Move 78: Into the Void
In March 2016, a South Korean grandmaster placed a single stone that broke the world’s most powerful AI. What happened next wasn’t a malfunction. It was a revelation.
SECTION I
The Match That Changed the Conversation
In March 2016, in a conference room at Seoul’s Four Seasons Hotel that felt more like a coliseum, Lee Sedol — one of the greatest Go players in a game with 2,500 years of recorded play — sat down across from a machine. Not across from it, exactly. Across from a screen, a proxy for something with no body, no nerves, and no career on the line. DeepMind built AlphaGo, and Google owned it, and it trained on thirty million human moves. It had then played itself millions more times, iterating at speeds no biological organism could sustain. By the time it sat across from Lee Sedol, it had consumed more Go than any human being ever born could absorb in a thousand lifetimes.
The match was billed as a civilizational test. Not just for Go, but for the broader question everyone in artificial intelligence was circling without quite wanting to say out loud: had the machine crossed some threshold? Was pattern, compressed and accelerated to this degree, equivalent to understanding? Was it equivalent to intelligence?
AlphaGo won the first three games, and the Go establishment, which had initially given Lee Sedol a comfortable edge in pre-match assessments, went quiet. But Game 2 produced the moment nobody had a frame for. On move 37, AlphaGo placed a stone on the fifth line—a shoulder hit that human players considered, at best, eccentric and, at worst, a beginner’s error. The two human commentators watching the game live assumed it was a glitch. One of them laughed, nervous. Lee Sedol stood up and left the room for fifteen minutes.
He wasn’t rattled by the move’s strength. He was rattled by its strangeness. Move 37 made no sense in the game’s grammar as humans had ever played it. It violated the aesthetic logic, the accumulated heuristics, the unspoken agreements about what good Go looked like. Yet over the next hundred moves, it proved precisely correct. AlphaGo had assigned it a 1-in-10,000 probability — had seen what no human teacher had ever taught, because no human had ever played it. It had found something genuinely new inside the map it had built from thirty million human choices.
This is the moment the popular narrative froze. The machine beat the grandmaster with a move no human would ever play. The story wrote itself: AI sees further, thinks differently, has transcended us. Depending on your disposition, that was either the beginning of a golden age or the beginning of an existential reckoning. Both camps missed the more important thing that happened three days later.
Game 4 is the game the AI industry would prefer you to forget. Lee Sedol played move 78, and the most powerful game-playing system ever built started to come apart.
Lee Sedol won Game 4. He is the only human to have defeated AlphaGo in competitive play. He did it with a single stone — move 78, a wedge placement in the lower right corner of the board — that the Go community immediately named “the Hand of God.” AlphaGo’s calculated win probability, which had hovered near 70 percent in its favour, began producing nonsensical readings. The system started playing moves that experienced observers described as panicked, desperate, confused. It lost the thread. It lost the game.
Move 37 told us that a machine, given enough data and enough compute, can discover things inside a known space that no human had found before. That is genuinely remarkable. Move 78 told us something harder to digest: a human being, operating from a form of intelligence not reducible to pattern recognition, can step entirely outside the known space —and when they do, the machine has nothing to reach for.
That asymmetry is what this piece is about.
SECTION II
What “Hallucination” Actually Means — And Why It Should Worry You More Than Superintelligence Does
The word most people used to describe AlphaGo’s behaviour after move 78 was “confused.” Some said, “panicked.” In contemporary AI discourse, we would call it hallucination — the system producing confident outputs that are disconnected from reality. But that framing, while technically accurate, obscures what actually happened mechanically, and the mechanics matter if you want to understand what we are actually building when we build these systems.
AlphaGo was not confused in any experiential sense. It has no experience. Move 78 put the game in a region of possibility that AlphaGo’s model had no reference points for. Every move AlphaGo calculated was evaluated against a value function — a learned estimate of how likely a given board position was to lead to victory, built from the enormous library of games it had processed and the millions of self-play iterations that followed. That value function was exquisitely calibrated to the space of Go positions that humans and AlphaGo had ever produced. Move 78 pushed the game outside that space.
The technical term is distributional shift. A model trained on a distribution of inputs — in this case, the full recorded history of elite Go — will produce reliable outputs when the inputs it encounters resemble that distribution. When inputs diverge significantly from the training distribution, the model doesn’t know it’s lost. It keeps producing outputs. It keeps assigning confidence scores. But those scores are now unmoored from any valid reference. The system was not panicking. It was doing exactly what it was designed to do, in a context where doing exactly what it was designed to do produced catastrophic results. That is a more unsettling failure mode than panic, because panic at least signals awareness of a problem.
