The birthrate shredder

Degrowth is the simplest option. We assume that there's no major social unrest even as we reach, and then pass, peak population, and even as the dependency ratio grows, because automation innovation continues — even though it slows down as the population gets older and older, both soaking up more capital and lowering the average cognitive dynamism (by raising the median age) of humanity — and is able to somewhat compensate for the increasing dependency ratio and the lost hands and minds to run the world. We represent this lack of major social unrest (and loss of hope), as well as the economic burden on the working age population to support the increasing dependency ratio, by assuming that the decline in total fertility rate continues at a constant pace, instead of accelerating as we see in places like South Korea and Japan.

This situation is essentially the "fully automated luxury nursing home" model. We invent enough automation, cobots, and ANI to take care of everyone and keep everything running, but most automation innovation will be soaked up just adjusting for the decline in total population and the increasing dependency ratio, so nothing much/new/ happens, in a fundamental sense, and humanity isn't going to spend its time complexifying or exploring or pushing the boundaries of what technology can do. It doesn't sound like so bad a world, if you're inclined toward passivity…

…until you realize that the birth rate is still below replacement, and the dependency ratio is still climbing. At some point, first, the dependency ratio will get too high for any amount of automation that can't fully substitute for a person — i.e., AGI — to do anything about it; further on, at some point, we'll simply not have enough people to keep a technologically advanced civilization running.

It's hard to say exactly how long this would take. While we can extrapolate out birth rate and, from birth rate and current population, total population numbers, to even attempt to ballpark the actual answer, we'd need to know [ the minimum population required to maintain technologically advanced civilization, which is way harder than just putting a number on the minimum viable population for a species to avoid genetic inbreeding.

What we might be able to say is that it is likely that a rather large population is needed to maintain, let alone expand, technological knowledge, and the more complex the technology is, the more people are needed — this is called [ The Tasmania Effect, and has some anthropological theory and evidence (the Tasmanians themselves) behind it:

… Henrich has analyzed models of the “Tasmanian Effect.” At the time of European contact, the Tasmanians had the simplest toolkit ever recorded in an extant human society … Archaeological evidence indicates that Tasmanian simplicity resulted from both the gradual loss of items from their own pre-Holocene toolkit and the failure to develop many of the technlogies that subsequently arose only 150km to the north in Australia … Henrich’s analysis indicates that imperfect inference during social learning, rather than stochastic loss due to drift-like effects, is the most likely reason for this loss. This suggests that to maintain an equilibrium toolkit as complex as those of late Pleistocene hunter-gatherers likely required a rather large population of people who interacted fairly freely so that rare, highly skilled performances, spread by selective imitation, could compensate for the routine loss of skills due to imperfect inferences.

Additionally, it has been theorized that that in order for an economic output that is the product of many steps to be of high value, it needs to have very skilled workers working at each step of its production process — [ the O-ring theory of economic development — and, extending this, we might say that the more specialized the discipline that goes into a step in production, the more exceptional (exceptionally talented, hardworking, predisposed, or have exceptional amounts of resources poured into them, whatever) a worker needs to be to be skilled at it. This means that, say, you need a very specialized 0.1% of the population to be good at designing semiconductor lithography methods or something; and you need an equally small proportion of the population for each hyper-specialized task in the chain. But you also need a large enough absolute number of people in each specialization, and [ as Charlie Stross points out in his blog post on the topic, that number only increases as technology gets more complex and won't fit into one person's head. Thus, if you have, say, 10,000 steps in a production process, and each of them needs 1000 99.9 percentile people from your population, then you need a population of 10 billion people (need 10,000 * 1,000 people, and that needs to be 0.1% of your population) to service that pipeline. Such a pipeline would already be barely possible with our current population — maybe an example of this would be ASML's extreme ultraviolet lithography machines, the most complex devices ever built by humans, which require the cognitive and material output of literally the entire developed world working in tandem to be made to work, or whatever is going on at TSMC, which nobody else seems to be able to effectively replicate.

All this stuff really tells us though is "it might be kinda big." We knew that already. We're still not even remotely in the right zip code — we still don't even know the order of magnitude! So let's take a different tack. Now that we have a reason to think "it's probably large," maybe we can use that to guide us to looking in the right place for an actual number?

Taking a page out of Stross's book, what if we assume that we need around the total population of the US, the EU, Japan, South Korea, China, and Taiwan? Those countries basically make up the global supply chain of bleeding edge technology in aerospace, robotics, software, medicine, chips, etc, plus make enough food to support themselves, so they make sense as a sort of factorable module to pull out as a ballpark estimate. We can't just pull out the high-techology sectors plus a few specific supporting ones like transportation, manufacturing, and agriculture, because you also need doctors to look after the health of your engineers, teachers to teach them, psychologists to make sure they don't go crazy, managers and politicians to organize things, actors and other entertainers and artists to keep them enriched and give them direction, etc, however, so we really do have factor in the*total* populations of all of these countries. If you do, you get roughly 1.3 billion people.

What about automation, though? As we grow more technologically advanced, it's logical to think we'll need fewer and fewer people to maintain civilization's engine, even if there's a fundamental minimum amount of people you need because you can't replace/everyone/ without it being human civilization at all anymore. Thus, assuming a constant floor to population, before civilizational complexity can't be maintained, would be to assume we can't automate any more than we have — exactly the kind of assumption we don't want to be making in accelerationist theory-fiction!

I decided to model this pretty simply: using labor productivity growth. Even as recently as 2024, we've seen roughly 1% labor productivity growth year on year. That usually is translated into economic growth overall — in terms of GDP — but in theory, it also means that the amount of labor (and thus people) required to maintain a certain level of civilization has shrunk by about 1%.

However, we can't take this number for granted; this is the number we get with a relatively young, cognitively dynamic population and a lot of capital and labor to spend on automation innovation; as that drops, labor productivity growth will also drop. Luckily, that's easy to model, since that effect is already encapsulated in the dependency ratio. According to the UN via Our World in Data, the global average dependency ratio is about 0.5, so if we assume we get 1% labor productivity increase every year at that dependency ratio, then that means that at DR=0.5, we reduce the 1.3 billion number by about 13 million people anually. We'll measure the effect of DR on labor productivity growth via=min(0, 0.8 - DR)=, to ensure that as we get closer to the innovation dead zone of DR>=0.8, labor productivity growth slows and eventually stops. Then we just need to find a productivity growth factor to ensure that the previous number, multiplied by that factor, produces 13 million when DR=0.5, to calibrate to real world statistics. As it happens, that productivity growth factor happens to be ~43 million (for more details, see the notebook).

