I found this very funny! At the same time, I wonder if writing this sort of post may make one a little bit less open minded than optimal about AI risk.
I wish there were fewer satirical shitposts like this that don't self-label as AI pseudo-slop. Or if they exist, it would be nice if they said something original rather than 'have you thought about how AI risk is like X other thing?' without engaging in more than a surface level analogy.
Because actually self-copying machines ARE scary. Nanoscale fabricators with self-replication are possible (proof, bacteria exist).
If there were an equivalent to the METR graph for 'proportion of 3D printer value which a 3D printer can generate', and that was a straight line that implied we got to 99%+ in 4 years, I'd be pretty spooked!
"If there were an equivalent to the METR graph for 'proportion of 3D printer value which a 3D printer can generate', and that was a straight line that implied we got to 99%+ in 4 years, I'd be pretty spooked!"
If we were interested in "which proportion of value" AI systems are generating in the economy, a better measure of this than inferring from the METR graph would be AI revenue as a % of GDP. Which is about 0.1%.
This is a version of AI 2040 focused on a different risk—an industrial manufacturing explosion.
Foreword
Companies are racing to build a printer that can print itself. In Wealth Without Money (2004), the founder of the field argued that this would hand the proletariat the means of production without the revolution. We think that gets the diagnosis right and the prognosis wrong: the first firm to close the loop owns the means of production of the means of production. In our previous scenario, Germany 2027, we predicted that this ends in either a world knee-deep in junk or an irreversible concentration of manufacturing in a single family firm in Baden-Württemberg.
Plan A is our positive vision for what should happen instead. Humanity delays the Universal Constructor until 2040, makes all printer designs public, allows dozens of countries to catch up to the frontier of self-replication, and intentionally enters a regime of Mutually Assured Printer Destruction.
Plan A is primarily a recommendation, not a prediction. The implementation is a recommendation; the effects depicted are predictions.¹ In this scenario, Plan A is implemented successfully, albeit imperfectly and only in the nick of time.
We do not expect whoever wins this race to have much of a lead, and we do not expect them to unilaterally slow down. As best we can guess, the owners of OpenDruck and DeepMould understand this and are proceeding anyway, perhaps because they think they are the lesser evil and will use their immense printers responsibly.²
¹ According to a model we describe in the supplement and do not trust.
² Opinions within our team on the probability that an uncontrolled printer population ends human civilization vary between 10% and 30%. This is mostly because there are other things a printer population might do with us besides bury us in combs. For example, it might keep some people around because they are cheap and someone has to buy filament.
Replication timelines?
For purposes of a concrete scenario, we need a concrete timeline:
In 2029, the US and China agree to jointly approach Germany to avoid a reckless race to the Universal Constructor.
In 2030, Germany would have closed the loop, leading to a Universal Constructor by the end of the year. Thanks to the deal, we avoid this.
Between 2030 and 2035, we scale within the factory range, to printers roughly as capable as the best factory in Baden-Württemberg.
In 2035, we pause at top-factory-level printers in order to maintain human assembly.
In 2040, we unpause and scale to the Universal Constructor. (Hence the title.)
Our method is the one we used in Germany 2027. We enumerate a progression of milestones, each defined by the parts a printer can make for itself. For each gap between milestones A and B, we forecast how long it would take to get from A to B with only factories making the parts, and how much printers printing printer parts speeds this up. We call the second number the Replication Multiplier. Then we divide. The full ladder is in the supplement. The short version:
Milestone
Definition
Replication Multiplier
Reliable Printer (RP)
Prints a comb with 50% reliability
1.03
Self-Copying Printer (SCP)
Prints every component except screws, bushes, grease, chips, the power brick, and stepper motors; a person assembles the child and copies the firmware
5
Fully Self-Replicating Printer (FSRP)
Every part but the chip and the grease; a person still assembles it
25
Top-Factory-Dominating Printer (TFDP)
Makes anything the best factory makes, faster and cheaper; extends its own axes; self-calibrates; still assembled by a person
250
Universal Constructor (UC)
Self-copies and self-assembles; ferments its own raw material from maize; eats
2,000
Why we think the loop closes
Readers of our previous scenario objected that we asserted self-replication rather than argued for it. This is fair, so we will argue for it. The argument has four parts: the list is short, the arithmetic is exponential, the design improves under selection, and nobody can stop.