This is not a bug that gets patched in the next version. It is not a limitation of AlphaGo specifically — it is a structural property of how these systems are built. Every large language model, every image recognition system, every recommendation engine operates on the same principle: it learns a function that maps inputs to outputs by finding patterns in historical data, and it applies that function to new inputs under the assumption that the new inputs resemble the old ones. When that assumption fails, the model fails — not visibly, not loudly, but in the most dangerous way possible: confidently and incorrectly. You can’t build your future by pulling the block that built the past.
AlphaGo didn’t lose Game 4 because it was outmatched. It lost because it was out-framed. Lee Sedol didn’t play a better move within the game AlphaGo understood. He played a move AlphaGo had never conceived of.
Here is where the argument sharpens. The popular concern about superintelligence is that it will be so capable that it will optimize so effectively for whatever objective we give it that it will overwhelm our ability to course-correct. That concern is real and worth taking seriously. But Game 4 surfaces a different and less-discussed problem: a system that is simultaneously too confident and too bounded. Too confident because it keeps producing outputs even when it has no valid grounds to do so. Too bounded because all of those outputs, however sophisticated they appear, are interpolations from a closed library of prior experience. So doesn’t it stand to reason that effectively optimizing the processing of information we already know will trend toward output certainty by reverting to the mean? What other destination would it be to process thoughts, discussions, and research from the past without the ability to investigate and propagate new information? Are humans the rung or the ladder?
An AI system cannot encounter something genuinely new and recognize it as such. It can only measure new inputs against old patterns. If the new input is close enough to something in the pattern library, the system functions well. If the new input is far enough outside the library — if it represents genuine novelty, not just variation — the system has no flag to raise, no alarm to sound. It simply continues, producing outputs that look exactly like its correct outputs, because it has no mechanism to distinguish between the two cases. Lee Sedol’s move 78 was, in the most technical sense, a Black Swan: an event outside the model’s reference class, which the model therefore could not assign any meaningful probability to, and which the model’s architecture gave it no way to even detect as anomalous.
Nassim Taleb built a career on the observation that we systematically underestimate events outside our historical reference class. What AlphaGo demonstrated in Game 4 is that AI systems don’t just underestimate those events; as we have been discussing, they are architecturally blind to them. Even if the system has been populated with every game that has ever been played, it could be argued that it is operating within a vacuum. They can be extraordinarily powerful within the territory their training data covers. But they carry no map for the territory beyond the map.
And here is the question that should follow every breathless headline about what AI can do: who is responsible for knowing where the map ends? If critical thinking gives way to cognitive atrophy, you are closing the door to the hallways of chance and discovery.
SECTION III
The Third Kind of Knowing: Black Swans, System Collapse, and the Move That Has No Name
A question is buried in what Lee Sedol said after Game 4, and almost no one has asked it. He was asked what he was thinking when he placed move 78. His answer: that was the only move he could see. There were no other moves. He couldn’t think of any other placement. It was the only option, so he played it.
Read that carefully,, and you realize it doesn’t fit either cognitive category we use to describe intelligence, human or artificial. It was not a calculation. Lee Sedol did not evaluate move 78 against a field of alternatives, assign probability weights, and select the highest-scoring option. But it was also not random. He did not place the stone arbitrarily, without knowledge or intent. It was something else entirely — a move that arrived—not chosen from a menu and not generated by noise. Arrived. The product of thirty years of embodied expertise collapsing, in a single moment of constraint, into one irreducible perception: here, and nowhere else.
That distinction matters more than anything else in this piece, because modern artificial intelligence is built on two cognitive primitives: calculation and randomness. It can do both with extraordinary power. What it cannot do — what it has no mechanism to do — is produce what Lee Sedol produced at move 78. The reason it can is not a limitation of processing speed, data volume, or model size. It is a structural impossibility that follows directly from what these systems are and how they are built.
To understand why, you need two thinkers. One mapped the boundary of what models can know. The other mapped the boundary of what minds can do. Together, they define the ceiling that move 78 broke through.
Nassim Taleb spent years as a derivatives trader before he became a philosopher of risk. That path matters, because he was not writing theory from a library. He wrote from the perspective of a man who had watched confident quantitative models produce catastrophic wrong answers in real markets, with real consequences. He called them Black Swans, high-impact events outside a model’s reference class, and it is not merely a metaphor. It is a precise technical observation: a model trained on a distribution of historical events cannot assign meaningful probability to events outside that distribution. It cannot even recognize that it has encountered one. The model does not know what it does not know.
We have already seen that a model operating outside its training distribution keeps producing confident answers. It cannot detect that it has crossed the boundary, because doing so would require knowing where the distribution’s boundary is, which would require observing inputs outside it, something it hasn’t done. The AI system has no critical thinking because it has no heartbeat powering a nervous system with a visceral connection to the world. This is Taleb’s insight applied to AI with surgical precision: the systems most confident in their outputs are the systems most vulnerable to genuine novelty, because they have no mechanism to distinguish novelty from variation. As Ben Affleck put it when discussing AI at the 2024 CNBC Delivering Alpha investor summit:
“Craftsman is knowing how to work. Art is knowing when to stop.”
