There is a better than even chance that somewhere in the last minute, your mind left this page without asking permission, and without telling you it had gone. Not for anywhere dramatic, just away: a half-thought about later, a pull toward the phone, a small worry that surfaced and sank before you noticed it arrive. You came back, the way you always come back. But for a moment you were somewhere your body was not, you did not decide to leave, and you may not have caught yourself leaving at all.
Multiply that moment by a lifetime and it stops being a moment. The measured figure is that the human mind spends close to half its waking hours adrift like this, and the number does not describe the distracted, the anxious, or the unwell. It describes the species, and it describes you, on an ordinary day, doing ordinary things. Stacked end to end it comes to about two decades: awake, and absent.
For almost all of recorded history, that drift was filed under private failings: laziness, weak will, a personal inability to concentrate. It is older and stranger than that. It is one of the first problems human beings ever wrote down about themselves, and across forty-five centuries of naming it, no one has solved it. Now it has become the most expensive problem in the world. I will make one claim you can check and hold against me: the machines everyone is celebrating do not just compete for your attention. They are quietly removing your reason to finish, and you will thank them while it happens.
01The constraint just moved
I have modeled complex systems like supply chains of semiconductors and energy grids, and the first lesson of that work is that a system’s output is set by its binding constraint, not by its most abundant input. When I modeled the U.S. semiconductor supply chain, the result was blunt: subsidizing fabrication barely moves output when advanced packaging is the constraint. Pour capital into the abundant stage and the shortage does not move. The constraint sits there, invisible, setting the ceiling, while everyone celebrates new capacity that cannot ship.
I ran the model many ways. The answer did not change.
The same analysis now applies to human cognition, and almost everyone is subsidizing the wrong input. For all of history, intelligence was the binding constraint on what a person could produce. Information was expensive, expertise sat in a few heads, and sustained reasoning was rare enough to be valuable. That constraint just broke. AI puts on a phone what recently required a trained profession: synthesis, code, analysis, drafting, design. The cost of a fixed unit of machine cognition has been falling by large multiples every year, and open models now trail the frontier by months, not decades. If that trajectory stalls, the clock on this argument stalls with it; nothing here requires the machines to be perfect, only cheap. And when an input goes from scarce to abundant, the constraint does not disappear.
It moves.
Where it moved is the real question, and the common answer, focus, is wrong in an instructive way. The new constraint is command: the ability to choose what your mind works on, and to carry that choice through to something finished. Choosing, and finishing. AI supplies everything else. It cannot supply either of those, and the evidence is already in that it is corroding the second one.
This constraint did not arrive with the smartphone. It is the oldest documented problem in the human record, and the long record of failing to solve it is itself part of the evidence.
02The oldest complaint on record
Treat the ancient texts as data, because that is what they are: independent observations of the same phenomenon by cultures that could not have copied each other, separated by oceans, millennia, and languages that would not be mutually translatable for thousands of years. In science, independent replication is the strongest form of evidence. Here is what replicated.
The ancients described it. Egypt, around 2400 BCE: the Maxims of Ptahhotep, among the oldest wisdom texts that survive, already distinguishes hearing from listening, sound arriving at the ear versus attention sinking in and taking root, and it stakes a person’s whole standing on the difference. India, fifth century BCE: the Dhammapada gives the problem its own chapter and its most durable images. The mind is “flickering, fickle, difficult to guard, difficult to control.” The wise person works it straight the way a fletcher straightens an arrow; untrained, it thrashes like a fish pulled from water. Two centuries later, Patanjali’s Yoga Sutras compress the entire discipline of yoga into one definition, in its second line: yoga is the stilling of the fluctuations of the mind. Not posture. Not breath. The quieting of mental motion itself. The Bhagavad Gita stages the difficulty as dialogue: the warrior Arjuna protests that the mind is “restless, turbulent, strong and obstinate,” harder to hold than the wind, and Krishna does not argue. He prescribes decades: steady practice and a loosened grip on outcomes.
China, fourth century BCE: the Zhuangzi teaches a “fasting of the mind,” stop listening with the ears, then stop listening even with the busy, calculating mind. Its portrait of mastery is Cook Ding, whose blade never dulls because after years of practice he no longer hacks at the carcass; the knife finds the gaps the joints already provide.
