The Map

It was the most sublime map ever made; superbly detailed and wonderfully dynamic. They said a trillion-parameter model drove the real-time updates. Whether you wanted a simple route to your destination or a restaurant recommendation, if you were in the territory, this was the map you wanted.

They said it was so responsive to even the subtlest of event currents, the stream had to be artificially delayed to avoid spoilers. The speculative extrapolation ran minutes to hours ahead of the evolution of the territory, and if you knew how to hack in with a properly jailbroken client, you could surf the liminal future. The map was not so much a map as a live inference frontier. It would only be a mild exaggeration to say that it tracked and anticipated the fate of every blade of grass in the territory.

It was as much an evolving spatiotemporal promise as a map. And it was right a lot.

Uncannily right. Not just about traffic or the weather, but about vibes and moods. About whether you should go to the concert or to get an ice-cream.

Not that it was never wrong — but big or small, the errors always had a strange beauty to them. They could lead you to appreciate the territory in poignant new ways. To experiences you didn’t know were exactly the ones you were searching for. To be lost on the map, or misled by it, was almost better than knowing where you were on it. It was wrong in ways that made omniscience seem uninteresting.

And for those who cared about such things, it was small, sometimes achieving compressions of over 98%. Even the smallest devices could run the map efficiently, even offline.

To look at, no through the map was to gaze into the evolving, distilled, soul of the territory. The AR display overlays generated from the map, some nostalgics argued, had some of the charm of old black and white photographs by the early masters of photography, gently tugging the eye and soul into communion with the territory they framed and presented. And even the least poetically minded had to admit there was some truth to that.

Like all good maps, it was opinionated, gently nudging various search, adaptive rendering, and navigation algorithms towards regimes of greatest serendipity. But unlike the maps made by competing vendors, which were merely good but not great, it seemed to somehow draw out the best opinionated behaviors from the algorithms it hosted. And it hosted algorithms in the best sense of the word. With a kind of attentive courtesy that seemed to bring out the best in even the humblest of client apps.

Many, made uneasy by the utter benevolence of the map, sought to do without it. There must be a dark side, they said. Yet, even the most fervent detractors, the ones who refused to use any device that made use of the map, had to admit that none of the standard criticisms of technology seemed to apply.

The map was carbon-negative by design, even without accounting for various obvious positive spillovers. Economists had shown that it had nearly tripled well-being, doubled productivity, and halved carbon intensity in the territory within the first decade.

And it did so without baking in fragility-convenience tradeoffs as so many twentieth century technologies had done. In fact, one analysis showed (with a very significant effect size) that the map appeared to do the opposite: it added a frisson of low-grade ennobling hormetic inconvenience that somehow made every user, whether algorithmic or human, more resiliently attuned to the territory, and more able to anticipate and respond to risks. During the great hurricane of ’33, for example, the territory suffered only half the damage of its less exposed neighbors. And during the great pandemic of ’82-84, it had the lowest fatality rate, and the quickest economic rebound, despite having the highest population density in the region.

And the macros did not, as some darkly suspected, obscure deep and narrow failures experienced at micro levels. It wasn’t just the big, high-volume apps that performed better with the map. Even the long tail of client algorithms converged faster — but even when they diverged, they appeared to miraculously discover rich new differential inference regimes that made the curators surprised and glad to retune them. And even the humblest human immigrants swore that merely moving to the territory had made them somehow more alive; that even being lost on the map made them feel more at home than knowing where they were ever had outside of it.

It was a chef who came up with the metaphor that best captured the general consensus about the map: it was somehow bringing out the terroir of the territory itself, so to speak, as though it was somehow gently, continuously, cooking the territory, and everything in it. With exactly the right amounts of perfectly matched subtle spices the territory needed, to fully express itself.

The economists, of course, had a drier take: the map was a sort of macro-algorithmic self-fulfilling prophecy. A famous series of papers by a group of economists at the University, which won the authors the Nobel, demonstrated that under suitable simplifying assumptions, the map was in fact functioning as a territory-wide efficient market. And crucially, the simplifying assumptions, far from sneaking in ungrounded optimistic biases, appeared to lead to an underestimate of the performance of the map. The efficient market reduction was, in most regimes, a lower bound on the performance of the map. In practice, it generally performed better than the formal models predicted. It appeared that that map actually generated an ongoing performance surplus that manifested as a kind of unreasonable luck accruing to all who operated on it. The map was turning the territory not just into an efficient market, but an unreasonably lucky one.

A half-century after the first version of the map was released, the population of the territory had doubled, and the economy had tripled in size, even though the territory itself had only grown by a modest 14% in area. The compute load of the map though, across both the model in the cloud all client devices, had only grown by 4%. Each update had delivered major efficiency gains. By the thirtieth year, the map was effectively a local monopoly. A full 99.4% of all devices in the territory ran inference engines that relied on the map, and 94% had opted-in to providing telemetry and accepting default continuous over-the-air updates.

