On linemen, robots, hands, and brains

The premise is quoted from his own earlier post: digital tools have made knowledge workers far more productive while labor productivity in the physical trades has stagnated, which he says puts it on track to become one of the biggest constraints on the global energy transition. A meeting the day before makes it concrete. Among about twenty-five utility executives responsible for transmission and distribution, one said his biggest problem was now finding linemen, having knocked on every door with no trained linemen left in his state. The framework, also carried over from that post, is a spectrum of robotic capability. At one end are discrete, repetitive, low-variability tasks that can be fully programmed and need no advanced AI at all. At the other are jobs in highly variable environments with more exceptions than rules, which also demand very fine object manipulation. His mental model there is a plumber, and linemen fall in the same category. That end is inconceivable without further AI advances; despite real progress in the field, autonomous vehicles among it, he puts robotic plumbers and linemen at least one giant leap away and says he personally believes several.

Why so far? Hands and brains. On hands he builds on a roboticist he quotes at the top: the space of words is bounded, the space of physical interaction is not. Even after collapsing each of the hand’s joints into a handful of discrete positions, the number of possible hand arrangements dwarfs the English vocabulary by an unimaginable margin, and that is before counting the hand’s touch receptors and its sense of its own position, which he declines to try to quantify.

On brains, two efficiency comparisons do the work and the second is the turn. Converting the food a child eats to adulthood into energy and taking the brain’s share, he puts the energy cost of training a human brain roughly four orders of magnitude below the reported energy cost of training a leading language model, while the child’s brain also learns bodies and emotions and acts throughout, doing training and inference at once. It is as frugal with data, learning language from roughly 150 million words where a model needs trillions. Then he extrapolates to hands, and that is where the answer comes from: even granting a robotic hand model the same data efficiency relative to a brain that a language model achieves, which he flags as extremely optimistic, training one would still take billions of hours. Hence humanoid robots have a long way to go, and he doubts they will beat purpose-built specialized robots within five years; the trades are the professions most insulated from AI in the labor market; and a young person unsure about desk work should consider becoming a lineman.