Have you ever considered how expensive your thoughts are? I mean, quite literally, how expensive is it — in U.S. dollars — to think?
It turns out the answer depends sharply on what you consider a “thought.” Is a thought just the ability to put words or numbers in a meaningful order? Or is a thought, more broadly, defined by the neuroscientific experience that happens when your brain tells your body to move? Or is it better to consider the maximum computational capacity of your brain?
The question of what makes something a thought is far more substantial than it may seem. Embedded in it is what distinguishes human from artificial intelligence. The account often used by AI apocalypticists and optimists who say that AI will replace every job is one according to which intelligence is defined by maximum computational output. But a better definition would be one that focuses on the defining characteristic of humanity: our hands.
A recent study suggests that the maximum computational capacity of our conscious thoughts is astonishingly low: we can think at a mere 10 bits per second.1 Granted, that estimate is not a consensus throughout the scientific community — after all, a single character of text has about one bit of information when considering context, and most of us can comprehend literature faster than two words per second.
But whether or not we accept that particular study’s estimate, the conscious brain is hardly the speediest part of the human mind. Rather, what is astonishing is the number of “calculations” one does subconsciously: picking up an egg delicately enough to not break it, typing out individual characters, each in less than 80 milliseconds, or tapping one’s foot on the ground. Man’s brain is capable of one quadrillion to one quintillion computations per second — which is about as many as the world’s fastest supercomputer. And yet, the brain runs on only 20 watts, which requires — in the not-so-cheap town of Cambridge — about 30 to 40 dollars worth of food per day. Compare this to the cost of the world’s fastest supercomputer, which demands almost $100,000 per day in order to run. That is a difference of over three orders of magnitude.
This is not to say that man is always more efficient than machine — AI has become a trillion-dollar industry precisely because this is not the case. But this has more to do with the fact that certain problems require far fewer computations on a circuit of transistors than on one with neurons. For example, when a man computes the sum of two four-digit numbers, he is subconsciously performing billions of computations — first in his visual cortex to map each digit to its representation, and only then actually adding up the numbers. A computer, meanwhile, only needs to perform fifteen computations (merely the addition of two numbers that can be represented in 14 binary bits). Human minds are not optimized for this type of rigid computation.
Indeed, the speed with which computers can manipulate numbers has been the basis of machine learning. If one has enough numbers, the theory goes, then complex behavior will emerge. And, so far, it has — at least in disciplines such as mathematics and computer science, in which rigid definitions constrain the possible set of outcomes.
But what happens when the output space becomes practically infinite, like in the case of a human hand’s subtle movements? Robotics is an incredibly difficult task. Sure, we have robots that can throw a baseball faster than a human, that can pick up an egg, and that can climb mountains, but are there any that can do all three? No, and that is precisely because being able to ambulate with several joints all rotating differently, while not breaking any internal machinery, requires far too many computations for a machine to do efficiently.
This is to say that there are tasks for which AI, in any reasonably foreseeable form, cannot compete with humans. The manipulation of objects is one of those tasks, and it is essential to remember this fact when discussing the role man plays in the economy and in society.
Discussion in the past few years surrounding AI and jobs has revolved around, to put it kindly, desk work. Perhaps unsurprisingly, writers and scientists worry that, because workers who spend 95 percent of their time typing on computers will be replaced by AI, this means the concept of “work” itself is ruined. But such fearmongering relies on a peculiar notion of what “work” really is.
For the past fifty or so years, the jobs that paid the best were “desk jobs” like writing and programming, and so — according to these same white-collar professionals writing articles and being quoted in them — if AI can do these jobs, surely it can do everything else too. But this “everything else” is precisely the thorn in AI’s side, or perhaps the glimmering sliver of hope for humanity. Desk jobs, ironically enough, since they are often described as “skilled labor,” require the fewest computations for an AI to simulate. It is incredibly easy for ChatGPT to predict the next word in a sentence one billion times. It is much harder for a robot to pick up a 100-micron-thick glass square with a pair of tweezers and place it perfectly in the middle of a centrifuge.2
It might be the case that a sufficiently advanced machine learning model could replace programmers or investment banking analysts in five years, but it would be astounding if there were a robot in 2040 that could work in a nanofabrication lab or crochet complex patterns. These jobs are not only technically difficult to automate. Replacing workers with machines in these industries also poses a fundamental economic problem: a robot with enough computational power to run around a nanolab and gracefully manipulate objects would require far more energy (and thus cost much more) than an equivalent human. Why, then, would any employer “hire” the robot?
They would not. In fact, man’s hands — perhaps his second greatest asset as a species behind his mind — are safe from the ravages of robotics, at least for the foreseeable future.
Thus, in some sense, AI will not so much “change the way we work” as return man to the way he used to work from the Paleolithic all the way up to the Industrial Revolution: with his hands, not purely with his mind. Man was always meant to work with the tangible string in front of him, not the squiggles on a screen.
In Paleolithic times, we did not type or write carefully; we did not even spend eight hours a day staring at an unmoving box. Rather, we carved stone tools, collected fresh berries, hunted down wild game, and spent our spare time storytelling and cave-painting. Up until the Industrial Revolution, the greatest citizens were warriors and artisans, both professions that required mastery of tools manipulated with the human hand.
And in the future, it will be the hand that dominates the mind yet again: researchers carefully manipulating E. coli colonies in a lab or plumbers feeling their way around corroded pipes. Even if AI bests us in cerebral tasks, it will never truly operate (physically) at the same speed as us.
Jieyu Zheng and Markus Meister, “The unbearable slowness of being: Why do we live at 10 bits?,” Neuron, 113, no. 2 (2025), https://doi.org/10.1016/j.neuron.2024.11.008.
This is a procedure commonly used in nano-fabrication, wherein a researcher will have to grab a 1 cm by 1 cm square of glass that is wafer-thin (without cracking it) and place it onto a small pole about 5 mm in diameter, centering it within about 0.1 mm. It’s a pretty difficult task to do.


![[Man Operating Brush Drawing Machine], Gilbert & Bacon, Albumen silver print [Man Operating Brush Drawing Machine], Gilbert & Bacon, Albumen silver print](https://substackcdn.com/image/fetch/$s_!C3wv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfa6ca98-2a8c-487c-8eb5-965ea910a363_797x1200.jpeg)