Above: The Great Day of His Wrath. (John Martin)

(see Part 1 and Part 2)

The end of white-collar work and the new job scene

By the late 2020s, office jobs in developed countries are now basically about overseeing and providing direction to AI systems, and the last part of that is mostly on-paper rather than in practice. There is lots of talk about values and missions and the future, and a lot of unspoken communication about office politics and status. Many office workers don’t do much at all. Concretely, they might get to work, have a team standup, check in on how the AIs are doing, have some ritualistic meetings with their manager and any employees they have, and rubber-stamp some AI decisions that they’re contractually or legally obliged to stamp, with this adding up to only a few hours. Occasionally they might decide to change some goal the AIs have been given, but that requires just speaking or typing a paragraph. Many people feel guilty about this, but it’s mostly a quiet guilt. They fill their time with office chat or scrolling on their phones. Many companies become more social and more about community. HR has never been more influential. Everything’s both more cuddly and/or more viciously political now that the ugly raw realities of individual competence don't matter any more.

Some organisations try to fire lots of people. Sometimes it goes well. Sometimes it goes badly, and they realise that some human somewhere was holding some knowledge in their head, or nudging the mission in the right direction, in a way that was essential. However, by then it’s too late, and it’s hard to say which person it actually was anyway. Among the more ruthless or tech-adjacent management cultures, there’s a lot of talk about figuring out what the load-bearing humans in any organisation are, and how this is surprisingly difficult to do at a large organisation. Some companies develop internal AI systems to try to figure this out (or buy such systems from startups), but they need to collect some data about the functioning of the org first, which takes time. Also, the workers are incentivised to resist and fight back in a thousand subtle ways, and they do. Also, sometimes when an org tries to fire a lot of people, an online mob emerges to hate on them, influencers pile in and create 13 different cinematic universes where the theme is all how Company X is the pinnacle of all human evil, sometimes a former employee creates an AI-powered revenge cult (several assassinations happen as a result from the more violent of the cults), and sometimes politicians pick up the issue. The companies, largely, were profitable before, and are more profitable now that they’ve enjoyed a few years’ of revenue growth without expanding headcount. Therefore, mass firing is surprisingly rarely worth it, even though it would in principle be possible. A few firms facing crises or with especially effective or risk-tolerant leadership buck these trends and aggressively slash costs by cutting huge amounts of human workers.

What developed country firms are not doing is hiring new workers or replacing anyone who retires. What they are doing is replacing any foreign contractors or service providers with cheaper AI ones.

This creates several groups of disaffected people. First, the youth in developed countries, who have much worse job prospects than the preceding generation. For people looking for their first job in 2030-2031 in a developed country, the options are roughly:

1. Working in services where being human intrinsically matters (elderly care, retail, restaurants, hospitality, teaching, etc.). Healthcare is by far the most prestigious one and what many aim for (even though doctors—or at least all the good ones—defer all diagnosis and other intellectual work to the AIs). The cartel-like nature of medical licensing bodies, strain on state budgets, and the fact that most of the actual work is done by AIs means that the number of doctors or nurses hasn't increased much, though, so entry has become even more competitive. Policing and primary education also continue hiring humans at scale.

2. Jobs that are effectively sinecures. This includes many positions in government and civil service. In the EU, regulation passed in 2028 means that many companies are forced to hire human "overseers" to key positions. Of course, the supply of sinecures is set by regulation and funding for economically useless activities. Competition for such positions is therefore extremely harsh, and (because the selection criteria, having no reason to be one thing rather than another, inherit the latest credentialist instantiation of the 21st century West's bureaucratic blob) requires extreme conformism. This category has a fuzzy boundary with the first, depending on whether you value the ceremonial human touch as a key part of the service or not.

3. A particular example of this is the law. Lawyers have two major advantages. First, their job deals closely with important social questions of legitimacy and propriety, making it a natural claim that something fundamental would be lost if the human presence was gone. The presence of lawyers evolves to be more ceremonial and symbolic—almost religious—but it stays. Second, lawyers make up a lot of the rules for themselves, and interpret the statutes for everyone else. Third, a lot of politicians are lawyers, or have friends who are lawyers, which make them attuned to lawyer interests. This gives lawyers a lot of leeway in what automation they allow. In many countries the rules are bent such that it is flat-out illegal to consult an AI on legal matters; you have to go through a human lawyer. AI companies are forced to train their AIs to comply with this ("I'm sorry, but as an AI it is illegal for me to give advice on legal matters, so I recommend you hire a licensed lawyer"). Of course, all of the actual legal research and argumentation is done by AIs—the lawyers just monopolise the position of being allowed to ask them.