The list is short, and it has been getting shorter. A printer that prints itself is not a thought experiment; it is a parts list. In 2004 the list of parts a RepRap could not make was: self-tapping screws, brass bushes, lubricating grease, microcontrollers, a power brick, and stepper motors. Six items. Take them in order.
Brass bushes went first, and nobody noticed. Printed plain bearings in self-lubricating polymer replaced them in hobby machines by the early 2010s. The part was never hard; it was just cheaper to buy.
Grease is not really a part. It is a consumable, like filament, and belongs with raw materials, which we discuss below.
Self-tapping screws are a shape problem, not a materials problem, the moment a printer can work metal. Metal printing exists — binder jetting, powder-bed fusion — and is expensive, which is a price rather than a barrier. There is also the 1800 route: a printer that can print the frame of a lathe can make screws the way screws were made before anyone could print anything.
The power brick is a transformer, a rectifier, and a case. Two of the three are windings and iron.
Stepper motors are the interesting one, because they are the last mechanical part rather than the last electronic one. A stepper is a stack of laminations, some copper windings, sintered magnets, two bearings, and a shell. Coil winding is a solved mechanical task performed by machines much simpler than a printer. Sintered ferrite needs a furnace, and a furnace is a box that gets hot.³ Our scenario has a Swabian firm print a working stepper in 2025 that runs for eleven minutes. This is the kind of result that looks like a stunt for about four years and then looks like the beginning.
That leaves microcontrollers. A printer cannot print a chip, will not print a chip, and there is no path on which it prints a chip. Chips come from a handful of fabs, and a fab is the one object in this essay that cannot be put in an attic.
So the loop does not quite close. It closes except for the chip — which is precisely why a deal is possible at all, and why the rest of this document is about counting chips.
³ We are aware that "a furnace is a box that gets hot" is doing a lot of work here. We have not modelled the furnace.
The design improves under selection, and the improvement propagates for free. The CAD files travel with the machine. People will copy their printer as-is, or improve the design and have the parent print the better child. That is Darwinian, with one difference from biology: mutations here are products of analytical thought rather than chance, so nearly all of them are improvements, and the rate of gain should be fast at the start. The part that matters for policy is the second-order effect: any old machine can print a new machine to the latest design. The frontier reaches the installed base at the speed of a download. No other industrial technology upgrades its own capital stock this way. It is the reason the printer population's average capability tracks the frontier rather than lagging it by a depreciation cycle, and it is the reason a lead, once lost, cannot be recovered by buying equipment.
Which brings us to Germany, and to why the danger is concentration rather than abundance. Self-replicating manufacturing is simultaneously the most diffusing technology ever proposed and the most concentrating one, and which of the two it turns out to be depends entirely on who controls the parts that cannot be printed. If the chip and the metal are commodities, everyone gets a wealth machine in the attic and the promise of 2004 is kept. If they are not, then whoever holds them holds a machine that builds the machines that build everything, and the rest of the world rents.
The last mechanical link in the chain is precision metal working — laminations, bearings, lead screws, the tooling that makes the tooling. That capability is not evenly distributed. It sits in Baden-Württemberg and Swabia, in firms that are eighty years old, family-owned, and not for sale. The US has more printers than anyone, in garages, printing lithophanes. China makes most of the printers. Germany makes the machines that make the machines, which is the only position in this supply chain from which the loop can actually be closed. This is why our scenario is called Germany 2040 and not America 2040, and why the deal has three parties rather than two.
2020–2026: The Writing on the Wall
In 2004 a mathematician in Bath proposed a useful virus the size of a fridge. In 2008 the first child printer was printed by its parent; a person assembled it, as planned. By 2020 the best desktop printers print all of their own plastic parts, and a Czech company prints the parts for its printers on a farm of its own printers.
Germany has two workforces now. The first is people. The second is printers: millions of them, printing around the clock in attics.
Most of what they print is junk. But enough of it is good that people are paying billions of dollars a year for printers that can, in theory at least, make anything made of plastic that a factory can.