He added that knowing when to stop will be very difficult for AI to learn because it’s taste. AI, he said, can write you excellent imitative verse that sounds Elizabethan; it cannot write you, Shakespeare.
Move 78 wasn’t a variation on anything in AlphaGo’s training data. It was a Black Swan on a Go board — an event the model could not recognize as anomalous because recognizing anomalies requires a map of the territory beyond the map.
Demis Hassabis confirmed as much the morning after Game 4. AlphaGo assigned move 78 a one-in-ten-thousand probability not because its calculations suggested it was a bad move, but because it determined no human player would ever choose it. It had built a model of human Go cognition from thirty million games, and that model told it the move was outside the distribution of human play. It was right. That was precisely the problem. The model’s accuracy in characterizing what humans had done became a catastrophic liability the moment a human did something the model had never seen. It was operating on a map of human possibility. Lee Sedol stepped off the map. And AlphaGo had no response for territory that, by its own reckoning, did not exist.
Daniel Kahneman’s framework is the second piece. In his account of human cognition, there are two systems of thought. System 1 is fast, automatic, intuitive, the part of the mind that recognizes faces, catches falling objects, reads social situations without deliberate effort. System 2 is slow, deliberate, analytical — the part that does long division, evaluates a contract, calculates a chess position. Kahneman showed that most decisions are made by System 1 and rationalized afterward by System 2. We believe we are calculating when we are mostly pattern-matching at speeds that feel instant. Pattern-seeking is not something the mind does; it is what the mind is. The tendency is so dependable that its failures have names: apophenia, meaning in random data, and pareidolia, faces in clouds, outlets, and toast. Michael Shermer called the general tendency patternicity. The point is not that we sometimes see patterns that aren’t there. We are built to see them first and question them later. System 1 draws us toward the pattern-seeking comfort zone, where our minds find inner peace.
In this framework, artificial intelligence is the most powerful System 1 machine ever built. It pattern-matches at scales and speeds no biological mind approaches. A large language model responding to a complex question is doing something that looks like System 2 reasoning — deliberate, structured, analytical — but is in fact an extraordinarily sophisticated System 1 operation: statistical pattern completion across a vast training distribution, producing the most probable continuation of the input sequence. It is not reasoning from first principles. It is recognizing, at enormous scale and resolution, what typically follows what. The appearance of reasoning is a property of the patterns it has learned, not a property of the cognitive process itself. Simply put, AI does not second-guess its output.
So why does all this matter? Kahneman’s framework, like Taleb’s, has a built-in boundary. System 1 fails at precisely the moments System 2 is needed — when the situation is genuinely novel, when the pattern library doesn’t apply, when the fast answer is not just wrong but confidently wrong in ways that compound. Kahneman documented these failures in humans. The difference is that a human being, confronted with the failure of System 1, has access to System 2. They can slow down, interrogate their assumptions, reason explicitly about what they don’t know. A machine learning system, when its pattern-matching fails, has no System 2 to reach for. It has only more of System 1 — faster pattern-matching in territory where the patterns have broken down. After move 78, AlphaGo did not slow down to reason about what had happened. It continued selecting the statistically most probable moves in a position its training data had deemed impossible. The result looked like panic. It was worse: confidence operating in a void. What we refer to as hubris.
But here is where the argument has to be honest about its own complexity. Because Lee Sedol’s move 78 was not pure System 2 reasoning either. He did not sit down, enumerate all possible moves, and calculate that 78 was optimal. He said there was only one move he could see. It arrived. That is the language of System 1 intuition, immediacy, the felt sense of necessity. And yet it was not the ordinary pattern-completion of System 1, because ordinary pattern-completion is exactly what AlphaGo had, and AlphaGo could not produce it.
What Lee Sedol produced at move 78 was something else entirely. Not fast pattern-matching or slow calculation, but something cognitive science has struggled to name precisely: the intuition of a master. Thirty years of embodied practice, compressed into a perceptual field so deep that the patterns he recognized were not patterns in AlphaGo’s training data — were not patterns in any data — but patterns emergent from the lived experience of having played, lost, adapted, suffered, returned, and played again across three decades of competition at the highest level. The move did not come from his analysis. It came from what he had become.
AI OPERATES HERE
System 1 at scale
Pattern completion across historical distributions. Extraordinarily powerful within the map. Structurally blind at the boundary.
AI CAN SIMULATE THIS
System 2 reasoning
Slow, explicit, deliberate. Evaluates alternatives. AI produces its appearance through pattern completion — it is not the thing itself.
