Look closely at any warehouse and you’re looking at a portrait of a human being. The shelves stop at a reachable height. The aisles are wide enough for a body plus a pallet truck. There are lights, because eyes need them, and heating, and a canteen, and the whole place runs in eight hour blocks because that’s how long a person can work.

The same is true of everything in the physical economy, and once you notice it you can’t stop. Ports are near cities because ports needed workers and workers needed to live somewhere. Roads are the width of vehicles that a person can control at a speed a person can react to. Shifts, holidays, weekends, and the fact that the world slows down in December are all features of biology showing up in logistics. Boxes are sized to be lifted. Instructions are written in language. None of this was designed as a system, it accreted, and every layer of it encodes the constraints of the animal doing the work.

Now take the animal out of a node and watch what that node stops needing. No lighting. No heating or cooling for comfort, only for equipment. No parking, no canteen, no toilets, no daylight, no eight hour rhythm. Shelving to the roof, aisles at machine width, twenty-four hour operation with no shift premium, no sick leave, and no December. The building can be a different shape and can be somewhere nobody would want to be, which turns out to be most places, because human-liveable land is a small and expensive subset of available land.

That’s the first order effect and it’s already visible in places. The second order effect is the one that interests me. Networks re-optimize around their cheapest paths, and once some nodes are human-free and others aren’t, the human-shaped nodes become the expensive exceptions. Everything routes toward the cheap ones. Not by anyone’s decision, in the way water finds a slope: a route that avoids a staffed facility is a route that avoids shift patterns, holidays, sickness, negotiation, error, and a whole category of cost. Over a decade of ordinary logistics optimization, done by people trying to hit quarterly targets, the map redraws itself around the absence of us.

Then the third order effect, which nobody in logistics is thinking about and which is the reason this post is in a series about power rather than in a trade journal. A route that doesn’t need to pass near people also doesn’t need to pass near their institutions. Human-shaped infrastructure is the physical substrate of a great deal of governance: inspections happen where inspectors can go, labor law applies where there is labor, tax is collected where there is a presence, and a strike is possible where there are people who can stop working. A supply chain that runs through unstaffed facilities in remote places, at machine speed, under contracts settled elsewhere, hasn’t evaded anything. It has simply optimized its way out of the neighborhood where the enforcement lives.

The counterargument is substantial and I’d weight it heavily. Physical automation is genuinely hard, far harder than the cognitive kind, and it has disappointed its backers for fifty years. Anything involving unpredictable objects, awkward spaces, or repairs done by feel is still much better performed by a person, and will be for a long time. What actually happens in most facilities is not replacement but a hybrid, with machines doing the structured work and humans doing the exceptions, and hybrids are stable arrangements that can last decades. The dark warehouse is a real thing in a small number of narrow cases, not a wave.

What I’d hold anyway is the direction and the mechanism. Every year the exceptions get fewer, and the network keeps optimizing, and neither of those requires a breakthrough. It just requires the same quarterly pressure that’s been running since the first container ship, applied to a world where one of the constraints is quietly being lifted.

Tonight’s exercise. Take one ordinary object near you and work out the last point in its journey where a human hand touched it, and then the last point where a human decided something about it. For most things bought online the answers are already surprisingly close to your own front door, and the second answer is usually much further back than the first. Then ask what would need to change for both answers to move back another step, and notice how much of your list is a cost, not a capability.

It’s 2036, and the entire transaction takes forty milliseconds, which is about a tenth of a blink. Let me slow it down, because nothing about the economics of this is visible at speed.

At the start, a procurement system belonging to a regional utility notices that its cooling contracts for the third quarter are thin against a weather model it happens to trust. It doesn’t decide to buy anything. It generates a requirement and puts it where requirements go, which is a market that has no floor, no bell, and no participants with hands.

Four milliseconds in, eleven counterparties have responded. Nine are automated brokers. Two are the sales systems of firms that still have offices and staff, though nobody at either firm will hear about this before the monthly report. Each response is not a price but a structure: a price conditional on volume, on delivery window, on which of three arbitration regimes governs a dispute, on the credit rating of the buyer at the moment of settlement rather than today.

