Sixty-Four Hexagrams, Sixty-Four Pages
— In the Age of AI, Where Does Judgment Live?
The fictional dialogue in this essay is written in Hong Kong Cantonese and presented with inline translations — because the essay's argument is precisely this: some things, once expanded, are already dead.
Prologue: Sixty-Four Pages
Eleven at night. Chan Sun was still editing page 41.
The client's question had only ever been one sentence: "This acquisition — can we pull it off?"
Chan Sun knew the answer on day one. The target's factory was genuinely impressive, the technology genuinely hard, the numbers genuinely good — but on the client's side, the integration team barely existed, and the integration budget had been conjured out of thin air. One sentence would have covered it: The asset is real. You can't handle it.
The problem was that one sentence doesn't justify a three-million-dollar advisory fee.
The partner patted his shoulder: "The client is paying three million. You can't hand them one sentence. We need a framework, a sensitivity analysis, scenario planning. Sixty pages minimum."
So Chan Sun opened AI. Forty minutes later, a sixty-four-page first draft was ready: market sizing breakdowns, a synergy matrix, three scenarios with nine sensitivity curves, two 2×2s, a Five Forces analysis. He then spent four days doing the one thing AI couldn't — making those sixty-four pages look like they'd taken three weeks of human labour.
The real answer was buried in a single line on page 64, in the appendix:
"The above valuations assume integration costs remain controllable."
Nobody would read to page 64. This was understood by Chan Sun, the partner, and the client alike.
Incidentally, the target company — the one with the "genuinely impressive" factory — had its main plant in the Greater Bay Area. Three production lines, running for over thirty years. Chan Sun had seen its asset number in the due diligence report. He had never seen the factory.
He certainly had no way of knowing that inside that factory was an old technician due to retire the following year, with a dog-eared Chinese almanac sitting on his toolbox.
His name was Uncle Cheung. He and Chan Sun would encounter the same hexagram, each at opposite ends of this story.
I. Two Arrows
Humanity has two arrows for dealing with complexity. They point in opposite directions.
One is called compression: taking a vast, chaotic, inarticulable world and distilling it into a handful of graspable symbols. In Cantonese, the single character mat (搣 — to pinch-tear) compresses action, force, contact type, and emotion into one syllable. Shakespeare carried the most meaning in the fewest words. A proverb passed down through three generations has behind it hundreds of unrecorded life tests. Compressive precision is precision-after-consensus — built not by definition, but by generation after generation of use.
The other is called expansion: taking a concept and splitting it into definitions, categories, conditions, metrics, and frameworks, then splitting again, subdividing again, until every possible ambiguity has been separately handled. Modern finance, management, policy, and corporate language all take this path. Expansive precision is institutional precision — it does not ask that you understand; it asks that you cannot argue.
Both arrows are called precision. But one folds complexity into symbols. The other spreads ambiguity into systems.
Chan Sun's sixty-four pages were the everyday form of expansive precision: the answer had existed from the start. The expansion was not for truth. It was for billing, for liability distribution, and for making the answer look sufficiently professional.
The farthest that compression arrow ever flew in the history of human civilisation happened three thousand years ago.
II. Humanity's Most Extreme Compression
At its base, the I Ching has only two symbols: an unbroken line and a broken line. Yang and yin.
Two symbols generate the four images. The four images generate the eight trigrams. The eight trigrams, stacked in pairs, generate the sixty-four hexagrams. The Xici Zhuan (Great Commentary) puts it plainly: "The Changes has the Supreme Ultimate, which generates the two modes, which generate the four images, which generate the eight trigrams." The information floor of the entire system is a binary code.
With this code, the ancients attempted to build a computable model of all change in the world: heaven and earth, advance and retreat, rise and fall, gathering and dispersal, crisis and turning point. Sixty-four hexagrams, six lines each, three hundred and eighty-four states — to describe infinity.
You might call this hubris. One man would not — Leibniz.
In the late seventeenth century, Leibniz invented binary arithmetic. A few years later, in 1701, the Jesuit missionary Joachim Bouvet sent him a diagram from Beijing: Shao Yong's Fuxi arrangement of the sixty-four hexagrams. Leibniz opened it and discovered that these sixty-four arrangements were structurally isomorphic with his binary numbers 0 through 63. He had not learned binary from the I Ching — he had invented it independently, and only then discovered that someone had built with the same symbolic logic three millennia earlier. He was so struck that he added a section on the Chinese symbols to his 1703 paper on binary arithmetic.
