The current simulations' lack of fidelity to real-world physics is hindering virtual training for a lot of manipulation tasks. An accurate but efficient simulation would produce abundant training data, allowing the realization of the ~full potential of existing robots.
But working on the simulation problem requires expertise in primarily programming and physics; jagged AI could contribute massively to progress here well before AGI comes.
I think this is very valuable to decompose manual jobs/tasks to see what we could already do with current robots with intelligent control. I think we could have big improvements in that control before AGI.
In this series so far we have investigated the consequences of automating labor with advanced AI systems and robotics. Part 1 found that physical growth would be unprecedentedly fast even if production technologies otherwise remain stagnant, and Parts 3 and 4 showed that even faster growth is possible with process and technological changes.
But all of this took for granted that labor could be automated. In this part I ask whether, given AGI, we would know what robots and machines to build in place of human labor. Manufacturing them is not the obstacle. In Part 2 I found that once we know what to build, making them at scale takes only a few years. The open question is whether we would know what to build at all, and I think that, given AGI, we probably would. Our "geniuses in a data center" could thus do more than sit around contemplating cancer cures and algebraic topology; they could also automate physical labor, once the necessary and not-too-hard to build actuators are constructed. That would kick off the kind of rapid physical growth discussed in previous posts.
An AGI that can do remote work can probably operate robots too
The term "artificial general intelligence" gets bandied about a lot, to the point it's at risk of becoming utterly useless. But there is, and always has been, a clear core concept that AGI points toward: an AI capable of doing the work humans can do. Authors have operationalized this in more precise ways for various purposes, but a system that can automate essentially all remote tasks is good enough for us. Astute readers may have noted that remote workers still exist (the author among them) and that current LLMs, smart as they may seem, cannot perform remote work (I've tried). AGI is also a much lower bar than superintelligence, which Bostrom (2014) defines as an intellect that "greatly exceeds the cognitive performance of humans in virtually all domains", or than Amodei's "powerful AI", which is "smarter than a Nobel Prize winner across most relevant fields".
Discussions of AGI focus on what the system can accomplish cognitively rather than on bodies or actuators. While sensible for isolating the core concept, this should not be confused with the claim that AGI cannot or will not act in the physical world. Humans can operate factories, drive cars, teleoperate robots, direct air traffic, perform surgery, and do numerous other "physical" tasks while being physically remote. If provided with suitable actuators, AGI would be similarly capable.
One could gerrymander these tasks out of the definition, so that AGI covers programming and paperwork but not machine operation. Such a system is not a logical impossibility, but I don't think it reflects the likely trajectory of technological progress. Remote cognitive work is hard. Doing all of it means mastering real-time control, spatial reasoning, physical prediction, and continual learning, and a system with those skills has what it needs to learn to operate most machines, much as a human handed the controls becomes passable within a few hours and proficient within months.
We can already see current LLMs developing rudimentary physical competency. They can provide high-level control over robots, watching the scene and choosing the next action, for example running a robot arm or flying a drone. [1] They do the same in analogous virtual settings, playing Pokémon slowly. And they show broader physical reasoning, from "vibe CADding" parts for 3D printing to programming robot dogs.
The most challenging case is work that requires dexterity and quick, fine-grained control of motion and force. It needs a fast control loop, since the large models supplying the high-level judgment are too slow to run it, and running that loop well takes a detailed understanding of how the actuator behaves. But neither is clearly a hard limit — a model running many times faster than a human need not find the speed a problem — and both are just what current systems are getting good at, with Physical Intelligence's π0, Gemini Robotics, and Figure's Helix integrating high-level reasoning with fast low-level control. Low-level hand dexterity, moreover, is mostly learned from very short interactions — a grasp or a fine adjustment, over in a second or two — which makes it a natural target for brute force, with a fleet of arms or a simulator running through vast numbers of them.
