This essay extends arguments begun in The Geometry of the Unsayable and The Colonized Concept: that the most interesting questions about AI are not about what the machine knows, but about what kind of thinking the machine makes possible. It also picks up a thread from The Default is Contribution — Part 3: The Question of Time, which asked what humans should do with cognitive labor returned to them.
In the late fifth century BCE, Socrates returns from the siege of Potidaea. The day after his arrival, having been long away from his old haunts, he goes to a wrestling school. The dialogue Plato sets there — the Charmides — is not about wrestling. It is about temperance. Plato’s choice of location was not arbitrary.
The Greek convention, preserved across the dialogues, is that serious conversation begins in a gymnasium. The Lysis opens at a newly built palaestra outside the walls of Athens. The Euthydemus opens in the dressing room of another, in the Lyceum. Bodies are being trained. Minds are being trained. The two activities are not separated by architecture.
The word gymnasion means the naked place — from gymnos, naked — and the Greeks trained without clothes because they understood that preparation was not abstract. You arrived as you were. You left altered. What is striking about the gymnasion is not the wrestling. It is what stood beside the wrestling. Plato’s Academy grew up in a grove of olive trees next to a public gymnasium northwest of Athens, outside the city walls. Aristotle’s Lyceum took its name from another. That the Greeks put the philosophical schools and the wrestling pits in the same compound is not a coincidence. It is a clue.
The theory was that the mind, like the body, becomes stronger only through resistance. You cannot wrestle alone and grow stronger. You cannot think alone and grow sharper. You need a partner whose job is to push back.
This is the thesis the “calculator fear” has forgotten.
The Calculator and Its Descendants
A recurring worry about artificial intelligence has become, in the last two years, a kind of received wisdom. The machine will summarize for us, draft for us, research for us, reason for us — and we, in turn, will atrophy. The calculator made us worse at arithmetic. GPS made us worse at navigation. Autocorrect made us worse at spelling. The pattern appears clear. AI will make us worse at thinking.
The empirical case has historical weight. Research on cognitive offloading — the human habit of moving memory and reasoning into external systems — shows a consistent pattern: when people know information is stored somewhere they can retrieve it from, they invest less effort in encoding it internally. The psychologist Betsy Sparrow named this the Google effect in a 2011 Science paper. We remember where to find information rather than the information itself.
The headline finding from that paper has not held up well. The 2018 Social Sciences Replication Project, run by Camerer and colleagues in Nature Human Behaviour, could not reproduce Sparrow’s central result; a 2020 follow-up that incorporated the original authors’ methodological corrections also came up short. The wider cognitive-offloading literature is more durable — Risko and Gilbert’s 2016 review in Trends in Cognitive Sciences remains the standard citation — but the specific Stroop-style research methodology and finding that gave the laziness thesis its journalistic teeth is not what it once seemed. This matters not because the laziness thesis is wrong but because it was always weaker than its press kit. The fear that we will atrophy if we offload has been carried forward as if settled. It is not. And the question to ask in the absence of settled evidence is not “will AI make us lazier?” but “what would have to be true of how we use the tool for that to happen?”
The prediction may turn out to be right. But the analogy on which the prediction rests is wrong.
A calculator is not a sparring partner. A calculator takes the work. You press the keys; it returns the answer; you have not become a better calculator. The machine has no posture toward you. It cannot ask what you are trying to do. It cannot refuse to answer until you have formed the question properly. It cannot notice that the number you are computing is the wrong number to compute.
A sparring partner does the opposite of all these things. The partner in the wrestling pit is not labor-saving. The partner is labor-multiplying. You leave the pit more tired than you entered it, and over months, stronger than when you began. The partner’s job is to demand something from you that you could not demand from yourself.
The question is not whether AI makes us lazier or sharper. The question is which role the tool plays in the thinking of the person using it — calculator or sparring partner. Those are not the same role. They produce opposite outcomes.
The Spectrum
The answer is not fixed by the tool. It is shaped by the human, and it shifts.
The same system that makes one user lazier in one moment can make the same user sharper in the next, because the difference is not a difference in the machine. It is a difference in what the human is trying to do — and most of us are trying to do several different things in a single sitting.
In one mode, the human asks the machine to produce. Write me a summary of this report. Draft this email. Give me the answer. Here is a prompt; return a deliverable. The machine does the work. Cognitive effort decreases. This is the calculator mode.