Today, your computer and the AI you use still run on binary. In other words: the oldest demonstration of human compression technology, and the newest machine of human expansion technology, operate on the same alphabet.
Here is a piece of history rarely mentioned, worth inserting — because it records the moment the direction between human and machine reversed.
Early computers had word lengths of 12, 18, and 36 bits — all multiples of three, because three bits made exactly one octal digit, and engineers liked reading memory dumps that way. This was the machine's aesthetics: mathematically clean, hardware-efficient. But this aesthetics had a problem — it couldn't fit humans. A 6-bit character had only 64 states: enough for uppercase letters and digits, with no room for lowercase. 36 divided by 8 equalled 4.5 characters; human text could never be cleanly sliced within the machine's word length.
In 1964, IBM's System/360 cut the knot: abandon multiples of three, define the basic unit as 8 bits, and call it a byte. Why 8? Because 8 bits could hold a complete human alphabet, uppercase and lowercase alike. In that moment, the machine bent to accommodate humans — the entire computational world's base architecture was redesigned so that human language could fit inside.
Sixty years later, the direction has reversed. Prompt engineering, as an entire discipline — what is it, really? It teaches humans how to expand their intent into a format machines can read. Context, role, constraints, output format — you must take something that could be said in one sentence and decompose it into a specification document before the machine performs well. This time, it is humans bending to accommodate the machine.
Same symbols. Both binary. Opposite directions. This reversal is the entire essay in miniature.
The Tao Te Ching takes the compression arrow even further: five thousand characters, covering cosmic operation, politics, warfare, self-cultivation. Its opening six characters — dào kě dào, fēi cháng dào ("The way that can be spoken is not the enduring way") — already declare the book's compression principle in advance: whatever can be fully articulated is not the thing itself.
These are not mysticism. They are a cognitive technology, and they have a design feature that modern institutional language entirely lacks. The next section explains. But first, back to Chan Sun — he has just submitted the report, and there is one person he has not yet asked.
Interlude I: One Sentence
Chan Sun's grandfather was ninety-two, living in a care home. On his bedside table sat a copy of the I Ching, so worn the spine was coming apart.
The old man had been a sailor in his youth. No university, no business school. The Saturday after submitting the report, Chan Sun visited him and mentioned the deal half-jokingly: "Grandpa, our firm is looking at an acquisition for a client. Big deal."
「講嚟聽吓。」 — "Tell me about it."
Chan Sun talked for ten minutes. Market share, synergies, EBITDA multiples — halfway through, he was boring even himself.
His grandfather listened, closed his eyes for a moment, and said one sentence:
「 即係人哋間廠好靚，但你哋班人接唔起。」 — "So the other side's factory is beautiful, but your lot can't handle taking it over."
Chan Sun went still. That sentence was the line on page 64 of the appendix. What he had spent sixty-four pages burying, his grandfather had said in one breath.
His grandfather opened the bedside I Ching and pointed to a hexagram. Lake above, wind below. Its name, two characters: 大過 — Dà Guò. Preponderance of the Great.
「棟橈。」 — "The ridgepole buckles." His grandfather said. 「條大樑彎咗——唔係樑唔靚，係兩頭太弱，中間太重。間屋愈靚，冧得愈快。」 — "The beam is bending — not because the beam is bad, but because both ends are too weak and the middle too heavy. The more beautiful the house, the faster it collapses."
The original text of the Tuanzhuan commentary: four characters — 「棟橈，本末弱也」 — "The ridgepole buckles; the base and tip are weak." The centre overpowered, both ends underpowered, the structure inevitably sags. Structural mechanics from three thousand years ago, which happened to be the due diligence conclusion of a 2026 acquisition.
Chan Sun asked: 「爺爺，你點知？」 — "Grandpa, how did you know?"
His grandfather smiled: 「我唔知。**係你自己頭先講咗畀我聽，你只係唔覺自己講咗。**我做嘅嘢，係幫你刪走嗰啲唔重要嘅。」 — "I didn't know. You told me yourself just now — you just didn't realise you had. All I did was delete the parts that didn't matter."