What the ancients described, the later traditions moralized: the wandering mind stopped being a difficulty and became a failing. Rome, 170s CE: Marcus Aurelius, running an empire from a war camp, writes himself the same order on page after page of his private notebook, give vigorous attention to the task in front of you, perform each act as if it were your last. For him the failure to attend is not a lapse of productivity. It is a moral failure. The Abrahamic traditions file the same verdict in their own vocabulary. Psalm 46 issues the cure as a command: “Be still, and know.” The Letter of James sketches the failure state in one line that has never been improved on: “A double minded man is unstable in all his ways.” Al-Ghazali, in the twelfth century, separates the outward motions of prayer from the inward presence of the heart, and rules that the same act performed with attention and without it is not the same act at all.
Only in the modern era did anyone stop prescribing long enough to define and measure. William James, in 1890, gave the faculty its lasting description: attention is the mind taking possession, in clear and vivid form, of one object out of several possible ones, and he went further, venturing that volition itself is nothing but attention. A century later, Mihaly Csikszentmihalyi named the rare opposite of the wandering mind: flow, the state of involvement so total that the sense of self dissolves and time bends, and he found across decades of research that the gap between ordinary and extraordinary output often reduced to how reliably a person could enter that state and stay.
Same observation, five thousand years, every literate civilization. That part of the record is settled. The interesting part is what they did about it.
03The same wrong cure
Because they did not just converge on the complaint. They converged on the prescription. Still it, said Patanjali. Straighten it, said the Dhammapada. Empty it, said Zhuangzi. Subdue it, said the Gita. Across every language and god and century, the cure was suppression: the wandering itself was the enemy, to be conquered by discipline. And on that single point, the entire record is wrong, uniformly, for four and a half thousand years. The error did not die with the ancients. It survives in every modern program to eliminate distraction and defeat procrastination, the same misdiagnosis in newer clothes.
Here is what they could not have known. The brain network that activates the moment focused attention drops, the default mode network, is not idle circuitry. It is where planning, imagination, self-reflection, and the recombination of ideas largely happen. The mind that loses the thread of a meeting is the same mind that solves the problem in the shower, connects two ideas nobody had joined, and consolidates the day into memory during sleep. A mind that could never wander could never invent. Evolution kept the wandering because the wandering pays.
That is why forty-five centuries of trying to suppress it produced discipline, ritual, and no solution. The campaign was aimed at the wrong target.
The variable that matters is not whether the mind wanders. It is whether you can choose. Wandering you selected is doing useful background work. Wandering imposed by a notification timed to fire at the worst moment is interference wearing the same clothes. The neurons fire identically either way; only the operator can tell the difference, and only if the operator is awake. That capacity, to roam when roaming serves you and to hold when holding serves you, and to know which mode you are in, is command.
04The spec sheet
The ancients could describe the constraint. They could not measure it. We can, and the measurements read like a specification nobody would sign off on.
Duty cycle: the “close to half” in the opening is a measured figure. In 2010, Killingsworth and Gilbert sampled thousands of people at random moments and asked whether their mind was on what they were doing. It was elsewhere 47 percent of the time, across nearly every activity. One study, secondary findings contested, but the magnitude holds across the literature.
Working memory: Miller’s 1956 estimate of seven items was revised downward by later work, much of it Cowan’s, to roughly four. The entire stage on which deliberate reasoning happens holds about four things at once.
Environment: Gloria Mark’s lab tracked attention on a single screen falling from about two and a half minutes in 2004, to 75 seconds by 2012, to roughly 47 seconds in recent measurements, while full recovery from a real interruption runs twenty minutes or more. Herbert Simon saw the shape of this in 1971, before the internet existed: a wealth of information creates a poverty of attention. Every device built since has deepened exactly the poverty he predicted.
These are just numbers, until you stack them. A processor off-task half the time, a four-item workspace, an environment that breaks focus in under a minute, and a recovery cost measured in tens of minutes. None of it is a defect; it is the architecture. But it is the architecture something else has now learned to exploit at industrial scale, and to weigh what is at stake, look first at what this same architecture, properly commanded, has already built.
05The faculty that built everything
Take anything human-made and ask, mechanically, what produced it. Two events, in sequence, inside a skull. First a mind wandered far enough to find the idea, the connection nobody had joined. Then a mind held still on that idea, against the pull of everything else, long enough to build it into something that outlasted the moment. Agriculture required watching seasons closely enough, for long enough, to plant before starving instead of after. Writing required holding a sound in mind long enough to bind it to a mark. The microprocessor rendering this sentence is the residue of decades of minds that refused to look away from problems measured in billionths of a meter.
None of this is mystical. Other animals attend; a hawk locks onto a rabbit. The distinctly human move is pointing attention at a chosen abstraction, something with no smell and no motion, a proof, a load-bearing arch, a legal principle, and holding it there across hours and years until something exists that did not. Language, tools, institutions, and science are downstream of that one trick. Civilization is not a warehouse of objects. It is compounded directed attention: millions of instances of a mind choosing a problem and staying with it, accreted into everything you can see from where you sit.