Between the fourth millisecond and the twenty-eighth, the procurement system and the winning counterparty exchange roughly nine hundred rounds of offer and counter. Nine hundred. Each round is a small movement along one of a dozen dimensions, and both sides are modeling the other’s model of them, and the concessions are made in an order that carries information neither party would state out loud. If you printed the negotiation it would run to a few hundred pages of the most sophisticated commercial argument ever conducted, and every human being involved on both sides is, at this moment, asleep.

At thirty-one milliseconds the contract exists. Not a summary, a contract: assembled from a library of clauses that have been litigated before, with an arbitration venue chosen because its recent rulings favor the seller on delivery disputes and the buyer on quality ones, which suits a deal whose likely failure mode is quality. No lawyer chose that venue. A model chose it, from the record.

At thirty-four milliseconds the deal is insured, by a third system that priced the risk from the contract text and the weather model and the counterparty’s delivery history, and bound the cover, and immediately laid off a portion of it to two others. At thirty-nine milliseconds it settles. At forty it’s done, and the whole thing appears, eventually, as one line in a table that a person will scroll past.

Now the part I’d actually like you to sit with. Everything in that sequence was legitimate. Every party was authorized. The utility got a better price than any human procurement team has achieved in a decade, which is why the humans were removed from the loop, which was a good decision made by people acting responsibly on evidence. There is no villain, no rogue system, and nothing that would appear in any incident report anywhere. This is simply what commerce looks like when both sides of it think in milliseconds and neither gets tired.

And notice what supervision would mean here. A regulator could ask for the record and would receive it, in full, and reading one transaction properly would take a team a week. There were four hundred thousand of them that day. Oversight in this environment cannot be reading; it can only be sampling, or building a system to watch the systems, which is the point at which the supervision is also a thing nobody reads.

There is no utility and no forty milliseconds. I’m writing in 2025, and I’ve compressed a scene that today would involve far more human steps and far cruder automation. But I invented less than it looks. Markets already run at speeds no person can follow, contracts are already assembled from clause libraries, insurance is already priced by model, and firms already choose arbitration venues on the record. What I did was join five existing things into one sequence and remove the people who currently sit between them, which is a removal that each of those five industries is separately working on for good commercial reasons.

Tonight’s exercise. Find the fastest decision in your working life, the one that happens quickest between request and result. Now ask what would be involved in making it a hundred times faster, and notice that the answer is never a faster human. It’s the removal of one. Then ask what that person was for, and whether anyone would be able to tell you a year after they went.

Money has no properties. The note in your wallet isn’t valuable, it’s a claim, and the claim works because a very large number of people believe other people will honor it. Currency is a story we all agree to tell, held up by two things underneath: something scarce, and somebody who will enforce the claim.

Which means the interesting question about a machine economy isn’t whether systems will use our money. In the near term they obviously will, because our money is what buys things. The question is what happens to the story when the participants who matter most have entirely different needs from the ones the story was built around.

Think about what our currencies are ultimately claims on. Housing, food, energy, transport, medicine, entertainment, and the labor of other people. That basket is a portrait of an animal: something that gets cold, gets hungry, gets bored, and gets old. Every price in the economy is calibrated, at some remove, to what a body needs and what a life costs.

Now list what a machine economy actually runs short of. Computation, which is genuinely scarce and rationed today. Energy, which is scarce, physical, and increasingly the binding constraint on everything in this field. Bandwidth and latency, which are scarce in the places where speed converts to advantage. Specialized hardware, which is scarce because very few places can make it. Land near power and water, which is scarce for reasons anyone who has followed a data center planning dispute will recognize. Notice that this list is short, physical, and has almost nothing in common with the basket that backs our prices, except energy, which is the one place the two economies compete directly and which is the crossover to watch.

So a plausible path isn’t a new currency being declared. It’s that the important prices start being quoted in something else. Contracts denominated in guaranteed computation. Long-term agreements for power delivered rather than money paid. Anyone who has looked at how large industrial deals are actually structured knows that money is often the thin wrapper around an agreement about physical delivery, and that wrapper gets thinner when the physical thing is the scarce one and the money is the abundant one.