One might worry that the tacit, hard-to-verbalize know-how of a skilled tradesman would be hard for an AI to acquire. But acquiring tacit knowledge is not peculiar to physical work, and an AGI able to do remote cognitive work will almost certainly have to be competent at it regardless. [2] Indeed, an AGI could acquire a vast amount of tacit knowledge, much as it has already acquired vast, even superhuman, knowledge by ingesting the internet. This could be facilitated by parallel training across many instances, if that proves feasible, and what it learns can be specialized and copied as many times as needed.
For these reasons, I expect that an AGI able to automate most remote work will also be able to learn to operate robots, potentially to a high degree of sophistication. And operated robots would be valuable enough that there would be every incentive to push further, whether by specializing the models themselves or by giving them access to specialized tools for fine motor control.
Automating physical production mostly requires automating physical industry
What is required to automate physical production? At the very least, you need the robots and computer chips that can substitute for human labor, and you need to be able to make those work automatically. The computer chips require silicon fabs, which in turn require refining the silicon and producing a range of machine tools, dies, and so forth. Likewise, the robots themselves require all sorts of parts that must themselves be made from more materials. We obviously need a lot of energy to power this whole economy, and a good deal of infrastructure like roads and bridges to move things around.
Using the US input-output tables and the estimated compute and robotics costs from Part 1 of this series, we can actually calculate what a very naive version of this self-replicating economy would look like. We do exactly what we do today, but replace the human labor with the equivalent chips and robots. What we find is an economy heavily structured toward production — toward construction, production and other machinery, electrical equipment, computers, and other heavy industry:
The same data lets us look at which jobs this economy needs to automate, occupation by occupation. Again it is dominated by the heavy-industry trades:
Consumer goods and most services have essentially vanished. The jobs people usually reach for when they debate whether AI can automate human labor — waiters, doctors, babysitters, elderly-care providers, teachers, lawyers — are simply not needed if what you want is to build an unlimited number of robots. The same goes for research: how easily science can be automated, and whether AI will make its own discoveries, matter a great deal for other questions about the post-AGI world, but it is beside the point for massively scaling production once we have the necessary designs. [3] What matters here are the heavy-industry occupations above, and whether the tasks they involve can be done by machines.
Half of the tasks need no body, and only a third need hands
What would automating this labor actually take? Using O*NET, we can decompose occupations into detailed work activities, and weighting each activity by the number of people who perform it gives a task-level picture of the whole core (the occupations this economy needs):
The category column shows our classifications of these tasks into one of five broad categories, depending on what is required to automate them:
Applying this across all of the core's tasks, the labor divides as follows:
More than half of the core's tasks are either not needed at all or could be handled by a remote worker. The remote work tasks are mostly the routine data collection, paperwork, and coordination required to run a plant, and they are not especially demanding. Measured against the US workforce as a whole, the remote jobs being replaced require somewhat fewer bachelor's degrees — and the ones they do need are mostly engineering — and far fewer graduate degrees.
A further 6% consists of operating machines like cars or cranes where the human supplies the perception and the control signals. That human can be replaced by adding sensors to the machine and feeding them to an AI that supplies the same signals. If an AI has intellectual capabilities similar to a human's, it should be able to learn to operate such machines as readily as a person does. Even the hardest tasks take humans at most several thousand hours to master. An AI with comparable intellectual abilities could learn much faster.
Even without AGI, we are already seeing considerable progress at automating such jobs. Waymo robotaxis operate in more than ten US cities and have clocked hundreds of millions of kilometers; autonomous haul trucks from Komatsu and Caterpillar are increasingly common in mining; Built Robotics runs autonomous excavators and pile-drivers on construction sites; and automated cranes shift steel and stack containers with no operator in the cab. But today's AI is not very smart, so these systems carry high software costs and operate only under limited conditions. Waymo, for instance, is slow to expand because it must map and build custom software for each new city it enters, work that would be unnecessary for a system with human-level intelligence.
Another 8% is perceptual: the human reads an instrument, measures a part, or looks at something and judges it. Given AGI, most of this needs no dedicated worker; simply place sensors and instruments where needed and collate the readings centrally for AIs to interpret. Where a reading has to be taken on the move, a cheap drone or quadruped carrying sensors would work.