MOVE 78 LIVES HERE
Intuitive necessity
Neither fast pattern nor slow calculation. The only move a particular consciousness, with a particular history, could see. Unrepeatable. Untrainable.
This third category — call it intuitive necessity, or embodied expertise under constraint — is what Taleb’s framework identifies as structurally invisible to probabilistic models and what Kahneman’s framework cannot fully account for because it was built to describe failure, not mastery. It is the cognitive territory that exists past both the map and the reasoning about the map: the place where a particular human being, shaped by a particular irreducible history, perceives something that no distribution of prior events could have predicted, because it is the product not of information but of formation.
Alexander Fleming did not calculate that the mould contaminating his petri dish was worth investigating. He noticed it—a noticing that required not only knowledge of bacteriology but a particular cast of attention, a tolerance for surprise, and a willingness to treat anomaly as signal rather than noise, cultivated over decades of laboratory practice. August Kekulé did not compute the ring structure of benzene. By his own account, given a quarter-century later, he dreamed of a snake seizing its own tail and woke up with the answer. This answer could not have been reached by any systematic enumeration of molecular possibilities. Ignaz Semmelweis insisted on handwashing before germ theory existed to justify it, because something in his clinical experience told him the connection was there even when he lacked the conceptual language to prove it.
None of these were random. None were calculated. All three were the product of a consciousness that had internalized a domain so deeply that genuine novelty became perceptible—not because it had seen it before, but because it had developed the capacity to recognize it as significant when it appeared. That capacity is not a larger dataset. It is not a better algorithm. It is the residue of a life of sustained attention, pressed into the shape of a particular human mind.
AlphaGo could not produce move 78 because move 78 did not come from information. It came from formation. And formation is not something you train into a model. It happens to a person.
This is the philosophical ceiling that Game 4 exposed — not the ceiling of AlphaGo specifically, but the ceiling of any intelligence built entirely from the record of what has already happened. Such an intelligence can find everything within the map of prior human experience, faster and more completely than any human being. Move 37 proved that. But the map has an edge. And at the edge — in the moments that matter most, the moments of genuine novelty, of crisis, of discovery, of the kind of creative leap that rewrites what comes after — the most powerful pattern-matching system ever built has nothing to reach for. It does not know it has reached the edge. It does not know there is an edge. It keeps moving forward, confident and lost, into territory that its own architecture has rendered invisible.
Lee Sedol retired from professional Go in 2019. He said he quit because of AlphaGo. Lee also mentioned he had seen an entity that could not be defeated, and that knowledge had drained the meaning from a lifetime of competition. It was a deeply human response to something genuinely new. The irony is that in the very match that broke his will to continue, he did the one thing AlphaGo could not: he reached past the map and placed a stone in territory no model, however vast, had charted. He didn’t know he had done something historically unrepeatable. He just knew it was the only move he could see. He took the first step into the void….
That is what we risk losing, if we build a world where the map is all there is.
SECTION IV
The Metaproletariat: What Happens When Institutions Destroy the Conditions That Produce Move 78
The philosophical argument in Section III is not an abstraction. It has a direct institutional expression, and that expression is happening right now, at scale, inside every organization that has spent the last decade optimizing human judgment out of its decisions.
Here is the logic compressed: the highest-value cognitive act — the one that operates where AI systems go dark, where the map ends and genuine innovation begins — is not calculation and not randomness. It is the intuitive necessity that emerges from formation—from decades of embodied, felt, consequence-bearing experience in a domain, pressed so deeply into a human consciousness that the right move arrives rather than being chosen. That capacity is not innate. It is not a personality trait. It is the product of a specific developmental process: prolonged exposure to real problems, real feedback, real stakes, and enough scar tissue that the nervous system starts to perceive patterns with no name in any dataset.
For AI, the past is the only lens into the future, but it is a record, not a scar. AI can predict what comes next; it cannot anticipate it, because nothing ahead of it is at stake.
Some would argue that we are drifting towards a metaphysical conversation. But in a literal sense, we are trying to understand, intuitively, what lies beneath the physics. We can know that something is there without yet knowing what it is.
Lee Sedol could produce move 78 because he had spent thirty years losing games that hurt, winning games that mattered, and returning to the board regardless. The formation preceded the move by three decades. You cannot shortcut it, simulate it, or hire it on a contract basis. It grows in a person the way bone density grows. Bone grows slowly, through resistance, over time, and only when conditions are right.
Now consider what the dominant organizational logic of the last twenty years has been doing to those conditions.