III. Compressive Precision Means Leaving a Hole for You
That line — "delete the parts that didn't matter" — is the core of compressive cognition, and its deepest difference from expansive systems.
The sixty-four hexagrams do not provide answers. A hexagram is a highly compressed structural model — "the centre overpowered, both ends too weak" — but whether this model maps onto your acquisition, your marriage, your body, or your country is for you to complete. The same hexagram read at twenty and at fifty yields entirely different things. The hexagram has not changed. You have.
This "you must complete it yourself" gap is not a design flaw. This gap is precisely where judgment grows.
You must personally decide what matters and what does not; personally align your own situation against that abstract structure; personally bear the consequences of a wrong judgment, and align better next time. What compressive language forces you to do is exactly this exercise — every act of interpretation is a set of weighted reps for your judgment.
Two misreadings need to be blocked here.
The first: compression is not vagueness. Compression gives you a skeleton and lets you fill in the flesh. "The ridgepole buckles; the base and tip are weak" is not a fortune-cookie slip — it is a structural-mechanics model that can be articulated, taught, and applied to any structure. True compression is depth rendered simply: deep, because it has been distilled from countless cases; simple, because it can be expressed in the fewest possible symbols.
The second: compression is not simplification. Modern business culture is full of things that look like compression — elevator pitches, one-minute briefings, executive summaries, three-point takeaways. On the surface, they too are "saying things in fewer words." But what they do is deletion, not encoding. A fifteen-minute presentation compressed into one minute: the deleted details are simply gone. Press further and it has no answers, because those things were genuinely discarded. By contrast, "the ridgepole buckles; the base and tip are weak" — take those four characters to an acquisition, to organisational management, to the rise and fall of a nation, and each time they expand into something different — because what was compressed was not detail but structure. Deletive compression is a report's corpse. Encoding compression is a seed.
This distinction connects directly to AI: AI is extraordinarily good at deletive compression — summaries, TL;DRs, bullet-point lists — a hundred times faster than any human. So if you think that producing daily executive summaries is practising compression, you are in fact practising something AI already does better than you. What the machine truly cannot do is distil from chaos a structural skeleton that can be reused again and again — the thing Uncle Cheung took thirty-eight years to achieve.
The design philosophy of expansive systems is exactly the opposite: its goal is to eliminate the gap. Every ambiguity is pre-decomposed, every situation pre-classified, every decision has a framework walking you through it. On the surface: safer, more professional, beyond dispute. In reality, it takes the act of "judging" out of your hands and distributes it across every layer of the institution — until, in the end, no one has judged, but the report is sixty-four pages long and everyone has signed.
When things go wrong, you will find that no one needs to take responsibility. Because no one truly judged.
Is there evidence for this? Yes. And the testing ground is in the asset number on page 12 of Chan Sun's report.
Interlude II: Forty-Seven Possible Causes
Three months after the acquisition closed, the integration team moved into the Greater Bay Area factory. Among the first to leave was Uncle Cheung.
Nobody fired him — he had been set to retire the following year, and the new owner offered an early-departure package, which he signed. Before he left, the new management assigned a "knowledge transfer engineer" to shadow him for six weeks, tasked with digitising his thirty-eight years of maintenance experience and feeding it into the group's AI knowledge base. In the integration budget, the project was categorised as "institutional knowledge preservation."
Day one. The young engineer brought a voice recorder: 「昌叔，三號線上個月嗰單故障，你係點判斷到係軸承問題嘅？」 — "Uncle Cheung, that breakdown on Line 3 last month — how did you tell it was a bearing problem?"
Uncle Cheung thought for a moment. 「聽聲囉。把聲唔啱。」 — "Listened to the sound. The sound wasn't right."
Four words that wouldn't fit in a knowledge base. The AI system required a decision tree. So six weeks became a three-hundred-page document: vibration spectrum analyses, sound-pressure comparison tables, seventeen bearing failure modes, discriminant flowcharts for each mode. The document passed every review. On launch day, someone in management said: "From now on, we won't need to rely on individual employees' experience."