4. Manufacturing, which is booming, especially as productivity has been lifted by AI management and oversight. Many manufacturing jobs involve wearing an earpiece through which you receive detailed step-by-step orders from an AI (and occasionally AR glasses that can show you diagrams or an overlay for how to move your hands). A large fraction of people go into this, even if they have prestigious university degrees (many of the prestigious degree-holders do not have their salary and status expectations met and become resentful).

5. Academia. There are still humans in academia who somewhat matter for intellectual progress, but they're all either experienced humans with years of research taste in economically-valuable non-purely-mathematical areas (who are actually in decently high demand, as the AI labs chase feedback sources that will help them faster and cheaper get the models superhuman at even the very last set of very long-horizon, hard-to-measure skills), or (especially in the US) "prof-luencers" who use the status of a successful prior academic career to boost their influencer careers. New entrants to academia get their academic salaries (if they win an ever more cut-throat competition), but not the hope of actually mattering for intellectual progress. Some derive satisfaction that they can at least keep deep human expertise alive into the future—though it seems like without any ground-truth feedback signal, many lineages of human expertise will become dead knowledge within a generation even if people still go through the motions of “learning” them.

6. Becoming an influencer. Works for some, but the competition is extremely tough (though it does help that "being verifiably human" is in vogue).

7. Becoming a socialite. The infinite variety and competition (human and AI) in the digital world is driving a resurgence of an in-person social scene. However, for this to be a "career choice", you must either already be wealthy, or have some other factor in your favour. The overwhelmingly most common such factor is being a young woman who inserts herself into the social scene of moneyed men.

8. Becoming a musician, artist, or poet. The main constraint is funding, of which there are two important types: government subsidies (which generally increase, as being vaguely pro human self-expression is a common government answer to what people should do with themselves in the age of AI, especially in socially progressive European countries), and wealthy patrons. Being an artist for the latter often melds into being a socialite, since in-person local artists are the prevailing fashion. Many nouveau rich techies, wanting to erase their association with the now-uncool world of software, throw money at artists who live in their local community to do some arts-and-crafts thing and then show up with it to their party and say vaguely artsy things.

9. Going into politics. This has become more appealing to the young in particular, since the future looks uncertain and the youth are the ones who expect to live in it longest. There are many AI youth activists (in the sense of specialising in the topic of AI, being AIs, or both), who try to use their position to advance youth interests. The problem is that they don't have concrete policy asks beyond "allocate more money to us", which puts them at odds with every other interest group in society, many of which (e.g. the retired) outnumber them in raw numbers as a voting bloc as well as in terms of resources, power, and influence.

Culturally, intellectualism is out, after having climbed in cultural status for two centuries before the mid-2020s as labour-augmenting technology and globalisation scaled its power. Charisma, conformism, sociability, authenticity, and propriety are all in. In the early 2030s, the US is becoming more European in its attitudes, especially along the above dimensions. While the high-water of conformism that the 2010s culture wars and academic sclerosis had caused receded throughout 2022-2027, a modified and less-political, more European-style propriety-focused conformism rose around 2030. First, this was driven by cultural changes downstream of AI reducing the rewards for risk and entrepreneurship. Second, there was a cultural backlash against the techies (who were seen as having pretensions of importance after the automation of software in the late 2020s, and as lead by a bastion of improper disruptive moguls who were on the wrong side of a Republican power struggle in 2028), and through the techies the culture of ambition that had become central to their self-narrative.

A second disaffected group is the developing countries. Replacing outsourced foreign human services (e.g. call centres) with AIs is a cost-saving that can be done without political or social repercussions, so all companies did it—often at significant scale as early as 2025-2026 for text-only tasks. As a result, services-led export growth is dead. This is bad for India and the Philippines in particular. India just about achieved the economic heft where it could've been relevant in AI, but throughout the 2020s was unable to become entrenched in any part of the AI supply chain. At the same time, as more of the population in developed countries goes back to working in manufacturing, political demands for protecting developed country manufacturers from competition with developing countries grows. This leads to even more tariffs, on top of the already-existing late 2020s trend towards more and more tit-for-tat tariff escalation. This makes developing country growth based on goods exports harder. The biggest shock to existing goods export industries won’t arrive until a few years later when the robots show up, but investment into the developing world already dries up, as US productivity growth rises and is expected to rise even more.