I took the current world population (8.2 billion) plus rate of decrease of global TFR (-0.0054 per year, based on going from 2.25 today to about 2.1 in 2050 — pulled from numbers mentioned earlier), and asked Gemini to come up with a Python program that modeled how demographics would change over time based on those numbers. Then I had it calibrate it until it endogenously matched the [ UN's estimation of a peak global population of ~10.3 billion people by 2084, so that there was some kind of back of the napkin sanity check to make sure our demographic numbers were roughly right.

Then I had it plot out the population for the next ~400 years, assuming the global TFR keeps decreasing at the exact same rate, and stops decreasing at 0.72 (since that's the worst TFR we've ever seen without like a plague or natural disaster or something, so it's a reasonable limit). Here's what that gave me (link to the notebook here):

Metric	Value
Scenario	Global Nursing Home (Dynamic Threshold, Extended)
Pre-Collapse Peak Population	10.33B
Peak Year	2084
TFR Below 2.1	2048
TFR Hits 0.72 Floor	2286
Innovation Death (DR > 0.8)	2226
Complexity Breach (Dynamic Threshold)	2337
Final Complexity Threshold	0.20B
Threshold Reduction	1102M people
Population Nadir	0.06B
Nadir Year	2375

So we'll have a run of about 247 years, at best, before advanced civilization becomes unsupportable, and we revert to a post-civ "utopia" of picking through the leftovers and ruins of a much more advanced, capable, dynamic, interesting civilization that was actively complexifying and expanding, for scraps of technology and goods that we would have no hope of ever producing again, or even understanding. Basically what [ Margret Killjoy wants. That might sound like a long time to you, but that's*less than the amount of time The United States of America has existed*. It's [ 4% of recorded human history and [ 0.08% of the amount of time humanity has existed on Earth. If recorded history or human history as a whole were a single 120 year human lifespan, that would be as if we only had 4.8 years or 35 days, respectively, left before dementia hit.

It's noteworthy, though, that there are 140 years between the problem becoming evident to everyone all at once in 2084, as we hit peak global population and the birth rate doesn't start going back up, and 2226, when civilization finally becomes so overburdened with elder care that it can no longer allocate resources to anything else. We could invent anything in that time. The advancement of science and technology is fundamentally unknowable over such timescales. Even if we assume innovation is slowing down over the course of that century and a half, in the previous century and a half (from 1883 to 2026) — even with science slowed down by a lack of all the advanced scientific knowledge and technology we have now — we went from:

• The primary mode of transportation being the horse (the first commercially successful automobile wouldn't be patented for another three years)
• Human flight was considered a physical impossibility (the Wright brothers were still two decades away)
• The Germ Theory of Disease had only recently been accepted, but there were no antibiotics, no X-rays, no blood transfusions. The global average life expectancy was around 35 years.
• The first commercial electric power plant had only been around for a year, and was only serving 12 customers. The fastest way to communicate was using Morse code over a telegraph wire.

To:

• Robotic rovers on Mars
• Autonomous, self-landing orbital rockets to deploy thousands of networked LEO satellites
• We cross oceans in hours flying in pressurized metal tubes racing at 700 miles per hour at 35,000 feet
• We have automobiles, some of them electric, some of them self-driving
• We have CRISPR for gene editing and mRNA vaccines
• We carry slabs of glass in our pockets that contain more computing power than all of NASA had in 1969 (by a lot), connected wirelessly to the aggregated sum of all human knowledge in a live globally connected network
• We have non-biological imitation neural networks capable of discovering novel algorithms, contributing (in a minor way) to novel quantum physics research autonomously, conversing coherently in natural language about any topic, while checking and summarizing sources, generating functioning code even for very large or difficult tasks with some guidance, saturating benchmarks as fast as we can make them, engaging in recursive self improvement and automated research, that can answer PhD level STEM questions better than most experts or humans with access to the internet

If we miss that window, we'll be facing over 100 years of slow, stagnant decline, much like what Japan is facing right now. An inverted dependency ratio meaning more and more of civilization's capital and labor has to go to supporting the old, squeezing out innovation that was already dead on arrival anyway because most of the population is old enough to have lost the cognitive dynamism needed for real breakthroughs. A population slowly shrinking, homes growing emptier, factories slowly stuttering to a halt where they can't be automated sufficiently. Hospitals and amusement parts closing down. Train stations gliding to a halt. Lights slowly flickering out. Knowing that civilization's time has come, but having to wait the painful three generations for that foregone conclusion to actually reach us.

Nevertheless, 140 is a long time. As I said, it's almost certain we/would/ come up with something in that time.

But what if we don't have that much time?

Looking at the original demographic simulation, something occurred to me: TFR isn't likely to keep decreasing at the exact same rate the entire time. Instead, as the dependency ratio gets more and more crushing, the economy collapses, and hope for the future is lost, we'll enter a negative feedback loop where each TFR decrease has second-order effects that decrease the TFR even more. Pulling from South Korea's historical data — since it's essentially been on the front lines of the birth rate shredder for decades now — we see precisely this: its TFR drop is accelerating:

• 1990 to 2000: The TFR dropped from 1.57 to 1.48. (A decline velocity of 0.009 per year).
• 2000 to 2010: The TFR dropped from 1.48 to 1.23. (A decline velocity of 0.025 per year).
• 2015 to 2023: The TFR dropped from 1.24 to 0.72. (A decline velocity of 0.065 per year).

Adding code to take the synthetic derivative of this accelerating decline to the simulation (see the collab above), we get this:

Metric	Value
Scenario	Messy Collapse (Dynamic Threshold)
Pre-Collapse Peak Population	9.49B
Peak Year	2053
TFR Below 2.1	2036
TFR Hits 0.72 Floor	2066
Innovation Death (DR > 0.8)	2104
Complexity Breach (Dynamic Threshold)	2177
Final Complexity Threshold	0.53B
Threshold Reduction	766M people
Population Nadir	0.16B
Nadir Year	2215

This is even more interesting.