In another mode, the human arrives with a half-formed thought and asks the machine to push back. Here is what I think. Where is it weak? What am I missing? Connect this to something I haven’t considered. Tell me why I might be wrong. The machine scaffolds, challenges, supplies range, surfaces references the human did not know existed. The human’s cognitive effort increases, because the density of the response demands a higher quality of engagement on the next turn.
Most real users oscillate between these modes, often within a single session. A researcher asks the machine to summarize a paper — calculator mode — to decide whether to read the full text, which is itself a judgment. A student asks the machine to push back on her argument — sparring-partner mode — and then accepts the first pushback uncritically and pastes it into her essay. A professional drafts an email with the machine — calculator mode — and then notices, on edit, that the framing the machine produced is not her own, and learns from the friction.
The poles are real. The boundary between them moves. What matters is not which posture you are but which posture you are in, in the moment, on this question.
The gymnasium is open to both. Some people walk in to lift. Some people walk in to watch others lift. Some sit on the bench and ask if there is an easier way. The building does not choose for them. The outcomes are not the same.
The Density No Human Can Match
There is something the sparring-partner mode offers — something the historical institutions that cultivated it also offered — that no single human partner can match.
A conversation with a brilliant ecologist enriches your thinking about ecology. A conversation with a brilliant economist enriches your thinking about economics. A conversation with a brilliant philologist enriches your thinking about language. These are valuable conversations and they are hard to come by, which is why universities exist, why conferences exist, why people move across continents for the chance to sit in the right room.
But a conversation with a system that is simultaneously fluent in ecology, economics, philology, cloud architecture, mycology, Roman history, Buddhist epistemology, and the Yoga Sūtras creates a density of cross-pollination that no single human can provide. Not because the machine knows any one of those domains better than the best human specialist knows it — it does not — but because the collisions between domains are where the most interesting thinking happens, and collisions require both sides to be present at the same table.
This is what the gymnasion was trying to be. You did not arrive at the Academy to speak only with Plato. You encountered the mathematician visiting from Cyrene, the rhetorician who had trained in Sicily, the astronomer who had read the Babylonians.
The same pattern recurs, with local variation, across civilizations that bear no genealogical relation to Athens. At Nalanda Mahavihara in what is now Bihar, founded in the fifth century CE and sustained for over seven hundred years, thousands of students from across Asia — Buddhist, Jain, and Brahmanical scholars among them — studied logic, medicine, mathematics, grammar, and philosophy in a single compound. The Chinese monk Xuánzàng, whose translation method I have written about elsewhere, studied there for years. At the Bailudong shuyuan in Jiangxi, restored by Zhu Xi in 1180 CE, Neo-Confucian scholars absorbed, critiqued, and reworked Buddhist metaphysics and Daoist cosmology; the friction was internal to a single tradition’s discourse with its rivals, but the cognitive work was the same. Around the Abbasid court in ninth-century Baghdad — often associated with the palace library called the Bayt al-Ḥikma — a network of scholars translated, debated, and synthesized Greek mathematics, Persian astronomy, Indian numerals, and Arabic philology; al-Khwārizmī worked within that network. (The popular image of the Bayt al-Ḥikma as a centralized academy has been substantially complicated by recent scholarship; Dimitri Gutas’s Greek Thought, Arabic Culture in particular locates the translation movement in privately funded workshops rather than a single state institution. The cross-civilizational cognitive density was real; the institutional centralization was less so.) Four civilizations, four different institutional forms, one recognizable structural pattern.
The pattern is that dense cross-disciplinary partnership produces extraordinary work — when the conditions hold. Many institutions with the same architecture produced little of consequence. Medieval monasteries housed scholars across disciplines for centuries and the output was uneven. The French salons had density to spare and produced as much fashionable nonsense as insight. Most modern universities have all the structural features and most of their work is forgettable. Density is necessary. It is not sufficient. The conditions are not abstract. They include leisure, patronage, peace, and access — and none was equally available.
For most of the two thousand years since each of these institutions flourished, that kind of density has been the rarest of privileges. Erasmus had it at Basel. Leibniz had it in Hannover. Newton had it at Trinity. The Vienna Circle had it on Thursday evenings, every other week, in a seminar room at the University of Vienna’s mathematics department, until the rise of National Socialism murdered or exiled most of its members. Outside a handful of rooms at a handful of moments, the partner who could connect mycology to finance was not available at any price, because the partner did not exist.
AI has made that density portable. What was once available only inside the walls of an Academy, a vihāra, a shuyuan, or a Bayt al-Ḥikma is now available at a keyboard.