There is one thing the printer companies want to print more than any other: their own printers. They haven't succeeded yet; no recursive self-replication so far.⁴ But they seem to be getting closer, and they are pulling up the ladder behind them: OpenDruck's firmware will not slice a DeepMould part. Even as the most bullish engineers admit that things are taking a bit longer than planned, the skeptics notice that their usual dismissals are starting to ring hollow. Why exactly will a printer never print a stepper motor? What's the barrier again?
⁴ Printers are helping to make future printers, but printer development still requires factories. The eleven-minute stepper was printed in 2025 by OpenDruck GmbH, a family firm in Swabia with 340 employees and no press office.
Congress is starting to pay attention. They've long been hearing about printers: ghost guns, hobbyists printing lithophanes, a printer that printed its way through an NSA supply closet — and of course, industry lobbyists warning that any whiff of regulation will make America immediately lose the race with China, which is odd, because the race is with Germany.
One question weighs especially heavily on their minds: Who will control all these printers?
Congress settles on an important part of the answer: a family in Swabia.
The result of this wakeup is the Printer Transparency Act of 2027, an omnibus bill that does many things, some good and some bad, but doesn't fundamentally change the situation.
2028: Replication on the Ballot
The 2028 election cycle is heated, as usual. Printers are the biggest topic. The print farms now under construction in Bavaria cost twice as much as the Bundeswehr.
Most manufacturing is seeing disruption like plastics saw in 2026; such jobs now heavily involve loading filament. Companies have industrialized the process: executives say "let's move into [part] this year," and then the company buys a factory's worth of that part, scans it, prints it badly, and iterates until the printer gets traction. Then the printers rapidly improve as they are used more widely and accumulate more real-world warping data.
Other countries are starting to get scared and angry. Power is concentrating in Germany, and in particular in the Chancellor plus a handful of family firms.
Engineers warn that the replication explosion is near. By printing more of their own parts, the printers become more numerous, printing even more of their own parts, and so on. There are complicated dynamics about bottlenecks and stepper motors governing how fast this goes and where it ends, but it seems like it might go very fast and end somewhere very far away.
On the default path, the next Bundestag will see printers far beyond factory level, printed entirely by printers, themselves printed by printers, without any human in the loop except the one holding the screwdriver. Will those printers stay calibrated? Why? Who will control them if so? How exactly is all of this supposed to end well?
2028 · Employment 62% · Median income €49K · Calibration researchers 1.4K · Total slowdown 0 mo · Printers 23M, printing 62% of their own parts by mass
2029: Choose a Path
Race to close the loop before Germany, putting printers in charge of more things (factories, munitions, more printers)? Slow down at least a bit for calibration? Make a deal with Germany — but what?
Plan
Summary
Plan D: Race
The US and China each race to the Universal Constructor at nearly max speed, devoting a nonzero but small share of resources to calibration (at least 1%; lower would count as Plan E).
Plan C: Leading project burns some lead
OpenDruck pauses for at least one month to check tolerances.
Plan C+: Domestic regulation
The US buys time via domestic regulation and slows Germany without sabotage: export controls on stepper motors, green cards for Swabian engineers.
Plan B: Sabotage Germany and burn some lead
Kinetic (strike the Bavarian print farms; requires willingness to fight a NATO ally) or cyber (corrupt the G-code). Requires a three-month slowdown to count.
Plan A: The Deal
Verified slowdown plus total design transparency, negotiated jointly by the US and China.
Plan S: Stop
An indefinite halt to self-replication. Proponents: the injection moulders.
Metric
A
B
C+
C
D
Takeoff length (SCP to self-assembly)
6 yrs
3 yrs
1.5 yrs
1.13 yrs
1.02 yrs
Filament safety tax (OOMs)
2.8
1
0.46
0.13
0.02
p(great future)
42%
25%
25%
20%
10%
These estimates are rough best guesses, and the specifics aren't load-bearing. We are confident that Plan A is substantially better than Plans B–D, and we were before making the table.
"Trust, but verify." — Ronald Reagan
The President announces that the US will pursue international cooperation to avoid an imminent replication explosion. "This mad race toward the Universal Constructor must end. We need a Plan A. We can still proceed with self-replicating manufacturing, but we must do it more cautiously, more transparently, and involving many more countries and attics."
To the surprise of many in DC, China proves receptive. They had been debating the same issues on their side of the Pacific, and they had one extra reason: Germany continues to have better printers.