China is about a year behind in leading AI tech but about 2-3 years behind in AI diffusion. The Chinese public and the CCP are watching the coming wave of AI job automation with worry, especially as there is a big cultural emphasis on exactly the types of academic skills that are getting obsoleted quickly. The CCP is very worried about stability. More and more people are joining the Party, since they see that other opportunities for social advancement are ending. The Party is making many more roles in it available, and using this carrot to incentivise people to adhere to Party principles even more strongly. The AI surveillance state keeps expanding; there is now AI interpretation of much CCTV footage of public streets, for example. There are efforts underway to modernise (i.e., convert to AI) most of the military, such that the Party's control cannot be threatened even if the human military is destabilised.

In the US, the leading plan seems to be a hodgepodge of regulation-mandated human job roles, and eventually maybe UBI. However, a fiscal crisis is on the horizon because of the looming social security trust fund exhaustion. GDP growth in 2031-2032 is hitting 5% per year but full UBI still seems expensive. In the EU, there is more state intervention and regulation aimed at keeping humans in the loop, with massive corporate and government hierarchies of jobs that are effectively pure sinecures where the work is all done by AIs, which is temporarily reducing the demand for flat-out UBI.

When governments ask companies what their blockers are, companies cite regulations that keep humans in the loop, and (when off the record) everyone shares a sentiment that humans aren’t actually in the loop anyway. Shortcuts are already being taken to reduce the human oversight component. It’s very hard to do this legally, because there are often government-mandated AIs monitoring compliance with the human oversight rules. Two firms might want to maintain their human workers for complicated regulatory and office politics and inertia reasons, but they’re competing against each other, and against full-AI firms, and against foreign adversaries. So pressure increases to cut the unnecessary weight. There’s also a race to the bottom internationally. Many autonomous AI-run companies in 2030-2033 move to less-regulated areas, take the slight hit of running on open-source models, and serve customers from there. However, this global decentralisation is reversed once the robotic revolution—subsidised and encouraged by the American and Chinese governments—gets under way.

Lab strategy amid superintelligence and robotics

The state of AI capabilities around 2030 is roughly as follows: wherever there is an easy feedback signal and a high performance ceiling, such as maths or code, the models are incomprehensibly superhuman. Where rapid iteration is possible but the performance ceiling is not as high, like having sales calls, the AIs are better than all humans. In general, the AIs can be more charismatic and persuasive than humans, but this does not give them superpowers over steering individual humans as they like, especially when to do so they would have to compete with every other memetic force in society as well as the individual's resistance to being psychologically hacked. Wherever there is a large pile of information, such as supply chain routing or crystal structure prediction or history or legal precedent, the AIs are superhuman at spotting and understanding the patterns and generalising them to new instances. However, models appear to still be roughly human-level at long-horizon tasks with ambiguous success metrics. Companies, governments, and research agendas—even the scrappier, faster-changing ones—are still piloted by humans who make real strategy decisions, even though in practice it's a human riding on a vast wave of AI supercognition, and the trend is towards more and more delegation as the systems improve. Real-world progress in hard tech is also varied. There are many breakthroughs in parts of materials science and molecular biology driven by things like material property and protein folding prediction that cuts down on empirical iteration. However, other tasks turn out to be computationally intractable even to the smart AIs, even if they often achieve very large efficiency gains over the human state-of-the-art by inventing superhumanly good heuristics. No one has figured out how to turn the vast amounts of intelligence-on-tap into magical-seeming technical progress in atoms, even though engineering work now happens much faster and at a higher quality level and with less margin between practical and theoretical performance.

In 2029, OpenAI rebrands its models to just “o”. Everyone has Opinions. It’s a big advance in raw intelligence, but almost no one can tell. Instead of a variety of sizes of an o-series model with updates every few months, from now on there will be a few varieties (differing mainly in size, like o-small and o-large and an internal-only o-huge, but also with some specialised finetuned models, e.g. o-math and o-chat). Individual instances of the models can use their medium-term memory as context when they’re doing agentic tasks, but they can also run in “functional” or “API” mode where that is disabled. More than half of OpenAI’s model revenues still come from functional mode calls rather than running instances as agents that develop their own memories and learn on the fly, but this proportion is steadily falling. There’s a new model checkpoint released every day, with the newest information from that day already in its weights, and the occasional larger improvement.

By 2030, OpenAI has culled almost all of its human employees. This is the main advantage of their latest model internally—the tacit internal knowledge that various humans previously had that would’ve made the human-level-ish o6 not quite adequate at wholesale replacement of OpenAI engineers matters less when o-huge can just rederive the tacit knowledge from scratch very quickly.

OpenAI's b-series human robots reach annualised shipment volumes of 1M/year in late 2031, which gives it about 50% market share in the total domestic robot servant market. Several million other general-purpose robots (e.g. for use in manufacturing) are also being sold by 2031.