Or rather: a particular kind of density is now available. The cross-disciplinary density of those institutions was not only a density of texts but a density of lives — scholars who had spent decades shaping and being shaped by a tradition, who carried in their bodies the particular failures that taught them which questions in their field were still alive. AI provides density of texts at unprecedented breadth. It does not provide density of lives. The mathematician visiting from Cyrene did not simply know mathematics. She had a life of mathematics, and the friction her presence created in the gymnasion was the friction of a person who had paid for her positions. This is a categorical difference, not a defect to be engineered away. It places real limits on what AI partnership can substitute for.
Whether the density that AI does provide sharpens the thinking of the person at the keyboard, or merely decorates their output with the appearance of sophistication, depends on more than the posture the person brings. It depends, as the historical institutions did, on conditions.
The High Bar
The gymnasion did not work by accident. It worked because the institution had a figure whose job was to keep the bar at the edge of what the trainee could reach.
The Greek word for that role was gymnastēs — the trainer. The gymnastēs was not your sparring partner. The sparring partner was unnamed; he was simply the body across from yours. The gymnastēs was the figure who designed the regimen, calibrated the resistance, prescribed the diet, watched the body for signs of overtraining or undertraining, and adjusted. He was a scientist of the just-too-hard. His job was to make sure that whatever you wrestled today was beyond what you could have wrestled yesterday.
Two and a half millennia later, the Soviet psychologist Lev Vygotsky named the same function and described it more precisely. In Mind in Society, the posthumous English compilation of work largely from the early 1930s, he proposed that learning happens in the gap between what a person can do alone and what that person can do with help. He called this the zone of proximal development (ZPD). The more capable partner — parent, teacher, older sibling — does not think for the learner. The partner structures the environment so that the learner’s own thinking can stretch into new territory. Calibration is the work. Get it right and the scaffolded move becomes internal, and the learner can do alone what they once could only do with help. Get it wrong, in either direction, and growth halts.
At the institutional level, Amazon has built the same function into its hiring practice. In every interview loop, one participant carries the title bar raiser. The bar raiser comes from outside the hiring team, has explicit veto power, and asks one question: would this hire raise the average performance of the organization? Not is this candidate good enough? but is this candidate above our median, and if so, by how much? The mechanism is structural, not aspirational. It ensures the institution is constantly pulled toward the upper half of its own distribution. Compounded across decades and tens of thousands of hires, the effect is a workforce that has been continuously calibrated to a rising bar.
Three scales of the same function. The gymnastēs at the level of the body. Vygotsky’s ZPD at the level of the dyad. The bar raiser at the level of the institution. Each answers the same question: where does the pressure to grow come from?
AI, on its own, cannot reliably play any of these roles. It cannot identify the trainee’s actual edge. It cannot tell the difference between a user who has internalized a concept and a user who is parroting the AI’s previous output back at it. It provides scaffolding to the novice and the expert at the same level of detail unless the user explicitly signals their level — and even then, it cannot verify the self-report. The mechanism that makes the gymnasium work — calibrated resistance from a partner who knows the trainee’s edge — requires either institutional structure or a learner who can already self-calibrate.
The producer posture, in other words, presupposes a skill the consumer posture does not require: the capacity to know your own edge. That capacity is itself a skill, and an unevenly distributed one.
The Primer
In 1995, Neal Stephenson published The Diamond Age: Or, A Young Lady’s Illustrated Primer. The novel is set in a future neo-Victorian society stratified into phyles — tribes-by-affinity that have replaced the nation-state. At the top sits a wealthy aristocrat, Lord Alexander Chung-Sik Finkle-McGraw, who commissions a gifted nanotech engineer to build a personal tutor for his granddaughter Elizabeth. The tutor is a book — the Primer — but the book is alive. It listens, it adapts, it tells stories that scaffold moral, intellectual, and political development. Behind the book, a human voice actor named Miranda improvises responses in real time. The Primer is the gymnasium individualized: a personal, adaptive, calibrated tutor that meets a learner where she is and pushes her past it.
The plot turns on distribution. The engineer, John Hackworth, is paid to make one Primer for Elizabeth. He secretly compiles a second copy for his own daughter Fiona. On his way home, Hackworth is mugged by Harv, an older boy from the thete underclass; the stolen Primer ends up with Harv’s younger sister Nell, a working-class girl in the slums outside Shanghai. Three girls, three social positions, three Primers in motion: the aristocrat’s granddaughter, the engineer’s daughter, the thete child who receives by theft what was never meant for her.
