The Threshold

Every previous wave of automation produced a version of the same argument, and the argument was always, in the end, right. The machines will take the jobs that machines can do. The jobs that machines cannot do will remain for humans. The boundary will shift, but a boundary will exist. Adapt, retrain, move up the value chain. The ladder holds.

The argument was right because it was describing a real structural feature of automation as it had existed up to that point. Machines could handle tasks that were high-volume, highly standardized, and performed in environments that had been specifically designed around the machine’s capabilities. They could not handle the broader class of tasks that required adaptability, dexterity in unstructured settings, and the ability to respond to variation that had not been anticipated by whoever built the system.

That structural feature is being removed. Not incrementally improved. Structurally removed.

Understanding why requires being precise about what the limitation actually was, and what is now dissolving it.

What Automation Has Always Required

#

A machine, to replace a human at a task, needs to perceive the relevant inputs to that task, decide how to act on them, and physically execute the action. Each of these three requirements, perception, decision, and physical execution, has historically imposed limits that defined the boundary between automatable and non-automatable work.

The perception requirement meant that automation worked well in controlled environments where the relevant inputs were predictable and standardized: the same component arriving at the same position on the same conveyor belt in the same orientation. It worked poorly in environments where inputs varied, where objects came in different sizes and configurations, where something unexpected could appear. The human eye and the human capacity to parse a novel visual scene without prior specific training on that exact scene was not something machines could match in general settings. They could match or exceed it in narrow, controlled ones.

The decision requirement meant that automation worked well for tasks where the logic of decision was explicit and bounded: if the measurement is outside tolerance, reject the part; if the sensor reads above threshold, stop the line; if the barcode matches, approve the transaction. It worked poorly for tasks requiring contextual judgment, tasks where the right action depended on factors that varied in ways the designer could not fully enumerate in advance. Human workers navigated this judgment gap continuously, often without recognizing they were doing so. The factory worker who noticed that a batch of materials was behaving differently and adjusted their technique was exercising a form of contextual reasoning that automation handled clumsily or not at all.

The physical execution requirement was the most concrete limitation. Robotic systems required purpose-built end effectors, grippers and tools designed for specific objects and specific tasks. The human hand, with its combination of strength, precision, compliance, and the ability to manipulate objects it has never encountered before, was extraordinarily difficult to replicate for general use. A robotic system optimized to pick and place a specific electronic component performed that task with superhuman speed and precision. Asked to handle a slightly different component, or a garment, or a piece of fruit, it either failed entirely or required extensive reprogramming and physical retooling.

These three limitations defined, for roughly a century, what automation could and could not do. And their combined effect drew a line that corresponded, with uncomfortable precision, to the global distribution of wage labor.

Where the Line Fell

#

The tasks that eluded automation were not random. They clustered at the intersection of low wages, physical variability, and contextual complexity.

Garment manufacturing is the clearest example. The global apparel industry employs somewhere between sixty and seventy-five million people, concentrated in South and Southeast Asia and parts of sub-Saharan Africa. It has been a primary first rung on the development ladder for every country that has run that ladder in the last half century. Bangladesh’s economic transformation over forty years is substantially a garment story. Vietnam’s manufacturing emergence is partly a garment story. The industry absorbed labor at the bottom of the global wage distribution and converted it, through the mechanisms the previous essay described, into capital accumulation and workforce development.

Why did garment manufacturing evade automation for so long when other manufacturing sectors automated heavily? Because fabric is a physically complex material. It deforms. It stretches. It bunches. Its behavior under a sewing machine needle depends on tension, grain, and thickness that vary even within a single piece. Assembling a garment requires handling an object that behaves differently every time you pick it up, in positions and orientations that are never quite the same, executing fine-motor adjustments that respond to the material’s behavior in real time. Robot systems attempted to solve this for decades. None succeeded at general garment assembly with the speed and adaptability that human hands achieved.

