翻訳待ち：AGI Will Set Off an Industrial Explosion

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

AI Frontiers Aug 11, 2026 Damon Binder, Senior Researcher at Coefficient Giving — August 11, 2026 AI systems now write so fluently that cheating at universities has become ubiquitous and “AI slop” is displacing human writing across the internet. They have begun producing original proofs on long-standing open problems in mathematics. Researchers at the frontier AI companies have largely stopped writing their own code, and the models’ hacking abilities are strong enough that the US government temporarily export-controlled Anthropic’s Claude Fable 5, days after its release. Ten years ago, every claim in this paragraph would have sounded like science fiction. It is easy to become numb to this progress. Today’s AI agents remain endearingly clumsy; watching Claude run a small store is a bit like watching a child running a lemonade stand. But we should not let such awkwardness distract from the trend. These systems have evolved from research curiosities to extremely powerful economic engines in 10 years, with no end in sight. I want to take that trend seriously and ask: if AI does get to the stage where it can do the cognitive work that humans do, what happens to the economy? With notable exceptions, economists have mostly declined to address this question; their usual maneuver is to deny the premise, as when a widely cited Nobel laureate’s analysis put AI’s contribution to US GDP at about 1% over the coming decade, by essentially freezing 2023 capabilities for 10 years. But the question is worth taking seriously, because automating cognitive labor releases what has always been the critical brake on physical production: no matter how cheap machines and tools become, you cannot manufacture new workers. In this piece, I will assume no additional new technologies and no recursive self-improvement to superintelligence. Yet, even with these conservative assumptions, I will show that the economy stands to be profoundly transformed. For the first time in history, physical production could be fully automated. The rest of this piece makes that case and considers the implications. I argue that AI capable of cognitive work could also do physical work (since building the machinery it would need is relatively straightforward) and that its labor would be cheap. I then turn to input-output tables—the US government’s accounting of what every industry buys from every other—to ask how fast a fully automated economy could grow, finding that physical output would double roughly every year rather than every few decades (as it does currently). Finally I ask whether anything could stop such an industrial explosion, and what the implications are for power at home and abroad. Why AGI Gives You Robots The term “artificial general intelligence” (AGI) has been endlessly used and abused, but it points to a clear core concept: an AI smart enough to do the work humans can do, at a competitive cost. You could hand it a task, as you would to a skilled employee or contractor, and expect the job done competently. Current systems, for all their rapid progress, are not close to AGI; my own attempts to automate my job have been unsuccessful. However, once AGI appears, we should expect it to be able to perform physical work through robotics. An AI system that can do every remote job can also operate a robot. Definitions of AGI focus on cognitive work, but there is nothing special about the physical world that confines intelligent agents to their data centers. Remote cognitive work is hard: doing all of it means mastering the real-time control, spatial reasoning, physical prediction, and continual learning required for such varied tasks as mechanical engineering and animation. A system with those skills has what it needs to operate machinery, if given actuators to work with. Progress in AI is already pulling robotics forward. The general-purpose methods that cracked language and vision turn out to be good at manual manipulation too. Large language models can increasingly supply high-level control: today they run robot arms, fly drones, and program robot dogs. Hobbyists have even wired them into homemade robots. Meanwhile, the same training methods, applied to recordings of humans teleoperating machines, are producing the low-level dexterous control that humans perform instinctively; today’s systems can fold laundry and assemble cardboard boxes, using ordinary cameras and simple grippers. Hardware is not the bottleneck for robotic automation. Industrial arms have demonstrated precision and force beyond any human arm for over half a century. But, because they could neither see nor think, every motion had to be programmed in advance, at massive cost. The other work-around was a human operator: remote manipulators have handled radioactive material since the late 1940s, built underwater structures using remote-controlled submersibles, and performed surgery. That remote control was worthwhile only when a person could not be physically present. But it does show something we’ve known for decades: a machine under competent control can do skilled physical work. Once control can come from AIs rather than people, it becomes worth building general-purpose machinery for them. Robots will not necessarily resemble humans. While the human hand is a remarkable instrument, robot hands are already fairly good, and certainly better than the split-hook prostheses—two rigid fingers on a cable with no sense of touch—used by amputees to farm, weld, and perform all manner of other tasks. With AI, simple actuators go far, and AGIs will have intelligence and patience to spare. Nor will the machines work at human workstations forever. Production demands dexterity today because every process was designed around human workers with hands. New possibilities open up when the workers are machines; just as today’s programming agents do not type at keyboards as their human counterparts do, robotics tools