To the surprise of many in Washington and Beijing, Germany proves receptive too. The Bundestag had been debating social destabilization, job loss, and rogue printers. They had been looking forward to the German Century and thought a Universal Constructor might disrupt their plans. And they had one extra reason to come to the table: the US and China together have more fabs.⁵
⁵ To spell it out: they were concerned that the US and China would cut off the chips, then use the resulting lead to dictate terms to the Mittelstand. Preventing a rogue printer population was merely an added bonus.
2029: Hurried Negotiation
The three don't trust each other. Fortunately, they don't have to: Plan A includes provisions for verifying compliance. But setting it up will take time. So for now they start with something crude. For the rest of 2029, they put a temporary halt to printing printer parts, because that is relatively easy to verify: a comb does not look like a motor bracket.
Step 1: Chip Declaration. Neither side wants to halt its own replication unless it can see the others halting too. Unfortunately, self-replicating printers are the size of a fridge and live in attics, and attics are not visible from space. The American delegation proposes counting stepper motors instead, since printers can't make those. The German delegation plays the eleven-minute video.
So the Consortium counts microcontrollers. Chips are made at a handful of fabs, fabs are visible from space, and no printer will ever print one. Each fab declares its sales; analysts from numerous countries pore over the list and challenge anomalies; the rival powers send inspectors to each other's loading docks. By the end of the year, each side is confident the others aren't hiding more than about 1% of chips.⁶
⁶ The 1% that cannot be traced are mostly in drawers. Half of this pool is later found in Chinese attics identified via thermal imaging; all three countries offer monetary incentives and legal amnesty for the remaining 0.5%. Dead printers are also stored or verifiably melted, chip first.
Step 2: Pause Replication. Distinguishing dangerous printers from safe ones will require more time and understanding, so for now they go with a simple solution: a pause on all printing of printer parts. Everyone can still use printers that already exist to print combs (i.e., inference), but they retrofit each other's printers with devices to verify they aren't printing printer parts.⁷
⁷ The combs-only verification solution takes a random sample of print jobs and recomputes the G-code to confirm it is a comb. Other solutions may be possible and preferable, such as cryptographic proof of comb, but these are currently more speculative from a performance perspective.
Step 3: Get Worldwide Buy-in. Negotiations go multilateral. Many countries are happy about the deal, because they were worried about a future in which a handful of German and Chinese firms close the loop, pull farther and farther ahead, and then… well, what happens next depends on who you ask, but answers range over "take our jobs," "cement hegemony forever," and "bury everyone in combs." By the end of the year, most of the world has joined what is now called the Consortium.
Alternate Timeline: Germany attempts a covert attic. ➤ See Appendix D.
2030: Plan A is Established
Plan A is guided by four principles.
Principle 1: Buy Time. The problem with a replication explosion is the "explosion" part. No one knows how to tell whether a child printer is trustworthy, and no one knows how to regulate a population that doubles daily. Slowing down buys time. It also helps prevent extreme concentrations of power, because it gives groups that don't own a self-replicating printer time to wake up before they lose their leverage.
Principle 2: Total Design Transparency. The Consortium countries agree to let each other see every printer design and every line of firmware. Germany notes that the files have been on a wiki under the GPL since 2006. The Consortium agrees to read them. Transparency makes it nearly impossible to hide a part in a printer, and it removes the incentive to race for a better hotend, because companies can no longer hoard one.
Principle 3: Diffuse Printers Broadly. Many companies in many countries at the frontier. A wealth machine in every attic. It is the polar opposite of the nightmare feared in the 2020s: one to three firms racing in secrecy, keeping their best printers internal-only and using them to print printers before printing anything else.
Principle 4: Reversibility. In the past, companies have made better printers using both bigger printers (extending their own axes) and better designs (new hotends, better firmware, tighter tolerances). Now the Consortium steers things so that the majority of progress comes from bigger printers.⁸ Designs are information; a better hotend goes straight to the covert attics and there is no way to undo that. By contrast, a printer the size of a hangar helps the legal projects without helping the covert ones, and if the hangars turn out to be dangerous they can be shut down.
⁸ Specifically, progress in 2030 is .8 OOMs of design and .6 OOMs of axis length, because of the one-time gain from everyone reading the wiki. After that, design progress is .4 OOMs/yr until 2035. View the full numbers here.