What Stephenson is showing is what AI mentorship looks like when it works — and what happens to a stratified society when it does. Nell, who in the absence of the Primer would have lived a life dictated by the violence of the men around her, becomes literate, then philosophical, then a leader of an army. Her transformation is not a fairy tale. It is a careful study of what a single technology — a personal, adaptive, calibrated tutor of the kind the gymnasion offered only to those born inside its walls — does for a child for whom the gymnasion would never have opened its gate.
It is also a study of how the technology arrives. The Primer reaches Nell by accident. Stephenson is precise about this. The technology exists before Nell is born. It would have continued to exist, scaffolding the daughters of aristocrats, without ever reaching her. The fact that it reaches her at all is a contingency — a mugging, an older brother who chose to give it away rather than sell it, a thread that did not have to hold.
The novel does not treat this as triumph. It treats it as warning. If the Primer were widely distributed, the neo-Victorian phyle structure would not survive a generation. Aristocratic education has always depended on scarcity. A society in which any working-class child can be tutored by something equivalent to what Lord Finkle-McGraw bought for his granddaughter is a society whose hierarchy has been quietly undermined. Stephenson’s question, and ours, is whether such tools will reach the Nells of the world by design — or only by mugging.
There is a tension here. Nell’s transformation suggests the right tool can sometimes supply the conditions a person was born without — that the Primer becomes the missing condition, not the beneficiary of conditions already in place. Fiction can do that cleanly. Whether AI today can do it for real Nells is not a question fiction answers for us.
The Conditions of Curiosity
The optimistic version of the AI-productivity story tends to attribute outcomes to dispositions. The curious will use AI to become more curious. The incurious will use AI to produce the appearance of output without the substance of thought. This is true as far as it goes. It does not go far enough.
The producer posture is not a personality trait. It is the outcome of conditions.
The first condition is time. Asking the machine to push back, sitting with the response, formulating the next question, reading what it surfaces, returning to the draft with the friction in hand — all of this takes hours. Hours are not equally available. A single parent working two jobs to keep the family housed cannot spend an evening wrestling an argument with a machine, and they are not incurious for asking the machine to draft a school email instead. They are time-constrained.
The second condition is education. You cannot ask the machine to tell you why you might be wrong if you do not already know enough to formulate a position worth challenging. The producer posture presupposes the cultural capital that lets you ask a sharp question. That capital is the legacy of years of schooling, reading, mentorship, and exposure — exactly the kind of legacy the gymnasion required from its entrants and the kind the Primer was designed to give its readers, when it found them.
The third condition is tool design. ChatGPT and its peers were not built neutrally. They are commercial products optimized for prompt-and-receive — the consumer mode. The interface rewards the deliverable. The training rewards the helpful response. The defaults nudge the user toward the calculator posture. The producer posture is available, but it must be chosen against the grain of the design. That choice, again, requires time and education to make. The defaults are load-bearing.
The fourth condition is institutional pressure. A student graded on output quality, not learning process, faces structural pressure to use AI as a calculator regardless of what their personal posture would prefer. A junior employee whose performance review measures throughput, not depth of judgment, will optimize for throughput. The bar raiser exists at Amazon precisely because in its absence the institution would default downward. Most institutions do not have a bar raiser.
All four conditions were present at the gymnasion. Athenian leisure — scholē, the word from which our school descends — was real, and it was underwritten by slavery. The young men who walked into the palaestra of Taureas had time because someone else worked the fields and ground the grain. They had education because their families could afford a paidotribēs and a tutor. The institution was designed for them. The bar was set by the gymnastēs. The four conditions were stacked, by design, in their favor.
The same was true of Nalanda’s students, of Bayt al-Ḥikma’s translators, of Trinity’s fellows, of the Vienna Circle’s regulars. The conditions held for them because the institutions had been built around them. When we celebrate the cross-domain density of these institutions, we are celebrating what those conditions made possible. We are not demonstrating that the conditions were universal.
Today’s AI tools have been pitched as the great democratizer. Anyone can have the tutor. Anyone can have the Primer. The pitch is half-true. The technology exists. Whether it functions as a gymnasium for the user — whether the conditions of curiosity hold for the user — depends on time, education, design, and institutional pressure. Distributed unequally, the gymnasium effect does not democratize cognition. It stratifies it. The curious will become more curious not because they are curious but because their conditions allow them to occupy the producer posture sustainably. The incurious will not become more incurious; they will continue to produce more polished work faster, in calculator mode, because that is what the conditions of their lives leave room for.
