The next industrial revolution and the “art of life itself”

The talk is adapted from his annual keynote at his firm’s meeting, and it hangs on Keynes’s 1931 essay “Economic Possibilities for our Grandchildren”. Keynes made two speculations about a century ahead. The first, that the average person would end up eight times better off, has come in almost exactly right. The second, that the resulting abundance would leave people bored and purposeless and needing fifteen-hour weeks to fill the time, was wrong, and Lubershane’s diagnosis of why is the engine of everything that follows. Keynes misjudged humanity’s aggregate elasticity of demand: he assumed we would take the gains from productivity as leisure, and instead we took them as consumption, and no ceiling on that appetite has yet been found.

He then splits industrial revolutions into two branches, increasing resource productivity and developing new capabilities, while granting the line between them is blurry, since radical efficiency can amount to a break from the past and a new capability usually needs decades of incremental efficiency before it is viable. The distinction earns its keep because the downstream effects differ: the first improves economic efficiency, the second creates new paradigms for how life is lived. On that reading the first revolution was mostly productivity, the second mostly new capability and the one whose products we actually love, and the third, computing and communications, a mixed bag he judges much less impactful than either. That is what sets up the question he puts to AI: which of the three will it resemble?

The lessons he draws carry the reasoning. From the first revolution: people resist technology that threatens their livelihood, which is why the Luddites left the impression they did; and a large gain in one factor of production unbottlenecks demand for the others, with consequences nobody can forecast. Cheap British fabric pulled demand for fiber across the Atlantic, which he traces both to a cotton harvest dependent on chattel slavery and, much later, to Levi’s, and his lesson is humility about second and third order effects. From the second: the technology we love most needs no tedious search for use cases because its benefits are self-evident, and he says plainly that the subtext is AI. From the third comes the sharpest tool in the piece. Two variables decide what automation does to employment, how elastic demand is and how complementary human labor remains, and both are hard to answer in advance. The electronic spreadsheet cut bookkeeping work sharply and, he argues, still created more jobs than it removed, and bank teller numbers actually rose as cash machines spread, in both cases because demand turned out to be enormously elastic and humans stayed useful alongside the machine. Mobile banking then did to tellers what the cash machine had not, because demand kept growing but a smartphone app has little use for a human. So the question is not how much work can be automated but how complementary the remaining human work is. His final lesson is that digital technology has consistently struggled to get beyond the computer screen, which is why productivity rose in desk work and in a few marketplace services and stayed flat in construction, manufacturing and infrastructure, and why transforming the physical world takes innovation in the physical world.

What makes him lean toward a fourth revolution is partly that AI is showing up in labor data and partly that it is not staying on the screen. In professional and business services, output and employment have nearly always moved together outside recessions; since ChatGPT arrived, he says, the gap has widened during an expansion for the first time in memory, output rising while employment falls. Recent college graduates, who have always found work more easily than the average worker, lost that advantage over the same period. And the same techniques are being embodied in machines, the physical AI he sees as the route by which AI reaches the sectors the last revolution bypassed. That takes him back to his first lesson, and he says he is beginning to worry that the response to this revolution could be much more disruptive than the brief Luddite rebellion, adding that the pushback against data center development, which he says seems to be accelerating across many parts of the US, may already be a form of it. The answer to his cotton question is therefore the electron, and the caution he attaches concerns the production process rather than the demand: chip efficiency improved enormously over a decade but the best chips have mostly stalled since around 2024, and nobody can say how much more efficient the algorithms will get, so the power demand curve cannot be projected far.