On his last day, Uncle Cheung flipped through two pages of the document, put his almanac in his bag, and left one sentence behind:
「冇寫漏。不過呢三百頁，講嚟講去都係嗰四隻字。你哋將佢拆到咁散，第時邊個負責砌返埋佢？」 — "Nothing's missing. But these three hundred pages, when all's said and done, are still just those four words. You've taken them apart this finely — who's going to be responsible for putting them back together?"
Eleven months later, Line 3 went down.
The technician on shift was two years into the job. He followed every procedure, entering fault symptoms, vibration data, and temperature readings into the system. The system worked correctly — it expanded a complete diagnostic report: forty-seven possible causes, each with an attached probability, reference chapter, and recommended inspection steps. Thirty-two pages, beautifully formatted, logically rigorous.
The young technician began checking each cause. With every one he checked, the system expanded another layer of inspection procedure. Three days later, at cause number nineteen, downtime losses were already six figures.
Someone lost patience and made a phone call.
Uncle Cheung arrived, walked up beside the machine, and stood there for a moment.
「唔係軸承。」 — "It's not the bearing." He said. 「係地腳鬆咗。你哋係咪郁過條線？」 — "The machine's footing has come loose. Did you move this line?"
Yes. The integration engineering had changed the layout. The line had been shifted three metres.
The repair took forty minutes.
Afterward, someone checked: was "loose footing" in the three-hundred-page document? It was. Page 214, number 31 of forty-seven causes. The system hadn't been wrong — it had expanded all possibilities, including the correct one.
But expanding all possibilities and judging which one is the answer are two entirely different capabilities. The former, the system completed in three seconds. The latter — only one person in the entire factory could do it, and he had already left.
Before leaving, the young technician chased after Uncle Cheung: 「你係咪真係聽聲聽得出？教我。」 — "Can you really tell by listening? Teach me."
Uncle Cheung stopped. He thought for a long time — as if searching for something he had never needed to put into words.
「唔係話你企喺度聽三十八年就得。」 — "It's not like you just stand here listening for thirty-eight years and then you've got it." He said at last. 「部機同間屋一樣，你要知邊度承緊重，邊度應該實但係鬆咗。承重嗰度太剛，兩頭太弱，把聲就會唔啱——唔使等佢斷，聲已經話咗你知。你日日聽，唔係聽把聲，係聽緊呢個結構。」 — "A machine is like a house — you need to know where it's bearing load and where something that should be solid has come loose. When the load-bearing point is too rigid and both ends are too weak, the sound will be wrong — you don't have to wait for it to break; the sound has already told you. When you listen every day, you're not listening to the sound. You're listening to the structure."
The young man did not notice that Uncle Cheung's sentence and a hexagram commentary from three thousand years ago described the same model.
The ridgepole buckles; the base and tip are weak.
Uncle Cheung had never read the I Ching in his life. He only ever flipped through his almanac for auspicious dates. But he had spent thirty-eight years in a single factory and arrived at the same place on his own — from countless breakdowns, he had compressed a structural skeleton he could articulate. This is the true face of "listening to the sound": not mysticism, not innate talent, but an unnamed hexagram.
The three-hundred-page document preserved everything that had been expanded. The skeleton — the structure behind those four words — never entered the system. Not because it was mysterious, but because no field in the system's forms could hold "where the load is."
IV. AI Is an Expansion Machine
Now place AI into this picture. See clearly why it cannot hold those four words.
The essence of a generative model is perpetual regression toward the statistical centre of its training corpus — outputting the most predictable, the most mainstream, the most expanded expression. I wrote in "How a Machine Helped Execute a Deletion That Had Already Happened": it takes a three-word Cantonese dismissal — gaau m dim (搞唔掂, "can't sort it out") — and expands it into a verbose paragraph of standard written Chinese; it takes compression layers that deviate from the corpus centre — Cantonese tonal topology, three-layer phonetic encoding like hang ga chaang (杏加橙, a tonal cipher for a profanity) — and "corrects" them as errors and noise. That essay dealt with one language's fate. This essay's point is: that was not Cantonese's exclusive misfortune. It is the machine's default direction.
AI performs three systematic operations on compressive language:
First, it cannot read true compression. Decoding compressive language requires things outside the corpus — bodily experience, communal tacit understanding, situational context, time. It can recite every commentary on every hexagram, but it has no "situation" to bring into collision with the hexagram — it has no life. It can read the eight characters of "the ridgepole buckles; the base and tip are weak," but it has never stood before a machine whose sound was wrong. The bottom half of compressive language is not in the text, and AI has only the top half.