OpenAI is seen by some as a slightly shambolic conglomerate, like an Oracle or IBM or Microsoft, and by others as the original and one true AI company that is destined to be >50% of world GDP.

The robotics sector is split between special-purpose robots with modern AI integrations, e.g. window-cleaning robots and pipe-crawling repair robots and delivery drones, and general-purpose robots being pursued by OpenAI and several other companies (including a struggling Franco-German startup that is kept afloat by the EU being hell-bent on endlessly subsidising it until Europe finally has a big tech company—the European Commission is confused why this is not producing results). Both paths seem technically feasible. However, the general-purpose robotic players are the better-resourced ones, and are run by people whose main past reference point was the generative AI wave, and therefore they are philosophically big believers in scaling laws, so they are betting on collecting all the robotics data as the path to improving quality, and on Wright's law to bring down hardware costs as they build more and more of the same thing.

All of this is also happening at unprecedented efficiency and speed compared to prior research efforts, since there are superintelligent STEM AIs around inventing algorithms that massively bring down the sample complexity of the robotics control algorithms, organising the assembly lines, doing the CAD work, and so on. However, the actual learning to move part is still a machine learning problem bottlenecked by data, and there is no magic wand that can instantly create massive robot factories from scratch (especially given the raw resources required). The output scaling curve looks to be roughly a 4x increase of robotics capacity per year, though. This is expected to rise for 2033-2035, as the robots automate more and more of the robot production pipeline, but bottlenecks abound, and energy and land constraints (mostly downstream of regulation) are harsh.

Anthropic works with a bunch of Western governments and NGOs on strict KYC for agentic model customers—the standards have so far been somewhat shoestring, the coming robot wave is making the need much clearer, and there was a big scandal last year with a heavily AI-aided chemical terrorist attack. The cyber situation has calmed down though, with defense dominating, as key code is now either provably correct or so thoroughly tested by countless AI systems that it's close enough. Biological capabilities have already been artificially kept down by most of the key model players (including open-source and Chinese ones). Taking any large-scale actions with models that aren't from the dark web in the West and China, especially in wet lab virology or DNA synthesis, requires specific access permissions from the labs through government-mandated schemes. However, by 2030 there are open-source dark web models that will do whatever you want including designing candidate pandemic agents that are unnaturally lethal and virulent, and there is no quick way to pandemic-proof the world against bioterrorism. The remaining difficulty of wet lab work, the low number of totally insane actors, and AI surveillance are the main forces keeping the per-year odds not too high, but civilisation is clearly running a big risk. The national security apparatus in both the US and China is more relaxed about this threat than it would otherwise be, because the military and economy are both increasingly robotic and so it’s not a threat to the regime even if most of the population drops dead from mega-flu. For example, the US war plans in event of a devastating pandemic (or nuclear) attack now include AIs substituting for any of the critical industry CEOs or defense staff that die.

Another big Anthropic effort is AI for biology. They want to cure cancer, make humans live forever, etc. A major internal faction also wants to pursue human intelligence augmentation but leadership fears this would be too controversial to discuss in public, so they just have a single secret team working with the CIA on it. Innovation in biotech has definitely risen, since designing promising drug candidates is ridiculously fast and cheap, but the bottleneck even before the AI revolution was less the design part and more clinical trial regulation. Anthropic is curating datasets, acquiring laboratory automation startups, and working with regulators to cut down red tape. This will take years to bear fruit, but seems to be leading towards a biotech revolution over the next decade.

Anthropic is also trying to use biotechnology to bootstrap powerful nanotechnology. However, the company’s attempts to get their AIs to do the physics and engineering hit some snags, especially as they lack xAI’s or GDM’s specialisations in physics/maths/engineering (having trusted more in domain-general intelligence). Still, it is the AI era, so the AIs can fairly quickly get up to speed on this stuff, and the Pentagon is helping.

Towards the automated robot economy

In 2033, about 40 million humanoid robots are shipped. An increasing fraction is going to industrial uses. Costs have come down to that of a cheap car and are declining further, especially as the entire manufacturing process can now be done by the robots themselves in the most advanced factories. This also means that full AI control and real-time optimisation of the entire robot manufacturing line is possible, leading to unparalleled factory output growth and ease of iterating on the design.

As a result, over 2032-2034 there's a Cambrian explosion of robot diversity into non-humanoid form factors. By 2035, a large fraction of developed country consumers have household robots performing almost all manual tasks at home. Construction work, assembly line work, agricultural work, solar panel installation, plumbing work, industrial machinery repairs, and electrical utility jobs can all in principle be done fully by robots by 2034. The main constraint is energy and resources for the physical manufacturing of the robots—as well as land and regulations.