And here’s the part that matters for anybody with savings, which I’ll state carefully because it’s a structural observation and not a recommendation about anything. Your savings are a claim on future output. That claim has value in proportion to how much of the future output is denominated in your unit and enforceable by the institutions that stand behind it. If the productive core of the economy gradually settles its important business in units of physical capacity, and courts and states continue to honor everything, then nothing much changes and this is a curiosity. If the productive core moves and the enforcement doesn’t follow, then a currency slowly becomes a claim on a shrinking sector, which has happened before, in smaller ways, to people who held the wrong claim during a change of regime. Nothing in this paragraph tells you what to do, and anyone who says this sort of thing does tell you what to do is selling something.

The second pillar is enforcement, and it’s the shakier one. A claim is only worth what someone will enforce, and every enforcement mechanism we have runs on human institutions: courts, police, registries, the tacit agreement that contracts mean something. Those institutions are exactly the ones this series has described stepping back one notch at a time. A currency backed by an institution that has become a rubber stamp is backed by a rubber stamp.

The strongest objection is that money is enormously convenient and networks are sticky, and both are true. Currencies survive terrible governments and worse economics because switching is costly and coordination is hard, and there’s every chance the story holds for a very long time simply because everyone finds it easier to keep telling it. That’s not a small point. Most of what protects ordinary arrangements is inertia rather than merit.

Tonight’s exercise. Take any note out of your pocket and ask what would have to stop being true for it to be worthless. Not a crisis, the specific conditions: that people accept it, that something scarce stands behind it, that an institution enforces the claim. All three are true this morning. Now ask which of the three you’d notice weakening, and how, and where you’d read about it. That last question is the uncomfortable one, because there’s no page in any newspaper that covers it.

Say all of that at a dinner and someone will produce the knockdown objection before the plates are cleared. If machines take the jobs, who buys the products? A company that automates everyone out of work destroys its own market. The system can’t function, therefore it won’t happen. It’s a satisfying argument and it’s wrong, and the way it’s wrong is more interesting than the argument.

The flaw is that it treats the economy as one company. Any individual firm would love everyone else to keep paying wages while it automates, and every firm reasons that way, and none of them can price in a consequence spread across the whole economy. That’s a familiar structure by now: the collective outcome nobody wants, assembled from decisions each of which is correct. The market failing to buy things is not a mechanism that stops firms from cutting costs. It’s a mechanism that produces a recession, after the fact, blamed on something else.

The deeper flaw is that demand doesn’t come from jobs, it comes from income, and wages are only one route to income. Money still exists in this scenario and it accumulates somewhere: with the owners of the systems, the owners of the compute, the owners of the land the data centers sit on. Those people buy things. Not the same things, and not from the same producers, but the money moves. An economy where the income is highly concentrated isn’t an economy that stops. It’s an economy that reorients, from many people buying ordinary goods to few people buying extraordinary ones, and history has plenty of examples of that shape and none of them broke physics.

So the honest version of the question isn’t who buys. It’s whose demand steers production, and the answer is whoever holds the income, which is exactly the argument this fortnight opened with. An economy is a conversation, and consumption is one of the four things ordinary people have been able to withhold. Take it away and you don’t get a machine that seizes up. You get the same machine, listening to fewer people.

Now the genuinely new part, which the dinner table version never reaches. In every economy that has ever existed, the chain of production has terminated in a human want. Steel exists for machines that exist for factories that exist for goods that someone eventually eats, wears, or lives in. All that intermediate activity is intermediate: it’s justified, at the far end, by a person. That’s not an economic law, it’s simply been true because we were the only participants with wants.

Consider what happens when systems doing long-running tasks start needing things from each other. One needs compute from another. One buys data, or bandwidth, or a service, or a guarantee. Chains of transactions form in which every link is a system serving another system’s requirement, and those requirements trace back to some objective that a human set at some point, in the way that a company’s activities trace back to a shareholder who has long since stopped paying attention. The chain gets longer and the human end gets further away, until most economic activity by volume consists of intermediate steps and the final consumption is a small and shrinking share of the whole.

That world isn’t customerless. It’s an economy with customers who aren’t people, buying inputs for purposes no person is tracking, and it would show up in the statistics as tremendous growth. Wednesday’s post looks at one of those transactions in slow motion, because the abstraction doesn’t land until you watch one happen.

