On linemen, robots, hands, and brains
Hands are still the hardest
“The challenge is that the space of ideas, semantic concepts, words… is actually constrained. There are fewer words in the English language than there are valid positions of just a single hand. The space of physical interactions in the world is unbounded.” - Brad Porter, CEO and founder of Cobot
Regular Steel For Fuel readers know that I have high hopes for robotics. In my recent “Ode to Physical AI”, I pointed out that labor productivity in the physical economy is on track to become one of the biggest constraints on the global energy transition:
“Across the economy, the biggest bottlenecks to growth are generally not PhD level bottlenecks. The digital revolution of the past 25 years has already made anyone working primarily with data & ideas — i.e. “knowledge workers” — tremendously more productive. As discussed above, LLMs are continuing this trend. During that same period, however, labor productivity in many “blue collar” fields has stagnated. In manufacturing and construction — which are the foundations of our ability to shape the physical world — labor productivity in the United States has been flat to declining for the past two decades.”
One quick example: Just yesterday, I attended a meeting of about 25 executives from the electric utility business who are responsible for transmission & distribution infrastructure. As we discussed various challenges to improving reliability and resilience in the industry, one member of the group stated that his biggest problem is now the availability of skilled labor. Specifically, his challenge is the limited availability of linemen: the guys who climb up utility poles to install and maintain electrical equipment. Paraphrasing this particular utility executive…
“I’ve knocked on every door. There are no more trained linemen available in the state.”
In my Ode to Physical AI, I introduced a framework for the role which robotics could play in addressing this type of labor shortage: The Spectrum of Robotic Capabilities.
At one end of the spectrum are discrete, repetitive tasks with minimal variability, and only the coarsest levels of object manipulation. These tasks can be fully programmed as a series of concrete steps. Picture a robotic arm in a factory, which spends its days picking up uniform pieces of sheet metal from the end of a conveyor belt, rotating them, and then stacking them in a pile.
This kind of robot doesn’t depend on fancy AI at all.
At the other end of The Spectrum, there are much more difficult forms of human labor to automate. These jobs typically require navigating highly variable, uncertain environments — those in which there are more exceptions than there are rules. Crucially, many of these jobs also require an extremely fine degree of object manipulation. My personal mental model for this type of job is a plumber; but most other skilled trades require similar levels of manual dexterity, and operate in similarly messy work environments.1 Linemen certainly fall into this category.
This kind of robot is simply inconceivable without additional advances in AI. While the field of robotics has made great strides in the past few years — for example, see autonomous vehicles — we are still at least one more giant leap away from robotic plumbers or linemen. I personally believe we’re several leaps away.
Why? The answer mostly comes down to hands and brains.
Human hands are extraordinarily dextrous, while human brains are extraordinarily efficient. This combination suggests that AI models will have an extraordinarily difficult time matching the capabilities of humans working with their hands.
Human hands are extraordinarily dextrous
I kicked off this post with a quote from Brad Porter, who is a highly distinguished roboticist. (Quite literally, one of his former job titles on Linkedin is “Vice President & Distinguished Engineer, Robotics” at Amazon.) Porter’s recent article highlights a notoriously difficult problem in the field of robotics: the challenge of applying deep learning techniques to the movement of a human hand.
As Brad points out, the number of degrees of freedom available to just one single human hand is unfathomably high. My EIP colleague Anil Achyuta got me thinking about this; and just to be sure, we crunched some numbers.
The human hand has 27 joints. Because the physical world is continuous, not discrete, each of those joints could theoretically be positioned in an infinite number of ways. Yet even if we simplify the world by collapsing an infinite set of positions into just ten distinct positions per joint, we still end up with 10^27 unique hand arrangements.
Compare that number (an “octillion”) to the number of words in the Oxford English dictionary. The OED contains about 500,000 words, but we use far fewer in regular conversation.
There’s more. According to the National Institutes of Health, the human hand also has about 17,000 touch receptors, which works out to approximately 40 receptors per square centimeter. Additionally, our hands have internal nerve clusters which contribute to “proprioception” (one of my favorite uncommonly used words in the OED) — which is the sensation of how our bodies are positioned. I’m not even going to try to calculate the number of distinct sensory combinations that we can perceive.
Try making an affordable robotic hand with 40 touch receptors per square centimeter.
Human brains are extraordinarily efficient
A typical human child requires somewhere in the ballpark of 15 million calories worth of food to mature into a fully grown adult (age 21). About 20% of that is allocated to the brain.
Hence, “training” a mature human brain requires about 3.5 megawatt-hours worth of energy input. Reportedly, training GPT-4 consumed about 50 gigawatt-hours of energy, which is about 14,000 times as much.
In short: Using about four orders of magnitude less energy than an LLM, any regular, standard-issue human brain can achieve the same linguistic and visual skills as Chat-GPT, with (mostly) superior reasoning capability. Given that extremely limited energy budget, the human brain also learns to manipulate hands… Plus all of the other body parts. It also learns to sense and manage emotions. And during this entire time, in addition to learning, the brain is also taking action — in the parlance of AI, a child’s brain is simultaneously engaged in training and “inference”.
To reiterate: A child’s brain accomplishes all of this, and more, with less energy than a single large wind turbine generates in an hour.
And it’s not just the brain’s energy efficiency that’s extraordinary — it’s also the brain’s “training data” efficiency. A child in a talkative household hears roughly 20,000 words every day, which comes out to roughly 150 million words on the way to adulthood. Meanwhile, the number of words that leading edge LLMs require for training is in the trillions.
What if we extrapolate this level of training data efficiency to the hand coordination problem? A human child gets 184,000 hours of practice using his or her hands before the age of 21. Let’s assume that a “robotic hand model” could achieve the same level of training data efficiency, relative to a human brain, as an LLM. (I believe this is an extremely optimistic assumption, given all of the challenges to modeling a robotic hand I discussed above.) Yet, even in that optimistic scenario, our robotic hand model would require billions of hours of training.
Because hands are so hard…
I can think of a few consequences:
Humanoid robots have a long, long way to go. I remain skeptical that humanoids will be more valuable, at least within the next five years, than specialized robots whose form factors are built-for-purpose.
“The trades” are almost certainly going to be the professions that are most insulated from the labor market effects of AI.
If you’re a young person reading this, and you’re not fully convinced that desk work is for you, consider becoming a utility lineman (or line-woman). The industry needs you!
Consider the variability that a typical plumber encounters on a typical household job site. For starters, he probably meets an upset customer (another messy human) with an ill-defined problem (e.g. “My sink is leaking, and I’m not quite sure where”). Then he enters a messy work environment — perhaps under a sink, or in a basement crawl space — and encounters a messy system, with components installed over the course of multiple decades. He typically needs to diagnose the problem with just a few observations. Then, finally, he needs to get to work on a solution, which probably requires a combination of agility, brute strength, and fine motor skills. For example, the plumber may need to snake his arm into an awkward space — while holding a wrench — in order to gently tighten a valve by just the right amount.”
Don't you think we can train linemen faster than we can develop and deploy robots to do their jobs?