The final 33% of tasks require movement and manual dexterity. To see where this hands-on work sits, we take each occupation the core needs and weight it by the number of workers times the fraction of its tasks that are manual. The top 15 occupations, which together make up 56% of the manual work, are as follows:
In the table, I have split the occupations into three broad classes. Factory jobs, accounting for half of the manual work, are those largely done in controlled factory environments. These are generally the most favorable for automation, since with enough care the factory itself can be designed to remove sources of variation, and it is where automation has already gone farthest. Field work, 32% of the total, is done outdoors in uncontrolled environments. This is tougher for several reasons: the tasks are unstructured and variable, and the machines must be mobile and robust to the elements. The remaining 18% are occupations that perform maintenance and repair, though they often install new equipment as well. These jobs are particularly difficult to automate because breakdowns are, by nature, non-routine, the work is constrained by whatever is already built around it, and it often demands dexterity in cramped or difficult surroundings.
Note that these classes are approximate. Factory workers spend some of their time repairing machines, and much of what maintenance workers do is routine servicing that could readily be automated. But the classes still give us at least a rough sense of what is involved in the labor and how easily it could be automated.
Existing robots can physically do the things needed to automate routine factory work
Factory work should be the easiest part. A factory is a controlled environment: we design the building, the machines, and the flow of parts, and automation has already gone far here; some plants even run with no one inside at all.
The workhorse of this automation is the industrial robot arm. These arms are capable of a wide range of motions, using wrist-mounted tools such as grippers, grinders, or welding torches with precision and force well beyond a human arm. Millions are installed worldwide, and they are becoming increasingly cheap. Yet the arms have been restricted to repetitive, pre-programmed motions in fixed cells, because they cannot intelligently sense and operate in their environment. Everything they do must be carefully planned in advance, with integration running multiple times the cost of the arm itself. In practice, robotic automation has reached only tasks that are high-volume and have very low variability.
An AGI operating these same arms would greatly expand the range of tasks that are economical to automate without requiring substantial changes to the arms themselves. The upfront engineering largely disappears, since the AI can work out the process itself, and plans can be changed on the fly rather than re-engineered. Exceptions also stop being a barrier: a process that runs correctly 99.9% of the time is fine, so long as the AI can recognize when something exceptional has happened and direct the arms to deal with it. In practice, the arms would probably run under a mixture of pre-programmed rote actions, which the AI may itself have programmed, and AI control at varying levels of intelligence. The most powerful models need not be running the arms most of the time, but can step in when necessary to correct problems and implement longer-term process fixes, eliminating the need for human oversight.
One task industrial arms have historically not been able to do is dexterous work, such as picking parts out of a bin they lie scattered in, packing goods of different shapes, fitting components together, or handling anything soft or floppy; these tasks are too variable to be reduced to pre-programmed repeated motions. This is the work that human hands still do in factories.
We have long had bodies physically capable of performing most of this work. Humans teleoperating machines has a long history in environments where using humans is not possible:
In the last few years, as AI has improved, there has been a surge of interest in building general-purpose manipulators and getting them to perform more complicated dexterous tasks autonomously. The method usually has two parts: humans teleoperate the machine through a task, many times over, and then a model is trained to imitate them. Since we are most interested in what the machines can physically do, the teleoperation datasets are very useful, as they show directly what is within reach of the actuators:
These datasets involve a fairly general range of tasks performed with the same actuators, and the actuators are usually much simpler than human hands. They cover a broad range of tasks relevant to light household and tabletop work. There are notably few examples of these manipulators using tools or doing more forceful work. I am not sure how much of this is a real limitation and how much reflects the researchers' interests, but it makes little difference here: in a factory you would not use a general manipulator for these anyway, mounting the tool on the machine directly and leaving heavier work to industrial robots.