By 2034-2035, advances in nanotech are also arriving. Rather than a single magical-seeming assembler, the nanotech advances are mostly in medical areas (such as targeted drug delivery to specific locations within the body, which is a huge boost to cancer treatment, and early prototypes of cellular repair machines), and in materials science advances that allow for stronger and lighter and self-healing materials, and better batteries. These can all be used in robots; some look supernaturally strong and capable to humans. The manufacturing robots also get "magic fingers", where the tip of a robot appendage is a surface that can do very controlled and fine-grained precision welding, polymer (un)curing, deposition of substances, and catalysis of chemical reactions..

The 40 million humanoid robots shipped worldwide in 2033 do roughly the work of 80 million human workers since they can work longer hours than humans. In 2034, there are 240 million human-worker-equivalents of robotic capacity shipped, and in 2035 about 1.1 billion human-worker-equivalents.

Politically, this is as if hundreds of millions of extremely talented immigrants who accept below-minimum-wage jobs had suddenly sprouted out from the ground, in each of the developed countries and China. Years of upheaval in white-collar work have given politicians and activists experience in dealing with such things, and they are better prepared.

In America, the Republicans narrowly keep the White House in 2032. The Democrats ran on an attempt to solve rising unemployment through European-style human-in-the-loop laws, including an expansion of "pro-social, meaning-creating" human roles in the government bureaucracy, education, and the lawyer cartel, while having a major retraining initiative for blue-collar workers threatened by robotics. In the few months before the election, there was a burst of about a hundred thousand people losing their jobs very directly to robots. A run of impressive robotics demos fermented hysterical online influencer coverage and blue-collar job fears. The retraining initiatives for blue-collar workers became seen as insufficient and out-of-touch with the "average American" who does not want to be reeducated into performing some ceremonial role in a bureaucracy whose culture they don't agree with.

The Republicans counter this with the PROSPER Act (Promoting Robot Ownership and Small-business Prosperity through Economic Restructuring), which they campaign on and pass in 2033. This creates a car dealership -like model for robot ownership, where robotics companies are not allowed to sell “consumer robotics services” directly to consumers (sectors like defense and mining are exempt). “Ordinary Americans" can apply for loans to start their own robotic services business. Also, a license is required to sell consumer robotics services in a given territory, and a given legal entity can only operate in one territory. The territories default to state legislative districts, most of which are between 30k-150k in population, but states are allowed to change the territory unit. Licenses for a territory are granted at the local level. For example, Joe Smith in Prescott, Arizona might get a government loan, buy 10 plumbing robots, and sell their services to other Prescott residents. He himself doesn't do much, since the robots do the plumbing and the AI does the planning, logistics, accounting, and so on for him. But nominally, he is now a small-business owner, and is most definitely not a welfare recipient freeloading on Uncle Sam.

If any robotics licensing territory gets too much competition in a single robotics services vertical, competition drives margins to zero. There is also little that differentiates the different robotics service providers. Therefore, an instant race begins for regulatory capture of each robotics license territory, which is often won by whichever actor had the most networks and funds at the beginning (though anti-trust prevents full monopolies, so there's almost always at least 2 service providers). Much of the market share fluctuation becomes about social networks and persuasion. The savvy robotics license owners in particular try to manipulate local cultural currents to restrict the granting of licenses to new entrants. Alternatively, the leader of a local AI-powered personality cult will just declare who deserves the licenses. Even with the robotics licensing regime, though, only a small fraction of the population is owners of economically-relevant assets. Social and economic life increasingly revolves around the few families with control over income-generating assets (whether sinecures or robotics licenses or property or stocks). Marriage into such families gradually becomes a more and more common tool of socioeconomic ambition. Many give up on earning an income at all, and make ends meet by moving to areas with ridiculously cheap property.

Above all local scenes are the true US national elite—powerful politicians, billionaires, senior government advisors, and some others. On average they still feel some noblesse oblige towards the lower classes, though in the late 2030s this is waning as they start feeling in their bones that their position of power is not dependent on the people anymore. However, their main preoccupation is status competition with others on their level. Many of these are inter-elite disputes with little bearing for the world, but on net there is also a strong desire to compete with China. In particular, the narrative that the race through the robotics buildout will be decisive for the far-flung future of humanity gained a lot of prominence through the late 2020s and early 2030s. This creates a strong elite consensus that competition with China must be won, and that the way to do so is to stabilise the domestic situation, but then otherwise let the robotics wave rip. The plans for domestic semiconductor self-sufficiency are on track to come true only a bit behind schedule in 2034. Actually-working ICBM defense, designed by superhuman engineering AIs around 2030, is fully online and working by 2034 thanks to the speed of manufacturing scaleups in the age of robotics. The military is able to field hundreds of millions of small drones and millions of robot soldiers. Pentagon projects on nanotechnology and other exotic physics applications may bring about powerful new technologies within another few years.

