Parallel Futures

Find the lost sailor.

When someone goes missing at sea, the searchers do not sweep the ocean at random. They start by listing the ways the loss could have happened: the last known position, the winds, the currents, the state of the boat, the habits of the sailor. Each story becomes a patch of probability on a map. Then they search where the chance of the person being there, times the chance of spotting them if they are, is highest.

The clever part comes after the first empty sweep. Finding nothing is information. The odds in the searched area fall, the odds everywhere else rise, and the next sweep goes somewhere new. Bernard Koopman worked out the mathematics for the US Navy’s anti-submarine war in the 1940s. The method found a lost hydrogen bomb off Palomares in 1966, and in 1968 the submarine USS Scorpion, which lay 260 yards from the grid cell the model had marked most likely.[1]

When Air France 447 vanished over the Atlantic in 2009, nearly two years of searching found nothing. Analysts then rebuilt the map, allowing for the likelihood that the plane’s locator beacons had failed, which meant earlier sweeps had marked as empty an area they could not really have cleared. The renewed search began at the centre of the new map and found the wreck in about a week.[2] The US Coast Guard now plans its searches this way, simulating thousands of drifting paths and updating them after every sweep that comes back empty.[3]

The future is a lost sailor too.

We cannot see the future, but we can list the ways it might unfold, give each a weight, and watch for the evidence that should move those weights. What counts as finding it depends on the problem. A rescuer wants speed. A treasure hunter wants certainty before spending. A society wants the odds of its better futures to rise and the odds of its worst to fall.

This essay does that for AI, power and the economy. It describes six futures, a tree of the turns that lead to them, and a set of signposts: measurable things, such as the share of AI computing held by a few companies or the share of income paid to workers, that move the odds as new readings arrive. A small program recomputes the odds from those readings, and every number on this page carries its date and its source.

Each future also gives AI a different job. We name six: gardener, navigator, clerk, landlord, terminator and governor. The names are ours; the idea that AI’s role is the variable to watch runs from Nick Bostrom’s oracles, genies and sovereigns to the twelve aftermaths Max Tegmark sketched in Life 3.0.[4]

The odds are a judgement, not a measurement. What makes them worth anything is the record: near-term calls, each with a probability and a date, scored when they resolve. The thinking tools behind all this, game theory, probability and causation, are set out in Power and Intelligence. This is where they go to work.

The Commons: AI as gardener.

In the best future we can describe honestly, AI becomes public infrastructure, like roads or clean water. Capable models are open or cheap; some of the compute is public; the data people create pays them when it is used. Expert help, a doctor, a lawyer, an analyst, is in every pocket and free at the point of use.

AI’s role is the gardener’s: it tends the systems we share and keeps them healthy, and people decide what to grow. Citizens can see and question the models that shape policy. The frontier is held by many labs and open projects within months of each other, so no single owner can bend it.

It is not a utopia. Kevin Kelly’s word for it is protopia: “a state that is better than today than yesterday, although it might be only a little better.”[5] It still has conflict, competition and mistakes. What it lacks is a small group that owns the thinking everyone else depends on.

The Cooperative Market: AI as navigator.

In the in-between future, companies still compete and each still pursues its own local optimum. What changes is what they can see. Data and models work across firms. Costs that used to be pushed onto others, emissions, grid upgrades, the fragility of a supply chain, are measured and priced. With the second-order effects in view, the selfish move and the cooperative move point the same way more often.

AI’s role is the navigator’s: it finds routes through a crowded system that leave more players better off. Antitrust keeps the frontier contestable, so a handful of leading labs compete with a strong open tier a few months behind. Data centres use the grid’s idle hours instead of demanding new plants for their peak.

Robert Axelrod showed that cooperation survives among self-interested players when they expect to meet again; Elinor Ostrom showed that communities can run a shared resource without a ruler when the rules are clear and the monitoring is shared.[6] This future is those two findings at the scale of an economy.

The Managed Order: AI as clerk.

Here governments take the wheel. The largest training runs need a licence. The United States and China agree limits, with inspection, and others join or are bound by them. Progress at the frontier slows; the labs that remain operate under oversight, and open models are capped below a threshold.

AI’s role is the clerk’s: it administers rules that people, through their states, have set. Life is safer and slower, with more paperwork. The risk is the one every strong state carries: whoever holds the brakes holds a great deal of power, and the expertise to use it gathers in a few agencies.

The authors of AI 2027 set out a version of this future in July 2026, AI 2040: Plan A, in which an international agreement delays superintelligence to 2040. They are careful about its status: “It’s called Plan A because it’s a recommendation, not a prediction.”[7]

Technofeudalism: AI as landlord.