The tasks missing from the datasets — starting a fastener, taping a bundle, splicing wire — already have machines doing the skilled part: automatic screwdriving stations catch cross-threading in the torque signature, taping heads hold the tape at a programmed tension, splices are made by ultrasonic welders, and peel-and-stick gaskets can be replaced with a robot-dispensed bead. These are specialized machines rather than dexterous manipulators, and a person typically still feeds and positions the work, so the tasks are not yet fully automated. But what is left to the hand is coarse handling of the kind the manipulators above already manage, and the hand requirement can be engineered out wherever the work can be brought to the machine.
The one real holdout is routing floppy things. Wire-harness assembly remains overwhelmingly manual (Trommnau et al. (2019)) — the cables take no fixed shape, the variants are nearly unique, and cheap labor has beaten fixed automation — and neither route has cracked it: machines cut, strip, crimp, splice, and tape, but none routes a harness end to end, and lab manipulators are only beginning to route single wires into clamps. Much of what makes floppy things hard is that they move in hard-to-predict ways, and predicting how things move is just what the latest vision and video models have been getting good at.
Autonomous performance of dexterous tasks is also progressing rapidly. A sense of the current state of the art can be seen in the "Humanoid Olympic Games", a set of manipulation tasks that Benjie Holson — a roboticist who spent eight years at Google's Everyday Robots — posted in 2025 and judged beyond current systems. Within a few months Physical Intelligence had demonstrated most of them with plain vision and simple grippers. Holson has since posted a harder set that remains undone. It is difficult to assess what the true hardware frontier is because interest is focused on doing tasks that are plausibly automatable rather than merely physical possible.
The remaining ingredient of a workerless factory is moving materials between stations, which is now the easiest part. Much of the flow can simply run on conveyors, and the rest can be handled by simple wheeled robots with basic manipulators, carrying items along routine routes and collecting whatever goes astray. These machines are mechanically straightforward; what AI progress added is the ability to navigate a busy building and operate in variable environments. Amazon already operates over a million such robots, which navigate by their own sensors and roam freely among the warehouse workers where they were once fenced off.
Once we have AGI, building new factories that do not require humans for routine operation seems fairly straightforward: intelligent arms at the fixed stations, general-purpose manipulators for the dexterous work, and mobile robots moving goods between them. That covers running the plant. Installing the equipment in the first place, maintaining it, and repairing it when it breaks is a separate and harder problem, which we take up below.
Aside: on the other hand
Robot hands, for all their progress, are still no match for the human hand. Rodney Brooks, one of the field's most eminent figures, argues that no robot hand has shown much dexterity in any general sense, and that today's humanoids will not learn it, because human dexterity rests on a rich sense of touch — some 17,000 mechanoreceptors in the skin of the hand — that current approaches do not even try to reproduce. I agree that the human hand is a remarkable feat of engineering, and touch is not the only thing robots have yet to match — no actuator yet pairs the strength, precision, and give of the human arm, or the force and dexterity of the hand, as well as we do. But the point is largely academic. You do not need hands like ours to do most tasks, and the cruder robotic manipulators we already have manage an impressive range of manipulation, as the datasets above show — and robot hands are getting better besides, with tendon-driven designs now attacking the force-dexterity tradeoff directly.
But you can also verify for yourself that not all tasks require fully functioning hands. What fraction of tasks can you do with your fists closed, or with only two fingers, or with mittens on? Some things become very difficult, or impossible, but most tasks can still be done, if slowly and clumsily. Nor is it just about what a healthy human can do. When people have their fingertips anesthetized so that they can feel nothing with them, they become much clumsier — gripping harder, fumbling fast movements — but can still pick up and hold objects, even a brimming glass without spilling. And some people live with permanent numbness of this kind, yet learn to manipulate things perfectly well by guiding their hands with their eyes.
People living with certain disabilities have no choice but to get by without hands. Those with amputated arms can learn to do an impressive array of tasks with body-powered split-hook prostheses — two rigid fingers on a cable, with essentially no sense of touch. They farm, weld, and work construction with them, and handle everyday objects with a dexterity that belies the simplicity of the device. People born without arms entirely can learn to use their feet or mouths for many tasks, like flying planes, competing in archery, and painting. Human intelligence enables us to overcome physical impairments both temporary and permanent; AGIs would possess the same skills, along with greater patience and a much greater range of actuators to choose from.