Consumer electronics assembly presents a similar profile. The population of components on a printed circuit board includes parts that are tiny, fragile, oddly shaped, and must be placed with precision into positions surrounded by other tiny fragile parts. The hands of workers in Shenzhen or Hanoi, trained to this work over months of practice, developed a capability that automated systems could replicate only for specific, high-volume components on purpose-designed lines. General assembly, the ability to handle a new product’s components with the same facility, remained a human skill.

Food processing, light manufacturing, warehouse fulfillment, the physical tasks of the service sector: all of these shared the same profile. Variable objects, unstructured environments, fine motor requirements, contextual judgment. All of them concentrated labor at the bottom of the global wage distribution. All of them provided the first rung.

What Is Being Solved, and How

#

Two things are being solved simultaneously, and the simultaneity is what makes this moment categorically different from previous automation waves.

The first is general-purpose machine reasoning. Foundation models, large language models and their multimodal extensions, have demonstrated something that the field did not expect to achieve so quickly: the ability to follow novel instructions, reason through problems not encountered during training, interpret ambiguous inputs, and make contextual judgments across an enormous range of domains without task-specific programming. This does not mean these systems think in the way humans think, and this series has examined that question at length. It means they have crossed a functional threshold for a large class of tasks that previously required the kind of contextual reasoning only humans could provide.

The significance for automation is not primarily in what foundation models can do at a keyboard. It is in what they enable physically embodied systems to do. A robotic system directed by a foundation model can receive natural-language instructions about a novel task, reason about how to approach it, recognize when its initial approach is not working, and adapt without being explicitly reprogrammed for each variation. The decision gap in the automation triad, the gap that kept contextually complex physical tasks in human hands, is closing.

The second is dexterous robotic manipulation at scale. This is the harder problem, and it has been harder longer. The breakthrough is not a single technology but a methodology: training robotic systems in simulation environments where they can attempt a manipulation task millions of times, fail, adjust, and develop capability through accumulated experience at a speed and scale impossible in physical training. Combined with new approaches to robotic hand and arm design that prioritize compliance and adaptability over rigidity and precision in narrow tasks, the result is systems that can handle variable objects in unstructured environments with an increasing range of facility.

The state of dexterous manipulation in 2024 is not the state it will be in 2027. The improvement trajectory is steep, and it is being driven by investment at a scale the field has never seen, because the economic prize at the end of that trajectory, replacing the sixty-plus million garment workers and the hundreds of millions of workers in analogous roles, is among the largest economic opportunities in industrial history.

The Form Factor That Is Easy to Underestimate

#

There is a specific feature of the current robotics wave that most technology coverage treats as a curiosity but that carries structural significance: the humanoid form.

Human environments were built for human bodies. Factories, warehouses, hospitals, restaurants, retail spaces, vehicles, offices: their dimensions, their tools, their furniture, their workflows, the entire physical infrastructure of productive human activity was designed around the capabilities and limitations of the human body. Doorways are human-width. Workbenches are human-height. Hand tools are shaped for human hands. Stairs were built for human legs. The cockpit of a vehicle, the layout of a kitchen, the shelving in a warehouse: all of it encodes assumptions about who is doing the work.

Previous industrial robots required their environments to be redesigned around them. The automotive assembly line is not a human environment with robots inserted into it. It is a robot environment, purpose-built to accommodate the specific capabilities and constraints of the machines that operate within it. Implementing that automation required enormous capital investment in facility design, not just in the machines themselves. The cost and complexity of that redesign was part of what limited automation to high-volume, high-margin manufacturing where the investment could be justified.

A humanoid robot requires no such redesign. It can walk through the door of a factory or warehouse built for humans, stand at a workbench built for humans, use tools designed for human hands, and navigate the space the way a human worker would. The physical infrastructure of the world is, from the perspective of a humanoid robot, already built. The deployment cost of automation drops dramatically when you do not have to rebuild the environment you are deploying into.

This is not a marginal consideration. It is one of the principal reasons why the current robotics wave is capable of reaching tasks and environments that previous waves could not reach.

The Convergence Is the Point

#

It is important to be clear about what is actually new, because individual pieces of this picture have existed for longer than the current discourse implies.