Second, it expands everything. Why? Because the statistical centre of internet corpus is inherently expansive text — documentation, tutorials, reports, listicles. Expansive language is overwhelmingly dominant in corpus because it was designed to be written down; compressive language lives between mouths and ears, between masters and apprentices, beneath the table within trades. What has been recorded is the tip of the iceberg. To regress toward the centre is to regress toward expansion. Forty-seven possible causes, three-thousand-word hexagram analyses, sixty-four-page due diligence reports — all the same operation.
Third, it requires you to expand before it will communicate with you. This was covered in Section II: System/360 was the machine redesigning itself to fit human language; prompt engineering is humans redesigning their language to feed the machine. The default direction of human-machine collaboration is humans accommodating the machine's expansive requirements.
The three operations together close a loop:
Elite culture rewards expansion (the more definitions and frameworks, the more professional it looks) → AI makes expansion zero-cost (sixty-four pages in forty minutes) → Expanded output exceeds what the human brain can process (no one reads to page 64; no one checks all forty-seven causes) → So AI is brought in to help read (summarise, analyse, expand again) → Human dependence on AI deepens → more tasks handed to AI → more expansion.
Each revolution of the loop raises complexity by one level, and reduces by one notch the opportunities humans have to complete a judgment themselves.
Note that there are no villains in this loop. Chan Sun did nothing wrong, the partner did nothing wrong, the knowledge-transfer project did nothing wrong, and AI did least of all — every party made the most rational choice from their own position. Just like the machine in "The Efficiency Trap": no villain required. Only everyone being reasonable.
V. What Is Lost Makes No Sound
The most insidious quality of this loop is that what it destroys does not cry out.
Judgment is not an asset. It is a muscle. Without use, it does not explode or trigger an alarm — it quietly atrophies. A person who has spent ten years without ever personally deleting their way from chaos to a key point will not know they have lost the ability to delete; they will simply feel that everything is very complex, and thank goodness for AI.
The two-year technician was exactly this. He was not stupid. He was hardworking. His "experience" was learning how to ask the system for more detailed expansion. He had no opportunity to learn what Uncle Cheung knew — not because he couldn't, but because the system was too good: so good that no one needed to stand beside the machine anymore. Judgment was not replaced. Judgment was immunised — the environment became one in which it had nowhere to grow.
And this is not happening in just one factory.
Every language has its own compression layer, and every compression layer is queuing up. Japanese honorifics compress three tiers of social hierarchy into a single verb inflection — AI translates the surface but not the precise shade of distance. German compound words fold entire concepts into one word. Indigenous languages compress millennia of ecological knowledge into place names — a name tells you where to find water, when the fish arrive. All of them deviate from their respective corpus centres. All will be "corrected" into expansive standard expression.
Not only language. A doctor's "clinical intuition," a veteran detective's "something's off," a chef's "the wok has enough heat," an engineer's "this code isn't clean" — every industry's core judgment is a product of compression, none of it fits into a knowledge base, and all of it is gradually being enclosed by expansive processes, metrics, and AI-assisted decision systems.
Cantonese was simply the first to hit the wall — political erasure and corpus bias, two blades falling simultaneously. But the reason it hit the wall is worth everyone remembering: not because it was backward, but because its compression rate was too high — so high that the machine treated it as noise.
Grandpa's ability — listen to you for ten minutes, delete down to one sentence. Uncle Cheung's ability — stand for a moment, hear the structure. Neither registers on any KPI. So when they disappear, no dashboard will sound.
A capability that reads as zero on every metric and as everything at every critical moment — the fate of this capability is what this essay is truly worried about.
Finale: The Ridgepole Buckles
Six months later, the client issued a profit warning: integration costs had tripled, more than half the core team had left, and the acquired assets were written down heavily. The Line 3 shutdown was classified in the announcement as a "one-off operational disruption."
No one mentioned page 64. No one mentioned page 214.
Chan Sun's firm was entirely unaffected — quite the opposite. The client's new CEO commissioned them to produce an "Acquisition Failure Root Cause Analysis." This engagement: four million, one hundred and twenty pages required. AI generated it in an hour.