There is, in any case, no need to restrict ourselves to our current hand-based manipulations. There are numerous other ways to grip and move in the world, and ways to arrange production tasks to not require manual dexterity. Nor do the tools humans wield with their hands need a hand to wield them: a machine does not grip a hammer or hold a welding torch the way a person does, but instead just mounts the tool directly. I do not know if or when robot manipulators will manipulate objects as gracefully as human hands can, but for automating physical production the issue is largely moot.
The construction work still done by hand is coarse and heavy
The autonomous economy needs to build factories, power plants, mines, roads, and the other infrastructure of heavy industry. Much of this is structurally simple work, and it can be standardized to a much greater degree than housing or office construction, with the same few designs repeated across many sites. The structural shell is the straightforward part; fitting out the inside is more involved, especially in equipment-dense buildings like process plants and semiconductor fabs, and that is the installation work we turn to afterwards.
The early steps are the groundwork, the foundations, and the shell: clear and grade the site, trench and lay the buried services, pour the slab, erect the frame and envelope. This work is already heavily mechanized. Industrial sites are chosen for convenience, flat and next to transport; excavators and dozers cut them to grade, concrete is delivered by truck and pump and struck off by laser screed, the steel frame arrives from a fabrication shop, and big-box walls are commonly cast flat on the slab and tilted up. The human labor that remains sits in steps like trench work, laying and joining the buried pipe and cable, tying rebar where it is still used, and the ironworker's job of catching, aligning, and bolting the hanging steel.
The rigorous way to settle this would be to take each task in turn and grade it — from already demonstrated, through plausibly automatable, to truly hard — and for the hard cases, ask whether a different construction method avoids them altogether. I have not done that here. But on the whole these tasks just look like the kind of thing automated machines can do: the work is coarse, the pieces are large, and the forces are supplied by the machine, so given AGI, building robot bodies for it seems eminently possible.
There is no off-the-shelf robot for this work, but only because there has never been a reason to build one — without a controller smart enough to run it, a general construction machine would just sit idle. The job is to get an actuator to the right place and have it do its operation reliably, and we already have both halves: an arm places a tool with precision and force beyond a person's, and plenty of machines move around a site. The rough sketch is easy to draw. Groundwork might run on something like a small demolition robot with arms and swappable actuators; work at height on a boom lift of the kind that already carries workers up, with arms in place of the basket; sensing and spotting on drones and small robots around the site. Some of these exist already, from autonomous trenching machines to self-propelled fusion welders; others, like arms rugged enough for weather and dirt, would have to be engineered. There is real work in choosing the right sizes, actuators, and configurations and making them reliable, but it is ordinary engineering, not a barrier in principle. Machine-mateable fittings, as in subsea construction, would make it easier still.
That covers the shell, the coarse, heavy part. The work inside — running power, pipe, and duct, then setting and connecting the equipment — is finer and more variable, and it has more in common with maintenance and repair than with pouring a slab or raising a frame. We take it up with that work next.
Somewhat-humanoid bodies can probably do most installation, maintenance, and repair
Installation, maintenance, and repair work is much less repetitive than factory and construction work, and the tasks also often involve some degree of dexterity, which makes them among the most challenging things to automate. Weighted by labor, the tasks in this bucket collectively come to about 6% of all labor. The largest, by share of the bucket, are:
The variation within each of these tasks, and between them, is substantial, and larger than in the previous categories. It is hard to go through the list and determine, for every instance, whether the task is within the reach of existing machines, or whether it could be automated around.
The list of tasks is broadly familiar, but each comes up under varied circumstances, and the job begins with reaching the problem and working out what it is. This is much of why the work is hard to automate, and it is exactly the part that matters least here: recognizing and assessing problems is cognition, squarely within the remit of an AGI. The remaining question is whether a body can physically get to the work and do it.