need not be gripped and triggered by fingers—they could be mounted directly on a robotic arm instead. Machine Labor Will Be Cheap Automating labor is not free. But the computer chips and robotic actuators that replace workers are simply more capital goods for the economy to produce—and they’re not particularly expensive ones. A humanoid robot, broadly speaking, uses parts similar to those of a car: metal, motors, batteries, electronics, and sensors. At automotive production volumes, it should cost tens of thousands of dollars; Unitree already sells its child-sized G1 humanoid for around $13,500, and Tesla is reportedly aiming for $20,000 for its full-size Optimus. If a robot costs $30,000 and, working around the clock, can substitute for a single human worker who costs $30 an hour, it could pay for itself in six weeks. AI cognitive labor is generally significantly cheaper than the humans it replaces. It is hard to price the future AI cognitive labor involved in an industrial explosion, because it does not exist yet. However, on the tasks AI systems can already do—like transcription, routine translation, or writing one-off scripts—they are typically far cheaper, per unit, than the people they replace, and the cost of a fixed level of AI capability falls rapidly year over year. Even if the first AGI arrived merely cost-competitive with human workers, within a year it would cost a fraction as much. Input-Output Analysis How fast would the economy grow once human labor is automated? At first blush, this looks hard to answer. Standard growth models assume that labor and capital substitute for each other, treating each as a single dollar-denominated aggregate—barristers and bricklayers combined into “labor,” bulldozers and bridges into “capital.” Such abstractions are poorly equipped for a world where labor is unnecessary and capital reproduces itself, and classic concepts like GDP can become profoundly misleading. Input-output analysis helps forecast how an automated economy would grow. Input-output analysis takes a different approach, recording what each industry physically needs from others. The method was developed by Wassily Leontief, in the 1930s, to study the relationships between industrial sectors. In 1945, Leontief used his tables to project the United States’ 1950 steel requirements to within a couple of percentage points. Military planners, concerned about war with the Soviet Union, took up the method to study how fast industry could remobilize. Input-output analysis tracks what everyone buys from everyone else. The basic idea is simple: ask every business what it buys and from whom, then aggregate the responses into tables that track what each industry buys from others. An entry for aluminum smelting might record the electricity, ore, and machining required to produce aluminum, while an entry for aircraft manufacturing records how much aluminum is required. The Census Bureau runs this survey every five years, and the Bureau of Economic Analysis assembles the results. The tables take years to build, so the most recent set covers 2017. Companion tables record the equipment and structures that each industry holds, from machine tools to chip fabs. Entries are recorded in dollars, but they are detailed enough to closely follow the underlying kilowatt-hours of electricity, tons of steel, and other physical quantities. Most output is ultimately consumed; the rest is reinvested, either adding to physical capital or replacing what has worn out. Once robots can replace human labor, the economy will no longer be bound by a fixed workforce. It can produce every input it needs, including the “workers” themselves, so output can be fed back into building more capacity, and growth compounds. John von Neumann worked out how to compute the maximum rate at which such an economy could expand. Combining his analysis with the production data in the tables tells us how fast an autonomous industrial sector could grow. Input-output analysis measures economic growth in physical output, not value. This approach to forecasting growth tracks how much stuff can be produced. I make no attempt to convert this into GDP, because an industrial explosion would upend the prices any such conversion relies on: when machines produce everything, including more machines, goods become exponentially cheap. Physical output is also the better guide to what is at stake. Military power, human employment, and material abundance all depend on how much gets built, rather than on its dollar price. How Fast Could an Autonomous Economy Grow? Using the US government’s 2017 input-output tables, which track 402 industries, I find that a fully automated economy using US production methods could double its output roughly every year; input-output tables for other advanced economies give comparable results. The robots and computer chips needed to automate production represent only a small fraction of total output; even a tenfold increase in their cost would not impact the growth rate significantly. Construction lags can be incorporated too; even with these delays included, the economy still doubles in well under two years. Projected physical output of four key industrial sectors in the years after AGI arrives, relative to today’s level. Each scenario assumes that new capacity takes zero, six months, or a year to build before it starts producing output. Solid lines hold human consumption at today’s level and reinvest all output above that level into industry; dashed lines split that output 50-50 between industry and human consumption. A real buildout would differ somewhat from these projections: it would be slower initially as AI capabilities diffuse throughout the economy, but accelerate later as production methods adapt to robot workers. Source: author’s calculations from the 2017 US input-output tables, as described in Part 2 of the author’s memo series. ‍Why does a fully automated economy accelerate so much faster than the current one? The main [truncated for AI cost control]