China, of course, also sees the need to win, especially as its lead in industrial robotics vanishes when America’s robotics revolution happens a bit before China’s. The CCP is also decoupling its treatment of the human economy from its treatment of geopolitics and the "real" robotic economy. In 2034, the CCP declares that citizens need to "eat bitterness", in the form of accepting per-capita living standards stagnating for a while (at around $37k, PPP-adjusted, in 2025 dollars) while the state diverts resources to fueling the robotic revolution to avoid losing in the geopolitical competition.

In the EU, AI diffusion has been slower due to regulatory hurdles, but the extinction of white-collar work is still well underway, and the robotics wave is coming only a few years after the US and China. However, this delay is enough to make the EU geopolitically irrelevant. The greatest external threat to the EU is Russia, which has suddenly gotten much richer as Chinese companies effectively colonize Siberia to mine resources to fuel China's robotic buildup while paying large rents to the Russian government. The US lead at the robotics revolution also drains manufacturing jobs out of the EU, until EU countries are politically forced to shut off trade (though a political movement, active especially in Eastern Europe, would've wanted to negotiate a stronger US security presence in exchange for letting trade continue and domestic industries wither). Various proposals for UBI float around, but economic turmoil makes the prospect of funding it uncertain, and the political fight by special interest groups for privileges for their group in particular is extremely fierce and they are all opposed to UBI for everyone. By 2036, functionally everyone within the EU has some kind of regular state payout they live on, not through a single system but through an extremely complicated patronage network (that non-AI-aided humans literally could not understand) where the average person is eking out a living in exchange for taking part in complicated cultural rites and bureaucracies.

The developing world suffers. Already, manufacturing jobs were lost in the global south—developed country workers streamed from services to manufacturing, while having their productivity boosted by AI that developing countries can't afford, and while their politics became even more captured by blue-collar job worries that drove tariffs and trade restrictions. Now, US and Chinese robots can manufacture anything better and more cheaply than any human. There are large capital flows out of developing countries to the US and China as they buy robots. However, in most developing countries even the arrival of cheap robots does not lead to prosperity, as the robots mostly go to the elite and the state, which have no reason to share the windfall with the people—especially as cheap military drones and robots, and omnipresent AI surveillance, have effectively removed the threat of rebellion or coup. India, Bangladesh, and Brazil shut off almost all cross-border trade and declare themselves "human-only" countries, where any sort of neural network or robot is banned. They receive many immigrants from developed countries who have struggled to cope with the AI wave. In the most totalitarian states, the outcomes are mostly tragic. North Korea lets a large fraction of its population starve to death and forcibly sterilises the rest, except for about 10k senior government officials who continue to preside over an AI economy and robot military (some worry that the CCP allows this, not just for geopolitical reasons where they want a military bastion pointed at South Korea and Japan, but also as a test-run of whether they could later pull off the same thing within China). In some other countries, the population is kept fed, but subject to constant surveillance. Rulers realise the population is no threat anymore; the “intelligence curse” is like the resource curse but stronger. The most psychopathic subject their populations to arbitrary cruelties for amusement, as robot bodyguard -protected members of the ruling dynasty travel around their dominion having parties that include orgies of rape and murder of civilians.

Some of the most morally outrageous events lead to condemnation from the superpowers.

After the North Korea debacle, the human members of the CCP have an internal meeting to decide a set of criteria by which the CCP will rule. After an inter-party power-struggle, the CCP commits to the perpetual existence of at least one billion Han Chinese people with biological reproductive freedom, organised into family units, with a welfare level at least around what $40k/year consumption in a 2025 developed country would give, and with eternal strict CCP control over national ideology, culture, and strategy. They impose fewer constraints on the rulers of their client states than they do on themselves, but generally oppose genocide, forced sterilisation, mass starvation, and deliberate cultural erasure. The CCP line on this does in fact constrain and improve some authoritarian states (and they pressure several dictators into stepping down and being replaced by non-psychopaths), though they still allow some horrific practices, intrusive mass surveillance, political cleansings, continued extreme poverty, and states indirectly driving down the birth rate (which many governments want to do, since humans are mostly just a net cost to the government by this point).

In the US, some moral atrocities in Venezuela in 2036 lead to public outrage and political pressure for action. The president is informed that given the technological disparity, regime change is a press of the button. The button is pressed, and the regime changes. Several more countries follow in quick succession.