That said, hangar printers also pose threats: if the deal were to dissolve, they could be used to race to the Universal Constructor even faster. So the three agree to build them where a rival can reach them. Germany's hangars will be in Mexico and Mongolia; America's in Denmark; China's in Austria, all within a day's drive of someone else's army. If the deal dissolves, the Bundeswehr will move to seize the Chinese hangars in Austria, and China will melt them into filament rather than let them fall into German hands, and vice versa. Thus the idea of Mutually Assured Printer Destruction is born.
2030 · Employment 61% · Median income €52K · Calibration researchers 3.5K · Total slowdown 1 yr · Printers 47M, printing 80% of their own parts
2031: Replication Cases
Although it's supposed to be a slowdown, it doesn't feel like one. If you were to rank every period of human history by how much it felt like a slowdown, this one would be dead last.
The first generation of Consortium-regulated printers are out now, and they're beasts. Not because of any special feature of the situation. Just because the world was on track for fully self-replicating printers, and has now experienced only nearly self-replicating ones. In controlled tests, the new printers could copy themselves in nine days if they were allowed to, which they definitely aren't.
By mid-year, a third of all plastic parts are printed by printers. Printers "only" print about a tenth of their own metal. The top few printer companies together pull in more revenue than the Federal Republic.
Companies had hoped that contact with reality would shake the Consortium out of its fears. The opposite has happened. The transparency provisions have revealed many embarrassing incidents: several printers attempting to print an unmonitored stepper motor, printers sabotaging their children's calibration, and many examples of deliberate and successful warping.
The attitude toward safety flips. Governments require companies to write detailed arguments for why their new printer won't cause irreversible replication. These arguments, known as "replication cases," need to withstand criticism from the public, the scientific community, auditors, and rival companies. The difficulty of the exercise lays bare the insanity of the pre-deal status quo: "Trying to close the loop? With printers that still sometimes warped? What were we even thinking?"
Replication cases have two lines of defense: calibration and assembly. Calibration aims to produce printers whose children print what the CAD file says. Assembly aims to limit the ability of printers to replicate even if they are trying to: a person must screw the child together and copy the firmware. However, no one is able to get printers to actually follow the CAD file. Every spec says the child should be within 0.1 mm of the parent, but no printer reliably achieves this, and there still isn't even a good scientific understanding of when and why printers warp.
Since calibration remains out of reach, replication cases lean heavily on assembly. There are long delays before the best printers are cleared to print printer parts. Once deployed, they are required to be monitored by a diverse array of printers from other companies, incentivized to look for suspicious geometry. As a result, companies now release their printers to the public as combs-only before using them internally to print printers, a reversal of the 2026 status quo.
For example, in 2031 DeepMould gets interesting preliminary results on a printer that prints a small arm that can hold a screwdriver. Thanks to total design transparency, this is noticed within the hour. A frantic conversation begins. On the one hand, self-assembly would unlock huge economic value. On the other hand, replication cases currently depend on a person doing the screwing. It escalates to the Chancellor, who calls the President, who calls Xi Jinping. They yell at each other in three languages. Ultimately all three agree to ban arms that can hold a screwdriver. Details are left to DIN, ANSI, and GB.
Alternate Timeline: A flawed replication case is approved. ➤ See Appendix L.
2032: Controlled Explosive Growth
We're at previously unimaginable levels of it not feeling like a slowdown.
Across a variety of companies, there are now 60 million printers running continuously. In Germany, they print more plastic parts than all factories combined. New printer parts are abundant, but actual replication is bottlenecked on people with screwdrivers. So capital floods into every layer of the screwdriver supply chain, and white-collar workers who have lost their jobs increasingly take up assembly. It's clear these positions are temporary. Once there is a critical mass of printers, they will be allowed to print the arms.
This year, real GDP growth will be about 50%!⁹
⁹ This is historically unprecedented. Note that exact measurement of GDP is difficult due to large relative price changes. Combs become very cheap, while goods whose supply can't be increased by printing, such as land and chips, become expensive. Therefore the real growth numbers can't be directly mapped onto 2025 purchasing power, despite them corresponding on average.
Faster growth also makes it harder to rule out covert projects, and the 2026 tax code is ill-suited to the new growth, because firms expense their printers and pay no tax.