Harry Braverman saw it coming in 1974. His analysis of industrial capitalism — still the sharpest account of what happens when management ideology meets the labour process — identified a pattern he called deskilling: the systematic decomposition of complex work into simplified tasks, the separation of conception from execution, the transfer of knowledge and judgment from workers to management and then to machinery. Braverman was writing about factory floors and typing pools. The principle has migrated up every layer of the knowledge economy, and the arrival of AI has not changed its direction. It has simply accelerated it to the point where we can no longer pretend it isn’t happening.
In philosophy, intrinsic value is the worth something has in itself, as an end, not because of what it gets you. Happiness, knowledge, beauty, or a person’s life are classic candidates. It contrasts with instrumental value, which is worth as a means: money matters for what it buys, a hammer for what it builds. The test is to ask, “Why does that matter?” until you hit something where the only answer is, “It just does.” That stopping point is intrinsic value. The problem arises when instrumental value becomes a full-court press, and with the wealth extraction and rise of a powerful oligarchy that come with it.
What algorithmic management does to a workforce is precisely what distributional shift does to a model: it removes the conditions under which genuine novelty can be perceived. A driver who follows GPS instructions for ten years and is then dropped in an unfamiliar city without signal is not supported; they are deskilled. A claims assessor whose recommendations are overridden by a scoring model every time they diverge from the algorithm’s output eventually stops making independent assessments. The judgment atrophies because the feedback loop that would strengthen it has been severed. Formation requires friction. Remove the friction, and you remove the formation, and with it, any future capacity to produce the move the algorithm cannot.
Deskilling doesn’t just eliminate current judgment. It eliminates the developmental conditions under which future judgment could have grown. You are not losing a capability. You are losing the process that makes the capability possible.
David Autor’s task-based framework for labour economics gives this structural precision. Autor’s research shows that automation displaces routine tasks—both manual and cognitive—while complementing non-routine ones. The standard reading of Autor is optimistic: automation handles the predictable, humans handle the unpredictable, everyone moves up the value chain. The problem with this reading is that it treats non-routine cognitive capacity as a given: a stable reservoir of human capability that persists regardless of what happens to the routine-task layer beneath it. That assumption is wrong. Non-routine cognitive capacity — the kind that handles genuinely novel situations — is not separate from routine task experience. It develops through it. The junior analyst who learns to read a balance sheet by doing a thousand of them is not just building routine competence. They are building the pattern library that, under the right conditions of constraint and pressure, will one day produce a non-routine insight no model saw coming. Remove the thousand balance sheets and you remove the prerequisite for the insight.
Erik Brynjolfsson and Andrew McAfee were right that the second machine age creates enormous wealth. They were less clear about the distributional question: who develops the non-routine cognitive capacity that survives automation, and under what conditions? Their answer — invest in education, retrain workers, help people move up the skill ladder — is the right prescription at the wrong level of analysis. The question is not whether people can learn new skills. The question is whether the organizational structures being built around AI allow the kind of deep, consequence-bearing, longitudinal formation that produces genuine judgment. Most of them do not. Most of them are designed, explicitly or implicitly, to prevent it. Formation is slow, expensive, unpredictable, and resistant to the kind of standardization that makes organizations legible to their owners. The push towards a perceived certainty of outcome will stagnate non-routine cognitive input: the thinking, and the subsequent act of discovery, that happen within the realm of human consciousness.
Yanis Varoufakis named the endpoint of this logic technofeudalism: a system in which the owners of algorithmic infrastructure extract value from an increasingly undifferentiated labour force stripped of the productive capacity to generate genuine economic alternatives. The feudal analogy is precise in one specific way: in feudalism, the serf’s dependence on the lord was not incidental but structural. The relationship was produced and maintained by deliberately restricting the conditions under which independent capacity could develop. Algorithmic management does the same to knowledge workers. It is not merely that the work becomes more monitored and less autonomous. It is that the conditions under which independent judgment could grow — the latitude to be wrong, to experiment, to accumulate scar tissue and learn from it — are systematically removed in the name of efficiency, consistency, and risk management.
This is where the move 78 argument becomes an organizational argument, not merely a philosophical one. The question is not whether AI will eventually be able to simulate the intuition of a master. That is a question for a different decade. The question is whether the humans working alongside AI systems today are being allowed to develop the formation that would make them capable of move 78 when it counts. And in most organizations currently deploying AI at scale, the answer is no. They are deploying AI in ways that accelerate task completion while simultaneously eliminating the developmental friction that builds the judgment task completion cannot replace.
The organizations most aggressively automating human judgment today aren’t just cutting costs. They are removing the seedbed of the only competitive advantage that cannot be reproduced from a training dataset.
I have spent more than twenty-five years running field operations. The best operators I worked with—the ones who could walk into a failing job at two in the morning and immediately see what everyone else had missed—weren’t the ones with the most credentials or the most efficient dashboards. They were the ones who had been allowed to fail, at real cost, in real situations, enough times to develop a perceptual field that no process document could replicate. That field was not their personality. It was their formation. And it took years of friction to build it. “Knowledge does not advance practice. Rather practice advances knowledge” (Pascale et al., 2010). In any business, the field is where practice doesn’t just make perfect; it makes sense.