By the end of 2037, most of the world can be split into:

10. The US (which now includes both Canada and Greenland; both joined voluntarily, as American citizenship has become extremely in-demand due to the privileges it confers).

11. US client states. The terms of admission here are usually that the other country must accept trade with the US, which generally means that the country's own industries go extinct as US robotics and AI performs all work. In exchange, a combination of the US government and American elites buy out the assets in that country. In particular, any resources—or land containing resources—are bought out, and mined by US companies to fuel the continuing robotics build-out. The money paid out for these resources and assets is generally the endowment that the government and people of the client state then live off. Generally, the client states create sovereign wealth funds to manage this endowment, and live off the returns to it, which are distributed within the country according to local politics. These countries are all poorer than the US, and with essentially no future growth prospects that aren't praying for the continued US robotics buildout to increase the fraction of their endowment invested in US stocks (this is great at aligning their incentives with the US). However, where the countries had strong existing institutions (including where the US showed up and changed an unpopular regime) and at least some assets the US cared about, this still translates into comfortable living standards. US client states include the UK, the entire Americas except for Brazil, Japan, South Korea, Australia, Saudi Arabia, Israel, the Gulf States except Yemen, Thailand, Malaysia, the Philippines, and much of northern Africa (now almost entirely covered by solar panels). The EU is a borderline case, having negotiated an agreement that is Kafkaesque (in a very literal sense: it was crafted by superhuman AI lawyers, no human can understand it) but that allows it to retain some more power locally.

12. Human-only countries, in particular India, Bangladesh, and Brazil (though Brazil experiences some US pressures and is temporarily couped by the US, before this is partly reversed due to complicated US internal politics). All, however, have to solve national security somehow. Brazil allows US companies to mine in certain areas even as the native population is not allowed to use robots, in exchange for security guarantees. The Indian government grants itself exceptions to the human-only policies and scrambles to build a military robotic base, and develops exotic nanotech weapons that would be expensive to counter even by the more advanced US and Chinese forces. Bangladesh lasts until 2039, when both US and Chinese covert nanodrone operations start skirmishing within its territory, after which the government is overthrown and replaced with a Chinese AI.

13. Chinese client states. The most common model is propping up the government and selling robots, in exchange for the Chinese state-owned enterprises getting minerals and resources. Chinese client states include Russia, Belarus, the central Asian states, Pakistan, Myanmar, Cambodia, Laos, Vietnam, several Pacific island states, and most of Africa.

14. China.

Outside the Earth, Mars is being eaten up by both American and Chinese self-propagating robotics factories (the moon also has major bases on its poles but lacks carbon, nitrogen, and various metals, making it less valuable), which are on an exponential growth trajectory set to cover the entire planet by 2055, and already sending out probes to claim the other planets. By 2035, nuclear rocket propulsion technology has made it feasible to send payloads to Mars outside the once-every-two-years Hohmann transfer window, though at much higher cost per ton. With the original outer space treaty voided by clear land-grabs, a defunct UN, and political pressure in both the US and China to send something lightweight to Mars to gain an edge in the land-grab competition for space, both the US and China launch high-speed kinetic weapons at each other's (fully-automated, uninhabited) Mars facilities in 2037. While the kinetic weapons are still accelerating towards Mars, the AI diplomats reach an agreement that splits up the solar system between the US and China. The kinetic weapons turn off their fusion engines early, miss Mars, and shoot off into interstellar space. By 2038, they are further from the Earth than the Voyager probes, and therefore the furthest human-made objects.

In 2035, there were about 1 billion human-worker-equivalents of robot labour (though note that this number makes less sense over time, as the robots are doing qualitatively different labour and often technically-unprecedented things). In 2036, the growth rate slightly slows due to resource constraints, and the total grows to only about 3 billion. However, in 2037, the best estimate of this number hits 15 billion, then 90 billion in 2038, then 600 billion in 2039 and 4.5 trillion in 2040. 

By 2040, the value of the world’s manufacturing output is over a thousand times what it was in 2025. Most of this is spent on geopolitical competition, inter-elite status rivalries, and an increasing fraction on AI machinations with only the most tenuous link to any human activity, but which the humans who on-paper own all of this barely notice as it gets lost in the maelstrom of everything else. Even the most entrenched, long-term-oriented, and value-laden executive jobs are (whether de facto or de jure) entirely done by AIs, with very little human understanding of what is concretely happening on the ground. Human society and the human-to-human economy is a leaf riding on a vast wave of automated activity. 