Industrial robots have existed since the 1960s. Computer vision has been a research field for decades. Natural language processing had significant milestones well before foundation models. Humanoid robotics has been a research pursuit since the 1990s. None of these, individually, crossed the threshold this essay is describing. What is crossing the threshold is their convergence into integrated systems capable of doing things that none of the components, separately, could do.

Foundation models providing general-purpose reasoning and instruction-following. Computer vision systems that can identify objects, assess their state, and track their position in real time across the enormous variety of objects that exist in unstructured environments. Robotic hardware providing physical embodiment with improving dexterous capability. Simulation environments providing training at scale. Declining hardware costs driven by manufacturing volume: a humanoid robot that cost a hundred thousand dollars to produce in 2023 is on a cost curve that leads somewhere near ten thousand dollars within a decade. And edge computing providing the local processing capacity to run these systems without requiring constant connection to remote servers.

The integration layer matters as much as the components. A robotic system that can perceive variable inputs, reason about how to act on them, and execute the action with appropriate physical dexterity is qualitatively different from any individual component of that description. Convergences are threshold events. Below the threshold, the components exist but the capability does not. Above the threshold, the capability exists in a form that changes what is possible.

We are crossing the threshold.

Why This Time Is Different

#

The previous automation waves that produced the “adapt and move up” argument were waves that automated specific tasks in specific environments. They required substantial capital investment in purpose-built systems. They left enormous categories of work untouched because those categories had the profile that automation could not address: physical variability, unstructured environments, contextual judgment. The workers displaced by one wave of automation had somewhere to go, either within the same economy in the tasks that automation had not reached, or up the value chain as accumulated capital funded better education and higher-skill industries.

What is different now is the profile of tasks being reached.

The tasks that are now being automated are not the next set of highly standardized, high-volume, controlled-environment tasks. They are the tasks that eluded automation precisely because they had the profile automation could not handle. They are the tasks that the global south’s labor force provided most competitively. They are the tasks that constituted the first rung of the development ladder for every country that climbed it in the last half century.

And the cost curve is closing faster than the planning horizon of the factories that were supposed to be built.

A humanoid robot, amortized across its operational life, is already approaching cost parity with manufacturing wages in much of South and Southeast Asia for standardized tasks. For the tasks that the current generation cannot yet handle reliably, the next generation will. The trajectory is not a prediction about distant futures. It is a projection from current development rates across a planning cycle that factory investment decisions are made within.

The previous waves left a boundary between what machines could do and what humans needed to do. The boundary shifted over time, but it existed, and the tasks on the human side of it provided economic roles for the populations that needed them.

The current convergence does not move the boundary. It dissolves it, over the class of tasks that kept hundreds of millions of people employed at the base of the global productive economy.

The Structural Claim

#

This essay has been building to a structural claim, and the claim is this: what is happening is not a faster or larger version of previous automation. It is a different kind of event. Previous automation optimized within a system whose fundamental architecture remained stable. This one changes the architecture.

The architecture that is changing is the relationship between human labor and productive value at the global scale. For two centuries, that relationship provided the mechanism through which countries developed, through which populations moved from poverty to participation in modern economies, through which the gains of industrialization distributed broadly enough to create the consumer classes that sustained industrial production. The mechanism was imperfect, often brutal, and unevenly rewarding. But it was a mechanism, and it worked.

The convergence being described in this essay is removing the mechanism.

Not from all tasks. Not immediately. Not uniformly across geographies. But from the class of tasks that made the mechanism available to the populations that most needed it, and on a timeline that is measured in years and decades, not generations.

The essays that follow this one examine what that removal means at the civilizational scale: which development paths have been foreclosed, which populations are most exposed, what frameworks might survive the transition and for whom. Those essays cannot be properly read without the structural argument this one makes.

The threshold is the beginning of the analysis. Everything that follows depends on understanding what has been crossed.

The Approximate Mind is a philosophical essay series examining how artificial intelligence transforms human work, identity, development, and society. Parts 66-68 examine the civilizational consequences of the transition described in this essay.