The ledger has two sides. Wandering supplies the ideas. Holding does the building. The species that could do both on command built the world, which is why it should alarm us that the newest tools are aimed, with precision, at the second half.
06Start inflation
Feeds and slot machines take your attention. That threat is old, well named, and morally legible: they are thieves, and we have a vocabulary for thieves. What is arriving now is different in kind, and harder to defend against, because it does not steal your hours. It removes your reason to finish them.
Ask a question and a complete-looking answer appears before you finish forming the thought. The essay drafts itself. The analysis assembles itself. The code writes itself to roughly working. In that exchange, the hardest and rarest half of command, the staying-with, the carrying-through, is lifted off you and performed by the machine. What remains for the human is the starting.
Starting was never the problem.
Look at the evidence you already own. There is a folder on your computer you do not open: the half-built app, the abandoned draft, the analysis that got to almost, the course you stopped at lesson three, the repo with one commit. Ten years ago that folder grew slowly, because starting anything cost enough that you thought before you began. Now starting is free, and the folder is filling faster than it has ever filled. Nobody warned you, because it does not feel like a problem. It feels like productivity.
Call the result start inflation. When the cost of starting collapses, starts flood the market, and like any inflated currency they buy less and less. A prototype is no longer evidence of anything. A draft is no longer a commitment. Twelve open projects signal nothing except access to the same tools as everyone else. Meanwhile the scarce asset, a thing actually carried to done, appreciates by the month. Finished work is becoming the only hard currency left, at the precise moment the machines are training everyone to hold soft currency in unlimited amounts.
This is not introspection; it fits a measured pattern. Offloading a cognitive faculty to a tool degrades the faculty. Habitual GPS use tracks with worse spatial memory, and a small longitudinal follow-up pointed at the use coming first and the decline following. Sparrow’s 2011 study showed that when people expect information to stay retrievable, they remember where to find it instead of the thing itself; the mind stops storing what it trusts the machine to hold. And the trade has now been measured next door to finishing itself. In a 2026 randomized trial, fifty-two engineers learned an unfamiliar code library, half with an AI assistant and half without. The assisted group scored 50 percent on a comprehension test against 67 for those who worked it through by hand, and the widest gap fell on debugging: the skill of recognizing that a thing is wrong and carrying it to actually working. Participants who used the assistant to deepen understanding did fine. Those who used it to skip understanding did not. Practitioners have named the shape of this from the inside, the 70 percent problem: the machine carries a task swiftly to roughly its first seventy percent, and the final, comprehension-dependent stretch is exactly where the assistance thins and the human skill, if it was ever built, has to take over.
Navigation, then memory, now comprehension. Completion is simply the faculty next in line, and it would be strange if the capacity to carry a thought to its end were the one exception to a pattern everything else has followed. The mechanism is not dramatic. It is a generation learning, one effortless start at a time, that finishing is optional. You will not experience it as loss. You will experience it as help. That is exactly what makes it dangerous.
And finishing is only half of it. The same abundance corrodes choosing. When every direction returns a plausible-looking answer in seconds, the cost of exploring collapses along with the cost of starting, and the mind is flooded with viable-seeming paths it never has to commit to. Choice used to be forced by scarcity: you could only afford to pursue one thing, so you chose. Now nothing forces the choice, and a person can spend a week in motion without ever having decided anything. Command is choosing and finishing. The machine is quietly dissolving both ends of it.
07The strongest objection
There is a serious objection, and it deserves its strongest form. Perhaps completion is the wrong virtue for this era. Perhaps the age that makes starting free rewards a different person entirely: the one with taste, who spins up forty drafts and selects ruthlessly, and who beats the artisan hand-finishing a single memo. Volume plus judgment is the new craft, says the objection, and reverence for carry-through is nostalgia.
The true part first: generating many options and choosing well is a real and increasingly valuable skill, and the purist who refuses to widen the field will often lose to the one who does. But look at what the objection describes. Choosing among forty drafts and shipping one is completion, performed at a higher level of abstraction: the selection, the judgment, the carrying of one chosen thing across the line. The person who generates forty and ships one has finished. The failure mode is the person who generates forty and ships none, who mistakes the volume for the work, because making the cut is precisely the part the machine cannot do. Orchestration is not the opposite of completion. It is completion wearing the clothes of the new tools, and it is every bit as scarce.