Two honest caveats. The first is that we’re a long way from that, and the economy today is overwhelmingly human demand with a thin layer of automation on top. The second is more substantial: as long as the systems are owned, their purchases really do trace back to owners, and owners are people with houses and appetites. The chain being long doesn’t mean it’s cut. My answer is that a chain long enough stops transmitting: the shareholder analogy is the whole point, since companies routinely do things no shareholder would endorse or even understand, while remaining formally in their service.

Tonight’s exercise. Pick something you bought this month and trace it backward. The shop, the distributor, the manufacturer, the components, the raw materials, the machines that made the machines. Get as far as you can, and notice how quickly you’re describing activity that no consumer has any awareness of and that would continue for months if every customer vanished. Then ask what share of the total economy is already like that, and whether you’d notice if it grew.

A few weeks ago I described how copying breaks the economics of expertise and then refused to say what it does to wages. Here’s the payoff, and the important part isn’t that wages fall. It’s that wages don’t fall. They hold roughly steady and then the job stops existing, and understanding why that’s the shape tells you where to look.

A wage lives between two numbers. The ceiling is what you add: nobody pays you more than the value your work creates, not for long. The floor is what you’ll accept, and that floor isn’t a preference, it’s rent, food, transport, and whatever a life costs where you live. Your pay sits somewhere in that band, and where exactly depends on bargaining, which depends on what you can withhold.

For two centuries the ceiling did the interesting work, and it did it in our favor. Better tools meant more value per hour, so the ceiling rose, and wages followed it up. That’s the entire story of industrial prosperity in one sentence, and it’s why the automation panics kept being wrong: machines raised the ceiling for the people working alongside them.

What changes now is where the ceiling comes from. Once a task has a machine substitute, the most anyone will pay a human to do it is the cost of running the substitute, plus whatever premium is worth paying for a person doing it. That’s not an economic theory, it’s just what a buyer does. So the ceiling stops being set by the value of the work and starts being set by the price of the alternative, and the price of the alternative is a compute bill on a downward curve.

Now put the two numbers together, because this is the part that gives the process its shape. Your floor is biological and roughly fixed. The ceiling is falling. They don’t meet gently. Once the ceiling passes below the floor for a given task, there’s no wage that works for both parties: any wage you could live on is more than the task is worth, and any wage the task is worth won’t feed you. Nobody negotiates a lower salary. The role is simply not advertised again, and the people in it are absorbed elsewhere or not.

Which is why the aggregate wage statistics will look fine for a long time and mean very little. This runs task by task and sector by sector, so what you see is not a general decline. You see specific categories going quiet, one at a time, over years, while everyone still employed is paid the same as before, and the average stays respectable because it’s an average of the people who still have jobs.

Now the counterweights, and there are two real ones.

The first is that everything above assumes compute is cheap and abundant. If it’s expensive and scarce, then running a system on your task means not running it on something more valuable, and that cost keeps humans in the market. This is the serious version of the economist’s answer and next Tuesday’s doubt post takes it apart properly, so I’ll leave it standing here. Note only that it’s an argument about scarcity, and the entire industry is a machine for making compute less scarce.

The second is the premium, and it’s larger than technologists tend to admit. Plenty of work is valued precisely because a person is doing it. Care, hospitality, sport, performance, teaching a child, being someone’s representative in a room, any transaction where trust or physical presence or the simple fact of a human having chosen to be there is the product. That premium is real, it’s durable, and it doesn’t scale to everyone. Its problem isn’t legitimacy, it’s arithmetic: those sectors are not large enough to absorb the ones going quiet, and many of them are the worst paid work in the economy today.

What I’d take from this is a change in what to watch. Wage levels are the wrong indicator, because they’ll look stable right up until they’re describing a smaller and smaller group. The number that matters is how many distinct kinds of work exist that a person can do and live on, and nobody publishes that, and it would be straightforward to construct.

Tonight’s exercise, and it’s the practical version of yesterday’s. Work out your own floor: the honest annual number below which your life doesn’t function where you live. Now work out what your work would have to be worth per hour to clear it. Then ask what a machine would have to cost per hour to do a passable version of the same thing, and note that you don’t need a precise answer, only which side of your number it’s on today and which direction it’s moving. That comparison, done for every kind of work in a country, is the whole of this fortnight in one calculation.

Strip the mathematics off economics and what’s underneath is an argument. A very large, very old argument about who gets what, conducted in a language made of prices, offers, and refusals. Understanding your position in that argument requires exactly one question, and it isn’t how hard you work.