The PE-backed management logic that eventually acquired the business I had built would have optimized most of that formation out of the organization within eighteen months. That is not to say the PE owners were wrong. The legacy employee who becomes a hurdle rarely does so intentionally; they struggle to articulate a methodology that has become so reflexive and intuition-led that it is hard to explain. It becomes hard to transfer. So the playbook takes over. Standardize the process. Reduce variance. Replace judgment calls with decision trees. These are not irrational choices from a short-term financial perspective. But they are choices that, applied consistently over time, produce organizations full of people who are very good at executing within the known space and completely unprepared for the moment the known space runs out — which is precisely the moment when the AI system they have been relying on will have nothing to offer them either. Always keep in mind the power of being as nimble as a kayak and as powerful as a dragon boat.
The Metaproletariat thesis I have been developing is, at its core, an argument about what happens when the managerial layer is eliminated by AI and what is revealed beneath it. The conventional view is that AI destroys middle management and creates a barbell economy: a small class of highly paid cognitive workers at the top and a large class of precarious physical and service workers at the bottom, with it all held by a smaller class owning the algorithmic infrastructure. That analysis is structurally correct. But it misses the deeper problem.
The deeper problem is that the compression is not just economic; it is developmental. When the managerial layer disappears, it takes with it the organizational infrastructure through which formation used to happen. The junior analyst learned to see what the senior analyst saw by working alongside them, being corrected by them, inheriting not just their knowledge but their attentional habits, their tolerance for ambiguity, their instinct for when the model is wrong. That transmission was inefficient. It was slow. It resisted standardization. And it produced, over time, the conditions under which move 78 became possible for the people who went through it.
Remove the managerial layer, and you remove the transmission mechanism. What you are left with is a workforce that can operate the tools very well and perceive almost nothing that the tools cannot perceive. That workforce is not a partner to the AI. It is its dependent — capable of execution within the distribution the model has mapped, and helpless at the boundary. The Metaproletariat is not just economically precarious. It is cognitively stranded: cut off from the developmental conditions that would have made it capable of the one thing the algorithm cannot do.
Lee Sedol had thirty years to become the person who could see move 78. The question that should be keeping organizational leaders awake at night is not whether their AI system is accurate enough. It is whether they are building the conditions for the people working alongside it to develop the formation to know when accurate enough isn’t good enough and to see, in that moment, the move the model never conceived of.
SECTION V
What to Protect: A Structural Argument for the Conditions That Make Move 78 Possible
This is not an argument against artificial intelligence. Let that be stated plainly, because the moment you argue for limits on what AI can do, someone in the room will hear nostalgia, or fear, or the kind of reflexive technophobia that has been wrong about every previous wave of automation. That is not what this is. AlphaGo was a genuine achievement. The large language models (LLMs) reshaping knowledge work in 2026 are genuine achievements. The prescription that follows is not a call to slow any of that down. It is a call to understand what you are trading away when you deploy these tools without protecting the developmental conditions that produce the one cognitive capacity they cannot replicate. It’s the ‘when you know, you know’ feeling: the visceral sense of something impacting you in a way words can’t describe. That is the feel that seems to be missing.
The trade-off isn’t obvious in the short term. In the short term, replacing human judgment with algorithmic decision-making is almost always faster, cheaper, and more consistent. The dashboard looks cleaner. The variance goes down. The process becomes legible to its owners. These are real benefits, and anyone who dismisses them is not engaging seriously with the organizational pressures that produce these decisions. The problem is not that the trade-off is irrational. The problem is that it is invisible; the cost doesn’t appear on any balance sheet until the algorithm reaches the edge of its distribution and no one left in the building can see what the algorithm cannot.
By then, it is too late. Formation is not a resource you can stockpile and draw down. It must happen continuously, in real people, in real conditions, or it stops. A decade of algorithmic management does not leave you with slightly less judgment. It leaves you with a workforce that has forgotten what judgment felt like and an organizational structure that has no mechanism to rebuild it. The Go board analogy is exact: you cannot become the person who sees move 78 by watching someone else play. You have to sit at the board, in real games, with real consequences, long enough for the formation to happen to you.
So what does protection actually look like? Not sentiment. Not a policy of hiring humans to do what machines do better now. The prescription is structural: deliberate organizational choices about where to actively cultivate human formation, even at the cost of short-term efficiency, because the long-term strategic value of that formation exceeds any efficiency gain its absence would produce.