The simplest way to automate these tasks, and the one that has received the most hype, is a general-purpose humanoid robot. Work sites and equipment were built for human reach and human hands, so a body shaped like ours can go where a person goes, use the same tools and openings, and take over the work as it stands. If humanoid robots could physically and economically perform these tasks, there would not be much left to argue about.
Unfortunately, the software is still the weak link, so it is hard to see where the hardware's limits lie. Autonomous control is crude and teleoperation is clumsy, and tasks performed through either give only a lower bound on what the bodies can do. The weak software also holds back the hardware itself: there is no point building a body that software cannot yet put to use, so general-purpose bodies have scarcely been worth making.
Existing humanoids are nevertheless the best gauge of what general-purpose robots can currently do. A sense of the state of the art can be seen in a WSJ hands-on with 1X's NEO, in which a remote operator in a VR headset drove the robot through loading a dishwasher, fetching a bottle from the fridge, and folding laundry; an early NEO cooked a steak dinner the same way, and Figure and Unitree robots run through similar household chores and, in the latter case, basic athletics as well. These videos deserve substantial salt; to quote Chris Paxton's golden rule:
But Paxton's rule applies to the software, not the body. For our purposes a teleoperated demonstration is enough to show that the body is physically capable; and a body that can fold some shirts can presumably fold other shirts, even if today's models seem limited to red T-shirts on black coffee tables 50 cm from the floor.
Judging from what I have (and have not) seen, humanoid robots can physically do a wide range of light domestic tasks, albeit slowly and clumsily. Nothing heavy or forceful is on show — no power tools, no digging — and reliability is hard to judge from choreographed videos. The robots are cheap — Unitree's G1 lists at $13,500 and 1X's NEO at $20,000 — but a body that breaks down repeatedly would be a practical barrier. Still, these gaps look more like product choices than physics. The robots are aimed at homes, where safety matters more than strength, and making bodies that push hard and survive rough handling is ordinary engineering: industrial arms already ship with payloads up to 2,300 kg and in sealed foundry variants built for harsh environments. I am relatively bullish that somewhat-humanoid bodies could cover most of this work.
The tasks I am least sure about involve deformable materials, work done by feel, and force exerted in place, though these are hard to different degrees. Manipulating deformable things like stiff cable is a recognized open problem in robotics (Zhu et al. (2022)), but the difficulty splits in two: the intuitive physics, which an AGI would have much as we do, and perhaps the grasp-on-contact reflex touch gives us, which grippers and sensors could approximate. Free-standing force is a bracing problem — a human braced against something firm can push about three times harder than one standing free — but a machine can clip in, or clamp onto a dumb wheeled weight brought alongside. The stubborn case is work done blind, where touch must substitute for sight (Li et al. (2025)) — connectors behind flanges, fasteners inside cabinets — and today's grippers have almost no tactile sense. Sensing is improving, but this is plausibly where automation lags longest.
Still, such tasks are a minority of the bucket, and the work can be engineered so that less must be done in place and less needs dexterity at all. Prefabrication moves installation off-site into the factory, where whole power rooms are shipped factory-tested and dropped in as units; repair by replacement does the same for maintenance, pulling the failed module and repairing or scrapping it on a bench elsewhere. Either way a cramped in-place job becomes an accessible swap plus controlled work that can be automated like production, as aviation and data centers already do.
A new-built economy can push this much further by designing equipment for machines from the start. If cable runs are laid out for a machine to reach, threading wire through conduit stops being a task at all, and if modules locate on dowel pins and connect through machine-mateable fittings, alignment happens by design rather than by feel.
That leaves the work the bodies cannot yet clearly do and that redesign cannot remove, particularly feel-based work in blind and confined places on older stock built for human hands. Some of this will stay manual for a while, some equipment will be scrapped and replaced rather than repaired, and for the rest it may simply remain economical to employ humans. This potentially intractable portion is a small fraction of labor, and it would be subject to intense efforts to engineer around or remove it.
