An Ode to “Physical AI”

The post is not an attack on language models, and that setup carries the conclusion. He credits them with real transformation already in translation, software development, customer service and illustration, and with promise at searching and synthesizing across large document collections, and he expects products of that kind to spread through knowledge work by taking the drudgery first. What he declines to evaluate is the tier above. He quotes Sam Altman saying OpenAI is now confident it knows how to build artificial general intelligence as it has traditionally been understood, and is turning its aim beyond that to superintelligence, and answers that he finds such claims difficult to evaluate if not outright unfalsifiable; asked whether more compute gets there, his own reply in the text is a shrug. The turn comes next: humans are still a material species whose biggest problems, security and climate and disease among them, are material ones, and an army of digital agents confined to a screen can only address those obliquely.

That framing sets up the real claim, which is about where economic bottlenecks sit. Twenty-five years of digital progress have already made people who work with data and ideas far more productive, and language models continue that trend. Over the same period, labor productivity in the trades that actually shape the physical world has stagnated, and demographic change is compounding it: the postwar infrastructure spree ran on a bumper crop of young workers, while today’s pool is shallower and tightening immigration is set to shrink it further. He offers a sector-level instance rather than a proof: surveyed about what most worried them, developers and utilities building large solar portfolios put labor availability above every other constraint. From there he asserts, without testing alternatives, that the only plausible route to meeting energy needs and staying competitive is a leap in physical labor productivity.

Physical AI, a label he says the field is beginning to adopt, is his answer, and he defines it as models whose inputs and outputs plug directly into physical sensors and actuators. He is careful that it is not new. Since the deep learning era began around 2010, industrial operators have used it to read patterns in machine data better than people can, which is why expensive and hard-to-service equipment has been steadily sensored up. Reading machines, though, still deploys a fully human workforce, so the productivity argument pushes on to robots, which he orders along a spectrum of capability. At the easy end are discrete, repetitive, coarse tasks that can be fully programmed, where millions of arms already work. At the hard end is his plumber: an upset customer, an ill-defined problem, a messy decades-old system, diagnosis from a few observations, and then agility, brute strength and fine motor control at once. We are not approaching plumber. But he argues three trends have pushed the frontier hard in five years. Robots have moved out of factories into warehouses and into drones, both far less controlled environments. Machine learning advances have reached robotics from two directions, both through generative models given robot sensor data to ingest and through separate work on imitation learning by teleoperation. And a decade of autonomous vehicle spending has thrown off engineers trained at the frontier along with much cheaper sensing, chips and batteries.

The one place he breaks with the industry’s enthusiasm is form factor. He grants the humanoid thesis its logic, that a world built by humans for humans should suit a machine shaped like one, and that humanoids suit imitation learning. He then says he is not so sure, and expects more value from a wide range of bodies each engineered for its own tasks, on the analogy of evolution filling ecological niches. What he does not supply anywhere is a timeline or a cost for any of it.