08A prediction you can hold me to
If the mechanism is right, its signature should become visible at population scale within a few years, and it is checkable: a rising flood of started-and-abandoned work. Repositories of half-built projects, drafts never shipped, activity metrics climbing while completion falls. That folder, replicated a billion times. People closest to the tools are already bracing for it; the Google engineer who named the 70 percent problem expects the coming year to bring a deluge of AI-generated, half-finished output across code, writing, and media. Same forecast, reached from inside the industry rather than from this argument.
And here is what would prove me wrong. The prediction assumes the offloading effect outruns the countervailing forces: models that genuinely finish rather than approximately finish, norms that price completion back in, tools built to scaffold follow-through instead of dissolving it. If those win, the abandoned-work flood does not come, the comprehension gaps close, and this essay overestimated the corrosion. That is the test. I would rather state it than hedge it.
09The gap compounds
For most of history, the cost of a captured mind stayed local, because the tools were weak. A focused worker and a scattered one, each with a hammer, produce comparable piles of work. The lever was short, so the gap stayed small. What changed is not the mind. It is the length of the lever bolted to it.
Hand the same model to two people. The first asks a shallow question, skims the answer, opens three tabs, and leaves with a vague sense of having worked. The second holds one problem open for an hour, sharpening the prompt, catching the output’s errors, asking the better question the first answer exposed, and walks away with a finished memo, a working model, a real decision. Same tool, same subscription, results not in the same universe. The difference is not intelligence; both had effectively unlimited intelligence on tap. The difference is who completed, under conditions engineered to make completion optional. The tools converge toward sameness. The operators do not. The spread widens every year the lever gets longer. Organizations will split the same way: many will buy intelligence with one hand while preserving the interruption-saturated environments that destroy the conditions for using it with the other.
What the commanded operator gains access to, and the captured one never does, is the state Csikszentmihalyi measured. Flow does not assemble itself in fragments. It begins only after continuity has been defended long enough for attention to settle and lock, which makes it, too, gated behind command.
One asymmetry belongs in the ledger, because it complicates the story rather than flattering it, and because it is the answer to anyone whose response to all of this is “try harder.” Command is not evenly distributed, and it is not purely a matter of will. It correlates with circumstance: with freedom from precarity, with a door that closes, with work that permits an unbroken hour. The parent working two jobs does not lack discipline relative to the funded founder with a quiet office; they lack the structural conditions under which discipline can operate at all. The attention economy is regressive in exactly this sense. It extracts the most from those with the least power to defend against it, and it returns the compounding advantages of command to those who already had room to exercise it. Any account of this problem that reduces to personal willpower is not just incomplete. It mistakes a structural extraction for a character flaw, which is precisely how the extraction stays invisible.
10Nothing we have fixes it
When a pattern hurts enough people, a society medicalizes it, and for attention that has meant diagnosis and stimulants. Both are real; for people whose lives have been capsized by severe attentional dysregulation, medication is the difference between chaos and function. But it is not a general answer. Prescriptions have expanded far faster than human nature could have changed. Even where appropriate, stimulants reliably reduce symptoms while their effect on deep, self-directed engagement is modest and variable. And no molecule addresses an environment redesigned to exploit the baseline: in one recent large study, two thirds of adults showed compulsive smartphone patterns. A drug can adjust the chemistry of a brain. It cannot redesign a day.
Training runs into the ceiling the ancients already mapped. The practice is real; imaging work has found structural differences in the attention-related cortex of long-term meditators, though most of it cannot separate cause from selection. The limits are equally real: progress is slow, uneven, and fragile under stress, and what holds in a monastery buckles on an ordinary Tuesday.
Step back, because the structural point is the one that matters. In every other domain, humans refused to leave a critical limit to willpower. We did not ask muscles to strengthen by discipline; we built engines. We did not ask memory to hold more by trying harder; we built writing, then libraries, then databases. We did not ask the eye to see farther by squinting. Yet for the capacity that now binds everything else, the choosing and finishing that abundant intelligence cannot substitute for, the prescription in 2026 is what it was in 2400 BCE: discipline, ritual, self-control. The oldest problem is still open for a simple reason. It was named, moralized, occasionally mastered. Never engineered.
11The question
The actual question of the next decade is not whether the machines can do the work; on current trajectory they will do nearly all of it. Whether the human at the center can still choose a problem and stay to the end while every incentive points the other way. The wandering mind is not going away, and it should not; some of your best ideas will always arrive uninvited. What changed, permanently, is the size of what rides on whether you hold the choice or something else holds it for you.
Intelligence is becoming abundant.
Command is becoming decisive.
The next intelligent tool should not only answer us. It should help us remain answerable to our own intentions.
This is an argument, not a study; the errors are mine. Sources for every figure are below. I’m at @niraj_1508.