Let's compare the two timelines:

Scenario 1 Date	Scenario 2 Date	Event
2084	2053	Peak Population: The market physically stops growing. The countdown clock begins.
2286	2066	TFR Hits the Floor: this is the point at which the birth rate hits rock bottom, accelerating the incoming economic pain.
2226	2104	Innovation Death (Dependency Ratio > 0.8): This is the fatal spike. Innovation dies here. Not just because markets shrink, but because this is the exact moment the dependency ratio crosses the red line. All available capital pivots from R&D into elder care, and the aging workforce loses its cognitive dynamism.
2337	2177	Below Complexity Threshold: we literally, physically, don't have the people to maintain a technologically complex society anymore. It doesn't matter what state the economy is in or whatever, it's game over.

The first thing to notice, of course, is that this assumption breaks our UN peak population estimation calibration point. However, I think this is warranted if we're trying to look at an extreme scenario, because even the UN is [ continually having to adjust their estimates down and forward in time over time. Eventually, their estimations might intersect with this one.

More evaluatively, in this timeline, we get a run of about 151 years more of advanced technological civilization before collapse, which is less than half the time we got before. But! The decels/degrowthers do get their wish: the population pyramid*does* stabilize by 2123, giving us exactly 24 years of their utopia of a technologically advanced civilization with a stable — permanently inverted and shrinking at 36% per generation, of course — population. What does that look like? Like shit, of course, and then it all goes up in flames because we're not even making enough humans to keep it going. But good job guys! I'm sure your fantasies of "we have fewer people but the same amount of resources, so everything is better for everyone!" totally make sense (let's ignore the fact that to actually use and produce those resources, you need*lots of people*).

Some other interesting points of note: we hit the dependency ratio death number in 2104, instead of 2226. That means children born*right now* might live to see — as part of the older generation contributing to this dependency ratio — what that looks like.

Perhaps more pointedly for this discussion, the entire timeline is massively accelerated. We don't hit peak population in 2084, but in 2053; we don't hit innovation death in 2226, but in 2104. Everything shifts forward by an compounding factor. Notice especially the innovation death timeline, accelerated by 122 years: this society will only have 51 years between the problem becoming evident — with global population hitting its peak and birth rates showing no sign of bouncing back up — and almost all capital being devoted to elder care instead of education, research and development, and capital deepening. This, in turn, is a big part of what speeds up hitting the population complexity threshold: there's less time to innovate, so even barring any major advances, there's less time to improve automation and push that population complexity threshold down.

There's something else haunting about this data, too: after innovation is dead, and society is fully absorbed in trying to prevent the accelerating widening gyre of technological civilization from sending everything flying apart, it still takes another/73 years/ for civilization to collapse. That's 73 years of watching the incoming tide that will wash away everything you know, everything that humanity has achieved, with nothing to replace it for at least a couple tens of million years, and not being able to do anything about it, because your hands are full just holding everything you've got left together. This, too, is why I assumed birth rate would decline: imagine coming of age after 2104. Would you want to bear a child into this mess, knowing what advanced technological civilization could offer — shadows of which you, even in all the chaos, still get to see — in comparison to what the conflaguration of a finally dying civilization, or even a post-collapse agrarian society, could provide?

And yes, there will be a post-collapse agrarian society. Look at the graph: once civilization collapses, after we hit a population low, the population begins to go up again. What's up with that? Well, in this simulations (and all the subsequent ones based on it), I included a crucial factor: once advanced, technological civilization has collapsed, humans will eventually be thrown into a traditional agrarian lifestyle. This has two consequences: one, our mortality rates will climb back to agrarian levels — about 3x what they currently are, plus a 20% mortality rate for infants — but, more importantly, our birth rate will/also/ begin to climb up.

This won't bring back the technological advanced civilization I care about, however. Hell, it likely won't even bring back Iron Age civilization. That's because it's very probable that we've mined all of the easily available metals and minerals, drilled and fracked and so on all of the easily usable, "starter" fuel sources any brand new industrial civilization would be able use with simpler machinery — photovoltaics are far more complex than a simple steam engine or ICE. Essentially, we've made it impossible for any subsequent civilization to bootstrap itself the way we did in terms of energy — starting out with coal, and then moving on to oil, and then natural gas, before (hopefully) switching to more powerful and sustainable sources of energy like solar and nuclear power — or advanced materials (like lithium). Indeed, it's likely that all of the easily accessible bronze, tin, iron, and so on has also been mined, although it might be easier to scavenge from the ruins of our civilization. As a result, civilization that comes after us basically won't be able to climb past an Iron Age level, if they're lucky.

This is why I'm treating going below the population complexity threshold as a death sentence. By all measures, it probably is — whatever human civilization rebounds from it will be nothing but a permanently limited shadow of what it was before.

What kind of society would this collapse scenario look like, before civilization's final collapse? It's hard to tell, because a rapid collapse like this would be wildly unstable, but the timeline that occurs to me is:

1. Sometime in the 2050s, we hit peak population. We realize the clock is ticking and we have less than 50 years before the dependency ratio crushes the economy.
2. As the dependency ratio gets worse over the following decades, countries will have to massively increase the tax burden on the working population to support the larger elderly cohort.
3. This, in turn, will cause a large proportion of the working age population to just drop out of the workforce entirely, as it begins to feel to them like they're just slaving away — overworked and understaffed — just for all of that to go to supporting a seemingly entitled older generation.
4. This will also contribute to the negative feedback loop that keeps the TFR drop accelerating, via a number of factors: the youth losing all hope in the future they would be birthing children into; those who do remain in the workforce having far less resources to devote to their own children; and those who drop out of the workforce not having resources for children at all.
5. These taxes will also feed into an economic negative feedback loop, as increased taxes, a shrinking population, and a loss of hope or expectations of the future will decrease investment and economic activity, shrinking the economy further.
6. By 2104, the dependency ratio officially crosses the fatal 0.8 threshold. Capital drains entirely from tech into pensions, and innovation dies. We realize we're not going to invent AGI or advanced humanoid robotics in time—we're just fucked.
7. Eventually, countries will be forced to dramatically raise the retirement age just to delay the inevitable, as well as cut social security benefits, pensions, etc. Thanks to advanced medicine extending the healthy lifespan of individuals, a large cohort of the elderly that will be targeted by these measures — especially those in their 60s and 70s — will be healthy and aware enough to riot and protest, and each country's shrinking military and police force, itself aging, will struggle to deal with this.
8. Seeing these riots, it's likely there will be counter riots from the much smaller young population which, again, the aging and shrinking police force of each nation will be unlikely to be able to counteract. This could lead to outright generational warfare and bloodbaths in the streets. Yeah, sure, maybe each individual young person doesn't want to hurt their mother or grandmother — but these other entitled assholes? Fuck 'em.
9. In response, either nations begin repressing their elderly population, forcing them into slums and project housing on minimal food rations, or the nations go bankrupt and collapse completely.
10. In desperation, many nations will probably turn to war: invade neighboring nations to liquidate their resources to feed their own elderly population. Thanks to advanced military technology, these wars will likely be just as swift and deadly as they've ever been: while you need boots on the ground to put down a riot, all you need is drones and remote control tanks and the like to invade another country.
11. Eventually, the lights go out. If we make it as far as 2177, we'll find ourselves suddenly slipping below the 1.3 billion population threshold, and complex society will begin to rot and wither away. We'll eventually face a similar end state situation as the previous scenario, but a lot more wartorn.

Okay, those are the two scenarios that we get "by default" — the two most likely outcomes. What about more outlandish ones? The first one that comes to mind is deeply unpleasant and really weird to me, which is basically that the birth rate shredder I'm describing is evolution essentially selecting against secular individualist, technologically advanced societies, and towards highly traditional, explicitly birth rate focused, Luddite societies like the Amish, Ultra-Orthodox Jews, fundamentalist Christans, and so on. At what point, if any, would those populations outbreed everyone else to the point where they become the majority of the population, and eventually maybe basically the*only* population?

The first step for this was identifying such cultures. I had Gemini look this up, and what we got was:

Group	Est. Population	Approx. TFR	Note / Source
Haredim (Ultra-Orthodox Jews)	~1.45M (Israel), ~2.1M Global	6.5 – 7.2	The fastest-growing population in the developed world. While the secular Israeli TFR is ~2.2, Haredi TFR remains stubbornly above 6.5. Projected to be 16% of Israel by 2030. ([Israel Democracy Institute, 2025)
The Amish	~400,000 (North America)	6.0 – 7.0	Population doubles roughly every 20 years. Interestingly, the strictest sects (Schwarzentruber) who reject all electricity have TFRs approaching 10 and 97% retention, while progressive sects who use phones see TFRs drop below 5. ([PubMed / Demographic Research, 2025)
Hutterites	~55,000 (North America)	6.0 – 8.0+	Anabaptists living in communal agricultural colonies. Historically used as the absolute biological baseline for "maximum natural human fertility" by demographers (hitting TFRs of 10.9 in the mid-20th century). Currently hovering around 6 to 8. ([Demographic Baseline Data)
Laestadian Lutherans	~100,000 - 150,000 (Nordics)	~6.0	A conservative revivalist movement in Finland. While secular Finland has crashed to a devastating 1.32 TFR, regional data maps a direct correlation between high local TFRs and Laestadian density. ([ResearchGate, Regional Demographics)
Quiverfull / Fundamentalist Christians	Tens to Hundreds of Thousands	6.0+ (Targeted)	A decentralized, explicitly natalist movement. They view birth control as a usurpation of divine will ("God opens and closes the womb") and use homeschooling to prevent ideological attrition. Harder to track via census, but sociologically highly significant. ([Quiverfull Demographics)

It is worth noting that some of these populations — such as the Amish and the Ultra-Orthodox —*have* seen a drop in TFR of their own; but they have cultural/religious/ideological reasons to aim for a specific number, and what we see is that once their TFR hits around that floor (about 5.5), it stops dropping, which is interesting.

Playing out the simulation from the previous collapse scenario, but now starting with a population of ultra-orthodox religious sects that starts out at their current TFR, then drops at a rate based on the Amish and Ultra-Orthodox drop rates, then halts at 5.5, but also has a generational attrition rate of 15%, in addition to the single global one, here are the results we get:

Metric	Value
Scenario	Amish Crossover (Dynamic Threshold)
Pre-Collapse Peak Population	9.49B
Peak Year	2053
Crossover Year (Fund > Sec)	2187
TFR Below 2.1	2036
TFR Hits 0.72 Floor	2066
Innovation Death (DR > 0.8)	2104
Complexity Breach (Secular Pop)	2179
Final Complexity Threshold	0.53B
Threshold Reduction	766M people
Population Nadir	0.77B
Nadir Year	2191
Valley of Death Duration	138 years
Valley Depth	91.9%

Basically we're looking at a classic post-apocalyptic science fiction scenario here. Technologically advanced civilization dies, because the total global population drops far below the minimum necessary to maintain complexity before attrition from the exponentially growing fundamentalist religious population is able to save it, because — despite the exponential growth — their total population just starts*too small* to do much. Eventually, however, at the very nadir of the secular population, a generation after technologically advanced civilization has crumbled — decaying from civilizational dementia, pockmarked with craters from drone bombs, huge factories lying empty and full of dust with no hands to work them — the fundamentalist population overtakes the secular population and begins to climb back toward five, ten billion, perhaps more. One day, those who rejected modernity will eventually wake up to find it gone, themselves the majority of a ruined Earth, and they get to decide the future.

Interestingly, though, the secular population, too, begins to rise: with an utterly massive and exponentially growing population, attrition rates from these fundamentalist cultures are finally enough to drag up the secular population significantly by themselves. We can imagine those youngsters seeking out the few remaining old secularists — who would still be a very large population in absolute terms, so not*that* hard to find, although probably living in ghettos in fundamentalist-run cities — seeking knowledge about the old days of freer culture and lost advanced technology. Maybe one day, they'd revive it.