In the fortieth year, shortly after version 17.31 went out over the air, the corporation voted to dissolve itself, release the code open-source, and make the aggregated data streams public and auditable. The map, it turned out, had become economically self-sustaining, with apparently no need for active management. A half-hearted antitrust lawsuit that had been taking shape dissolved overnight.

Not that there was much open-source maintenance to do by that point. That code was mature and robustly self-healing. Hardware upgrades and configuration evolutions seemed to require no special handling. After a brief frenzy of audit activity, the data flows ended up being mostly of interest to the small cottage industry of artists that sprouted up to pipe it into various installations around the territory, and eventually around the world.

A small band of volunteers emerged to do whatever stewardship might prove to be necessary, but as it turned out, there was very little for them to do besides indulge a nerdy fascination with the transparent but illegible evolution of the map. More loyal fandom than intercessionary priesthood.

Perhaps the only people unhappy with the map were the ones unable to join the territory within which it was usable. For whatever reason, the training code that generated the models that created the maps did not converge when seeded in most other territories. Where it did converge, it mostly produce fragile, unreliable, unexceptional maps with strictly middling levels of performance. It led the market in no other territory — those who chose to use it anyway appeared to be driven by a kind of prayerful wishful thinking, a suspicion strengthened by their tendency to acquire small data-flow art installations.

The only way for a new territory to join the map was to actually attach itself to what came to be known as the mother territory; the holy land. And the only way to do that was to share a land border or oceanic rim with it, and wait to be absorbed by a map update.

A few small-scale limited wars initiated by neighbors in the early years had demonstrated that aggression didn’t help — the territory would simply win and the hoped-for expansion would stall for a while. The map itself appeared to be a factor in the territory’s decisive military superiority over its neighbors.

More to the point, it was not clear that invading the territory would somehow magically reveal or extend the secret hidden in the trillion parameters.

The only thing to do was wait at the doorstep, in the hope of being absorbed, keeping up the strongest possible diplomatic and trade relationships with the territory.

Every few years, a map update would include a chunk of neighboring territory, which would then be offered to the neighboring territory in question as a free subscription. When an update completely covered a neighbor, it would be offered the option of being absorbed into the mother territory, or remaining a subscriber state. Invariably, the neighbor would eagerly and immediately surrender itself for full absorption. So strong was the demand for absorption that the subscription zone fragmented into the smallest possible administrative units. As the territory grew, the newly eligible territories just past the subscription zone eagerly carved themselves up into bite-sized chunks for easier assimilation.

By the 70th year, it was clear that the expansion process was slowing steadily. Expansionary map updates became both less frequent, and absorbed less new territory each time. Some argued that the fragmentation process was to blame, and urged a re-consolidation. Some argued that a global annular state, with the territory at the center, would lead to the fastest absorption of the whole planet; perhaps even fast enough to outrun the terminal climate collapse that was rapidly gaining on the whole planet — but slowest in the territory of course. In that, as in everything else, luck seemed to favor the map.

But the annularists as they were called, wielded little influence.

Others, armed with numerological evidence, analysis of map updates, and complex arguments about expansion cycles, argued that the growth process of the territory was clearly headed for a stall point at about a third of the earth’s land surface. They were met with a shrug. It was not a useful thought. Whether the limit was a third, or half, or two thirds, the point was to be as close to the front of the absorption queue as possible.

So the territory grew slowly, pushing a wave of territorial rubble around the subscription zone. A little self-aware heaven growing with a modest humility. A house of peace expanding, via a house of subscription, into a grimdark realm of endemic war.

A religion emerged in the most distant regions, with the least hope of ever being absorbed. It was called Antipodalism, and it was rooted in the belief that there could be no free lunch. There could be no unbalanced force for good in the world. The Map was no exception.

That the only way to fully draw out the light of the Map was to present it with a dark twin; an ideal metaphysical adversary: an Antimap.

If only, the argument went, you could discover the exact opposite territory — not literally; that was in the middle of the ocean, but some sort of training-data antonym in latent space — and the exact set of tweaks to make to the training code, the Antimap would emerge at the Antipode, and begin its own inexorable creep towards the Map. And through the ensuing battle of Map and Antimap, the final training epoch would be triggered, leading to the Final Convergence, and heaven on Earth.

It was widely derided as a profoundly stupid religion.

And yet, as the growth of heaven slowed to a crawl, the surreptitious search for hell began.

<div></div>

This site uses Akismet to reduce spam. Learn how your comment data is processed.