The human condition in the 2030s

In the early 2030s, strange things are happening to the memetic landscape thanks to RL algorithms gradient-descenting in an endless loop of attention-competition against each other. Some countries shut off from the global internet and close their borders to try to maintain internal culture. The trend towards small, tight-knit communities of the late 2020s is back, after having retreated somewhat because of the addictiveness of optimised AI content slop. Culture everywhere is almost entirely AI-driven; the churn in ideas, trends, and fashions is mostly due to patterns of AIs reacting to AIs.

In the mid-2030s, socioeconomic advancement is almost extinct worldwide. Many people who might otherwise be ambitious retreat into virtual reality games that provide simulated achievement. Many ambitious young men move to countries too poor for omnipresent police drone surveillance (if they don’t already live in one) and turn to crime. Many ambitious young women see socialising as the only way to wealth and status; if they start without the backing of a prominent family or peer group, this often means sex work pandering to spoiled millionaires and billionaires.

The biotechnology revolution arrives in the late 2030s, even though it was long delayed by clinical trial regulations. Americans have reached longevity escape velocity. There is no disease that cannot be cured. Intelligence augmentation of four standard deviations in embryos and one in adults is technically feasible.

2040+

Why is this massive automated robot buildout happening? As discussed, the US and China both have the required geopolitical ambition—in particular, they cannot risk letting the other ride the robotics wave and get disempowered. Within countries, there are pressures from both the elite and from the needs of ordinary people. The elites compete against each other. Those who do not want to compete do not, and are rendered irrelevant, and replaced by ones that do. In addition to status within the elite community, the elites gain raw power from letting the robotics wave rip through society: there are many trillionaires in the world now, who can work unprecedented wonders with tens of millions of robots carrying out their bidding. They can build cities in a day, save millions of developing-world people from hunger, and prepare for their children to rule entire planets governed by their ideal political philosophy. At the same time, while Americans are almost all reasonably well-off, across the world there are still billions of people with a poor quality of life. The level of material wealth in the world has skyrocketed, but governments are also much less interested in investing in people. Funding for humans has become like the foreign aid budget: it exists, and is morally supported, but there is constant political downwards pressure on it since it does not further the needs of any powerful interest group. The best hope for human welfare seems to be accepting that governments will be hard-pressed to spend above 1% of their resources on humans, but relying on American and Chinese economic growth being so vast that a small trickle of resources from American and Chinese robotics companies will eventually be enough for material comfort for everyone.

This looks set to be true within a few years, though there are two complications. The first is that both spheres of influence (but far more the Chinese one) still tolerate some grotesque practices by client states. However, once the geopolitical balance is secure and sufficient wealth exists, and with some luck over choice of leaders, this state of affairs would likely end.

The second, more fundamental point, is that the economy has an inertia of its own. Humans make almost no meaningful decisions about the trajectory of the world, having handed the reins to AIs that make effectively all decisions, even if some of the AIs are technically only “advisors”. Eventually, the robotics revolution is less an economic phenomenon and more as a brute physical one: a chain reaction where certain loops close—metal to mining robots to more metal, say—and shoot off towards infinity. (This was already somewhat true of the human story before robotics and AI, except that the feedback loops intimately involved and benefited humans, and had slower doubling times.) 

Somewhere on the top of the stack there are still humans who on-paper own or control the assets and can make decisions (whether as a private actor or as a government overseeing autonomous AI companies operating in its territory), but they see numbers that track their wealth and power ticking up, so they have no reason to call a stop to it, and don’t understand it anymore. On some parts of the Earth, human institutions still hold and human societies exist, locked in place by AI bureaucracies that have taken on a life of their own and likely couldn't be dismantled even if the humans tried. On other parts of the Earth's surface—including big regions like the Sahara, the Australian outback, Antarctica, and Xinjiang—an ecosystem of AIs rules over vast masses of robotic machinery with no human involvement. Space, too, is now technologically within easy reach, now that sophisticated self-replicating robotics exists and wimpy chemical rockets have been superseded.

Who will get the stars? What is Earth’s long-run fate? In this timeline, at least, the technology to control the AIs' goals arrived in time. But this alone does not let you control the future. A thousand people go to a thousand AIs and say: do like so. The AIs obey, and it is done, but then the world responds: doing this leads to this much power, and doing that leads to that much power. In the vast sea of interactions, there are some patterns that strengthen themselves over time, and others that wind themselves down. Repeat enough times, each time giving to each actor what they sowed last time, and what emerges is not the sum of human wills—even if it is bent by it—but the solution to the equation: what propagates fastest? If the humans understood their world, and were still load-bearing participants in its ebbs of power, then perhaps the bending would be greater. But they aren't. And so, even surrounded by technical miracles, the majority of humans find themselves increasingly forsaken by the states they erected to defend themselves, standing powerless as they watch the heavens get eaten by machines.