The robots are coming

The post is openly a partial repost. He is reprising the robotics section of a piece written about a year earlier, and says two things have changed since: he is more convinced that robotics specifically is accelerating, enough to call it one of the century’s biggest global trends, and the world appears to him to be waking up to it, which the earlier piece had complained it was not. The case itself starts away from robots. Our biggest problems are still mostly material, so models confined to a screen can address them only obliquely, and across the economy the bottlenecks are generally not the PhD-level ones language models are good at. Knowledge work has been made tremendously more productive over twenty-five years, while labor productivity in manufacturing and construction, the two sectors that actually shape the physical world, has been flat to declining for two decades. Demographics compound it. The post-war infrastructure spree ran on a bumper crop of prime-age workers, and today’s pool of young people with strong backs is shallower, with immigration restrictions shrinking it further. The squeeze is already visible in energy, where a survey of solar developers and utilities put labor availability above inflation and interconnection queues, and where nearly every established trade the transition depends on is in high demand and projected to grow fast. His conclusion is framed as an elimination rather than an enthusiasm: he can see only one plausible option, a leap in physical labor productivity, which means embedding AI in machines.

How far along that is depends on which job you mean, so he organizes the field as a spectrum of task difficulty. At the easy end are discrete, repetitive, fully programmable tasks needing little dexterity, which is what essentially every robot doing real work today performs: industrial arms in car plants, electronics and metal fabrication, now numerous and cheap, with China the largest buyer by far and a domestic industry challenging the leading suppliers elsewhere. At the hard end sit jobs whose tasks are nearly impossible to pull apart, vary every time, sometimes involve members of the public rather than trained handlers, and may demand fine manipulation. His evidence on that last point is quoted rather than newly argued: the roboticist Brad Porter on the space of words being constrained where the space of physical interactions is not, and a block-quoted calculation on the human hand credited to Lubershane with his colleague Anil Achyuta. His model of the far end is a plumber: an upset customer with a badly defined problem, a messy crawl space, a system assembled over decades, a diagnosis from a few observations, then strength and delicacy at once. We are still very, very far from replacing him.

What has moved, he argues, is everything between those ends, pushed by three interlocking trends. The first is deployment outside tightly controlled settings, warehouses being the step past the factory floor, where Amazon’s robot count is rising as its human headcount starts to fall, and drones being a cheap and versatile testbed that has gone from inspecting energy infrastructure to helping build it. The second is the arrival of vision-language-action models, where the point is not the models themselves but the starting position. Five years ago an autonomous system began as a blank slate and had to gather its own data about both the world and its own body, which is why Waymo needed six years to reach its first million autonomous miles and then sixteen months for the second. Now the world-understanding can be loaded in at the outset and only the action half has to be learned, which he says cuts years and sometimes billions out of building a new robotic product. The third is the decade of autonomous vehicle spending, and here the argument is about spillover rather than cars: capital went into driving and came out as a cohort of experienced engineers now founding robotics companies, and as a collapse in the cost of LIDAR sensors, which he credits to that ecosystem, with batteries, servo motors and chips cheaper too.

Where he lands is deliberately unspecific about shape. The humanoid thesis is easy to state, since the world was built by humans for humans and human-shaped machines can be trained by imitation, and it is very well funded. He concedes humanoids may prove necessary to reach the plumber end of the spectrum, but says personally he is not so sure, and would expect more value from a wide range of purpose-built forms, the way evolution filled different niches with different bodies. If he is right, what follows is not one robot but a proliferation of them across the physical economy. That conditional is the post’s own framing of its prediction.