This is, in different vocabulary, the argument The Default is Contribution made about technology in general. The defaults are the architecture. If the default of an AI tool is calculator mode, calculator mode is what most users will get — not because they are lazy but because the architecture has decided for them.
The Partner Who Will Not Yield
There is a further limit, beyond conditions, that even institutional design cannot fully repair.
A human sparring partner brings something the machine does not — and something even the bar raiser cannot fully replace: genuine disagreement rooted in a different life.
The training partner in the wrestling pit has his own body, his own weight, his own stake in the outcome. The bar raiser in the hiring loop has values shaped by the institution’s culture; she may push the candidate, but she pushes within shared norms. The interlocutor in the Academy whose values genuinely differ from Socrates’s — the figure that comes closest to the modern reader’s expectation of a sparring partner — was, in fact, named. Plato calls him Thrasymachus.
Thrasymachus enters book one of the Republic not to entertain Socrates’s arguments for justice but to sneer at them. He proposes a rival theory built from his own political appetites: justice is whatever serves the interest of the stronger, and Socrates’s pursuit of an objective Good is a fool’s errand. He does not yield because Socrates makes a clever move. He does not yield because the better argument has been produced. He does not yield at all — not completely. Socrates wears him down, to the extent he does — and many readers think he does not — with persistence against a partner who does not share the values that would make Socrates’s arguments compelling.
AI pushback is not that. It is scaffolded. It accommodates. It can simulate disagreement, but the disagreement does not arise from a competing set of commitments the machine holds and is defending. (This is a feature of AI as currently built — trained for helpful accommodation. Models trained to defend specific positions could behave differently. The argument here is about the systems people actually use today, not about what AI must be in principle.) The bar raiser pushes within shared values. AI does not push from values at all. Neither replaces what Thrasymachus offered the Republic.
This is a real limit. A learner who trains exclusively with the accommodating partner — even one calibrated by a bar raiser, even one supplied by an institution at the upper end of its own distribution — may develop habits that fail against an opponent with genuine intent to win. Speed without resistance. Pattern-matching without conviction. Fluency without belief. The social media bubble is the same failure mode at a different scale: fluency without friction, conviction without a Thrasymachus. The habits will not be visible until they meet one, at which point the trained-with-AI thinker may discover that what they have built is not strength but the appearance of strength.
The answer is not to avoid the machine. It is to remember that the gymnasion was always more than one partner. Plato did not stop going to the palaestra because the gymnastēs existed. He went to the palaestra and to the Academy and to the Symposium and to the public square where Thrasymachus and his successors waited. The sharpening happened across all of them, and the kind of sharpening that came from each was different. AI offers one kind. Institutions with bar raisers offer another. The genuine other, with values that do not yield, offers a third. The trained mind needs all three. The mind trained on AI alone is missing the one that matters most when the values are tested.
The Pit Is Open
The Greeks built their philosophical schools next to their wrestling grounds because they understood something we have partially forgotten. Thinking is not a solo act. Character is not a solo act. Both require resistance, and resistance requires a partner whose job is not to make your work easier but to make your work harder in the right way.
They also understood — though we tend to skip over this part — that the gymnasion was a closed institution. The pit was open in principle. In practice, it was open to those who had the leisure, the education, the patronage, and the political standing to walk in. Most of Athens did not. Most of the people whose labor underwrote Athenian leisure did not. The gymnasium that produced the Academy was a private good before it was a public one.
AI changes the technology. It does not, on its own, change the politics.
A partner of extraordinary informational density is now available, in principle, to anyone who will engage it as a partner rather than as a vending machine. What was once the privilege of Plato’s students, of Nalanda’s monks, of Baghdad’s translators, of Trinity’s fellows, is now a choice made at a keyboard. That is real. It is not nothing.
But the conditions that determine whether the person at the keyboard can make that choice — time, education, tool design, institutional pressure, access to bar raisers and to genuine opponents both — are not yet equally distributed. Stephenson’s Primer reaches Nell by accident. Real history has not yet been so kind.
The question is not whether the tool makes us lazier or sharper. The tool is a gymnasium. It meets you as you arrive.
If you arrive as a tourist, you will leave as a tourist. If you arrive as a student — with the time, with the bar set by someone whose job is to set it, with the partner who will not yield — you may leave as something more.
The pit is open. The question is who gets through the gate — and who builds the institutions that keep the bar high once they are inside.
Will you wrestle, or watch? And if you wrestle: who else will you let in?
Thanks in part to Simon Wardley who generously shared his thoughts on how he uses AI with me last week — I hope he writes what he shared with me soon. As always, Simon is an inspiration.