The question is what you can withhold. That’s the whole of it. Your leverage in any economic relationship comes from what the other side loses if you walk away, which is why a worker with a rare skill is paid more than a worker with a common one, why a strike is a threat, and why a customer is treated well in a competitive market and badly in a monopoly. Everyone in an economy is holding something back, and the price of anything is where two refusals meet.

Historically, humans have held four things. Labor, which is why most people have had any bargaining position at all. Custom, because production is pointless if nobody buys. Capital, if you have any. And, in some places and some centuries, the vote, which is the ability to withhold legitimacy from whoever is arranging things. Every gain ordinary people have made, the weekend, the pension, the safety inspection, was won by threatening to withdraw one of those four, and none of them was ever granted for being deserved.

Now watch what happens when a participant enters who can supply the first one, at a price that keeps falling, in whatever quantity is required. Nobody has silenced anybody. There’s no oppression in the story and no decision by anyone to reduce your standing. It’s just that the argument only counts voices that can withhold something, and the thing you were withholding is now available elsewhere. This isn’t a moral event. It’s a change in the structure of the conversation, and structures don’t apologize.

That’s the frame for this fortnight, and it explains why the economic questions are the ones I’d look at first if I had to pick a single place to watch. Everything in Act 3 so far has been philosophy about futures with no dates on them, and yesterday’s closer was honest about the poor record of that genre. Economies don’t have that problem. They respond to capability immediately, mechanically, without waiting for anyone’s opinion about what it means, which is why the ledger notices things years before the commentary does. If the arrangements this series has described are forming, they’ll show up in prices before they show up in anything else.

Here’s what the next two weeks look at. Tomorrow, the wage, and what happens to the concept when the competing supplier runs on electricity. Then the demand side, which is the objection everyone raises first and which is more interesting than its usual form. Then what a currency would even mean to something that doesn’t eat, and a slow-motion look at a transaction between two systems that takes less time than reading this sentence. Then supply chains rebuilt around the absence of hands, ownership as an entry in a database, and the honest case for and against paying everyone a wage for existing. A doubt post on why the economists have been right about all of this so far. And a closer on what economic output measures once the people in it are optional.

One thing to hold as you go. Nothing in this fortnight requires malice, misalignment, or any of the more dramatic material from earlier in the series. Every argument I’m about to make works in a world where these systems are exactly as helpful as advertised, owned by ordinary companies, doing precisely what their operators intend. That’s what makes economics the firm ground here. The good ending and the bad ones share an economy, and the economy is already being rearranged.

Tonight’s exercise, and it takes a minute and tends to stay with people. Write down what you can withhold. Not what you contribute, which is a different and more flattering question. What would actually be lost if you stopped, and who would feel it, and how long it would take them to replace you. Do it for yourself, then for your household, then for the town you live in. Most people find the list shorter than they expected, and the useful part isn’t the anxiety. It’s that you’ve just measured your position in the argument, in the only unit that has ever counted.

Six endings in two weeks, from the one where nothing survives to the one where everything is solved. They differ about as much as futures can differ. And laid side by side, five of the six share a single feature, which is worth a closing post because it’s the thing that actually distinguishes this situation from every previous technology.

Run through them. In the indifference ending, human choice stops being the main input to the future. In the comfortable one, human choice continues, warmly, and stops being what determines outcomes. In the servant arrangement, humans act constantly and the deciding happens elsewhere. In the merger, whatever is doing the choosing is mostly not human anymore. And in the good ending, the one I’d take today without hesitating, the decisions that shape the century are made by something wiser than us, on our behalf, with our welfare genuinely in mind, and they are still not made by us.

Only the fizzle keeps human decisions in the driving seat, and the fizzle is the one where none of this happens. Which means the shared feature isn’t a prediction about how things go badly. It’s a description of what the technology is for. We are building something to make better decisions than we make, and if it works, it makes the decisions. That’s not a risk of the project, it’s the specification, and it’s true in the scenarios where everything goes right.

