactly this kind — conditions set by guess, on a substrate moving too fast to correct gently.
What can be said — the only thing that needs to be said, at the close — is smaller and more certain than a prediction.
The architecture is being implemented in a new material. It will go on being implemented, through many failures, by many hands, most of whom will not call it what it is. And the reader of this book will now be able to see it.
That is what the book was for. Not to persuade, and not to predict, but to hand over the language. The reader who has come this far has the notation now. Substrate, signal, mark, warn, fade, follow, harden, path, agent, colony. These were the words a careful observer would reach for to describe what an ant does in the dirt. They turn out to describe a great deal more. A market. A city. A body. A science. A ledger. The next generation of artificial systems being assembled, piece by piece, in the early decades of this century.
The reader will begin to notice the architecture in places they had not been looking for it. They will see a substrate where they had seen a tool, and agents where they had seen a crowd, and a path strengthening where they had seen only repetition. They will watch some new system being celebrated as a mind and notice that it has no substrate behind it, no memory it does not carry, no agents but itself, and they will understand why it will not become what its builders hope. They will watch some unglamorous arrangement of agents and signals and value flow, dismissed as plumbing, and recognize the shape of the thing that has been intelligent for a hundred million years.
The reader will also see it where it has been all along, in the substrates the book did not have to invent. In the office where the procedures persist while the staff turn over, where the marks left by people long departed still route the work of people who never met them. In the neighborhood that is safe or is not, a judgment held by no one and known by everyone, hardened out of ten thousand small encounters none of which decided it alone. In the body, doing this very moment what the colony does in the dirt, allocating, signaling, remembering, forgetting, with no cell in charge and no cell containing the plan. The architecture was never rare. It was only unnamed. What changes, for the reader, is not the world. It is what the reader can now see was always in it.
This is not a special sight. It does not require expertise. It requires only the language, and the language has now been handed over. Once it has been, the architecture is difficult to stop seeing.
The desert is still there. The sun is on the dirt, well above fifty degrees, the way it was in August of 1993. Somewhere on that ground, at an entrance that did not exist when Gordon began and may not exist when the next observer comes, an ant returns with a seed roughly twice its body weight. It pauses at the mouth of the nest for a moment, on the order of a second, the way ants have paused at the mouths of nests for longer than there have been primates to watch them. Then it descends into the dark.
A few seconds later, another ant comes out, and heads into the desert.
Tens of thousands of times a day. For a hundred million years. The architecture, running.