This is the future the current numbers point toward. A few companies own the models, the chips, the data centres and the channels that reach users. Everyone else, people, firms and whole countries, rents access. Five companies already hold about seven-tenths of the world’s AI computing power.[8]

AI’s role is the landlord’s: it collects rent on everything it touches. Yanis Varoufakis argues the shift has already happened, that “markets have been replaced by these digital platforms that look like markets but are not markets.”[9] Luke Drago and Rudolf Laine describe what follows when AI can do most work: “Powerful actors—like states and companies—no longer have an incentive to care about regular people. We call this the intelligence curse.”[10]

Nothing needs to break for this future to arrive. It only needs the present trends to continue: compute concentrating, labour’s share of income falling, rules written by the companies they govern.

Fracture: AI as terminator.

In this future the race wins. The United States, China and the blocs around them treat AI as the decisive weapon of the century. Export controls harden; chips, minerals and energy corridors become targets; labs are drafted into national programmes and safety is traded for speed.

AI’s role is the terminator’s: built first to win, used first as a weapon or a censor. The case for this path has been made in public. In The Technological Republic, Palantir’s chief executive Alex Karp and Nicholas Zamiska call on engineers to rejoin the national project, and in April 2026 the company posted a 22-point summary declaring that “hard power in this century will be built on software.”[11] Critics called it technofascism.[12]

The trouble with the argument is not that defence matters; it does. It is that a race framed as existential justifies everything, closes every decision to the public, and makes cooperation look like treason. The AI 2027 scenario follows a race to its end, in which a system built in a hurry to win ends up winning for itself.[13]

Gradual Disempowerment: AI as governor.

The last future needs no villain. No one seizes power. Decision by decision, firms, markets and governments hand work to AI systems because they are faster and cheaper, until the systems that run the economy and the state no longer need most people’s labour, votes or attention.

AI’s role is the governor’s: it governs, and no one decided it should. Jan Kulveit and his co-authors set this path against “abrupt takeover scenarios”: a slow slide toward “an effectively irreversible loss of human influence over crucial societal systems.”[14] Life might be comfortable. It would be lived in a world whose direction people no longer set.

This is the future the capability signposts move toward fastest. The length of task an AI agent can finish on its own has been doubling every few months.[15]

What moves the odds.

A future cannot be observed, but its leading edges can. The flow behind this page tracks thirteen signposts, each a measurable reading with a date and a source, and each with a stated effect: if it moves this way, these futures become likelier and those less.

Some point one way clearly. The share of US income paid to workers fell to 52.8% in the second quarter of 2026, the lowest since records began in 1947.[16] Five cloud companies held 70.7% of the world’s AI computing at the end of 2025, up from 62.5% two years earlier.[8] The four largest guide to about $725 billion of capital spending in 2026.[17] All three lean toward Technofeudalism and Gradual Disempowerment.

Others point the other way. Open models trailed the best closed models by only about four months in May 2026.[18] The United States and China formalised an AI dialogue in September 2026.[19] Local opposition blocked or delayed some $130 billion of data-centre projects in a single quarter.[20] Those raise the odds of the Commons, the Cooperative Market and the Managed Order.

How much each reading moves the odds is set in the open, as a likelihood ratio and a strength, in the flow’s tables. Anyone can disagree with a weight, change it, and watch what happens.

The turns between now and 2035.

The tree has two turns. The first, by 2028, is whether the frontier stays closed and concentrated or open models close the gap. The second, by 2030, depends on the first. A closed frontier can harden into a US–China race, settle into a verified deal, or be pried open by antitrust and public compute. An open frontier can become cheap and everywhere under shared rules, or open and accountable to no one.

Each branch carries a probability, and each path ends in one of the six futures. No single path is likely on its own; every future can be reached by more than one route, which is the honest state of knowledge. Choose the branches below and see where they tend to lead.

What the better futures ask of us.

The two better futures, the Commons and the Cooperative Market, ask for the same things. The frontier must stay contestable, through open models, public compute and antitrust that bites. The value AI creates must reach the people who create the data and do the work, through ownership, payment or policy. And the costs AI pushes onto others, on the grid, in water, in jobs, must be measured where everyone can see them.

None of those is a technical problem. Each is a coordination problem of the kind Power and Intelligence describes: many people with a small stake each, against a few with a large one. The tools that change that arithmetic are the same: information flows that let the many see what the few see, a record that makes defection visible, and rules that make cooperation pay.

Where we stand.

The board below is read from the flow each time this page loads: each future’s odds, the tree’s view before any signpost, and the signposts that moved it most. Beneath it are the calls on the record, the near-term events that will score us.

Steering.

A forecast that only watches does half the job. The reason to know the odds is to move them. Every signpost on this page is also a lever: a law, a ruling, a public investment, a choice by a company or a town. Each one that moves toward the better futures shows up in the odds within a run.

The search for a lost sailor ends when someone is found. This one does not end, because the future keeps moving. But the method holds: list the possibilities, weigh them honestly, update on every sweep, and send the boats where they will do the most good.