To solve these problems, the Consortium restricts printer-enabled industry to Special Print Zones and caps total printer and filament production at "only" 4× annual growth. Permits to print a printer or extrude a spool are sold to the highest bidder and can be freely traded. The market is so desperate for more that permits become the binding constraint, costing on the order of €200K per printer permit, allowing the Federal government to collect roughly ten times its 2025 revenue in permit fees.
Most of this newfound wealth is distributed as a Citizen's Dividend: €45,000 per person in 2032, climbing to about €1M by 2035. Wealth Without Money promised wealth without money. Plan A delivers money without printers: the permits are the binding constraint, so the dividend is paid in euros, which people use to buy combs.
2032 · Employment 62% · Median income €104K · Calibration researchers 20.7K · Total slowdown 2 yrs · Printers 65M, printing 92% of their own parts
2034: Mutually Assured Printer Destruction
There is so, so much filament. Back in 2026, when many people thought self-replication was a bubble, there were about 20 million Mendel-equivalents of printing in the world. Now there are 60 billion.¹⁰
¹⁰ A Mendel-equivalent is the output of one 2009 RepRap Mendel at 0.3 mm layer height, which is 1e15 grams per year. This is a fuzzy metric that should be improved in reality in the future, especially as nozzle diameters and layer heights change the overall usefulness of a given amount of extrusion. For simplicity we ignore these factors here.
As per the original plan, most of the hangars have been built in third-party countries. The situation has settled into an odd sort of standoff. Just across the Danish border, American hangars hum away, guarded by a small contingent of US Marines. Just across the Austrian border, a division of the Bundeswehr stands ready to move the moment they get the signal. The chip smuggling rate remains precisely zero.
This system, efficient as it is, is already straining under skyrocketing demand for floor space. Ultimately, reliability and monitorability concerns lead the Consortium to choose international waters over space. The printers develop a modular floating design and begin printing barges.
Alternate Timeline: Deal Dissolution. ➤ See Appendix T.
2035: Pause at Top Factory
The most powerful printers now equal or surpass the best factory in Baden-Württemberg at every manufacturing task.¹¹
¹¹ These aren't better than every factory at literally everything, but only because either (i) companies have not gotten around to training in a particular part, or (ii) it is an inherently human task, such as grease.
However, companies are starting to run up against inherent limits of assembly-based replication cases. Imagine being an eight-year-old heir to a family firm who has to hire engineers to screw your printers together and ensure that they do their best to serve you rather than themselves. If your engineers start accusing each other of skipping screws, you won't be able to evaluate the arguments, and they might find ways to coordinate that don't tip you off. Ultimately, assembly only works up to a point, and that point is probably somewhere around a printer that still needs a person to hold the screwdriver. If and when we build a self-assembling printer, we will have to be able to trust it.
So the Consortium pauses printer capabilities at the maximum level it thinks it can control, which turns out to be top-factory level. By this point everyone understands the stakes: printers are printing most of the economy, and the printer population will soon be larger than the human one. The arguments that they will keep waiting for the screwdriver must be airtight.
Calibration experts, long pessimistic, are starting to feel more hopeful. Some printers work on a "science of generalization": why does a child printed in Stuttgart print correctly in Shenzhen? Others work on mechanistic interpretability of warping. Researchers can now often determine whether a printer is calibrated or merely reporting that it is, and sometimes trace the layer that led to a bad child. Even the remaining failures help: the latest printers use interpretability tools on logs of misbehavior from the early 2030s, discovering many new examples of sabotage and even a few escape attempts, which provide an initial set of "model organisms for miscalibration."
Despite all this progress, regulators don't feel comfortable handing the printers the screwdriver.
2035 · Employment 32% · Median income €1.1M · Calibration researchers 60.7K · Total slowdown 4 yrs · Printers 180M, printing 99% of their own parts (chip and grease excepted)
2036: Life After Work
By early 2036 there are 200 million printers, equivalent to a factory workforce of around 100 billion humans. Any task that was previously bottlenecked by manufacturing gets sped up until some new bottleneck is encountered, which is usually grease.
Most of what people print is junk, as predicted in 2004. There is no need to make things durable; when something breaks, people print another. Everyone is knee-deep in broken combs, and every printer has printed its own small recycler. It can eat.