01 Protect the developmental layer
The junior roles that look inefficient — the analyst who takes twice as long as the model, the operator who makes calls the algorithm would override — are not inefficiencies to be eliminated. They are the seedbed. Formation happens in those roles, in those years, through the friction of consequence-bearing work. Organizations that eliminate the entry-level judgment role to save costs aren’t just cutting headcount. They are severing the transmission mechanism through which expertise propagates across generations of practitioners. The cost shows up fifteen years later, when no one in the organization can catch what the model gets wrong.
02 Design for productive failure
Formation requires being wrong at real cost. Failure is not just part of human development; it is the mechanism. Scar tissue is not a metaphor; it is the physiological residue of a nervous system that has processed real feedback, real consequences, and recalibrated. Organizations that have optimized away the conditions under which their people can fail—through excessive oversight, algorithmic override, or removing any decision with meaningful stakes from human hands—have removed the primary driver of formation. The prescription is not to manufacture failure. It is to identify the domains where productive failure is currently being prevented and deliberately reintroduce the latitude for it. The pathway to a goal unpacks opportunity and ideas that will lead to the next adventure, the next movement, the next discovery. Take penicillin (1928–1945). Fleming’s contaminated petri dish is the famous part, but he couldn’t purify or mass-produce the substance and largely moved on. The real breakthrough came a decade later, when Florey, Chain, and Heatley at Oxford worked out the grinding process of extraction, testing, and scale-up. The accident became a medicine through process.
03 Distinguish execution from formation
Not every task produces formation. Routine execution — processing claims within a defined parameter set, scheduling within an optimization model, generating reports from a template — produces competence at the task and nothing beyond it. The organizational question is not how much human judgment to preserve in aggregate, but where in the value chain the formation-producing work lives and whether the people who need to develop that capacity are being routed through it. AI should own the execution. Humans should own the work that forms them: messy, ambiguous, consequence-bearing problems with no clean algorithmic solution, leaving a residue of judgment in the people who work through them.
04 Map where your distribution ends
Every AI system deployed in your organization has a training distribution — a boundary beyond which its outputs are unmoored from valid reference. Most organizations have no map of where that boundary is for the systems they are running. The prescription is not to limit AI deployment. Instead, build the institutional practice of asking, for every consequential AI-assisted decision: what category of input would cause this system to fail confidently? And who in this organization is trained to recognize that input when it arrives? If the answer to the second question is no one, you have a structural risk that no dashboard will surface until the moment it materializes.
None of this is anti-technology. The centaur model — a human and AI operating together, each doing what they do best — is the right frame. But a centaur only works if the human half is a human who has been formed. A person who has spent their career executing within algorithmic parameters, never developing the judgment that comes from consequence-bearing autonomy, is not a centaur’s human half. They are a cursor. They provide the inputs the model requests and execute the outputs the model produces. The machine is not augmenting their intelligence. It substitutes for it, in a configuration that looks like partnership but functions like dependence.
The centaur model that actually survives the next decade of AI development is not the one where humans and machines divide the work. It is the one where humans have been formed deeply enough to know when the machine has reached its edge and to offer something in that moment that the machine cannot provide. That formation does not happen by accident. It does not happen in environments optimized for consistency and efficiency. It happens through decades of friction, failure, return, and accumulation. It requires organizations willing to protect it even when doing so is expensive, slow, and at odds with the metrics that make quarterly performance legible.
The question is not whether your AI system is good enough. The question is whether the people working alongside it have been formed enough to know when it isn’t — and to see, in that moment, what it cannot.
Lee Sedol retired in 2019. He gave a specific reason: he had seen something that could not be defeated, and that knowledge had made the game feel hollow. He was the best human player alive, and the best was no longer enough to matter. By any reasonable measure, his decision was rational. What would be the point of continuing?
But here is what that retirement actually means: read through the argument this piece has been making. Lee Sedol was the only human being in recorded history to demonstrate the capacity to produce move 78—to step outside the distribution of all prior Go cognition and place a stone in territory that the most powerful pattern-recognition system ever built had deemed impossible. In the most precise technical sense, he was the only person on earth who had proven that the boundary of machine intelligence was not the boundary of human intelligence. And he walked away from the board.
He was not wrong to. The incentive structures gave him no reason to stay. The field he had devoted his life to had been restructured around an entity his formation could occasionally, brilliantly, historically exceed — but could not consistently match. The rational actor exits. And in exiting, he takes with him the only thing the algorithm could not replicate, the only thing thirty years of the most demanding formation in the history of the game produced, and leaves it on the table, unclaimed.
That is the cost of getting this wrong at the individual level. Scale it to an organization. Scale it to an economy. The question is not whether the machines will eventually be able to do what Lee Sedol did. The question is whether, in the meantime, we are building the conditions under which the next Lee Sedol — the analyst, the operator, the clinician, the leader, the engineer — is being formed. Or whether we are building organizations so perfectly optimized for the known space that the people inside them never develop the capacity to see past it.