Conclusion: robots are probably not the bottleneck
Automating physical production means automating a long and heterogeneous list of tasks. Yet we know these are all within reach of ordinary humans using their ordinary physical and cognitive skills, aided by tools and machines. We should therefore expect that a cognitively intelligent AGI with the right bodies can also do these tasks. Indeed, about half of the tasks needed for physical production require no physical body, and when we look at the remainder, most of it can plausibly be automated, whether by adding sensors to existing machines or by deploying a (not infinitely) diverse range of robot bodies equipped with arms and variably (but, for the most part, not overly) dexterous hands.
Some tasks would resist. If we had to automate everything today, my guess is that a few percent would defeat us, above all the blind, feel-based work and the routing of floppy things discussed above, along with the most dexterous manipulation. A larger chunk could be automated only awkwardly, on equipment and in workplaces built for human hands. Both residues would shrink over time, as sensing improves and new plants are designed from the start for the machines that run them.
So suppose we had AGI tomorrow. How much would the lack of physical actuators delay growth? We should distinguish this from the delay of simply not having enough actuators. There are nowhere near enough robot bodies, industrial arms, sensors, and so forth to automate today's production. We already analyzed this delay in Part 2, and saw that it slows growth by several years, but not much longer. Here we are asking a different question: how much additional delay is there because we don't know what to build or how to build it?
As I already argued, most of the labor doesn't require sophisticated actuators, so I don't think this additional delay would be that substantial. We could build a range of robots with plausible levels of dexterity and strength, deploy them, and see where they succeed and where they are weak. This process would be greatly facilitated by the AGI operators themselves, who could tell us exactly what is going wrong and design new, better robots. The designs wouldn't need to be perfect the first time, and they would keep improving throughout a buildout that takes several years in any case. This is a great deal of engineering, but it is engineering rather than new science, and it can happen during the buildout rather than before it.
During this buildout, human labor would still be available to do the tasks that are hardest to automate. If automation cut the labor needed in construction to a tenth of today's, the existing workforce could support a tenfold expansion of the sector without a single new hire. Workers freed from jobs outside physical production would swell the pool further, nearly free to act as dexterous hands when needed. As Shulman (2023) and Davidson and Hadshar (2025) describe, AI direction could make them productive at unfamiliar manual tasks with little training. Plugging in cables in the dark may prove to be one of humanity's last comparative advantages in physical production.
If AGI instead takes until the 2030s, actuators should be even less of a bottleneck. Robotics is attracting enormous investment as AI finally becomes capable of dexterous work and of following general instructions. Humanoids in particular are in something of a hype cycle. Tesla talks of converting a car line to build a million Optimus robots a year, while Figure and 1X have opened factories with capacity for about ten thousand robots a year, with stated plans to scale to a hundred thousand within a few years. Chinese manufacturers are further along in volume, with Unitree alone shipping over five thousand humanoids in 2025. These plans deserve skepticism; I'm still waiting for the ten thousand robots Tesla promised for last year. But robot production is already scaling up rapidly, albeit from a low baseline, and if demand emerged for millions of robots a year, producing them would require only a fraction of the industrial capacity we currently use to make cars.
AGI arriving in 2036 could thus find a world with millions of robots already at work, and factories substantially more automated, supplied by a much larger industry making robot arms and bodies. An industrial takeoff from that starting point could be very fast, although still continuous with pre-AGI growth that was already quite fast.
A very cute DIY example can be found here ↩︎
I also doubt that as much of this know-how is truly missing from the vast text and video these systems train on as the objection assumes, but that is an argument for another day. ↩︎
Automation and designing the robots in the first place could require some R&D, but once these exist, production is just a matter of scaling up existing techniques and existing production. To be clear, I am not saying there would be no R&D in this world. There would be a bunch of improvements and efficiency gains to be had, so I expect there would be a lot of spending on R&D. But that spending would occur only insofar as it in fact made production even faster or more valuable over time. ↩︎