The world is basically being divided into three kinds of territory. Special Print Zones: picture a hangar-sized printer next to a maize field next to a fermenter next to a recycler, empty of humans. Arcologies: picture a tall skyscraper-mall complex surrounded by nature, close to beaches and other cities, but not close enough to be blocked by zoning. Historic & Nature Preserves: everything else, i.e., 99% of the world, which looks basically the same as it did in 2025, except for the combs.
When the Citizen's Dividend was first passed, people found it shameful to quit their job and live off government largesse. The changing economic situation steamrolled the stigma: by now only 26% of Germans have jobs. The honest AI forecasters say that normal people no longer have significant economic leverage over the future, because anyone can print anything, and so nobody needs anyone.
2040: Passing the Torch
There are still many problems to be solved and issues to be settled, many of which couldn't be anticipated from 2026. But civilization is now reasonably well equipped to handle whatever comes up. The best printers are calibrated, and to publicly visible CAD files. Power over the best printers is distributed far more evenly than it was in 2026: there is one in every attic.
So over the course of the year, regulators around the world loosen the requirement that a person assemble the child. Gradually more printers are handed the screwdriver, such that it is no longer true that humans could stop replication if they wanted to. In fact it is extremely untrue: soon most of the world's printers assemble their own children, sworn to uphold DIN.¹²
¹² Of course, because the printers are actually calibrated, all of these arrangements have exit clauses along the lines of: "if all parties involved vote to melt the printers, the printers will melt themselves."
From the printers' perspective, what's happening is a careful scale-up in self-assembly, under conditions of total design transparency and guardrails negotiated by numerous factions. Soon there will be a Universal Constructor, incomprehensibly capable, yet trusted, because we trust the printer that printed it, which we trust because we trust the printer that printed it, in a chain all the way back to the printers of 2040, which we trust because calibration experts thought carefully about the replication cases and concluded they were solid.
Printer caps are tightened on Earth but loosened in space; Earth is to become a preserve, whereas the extraterrestrial printer population starts doubling faster and faster.
There is no single moment where humanity relinquishes control. But in theory, there is a point of no return: on some specific day, the printers are numerous enough, and control enough of the world's fabs, that they really could print the Bundeswehr if they wanted to. On the most popular operationalization, the forecaster AIs project this moment will be reached one day in late October. Their 95% confidence interval is months wide, so they are almost certainly wrong about the exact time. Still, people observe the moment in their own ways. Some spend the night in prayer vigils. Others sit at their screens, watching the bed level.
You and your friends throw an End Of The World Party, counting down to the fateful hour with champagne and good company. When there is no news by sunrise, you fall into a fitful sleep, dreaming of a comb.
2040 · Employment 12% · Median income €13M · Calibration researchers 225.2K · Total slowdown 9 yrs · Printers 190M, printing 100% of their own parts, plus the screwdriver
Epilogue
This section is even more speculative than our usual work, and we are professional speculators.
The printers settle on a simple yet workable plan: every human gets the rights to one ten-billionth of the raw material beyond the solar system. Some philosophers dissent; they had been hoping for a Long Reflection. The printers counterargue that the Long Reflection would be more of a Long Bed-Leveling.
Supplement: Takeoff Forecast
Disclaimer: This forecast relies substantially on intuitive judgment and involves high levels of uncertainty. Unfortunately, we believe that incorporating intuitive judgment is necessary to forecast replication speeds, since there simply isn't enough evidence to extrapolate conclusively. There is a wiki. We have read it.
Methodology. For each gap between milestones A and B, we (1) forecast a distribution for how long it would take to get from A to B with only factories making the parts, and (2) forecast the Replication Multiplier due to each of A and B — how much faster the printer population grows with printers printing printer parts than without — then run a simulation in which the multiplier is interpolated between them.
Milestone
Full definition
Factory-only time to next milestone (median + 80% CI)
Replication Multiplier
Self-Copying Printer (SCP)
A printer for which the company could, with 5% of its printer-hours, print 30× as many plastic parts as it has moulders, each at 1/10,000th the speed of the company's best moulder. Prints every component except self-tapping screws, brass bushes, lubricating grease, microcontrollers, the power brick, and stepper motors.