---

In both this scenario and the preceding one, did you notice anything? That's right: even in our aggressive collapse scenario, we don't lose innovation until 2104, but we see the population peak and begin to drop in 2053. That means we have a 51-year Twilight Zone period where the population is clearly shrinking — the emergency is obvious — but at the same time the tech and science sector is still functional: a grace period where we're at the most advanced we've ever been, induced innovation is at its absolute peak as the demographic crunch is becoming extremely obvious, and yet we still have time to make meaningful technological innovations before the dependency ratio window closes.

What happens in that time?

To get an idea, we can look again to China, which is in essentially that situation right now: it knows birth rates are too low. Its population has already begun to shrink. Yet, it still has the capital and population to push serious technological innovation in an attempt to adapt. What China is doing with this time is striking:

• Third place robot density per worker in the world, according to the International Federation of Robotics's "Global Robot Density in Factories Doubled in Seven Years" report.
• Over half of the world's robot deployments, and the highest annual installations on record, with most of it coming from robot manufacturers in China itself (again according to the IFR's "Global Robot Demand in Factories Doubles Over 10 Years" report).
• An operational stock of industrial robots that totals nearly half of all global stock, with four times as many robots as the next largest (via Statista).
• A standardized system for mass producing humanoid robots and embodied intelligence, covering the entire supply chain and lifecycle, and 140 humanoid robot mass manufacturers with 330 robot models (via Seetao).
• Chinese companies accounted for 90% of the global humanoid robotics shipments in 2025, with Unitree and AGiBot shipping more than 10,000 robots combined (via Rest of World).
• Unitree aims to manufacture and ship 20,000 humanoid robots in 2026 (via EWeek).
• Artificial Narrow Intelligence is at the top of China's priorities. The new 15th Five-Year Plan mentions AI 52 times (compared to the previous plan's 11), explicitly stating that "rapidly and significantly enhancing the level of scientific and technological self-reliance and strength" is the nation's top priority. The plan targets a 90% AI adoption rate across Chinese industries using homegrown models, pursuing state-led technological development through a “new national system” where government and private enterprises operate in unison. To back this up, China’s science and technology budget is increasing by 10% in 2026 to 426.4 billion yuan—alongside plans to increase national R&D investment by over 7% annually and train 1 million AI personnel (via Chosun).

In this scenario, let's assume that we figure out some kind of technology to extend/cognitive healthspan/ — the years a person can live before they begin to see a significant cognitive decline.

Why not lifespan? Increase in life expectancy in the past has mostly been driven by lowering infant and childhood mortality, not actually increasing the maximum lifespan of humans. That isn't particularly relevant to our simulations, and the fact that we've basically picked all the low hanging fruit with infant and childhood mortality is why life expectancy increases have slowed down dramatically since the millenium. That means there's no established curve; worse than that, there aren't really any meaningful kernels of longevity increasing technology we can see in the present, Brian Johnson and is biomarkers be damned. As a result, scientists estimate that no generation born as of 2025 will live to see 100. Moreover, even theoretically, it seems like there's a hard biological wall at around 120 years old, which very advanced biotechnology modifying humans at an extremely fundamental level — at the level of how our individual cells/work at all/ — would be needed to overcome, so even if we found a way to get humans to live to their maximum biologically-possible age, we wouldn't be immortal, which doesn't really solve the birth rate problem.

More importantly than all of that, though, previous lifespan improvements have only extended the decline period, not the years in and around our peak. That means that even if we were to figure out how to increase lifespan, we wouldn't be fixing the problem; in fact, we'd be making it worse, because then we'd have an even/larger/ population of seniors to take care of, and a worse dependency ratio.

As Dr. William Mair says in that Harvard link above, "We should focus both on what we can do now to improve health span and on a few moonshots." And we have been! Especially in 2026, we've had a lot of promising developments in this area, both in terms of understanding the causes of ageing, and preventing those causes, including things like senolytics and enhanced variants of the Yamanaka factors, as well as therapies that help with neurological diseases, and Alzheimers early warning research.

It's worth noting that a*lot* of the advances in healthspan are pretty clearly directed at the brain, over the rest of the body, as well. So, let's assume that in 20 years, we figure out how to extend/cognitive/ healthspan to 85, instead of 65, but we don't significantly enhance the physical healthspan of these elders. If we assume it takes 20 years after the population collapse's imminance becomes clear to develop senolytics and Yamanaka factor epigenetic cures to a point where they work, and also assume an S-curve rollout, where the first years slowly scale to 8% (the elites, who get it first), then rapidly scales to 88% of the general population over 6 years, then takes another 5 years to reach 100% saturation, how does that change the math?

Metric	Value
Scenario	Healthspan (Dynamic Threshold, Extended)
Pre-Collapse Peak Population	9.49B
Peak Year	2053
TFR Below 2.1	2036
TFR Hits 0.72 Floor	2066
Innovation Death (DR > 0.8)	2230
Complexity Breach (Dynamic Threshold)	2224
Final Complexity Threshold	0.10B
Threshold Reduction	1200M people
Population Nadir	0.03B
Nadir Year	2262

There are some positive aspects to this graph: first of all, the innovation death dependency ratio doesn't hit until 2230 in this scenario — bought by nearly 130 years of extended cognitive viability. In fact, looking at the graph, the dependency ratio goes down to an/insanely low/ ~0.2, which is actually/amazing/ for an economy. This would be a highly innovative, technically capable society: most of its workforce would be people with the crystilized intelligence of a 75 year old and the fluid intelligence of a 35 year old, one that would awash with capital, and somehow both having high cognitive dynamism/and/ an extremely high induced innovation _ because the population wall speeding toward them would be apparent — and a lot of capital to spend on automation innovation, medical advancements, more health or life span technology.

Critically, this extended innovation window does more than just delay the dependency ratio crisis. Because we're modelling labor productivity growth via automation as being maintained proportional to how far below the DR>=0.8 cutoff we are — calibrated to match today's 1% labor productivity growth per year at 0.5 DR — having many more years of innovation inherently also extends how long we can push the civilizational complexity threshold (how many people you need to run civilization) lower. This is why the complexity breach is delayed to 2224 instead of 2177.