Move 78 was not a product of genius. It was a product of formation. Genius is unteachable. Formation is a choice — an organizational choice, a cultural choice, a leadership choice — that gets made or not made every day, in every decision about how work gets structured and who gets the latitude to be wrong in ways that matter.
The superintelligence debate asks the wrong question. It asks: what happens when the machine becomes smarter than us? The move 78 argument answers a different question, the one that is already live, already consequential, and already being decided without most of us noticing: what happens when we become less than what we were — not because the machine took something from us, but because we handed it over in exchange for a cleaner dashboard and a lower variance on the quarterly report?
The map is extraordinary. It is the most powerful map ever built. But the map has an edge. Someone has to be willing to walk to it. Someone has to have been formed well enough to see what’s there. And someone has to have built, deliberately and at real cost, the conditions under which that person could have become who they needed to be.
That is not a technology problem. It is a leadership problem. And it starts with knowing the difference between the two.
Structure before sophistication—every time.
SOURCES
References
Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. The Quarterly Journal of Economics, 118(4), 1279–1333. https://doi.org/10.1162/003355303322552801
Benfey, O. T. (1958). August Kekulé and the birth of the structural theory of organic chemistry in 1858. Journal of Chemical Education, 35(1), 21–23. https://doi.org/10.1021/ed035p21
Berger, C. (2024, November 19). Ben Affleck says AI can’t ‘replace human beings making films’ because it lacks ‘taste’. Fortune. https://fortune.com/2024/11/19/ben-affleck-ai-hollywood-filmmaking/
Braverman, H. (1974). Labor and monopoly capital: The degradation of work in the twentieth century. Monthly Review Press.
British Go Association. (2016). Google DeepMind Challenge Match – Lee Sedol v AlphaGo – match report. https://britgo.org/node/5458
Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and prosperity in a time of brilliant technologies. W. W. Norton & Company.
Chain, E., Florey, H. W., Gardner, A. D., Heatley, N. G., Jennings, M. A., Orr-Ewing, J., & Sanders, A. G. (1940). Penicillin as a chemotherapeutic agent. The Lancet, 236(6104), 226–228. https://doi.org/10.1016/S0140-6736(01)08728-1
Fleming, A. (1929). On the antibacterial action of cultures of a penicillium, with special reference to their use in the isolation of B. influenzæ. British Journal of Experimental Pathology, 10(3), 226–236.
Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.
Kasparov, G. (2017). Deep thinking: Where machine intelligence ends and human creativity begins. PublicAffairs.
Kekulé, A. (1865). Sur la constitution des substances aromatiques. Bulletin de la Société Chimique de Paris, 3, 98–110.
Kekulé, A. (1890). Benzolfest: Rede. Berichte der Deutschen Chemischen Gesellschaft, 23(1), 1302–1311.
Kohs, G. (Director). (2017). AlphaGo [Film]. Moxie Pictures; Reel As Dirt.
Metz, C. (2016, March 16). In two moves, AlphaGo and Lee Sedol redefined the future. Wired. https://www.wired.com/2016/03/two-moves-alphago-lee-sedol-redefined-future/
Metz, C. (2021). Genius makers: The mavericks who brought AI to Google, Facebook, and the world. Dutton.
NPR. (2016, March 13). After 3 losses, master Go player scores a win against computer. The Two-Way. https://www.npr.org/sections/thetwo-way/2016/03/13/470284113/after-three-losses-master-go-player-scores-a-win-against-computer
Pascale, R., Sternin, J., & Sternin, M. (2010). The power of positive deviance: How unlikely innovators solve the world’s toughest problems. Harvard Business Press.
Semmelweis, I. (1983). The etiology, concept, and prophylaxis of childbed fever (K. C. Carter, Trans.). University of Wisconsin Press. (Original work published 1861)
Shermer, M. (2008). Patternicity: Finding meaningful patterns in meaningless noise. Scientific American, 299(6), 48.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489. https://doi.org/10.1038/nature16961
Taleb, N. N. (2007). The black swan: The impact of the highly improbable. Random House.
Varoufakis, Y. (2023). Technofeudalism: What killed capitalism. The Bodley Head.
Visco, G. (2026). The metaproletariat [Manuscript in preparation].
Yonhap News Agency. (2019, November 27). (Yonhap interview) Go master Lee says he quits unable to win over AI Go players.
Zimmerman, M. J., & Bradley, B. (2019). Intrinsic vs. extrinsic value. In E. N. Zalta (Ed.), The Stanford encyclopedia of philosophy (Spring 2019 ed.). https://plato.stanford.edu/archives/spr2019/entries/value-intrinsic-extrinsic/