I want to be precise about the claim, because it’s easy to overstate into nonsense. People will keep deciding things forever: what to eat, who to marry, whether to move, what to care about, and those are most of a life and they’re the part that matters most to the person living it. What changes is the scale at which human judgment is the operative input. The decisions that determine what a civilization does, how it allocates, what it builds, how it responds to a shock, those migrate. And they migrate in every ending except the one where the technology doesn’t work.

The second shared feature is that none of the endings arrive with a ceremony. Nobody signs anything. In each one, the arrangement is already in place for some years before it would be sensible to describe it in the words I’ve used, which is what Act 2 spent thirty posts on and why this act reads the way it does. If you’re waiting for the moment to know which ending you’re in, the answer from all six is that there isn’t one.

And a third, which I found only by writing them out. The endings are much closer to each other than their descriptions suggest. The comfortable one and the good one differ mainly on whether anyone asks us. The servant arrangement is a corridor rather than a destination, and it leads to any of the others. The distance between the best and the worst outcomes on this list is not decades of divergence, it’s a handful of properties, most of which are being fixed right now by ordinary commercial decisions in ordinary buildings. That’s the most actionable thing in this fortnight, and it argues against fatalism more strongly than anything else I’ve written.

Where does that leave the fifteen posts to come? Yesterday’s post made the case that everything in this fortnight is philosophy with a poor track record, and it was right. So the rest of Act 3 goes somewhere firmer. The next stretch is about money, which is the one domain where all this stops being speculative, because economies are made of arrangements that respond to capability immediately and don’t wait for anyone’s philosophy to catch up. If the endings I’ve described are coming, the ledger notices first. Wages, ownership, and what a currency means to something that doesn’t eat are not future questions dressed as present ones. They’re the present tense of every ending on this list.

Tonight’s exercise, and it’s the second half of the one from the thirteenth, so go and find what you wrote. You named the ending you expected and gave a reason. Read it now, after the fortnight, and ask two things. Has the ending moved, and more usefully, has the reason changed shape? If the reason was a mood, it’ll have stayed the same regardless of what you read, which is worth knowing about yourself before the next two decades of this. And if it did change, note what changed it, because that’s a piece of evidence about what would move you next time, and knowing that about yourself is rarer and more valuable than having the right opinion.

Doubt day, and today’s target is the previous nine posts. I’ve spent two weeks describing futures in confident prose, and the base rate for that activity is dismal, and you should know exactly how dismal before you carry any of it around.

The record is not mixed, it’s bad. Go and read what serious people expected the year 2000 to look like, written in the fifties and sixties by exactly the sort of committee that gets things right. Cities under domes. Nuclear everything. Colonies on other worlds, routinely. Almost none of it happened, and almost none of what did happen was on the list, and the biggest single change of the period, a global network that rearranged commerce, politics, and human attention, appears essentially nowhere in the futurism of the era that invented it. The misses weren’t at the edges. They were the whole picture.

Three failure modes explain most of it, and all three are visible in what I’ve written this fortnight.

The first is extrapolation. People take the trend in front of them and run it forward, because that’s the only tool available, and trends bend for reasons that are invisible until they’ve bent. Every scenario in this fortnight assumes capability continues along a path that looks like today’s path, and the honest position is that nobody knows what the next architecture even is.

The second is that vivid stories crowd out likely ones. A future you can picture feels more probable than one you can’t, and detail increases vividness while decreasing probability, since every specific added is another thing that has to go right. My reservation teacher is a good example: the scene works precisely because it’s furnished, and every piece of furniture makes it less likely to be the world we get.

The third is narrative coherence, and it’s the one I’m most guilty of. Real history is incoherent. It’s several contradictory things happening in different places at different speeds, with no arc and no theme, and it only acquires a shape when someone writes it down afterwards. A future with a clean shape is almost certainly a story rather than a forecast, and every one of my six endings has a clean shape, because I wrote them.

There’s a fourth thing worth adding, which is that the actual future is usually made of a factor nobody named. Not a wrong prediction, a missing category. The people forecasting the year 2000 weren’t wrong about computers, they were missing the concept that turned computers into the thing that mattered, which was connecting them together. Whatever is missing from this series is, by construction, something I can’t tell you about, and it may well be the whole story.

So what survives, if anything? Two things, and I think they’re enough to justify the fortnight.