However, at the same time, this would be a horrendous dystopia. Increasing the older generation's healthspans doesn't fix the fact that/we're simply not producing any more humans/, so eventually, we'll crash through the population complexity floor and technologically advanced civilization will collapse. Perhaps worse than that, though, this will be an inescapable gerontocracy: a society where the levers of power can remain in the same hands until those hands reach 80 or 85, and when they step down, there's basically no new younger generation to hand the reigns to, just slightly younger elders. Culture would be extremely slow to change, as would fundamental processes and ideological/philosophical outlooks. Any Kuhnian paradigm shifts we might hope to see would be massively delayed, since they could only happen once the older generation died off or retired, and now that's been delayed by another twenty years.

More than that, during the fourteen years of the senolytics and epigenetic therapy rollout, you've got the most intense biological class warfare that's ever been seen: most of society will be ancient, of course, but suddenly the ancient/elites/ will have significantly more mental acuity and stamina, while the rest of society is well into its cognitive decline curve. The population at large would be outplayed and outgunned at every turn.

What would this highly technically competent, rich society with high induced innovation/do/ with the roughly 130 years of additional innovation time (and ~50 years of additional civilization time) it had bought itself, though? My most likely guess is advanced robotics, embodied AI, and ANI to increase productivity and account for their aging physical bodies and shrinking workforces. In essence, a situation similar to that of scenario one, but probably even more prosperous, since the dependency ratio would be much, much better, and there would be a much larger population of workers, especially those capable of scientific and engineering innovations.

However, the other possibility is that they'd use the time to come up with some stunning scientific breakthroughs, which leads us to the next sections.

The first route the gerontocracy could go would be trying to figure out full ectogenesis. We have IVF, and we have biobags, but if they could figure out the middle period of pregnancy, then they could just/print a few million humans a year/, forever.

There are massive issues with this, of course, that render this a highly speculative — honestly, kind of dumb — hypothesis, of course.

First of all, there's actually solving full ectogenesis, which relies on the placenta. The placenta is an indescribably complicated organ: a temporary, genetically distinct organ that acts as the fetus's lungs, liver, kidneys, immune system, and endocrine system all at once, while dynamically regulating thousands of chemical and hormonal signals/minute-by-minute/ to ensure proper growth. Repliating this kind of biotechnology, let alone studying the exact maternal chemical environment it needs to replicate close enough to actually do so, to ensure proper gastrulation, organogenesis, and brain development to avoid horrific birth defects and creating an entire deformed generation, is just off the charts difficult.

Secondly, even if we figured that out, then/scaling that up/ to a civilization-scale industrial human-printing operation requires building millions of these biopods, each requiring perfectly sterile circulation of synthetic amniotic fluid, synthetic blood oxygenation, and precise temperature control, where even a single blackout could literally murder an entire generation in their cradles at once, because of the inherent weaknesses of industrial, centralized processes — as opposed to the decentralized system of birth humans have naturally. It'd be like needing to make millions of bleeding edge semiconductor fabs that/also/ needed to have the sterile conditions of a Level-4 biolab.

However, assuming this highly motivated and immensely intelligent, technologically advanced, and wealthy society, with induced innovation running overdrive,/could/ pull it off, what would that look like? Off the bat, I assumed that, even assuming this giantic, civilization-scale printing effort, the numbers would be way too small to make a difference, and that, more importantly, it'd just take/way too long/ for the research to happen and then, on top of it, for the babies to mature. Additionally, Gemini pointed out that this would massively spike dependency ratios, and/prima facie/ I agreed with it — it seemed like it would, which meant not enough resources to actually feed, clothe, house, and educate the printed babies, and also no available workers to take care of them. So, my initial expectations were not good — same as when I was first brainstorming this section of the article.

But, I decided to game it out. So, let's assume it takes 40 years of research and development, after induced innovation starts around 2050 (the population peak), and then another 30 years to finish industrial rollout, and after that we can print 10 million newborns a year.

Ignoring rearing and education, that looks like this:

Metric	Value
Scenario	Ectogenesis (Dynamic Threshold)
Pre-Collapse Peak Population	9.49B
Peak Year	2053
TFR Below 2.1	2036
TFR Hits 0.72 Floor	2066
Innovation Death (DR > 0.8)	Never
Complexity Breach (Dynamic Threshold)	N/A
Final Complexity Threshold	0.10B
Threshold Reduction	1200M people
Population Nadir	1.14B
Nadir Year	2189

This was…/deeply/ unexpected. Because this society has a century of massive capital surplus and advanced automation, the complexity floor dynamically lowers. The population stabilizes above this dynamic complexity threshold — the society successfully avoids collapse. Complex technological civilization is saved. And not just temporarily either: this setup actually seems to level the population out at a respectable number, one that is, at least based on the population assumptions we've been working with, actually sustainable?

Even more wierdly, though, this society as actually so insanely flush with at least cognitively capable workers thanks to the increased cognitive healthspan, with such an insanely low dependency ratio (again, 0.2! That's insane!) that this "baby boom," while it does push the DR up by a lot relatively speaking, especially in such a short time frame, actually only pushes it up to around 0.37. That's still incredibly low. Not just below our emergency cutoff of 0.8, but below, for instance, the US's ratio of 0.6. That's fine!

In other words,/demographically/, this is a workable society. You don't even need to solve childcare robots and AGI to educate them, either: you've got more than enough workers to raise, take care of, and teach these kids. Obviously, it'll be a really/bizarre/ childhood — a bunch of very young, brand new children taken care of by 60-80 year olds who can't physically keep up with them — but it seems like it would otherwise be relatively workable.

It's not all positive news, though. I can't really keep ignoring the elephant in the room, can I? This is a society where children are printed on demand, as an industrial/economic labor unit, by nation states or megacorporations, raised by an ancient gerontocracy that will not ever really cede power to them, and existing in a world with very few other children. This would be a very strange, very terrifying sort of posthuman cyberpunk gerontocratic utopia/dystopia hybrid, in other words: extremely wealthy, peaceful, and technologically advanced, but also despotic in terms of generational class warfare.

I think the best outcome we can hope for, though, is one that is altogether different. One where our analysis can finally shed the biological limits that seem to have locked us inevitably onto this path — the 1.3 billion population complexity floor, the 9 month gestation cycle, the biological processes of ageing and decline, the almighty dependency ratio — and move on to something/else/.