The week the report was delivered, Grandpa passed away.
Clearing out his belongings, Chan Sun took back the bedside I Ching. He turned to the page for Dà Guò and found a folded slip of paper tucked inside, in his grandfather's handwriting. Four characters:
「棟橈。睇實。」 — "The ridgepole buckles. Watch carefully."
The date on the slip was two days after their conversation — a full fifteen months before the deal was signed.
What one hundred and twenty pages of root-cause analysis could not say, four characters had said fifteen months in advance.
That evening, Chan Sun tried to read the Dà Guò hexagram properly. He found he couldn't understand the original text, so out of habit he opened AI and pasted in the hexagram commentary. Three seconds later, the screen scrolled out three thousand words: historical context, line-by-line analysis, modern applications, five insights, three action recommendations, and a table.
He stared at those three thousand words and suddenly understood something: not one word was wrong, but what they collectively did was fill — completely, tightly, leaving no room — the gap his grandfather had left for him. The space he was supposed to walk into himself.
He closed the app. He sat with those four characters for a very long time.
From this moment on, he was finally reading the I Ching.
VI. The Opposite Direction
Now the essay's claim can be stated plainly.
The mainstream question of the AI age is: "How many more new tools do people need to learn to keep up?" This question itself is already standing on the expansion arrow. This essay proposes a question pointing the other way:
When the machine has taken expansion to its extreme, should humans retreat to the compression end?
This is not nostalgia. It is a cold calculation of comparative advantage: expansion — the machine does it ten thousand times faster than you; competing with it on expansion is a dead end. Compression — deleting down to the one sentence from the chaos, aligning your situation against a structure from three thousand years ago, making a judgment in the gap where no framework leads you — the machine cannot yet do this, and its training direction is fundamentally not headed that way.
So what needs preserving is not only Cantonese, though Cantonese was the first to be struck by both blades at once. What needs preserving is the compression layer inside every language: a single word that carries an entire paragraph, a proverb that carries an entire report's wisdom, a hexagram that carries the structural intuition of a one-hundred-and-twenty-page root-cause analysis. Each time one of these layers disappears, one more gym where humanity trains its judgment shuts down.
So how, concretely, does one practise? This essay is not selling courses, but Uncle Cheung and Grandpa have already demonstrated the entire method. It can be deleted down to three steps:
Step one: find your industry's hexagram. Uncle Cheung spent thirty-eight years compressing an entire factory into "where is the load, where has something loosened." Grandpa compressed an acquisition into "the ridgepole buckles." Your field has the same three core variables, the same one structural formula. The way to find it is not to read more but to delete more: take that hundred-page process document in your hands and force yourself to delete until only one sentence remains that is still true. If you can't, you don't yet truly understand.
Step two: guard the gap. AI can help you expand all possibilities, but that last moment — "which one?" — do not outsource. Which of forty-seven causes, which sentence in sixty-four pages is the real one — those three seconds are your only weighted reps. Let AI choose for you once, and your muscle atrophies by one notch; and when things go wrong in the future, you will find you can't even judge where the error was.
Step three: reverse the operation. The mainstream use case is to give AI one sentence and have it expand into three thousand words. Practise it backwards: let it generate three thousand words, then delete back to one sentence yourself — not by asking AI to summarise, but by your own hand. When you're done, ask yourself: if I had only three seconds, could I use this one sentence to make a judgment? If yes, AI is your tool. If no — you are not using a tool. You are slowly becoming part of one.
Three thousand years ago, someone wrote down the changes of the entire world using two symbols, then handed the right of interpretation to every generation. Three thousand years later, we built a machine that computes using the same set of symbols, and what it is doing is taking that right of interpretation back, one notch at a time.
Same symbols. Opposite direction.
Which side you stand on is the last judgment you still have the power to make.
This essay is part of the "Lucid Record" series. Its companion essay, "How a Machine Helped Execute a Deletion That Had Already Happened," records the specific fate of compressive language under AI's corpus gravity. "The Efficiency Trap" records the expansive loop's operation in the labour market. Three essays together: one on language, one on labour, this one on direction.
Lucid Record · Nebula Walker · mythogenengine.com