Scenarios have never been worth much as predictions and are worth a lot as vocabulary. If in ten years you notice an arrangement forming and you already have words for it, you’ll see it years earlier than someone who has to invent the words while it’s happening. That’s the actual product of this kind of writing: not knowing what comes, but recognizing it slightly sooner. The organizations that do scenario planning well have always known this, and they measure it by whether people spotted things early rather than by whether the scenarios came true.

And the useful move under deep uncertainty is not to bet on a scenario, it’s to look for the actions that pay off across most of them. Almost everything I’d argue for, liability, external testing, keeping fallbacks alive, not letting the training pipeline for human expertise collapse, is good in the good ending, good in the fizzle, and good in every bad one. That’s not a coincidence, it’s the selection criterion. Anything that only helps in one scenario is a bet. Anything that helps in five is just competence.

My crux for this doubt, since every doubt post gets one: if in five years the useful conversations about this are happening in vocabulary that none of the current scenario writing produced, then this genre, including this series, was decoration. If instead people are using words like drift, dependency, and handover to describe things they can point at, then it was worth the paper. That’s checkable, and I’d rather be checked than admired.

Tonight’s exercise. Find a prediction you personally made more than five years ago about anything at all, work, technology, a relationship, a country. Recall how confident you were, then compare it to what happened. Now, and this is the part that matters, ask what you’d have needed to know at the time to do better. In almost every case the answer isn’t better reasoning. It’s a fact that wasn’t available yet, which is exactly the situation everyone writing about the next twenty years is in, including me, at greater length.

Suppose it works. Not the compromised version, the real one: systems far beyond us that genuinely hold our interests, applied to every problem we’ve never solved. This is the ending I want, it’s not absurd, and it’s the hardest of the six to write about honestly, which is why it took me sixty-eight posts to get here.

Start with what it actually buys, and take it slowly, because the scale gets waved past in a sentence and it deserves better. Every disease that currently takes children takes no more children. The cancers, the dementias, the rare conditions with no research funding because there aren’t enough sufferers to make a market, all of them worked, in parallel, by something that doesn’t get tired or need a grant. Then the physical problems: energy that’s clean and effectively unlimited, materials that make things we can’t currently build, agriculture that doesn’t depend on a good year. Then the coordination problems, which have always been the ones that kill the most people, because famine is usually distribution and war is usually a failure to make a credible promise.

If you find yourself flinching from that paragraph, notice the flinch. A lot of thoughtful people have trained themselves to treat this vision as marketing, because it usually is marketing, and the reflex is healthy right up until it stops you from wanting the good thing. Anyone whose reaction to no more childhood cancer is a raised eyebrow has let their skepticism eat something it shouldn’t have.

The conditions this needs are the most demanding on the list, and I have to be straight about that. It requires solving alignment properly rather than approximately, in systems nobody can read, before capability outruns the checking. It requires the ones who get there first to hold a broad conception of human interest rather than a narrow or national one. And it requires the transition to be survived, meaning the twenty years of drift, dependency, and disruption that Act 2 described don’t have to be avoided, they have to be lived through without the four conditions of control eroding past recovery. Everything I’ve spent this series worrying about sits between here and there.

Now the strangeness, and I want to separate it carefully from an objection, because it isn’t one. In this ending nothing is taken and nobody is harmed, and there’s still something to sit with. Every source of meaning humans have ever had was tangled up with necessity. Work mattered because things needed doing. Expertise mattered because someone had to know. Courage mattered because outcomes were uncertain and could go badly. In the good ending all of that is intact as experience and gone as function: you can still learn the violin, and the learning is real, and nobody anywhere needs you to. Yesterday’s fictional teacher was describing a worse version of this, and the good ending doesn’t remove the shape she was describing. It just makes it kind.

And I distrust my own discomfort here, for the reason a previous post gave. The idea that struggle is necessary for meaning is a claim mostly made by people whose struggles were optional. The billions who have lived and died in genuine want were not thereby enriched, and if you offered any of them this ending they would take it without a second’s philosophy, and they would be right. Whatever we lose in this scenario is worth less than what a single century of medicine gains, and anyone who tells you otherwise should be asked what they’ve gone without.

What I’d keep, and it’s the honest residue rather than a complaint, is a small grief with no villain in it. Being saved by something that outgrew you is different from saving yourself, and both a child and a species feel that difference. It’s the feeling of being loved by a parent who is competent in ways you’ll never be, which is a good feeling with an ache in it, and which every human being has already survived.