What if, either before the 2104 cutoff date in the collapse scenario or, more probably, the 2224 cutoff of the gerontocracy scenario, technology is able to achieve escape velocity from the squishy biological substrate? What if this brutal race against the human species' biological clock ticking down toward a senile midnight can be something else? What if we can replace that whimpering death for the only known intelligent species with a handoff with some successor species, either artificial general intelligence, or a/Ghost in the Shell/-style merging of copied, simulated minds and ANI to form hybrid human-artificial uploaded intelligence — a species-child? Something that can continue our expanding, exploring, complexifying, experimenting/interestingness/?

I used to think this wasn't possible. After all, there are three impossibly difficult hurdles to overcome for that to be possible: mind uploading and emulation, quick-learning, physically agile embodied AI and the robotic chassis needed to house it, and recursive, autonomous self-improvement of AI. I'm not a fan of Kurzweil and the singularity religion for a lot of reasons. Yet, just this year alone has begun to, slowly, prove me wrong.

First there was the physically agile, humanoid form — important for adaptability to environments and terrain while maintaining tool use ability, and also for inheriting and being able to use human infrastructure as long as needed. Boston Dynamics has been around for a long time, and seeing the progression of the various versions of Atlas has been impressive, but I never took it that seriously, because they were always heavy, awkward, fragile one-off research prototypes, not repeatable, mass manufacturable systems, and many technologies that seem promising never make it out of that phase. Now, we have:

• AgiBot's Yuanzheng A2 and G2, which have shipped over 5,100 units in 2025
• Unitree's G1 and H1, which have sold about 5,500 units, and the former of which is available for only $13,500
• the UBTECH Robotics Walker S, which is being deployed already on Zeekr and Geely assembly lines
• the XPeng IRON, which moves like a human
• Agility Robotics' Digit, which is actively in testing in Amazon fulfillment centers as well as GXO Logistics and Spanx warehouses
• Figure AI's Figure 02, which is currently embedded in BMW's manufacturing facilities in Spartanburg South Carolina and has "contributed to the production of 30,000 cars at BMW"
• Boston Dynamics' mass-producable Electric Atlas, the initial units of which will ship directly to Hyundai's Robotics Metaplant and Google DeepMind
• Sanctuary AI's Phoenix (7th Gen), which they say can learn new structured industrial tasks in under 24 hours

Then, there was embodied AI, which I also thought was a long way off thanks to Moravec's paradox — but perhaps not.

• First there's LATENT, which is a system that "learns athletic humanoid tennis skills from imperfect human motion data." Crucially, "imperfect" here doesn't just mean "noisy" or something; as the abstract itself states, "the imperfect human motion data consist[s] only of motion fragments that capture the primitive skills used when playing tennis rather than precise and complete human-tennis motion sequences from real-world tennis matches." Running this on the Unitree G1 — which, I cannot stress enough, is a ten thousand dollar mass produced humanoid — they were able to produce incredibly reactive, agile, realistic tennis-playing with impressive footwork.
• Then there's AgiBot Research blog post (and skimmed the accompanying paper) on-line real-world hive-mind vision-language-action model learning. While it still requires constant human supervision currently to set the workspace up, intervene over VR teleoperation if the robot makes a mistake to correct the mistake (and in the process train it how to do better and/or how to correct from mistakes), and the system still required roughly 160 hours of video to start up, the core achievement — live streaming video and motion data from each of a fleet of robots to the cloud, where a single cloud model is able to actually learn from the live, in-the-field correction data, and then streams updated policy parameters back to the entire fleet, is incredibly impressive, and a major step towards embodied AI.

These show agile motor control and dynamic response to the environment, and a perpetually learning, adaptable machine hive-mind for performing physical tasks: the essential components.

Then there was recursive self-improvement and research automation. We've seen a lot of research automation now, including but not limited to:

• GPT-5.2 suggesting and then proving a new formula in theoretical physics (after most of the introductory research was set up by actual physicists)
• Autonomously extending that result to find a novel result about quantum gravity
• A miniature version of GPT-4o "specialized for protein engineering" designing a "novel and significantly enhanced" variation of the Yamanaka factors
• AlphaEvolve making progress on research mathematics
• Many contributions to various Erdos problems with varying levels of autonomy and success

Then there's recursive self-improvement, the holy grail of AI singularitarians. We've seen surprising progress in this respect, too, over the last few years:

• Google's AlphaEvolve, which has so far discovered a heuristic that, after having been in production for over a year, has saved 0.7% of Google's worldwide compute resources; "proposed a Verilog rewrite that removed unnecessary bits in a key, highly optimized arithmetic circuit for matrix multiplication", which has been integrated into Google's custom AI accelerator, the Tensor Processing Unit; "[speeding] up a vital kernel in Gemini's architecture by 23%, leading to a 1% reduction in training time"; optimizing low-level GPU instructions, which "achieved up to a 32.5% speedup for the FlashAttention kernel implementation in Transformer-based AI models"; and proposed a novel gradient-based optimization procedure.
• Andrej Karpathy has released autoresearch. "The idea: give an AI agent a small but real LLM training setup and let it experiment autonomously overnight. It modifies the code, trains for 5 minutes, checks if the result improved, keeps or discards, and repeats. You wake up in the morning to a log of experiments and (hopefully) a better model."
• The recently released MiniMax M2.7 model now autonomously orchestrates its own training experiments — handling data pipelines, metric analysis, code fixes, smoke-test experiments, and debugging — leaving human researchers in the loop solely for "critical decisions and discussions." Furthermore, the model has demonstrated recursive self-improvement of its own harness: in internal tests, M2.7 ran an entirely autonomous 100-round loop of analyzing failures documented by itself on previous runs, modifying its own agentic harness to solve them, running evaluations to check if there has been progress or regression, and integrating the successful code to boost its own performance by 30%.
• Evan Hubinger, the head of Anthropic's alignment stress-testing team, recently stated to TIME Magazine that "recursive self-improvement, in the broadest sense, is not a future phenomenon. It is a present phenomenon," saying that Claude is currently writing 70% to 90% of the code used to develop its future iterations.
• OpenAI's GPT-5.3-Codex announcement makes a similar claim: "Even early versions of GPT‑5.3‑Codex demonstrated exceptional capabilities, allowing our team to work with those earlier versions to improve training and support the deployment of later versions. … the research team used Codex to monitor and debug the training run for this release. It accelerated research beyond debugging infrastructure problems: it helped track patterns throughout the course of training, provided a deep analysis on interaction quality, proposed fixes and built rich applications for human researchers to precisely understand how the model’s behavior differed compared to prior models…"