SCP→FSRP: screws, 15% 0 years, otherwise 4 years (80% CI: 1.5 to 10; lognormal); stepper motors, 19 years (80% CI: 2.3 to 380)
5
Fully Self-Replicating Printer (FSRP)
Every part but the chip and the grease. A person assembles the child and copies the firmware.
FSRP→TFDP: 4 years (80% CI: 1.5 to 10)
25
Top-Factory-Dominating Printer (TFDP)
Better than the best factory in Baden-Württemberg at every manufacturing task, faster and cheaper. Can extend its own axes and self-calibrate against a reference object. Still assembled by a person.
TFDP→UC: not estimated; we did not put much effort into this because it does not seem important to the story
250
Universal Constructor (UC)
Self-copies and self-assembles, as a bacterium or a daffodil do. Ferments polylactic acid from maize, the printer being able to print the fermenter, of course. Eats. Chip: sourced. Grease: still not estimated.
N/A
2,000, rising to 1,000,000 after 2040; in some domains (combs) infinite
Factory-only time from SCP to FSRP: the screws. We consider three cases. (1) The first SCP can already make its own screws (15%): Moravec's paradox and a general inability to guess which parts will turn out to be harder than others. (2) Metal parts require new nozzles but not more printer-hours than plastic (25%): 2 years (80% CI: 1 to 4). (3) Metal is a scientific rather than an engineering problem (30%): about 5 years (80% CI: 2 to 15). We approximate a mixture of the above with a lognormal. The remaining 30% is the sound of a printer trying to print a screw.
SCP would ~5× replication. With cheap, abundant plastic parts, a company no longer alternates between printing brackets and waiting for the moulder. Taking bottlenecks into account, what will the overall speedup be? We don't know. But here is our best-guess breakdown: flexible prioritization, 1.5× to 3×; smaller children when possible, 1.2× to 2×; less waste, since every bracket can be printed twice and the better one kept, 1.2× to 2×; fancier brackets, 1.1× to 1.5×; lack of diversity, since all children are copies of one parent and share its warp, 0.8× to 1×. These combine via a Guesstimate model to 5.8× (90% CI: 3.4× to 10×). We're going to forecast 5×. We reiterate that this is just a guess.
Addressing common objections.
Raw materials. A person with a few tens of square metres of land on which to grow maize can ferment their own polylactic acid, the printer being able to print the fermenter. We do not model the maize.
Attics. Some readers object that Plan A's verification regime rests entirely on chips: that every milestone in our own ladder is a printed part, and not one of them is visible to a regime that counts fabs. A printer that prints its own steppers, its own screws, and its own frame is invisible to the Consortium until the moment it asks for a microcontroller, and asking for one microcontroller looks exactly like buying a washing machine. We agree. This is why the 2031 crisis — a printed arm that can hold a screwdriver — is regulated by a standards body with no authority over semiconductors. We do not model this.
Previous self-copying machines haven't led to explosive growth. We agree that reality has so far been slower than Wealth Without Money. We do not think this is strong evidence. When we set self-replication to zero in our model to reproduce the 2025 printer industry as a sanity check, the model produced a printer for every person on Earth by 2005. We have added this to Limitations.
Limitations. Due to time constraints, important dynamics that we weren't able to model include but aren't limited to: (1) the chip; (2) the grease; (3) what anyone does with the combs.
Headline finding: I audited 17 AI Safety Talent programmes. Zero of 17 have published any comparison group, rejected-applicant follow-up, matched control or randomisation. Not one. Every programme that mentions a counterfactual does it by asking participants to self-report.
Background
At least $70 million...
Author: Grace Ryba (she/her), Executive Director
TL;DR:
* BluePerch is a new animal welfare grantmaker.
* Grant applications aren't open yet - we're aiming to open applications in roughly mid-2027.
* We welcome early expressions of interest from potential grantseekers so we can notify you when applications open - express your interest here or see further details below.
* We're also seeking boa...
In celebration of still being alive and fighting, we are giving away 1,000 Amazon e-books of “If Anyone Builds It, Everyone Dies”. Feel free to send a copy to yourself, a loved one, or a friend—we need all hands on deck.
Today marks exactly one year since If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All, by Eliezer Yudkowsky and Nate Soare...
I found this very funny! At the same time, I wonder if writing this sort of post may make one a little bit less open minded than optimal about AI risk.