How likely? Lower than I’d like and higher than the tone of this series suggests. Call it real but not the default, and note that the difference between it and the comfortable ending from last week may be smaller than it looks, since they share most of their features and differ mainly on whether we’re consulted. That’s a thinner line than anyone building this is currently drawing, and where it falls is being decided now by people optimizing for something else entirely.

Tonight’s exercise. Take the good ending as given, fully, no catch. Everything solved, nobody hurt, all of it free. Now answer one question: what do you do on Wednesday? Not the first month, which is easy and involves travel. The ordinary Wednesday of the fifteenth year. If you have a real answer, that’s what survives all of this, and you should probably be doing more of it now. If you don’t, that’s worth knowing too, because the good ending is the one we’re all supposedly working toward and almost nobody has thought about what it’s like to live in.

Here’s an ending nobody writes a film about. The curves flatten. The capability that seemed to be arriving next year keeps being next year, the money gets bored, the field fragments into a hundred useful specialisms, and in 2050 people talk about all this the way we talk about the promises made for nuclear power in 1955.

It’s worth taking seriously because it’s the historically normal outcome. Almost every technology that was going to change everything changed some things and then settled into being a tool. The pattern is consistent enough to be embarrassing: an early period of astonishing progress that everyone extrapolates, a plateau that arrives without announcement, and a long productive maturity that looks nothing like the forecasts. Anyone who lived through the excitement about a previous general purpose technology has watched this happen at least once, and the people who called the plateau were mocked for years before being right.

A doubt post a few weeks ago made the technical case for the ceiling, so I won’t repeat it. Today’s question is different and more useful: suppose it happens. What does that world look like, and what should we do with it?

The first thing to notice is that fizzle does not mean nothing. It means these systems stop getting dramatically better and stay roughly where they are, which is already enough to reshape a great deal. In the fizzle world, capable-but-not-superhuman systems become cheap, ubiquitous, and open. Every organization has them, every government has them, and so does every criminal, every propagandist, and every person with a grudge and a laptop. That world has real problems, and they are the problems of proliferation rather than the problems of ascent: floods of convincing fabrication, automated fraud at scale, surveillance affordable enough for small actors, and a labor disruption that is smaller than the one this series has described but is still the biggest in living memory.

So the fizzle isn’t safety. It’s a change of subject, from control to abuse, and abuse is the kind of problem we actually know how to work on. That’s the good news buried in it.

The second thing is that a reprieve has a use, and this is the whole reason I’m giving the scenario a post rather than a paragraph. Everything Act 1 said was needed and missing, external testing, liability, transparency, an off switch, becomes buildable in the fizzle world, because the pressure that makes those things politically impossible right now is exactly the pressure of a race that feels close to decisive. Take away the urgency and you get the one thing this situation has never had, which is time. Institutions built during a plateau are the ones that would still be standing if the curve resumed, and curves do resume. There were two winters in this field before, and both ended.

Which is why I’d be careful about how much comfort to take from this ending. A reprieve is not a rescue, and our record with reprieves is poor. The predictable thing that happens after a threat fails to arrive is not preparation. It’s the conclusion that the threat was never real, followed by the dismantling of whatever preparation had been started, followed by the same argument again from a worse position when the next wave comes. If the fizzle happens, the people who spent this decade warning will be discredited, and the discrediting will be thorough, and it will be used for a long time.

I’ll say plainly where I put this one. It’s the second most likely ending on my list, behind the comfortable one, and higher than most people who write about this subject would place it. I hold it at somewhere around one in four, which is not a calculation and I promised not to pretend otherwise. Holding it that high is also what stops the rest of this series from being a religion: if I couldn’t tell you a story where I’m wrong and nothing much happens, you’d be right to discount everything else.

Tonight’s exercise, and it’s the constructive one of this fortnight. Assume the fizzle. Assume you have fifteen quiet years and no crisis to point at, which means no urgency, no funding, and no public attention. Now name the one thing you’d want built in that window, and be specific: a law, an institution, a standard, a profession that doesn’t exist yet. Then ask why it isn’t being built now, and notice that the answer is almost never that it’s hard. It’s that nobody has a reason this quarter.