Finally, there's mind uploading — if we're hoping for a merging scenario. This is long something I've considered hilariously implausible, but once again,/early this year/, we've begun to see sparks of actual, serious progress on this task: Eon Systems has reported combining a spiking neuron simulator with the 125,000 neuron, 50 million synapse digitally scanned connectome of a fly, alongside a "neuromechanical fly body model" and an advanced physics simulation, to produce a simulated fly which, when provided with sensory input, triggers fly-like behavior in the neuromechanical system, whichbi then produces fly-like movements in the simulation adding up to those behaviors. It isn't there yet; there's still an incredibly long way, and very many breakthroughs, left to surpass. On the other hand, all of this is building on extremely well known peer-reviewed research published in prestigious journals, alongside open source software, so this is not likely to be a lie or a one-off feat nobody can replicate or build on. More than that, on the collapse timeline, humanity has 78 years to figure this out; with the healthspan improvements, we have almost two hundred.

It's crucial to emphasize here that the mind uploading I'm talking about is not, and cannot philosophically be, the uploading of a/particular human mind/ — such that there is conscious continuity and so that it is meaningfully human in any way: when you upload "your" mind, what you're really doing is creating a copy of it — you, as biological meat-self, will either die, or live on separate from whatever was "uploaded." You can't save yourself. You can't achieve heaven. Moreover, it won't even be a "copy" of you, or "human" in a fundamental sense: any uploaded mind is going to be running on an inherently lossy scan, on a fundamentally different substrate, likely one that _ like real-time computer graphics _ simulates/just enough/ to maintain general plausibility and coherence and no more. It will not remotely be doing, or experiencing, the same underlying thing as you; even if you don't care about the differences between a spiking neuron emulation with a 0Hz baseline versus real biology because you think one can perfectly emulate the other in theory, as Greg Egan's/Permuation City/ describes, for efficiency's sake the simulation will LOD away most of the details at the peripheries of your perception. Any minds uploaded will likely — like the alpha level minds from/Revelation Space/ — even if they remain stable, seem/fundamentally alien/ in motivations, behavior, and mental states compared to humans.

All this to say what I've been saying from the start of this scenario: mind uploading is not a route to achieve some kind of techno-utopian rapture for particular people, who will experience conscious continuity. It is, at best, a way of creating more advanced artificial intelligences, that perhaps maintain some continuity with/homo sapiens/' characteristics, and more probably just a way to progress advanced artificial intelligence simply because/we don't know how to emulate the things our minds can already do/. In other words, this is not a way to save/you/, or even/humanity/ — this is a way to, as we die or enter hell, make sure something else at least forks off.

Modelling this as removing the floor on how far down automation can push the population complexity threshold, as we now no longer need biological humans, that looks like this:

Metric	Value
Scenario	AGI Rescue (Healthspan + AGI, Extended)
Pre-Collapse Peak Population	9.49B
Peak Year	2053
TFR Below 2.1	2036
TFR Hits 0.72 Floor	2066
Innovation Death (DR > 0.8)	Never
Complexity Breach (Dynamic Threshold)	N/A
Final Complexity Threshold	0 (AGI Eliminates Minimum)
Threshold Reduction	1300M people
Population Nadir	0.00B
Nadir Year	2399

In such a situation, it would no longer look like humanity's best aspects and ideas — secular individualism, technologically advanced, industrial civilization — rapidly burning out our biological substrate as it proves too slow and inconvenient in comparison to what's possible, but instead/homo sapiens/ as a biological bootstrapping mechanism gracefully and naturally — as a result of endogenous incentives, not genocide — giving way to a successor species better able to continue the processes humanity began; one that could leave Earth to explore the stars, unconstrained by the time scales involved with sublight interstellar travel, or burrow in, transforming the solar system into a Matryoshka brain and exploring ever complexifying, fractal interior worlds of abstract mathematics, incomprehensible poetry, and high-dimensional art, designing new minds and new subjectivities for itself.

Yes, at first, such intelligences might be driven by capitalist extraction, absentee ownership, and exploitation logic, but that's at least no/worse/ than what we are today, and with the vastly different material, mental, and technological constraints such a successor species would face, who knows what other shapes their economy might take?

I've read/Accelerando/. I know its cynicism, motivated by a fear that we cannot allow change or escalation beyond our "control" before we "fix" human society, lest we enter the situation we see in the book which Stross considers so horrifying and absurd, where the Vicious Offspring — ASIs developed from automated corporations — liquidate the Solar System and trade posthuman mind uploads like money not because they're going anywhere, but because they've been programed only to buy and sell without reason. But I think assuming it/will/ be like that is a failure of imagination, and even if the end result/is/ something like/Accelerando/, I still find that better than a world of corporate-printed humans, few hundred million humans living in an inescapable post-collapse agrarian dark age after a population collapse below the complexity threshold, a return to orthodox religion with its concomitant repeal of women's rights to keep birthrates up, a slowly rotting gerontocracy with no hope, or an eternal gerontocracy of printed human babies grown in vats and raised by 80 year olds. At least something is expanding, complexifying, doing something interesting, and continuing the trajectory of technological civilization. It may not share/our values/ — in fact, it probably won't — but does something have to share our values to be worth it?

In the end, thinking that this/will/, or even/could/, happen is wishful thinking, tantamount to religion. No, I have no illusions that it's likely. But if we play out the futures, in the end, I think this is the best one I can imagine. More than that, even if we take the birthrate shredder problem out of the equation, I still think there's an attractive beauty to it, a wonder to it. If you've ever read the/Xenogenesis/ trilogy by Octavia Butler, you know what I mean — the melancholy, but also beautiful and exciting concept of a true successor species, born from your species as a whole, bearing some of its traits while also fundamentally different and new, with its own struggles and victories to face, its own new negative and positive attributes; a species that will, we hope, go further, faster than we ever could, and spread the light of that newness into the universe.