But it is about the thing this publication is named after, so stay if you can stand it. My brother Daniel died in September 2021. I’ve spent four of the five years since writing a book about him. It’s around 300 pages and it’s almost finished. Before I even consider an audio book version I have to make a decision about AI voice models, and I’ve ended up working that decision out with the help of a Chinese philosopher who has been dead for roughly 2,300 years. That’s the post.

Subscribe now

A plastic table with holes drilled through it

In late 2018 I converted a room under my house into a recording studio. Calling it a studio is generous. The previous owners had roughly constructed a subfloor space that was supposed to be a bedroom and never quite got there, and I filled it with a foldable plastic table my parents had bought me, four cheap office chairs, a mixer, and a PC I had lying around. I drilled holes through the table so the mic cables could run underneath, then threw a tablecloth over the top so nobody on camera would work out what was under there. The chairs broke almost immediately and stayed in service for two years anyway.

The microphones were the best I could afford at the time, because audio matters and some part of me already knew it was going to matter more later. I still have them.

Daniel called the room “Mum’s basement” on air. It was my house. I never bothered correcting him.

The first episode of Invert the Why went out on 20 January 2019. Daniel named it, after the control setting he used in every game he ever played, where you push forward to look up and pull back to look down. Most people never touch that setting. He swore by it. We recorded 69 full episodes over two and a half years (and one 20 minute intro that I still count separately, don’t ask), the first thirty-odd of them in that room.

What the files turned out to be

In 2019, a recording like ours was a fixed object. Four idiots, four microphones, one small room with hard walls, everybody bleeding into everybody else’s channel. If you wanted one voice on its own you were out of luck, and you certainly weren’t getting it out from under an office chair that creaked every time its occupant laughed.

That has changed, and mostly in the last 18 months or so. Software will now work out who spoke when, strip the room and the reverb and the chair off a track, and pull a single voice out of a mix where people are talking over the top of each other (Adobe’s Enhance Speech and iZotope RX both do good work here, and the ElevenLabs isolator is the one I keep coming back to). That last part still falls over on our worst overlaps… but it gets there more often than it doesn’t. Direction of travel rather than finished product, and the direction isn’t in doubt.

None of this generates anything. It recovers what was always on the tape and was previously sitting underneath three other idiots talking over the top of it.

Which means I can make an audiobook where I read my chapters and Daniel reads his.

The book already half does this on paper. Every one of his stories is tagged with the episode it came from, and there’s a QR code at the back pointing at the feed, because I wanted readers to be able to go and hear him tell it properly. His words sit indented and italicised so you can always see the exact point where I stop and he starts. In audio that indent stops being a typographic convention and becomes a cut to tape. His timing. His soundboard buttons. His inability to finish a sentence without three asides, a callback, and a loose sprinkling of profanity.

The thing I keep not doing

The other option is the one people assume I’d take. Clone the voice, feed it the manuscript, have Daniel narrate the entire book including the 51,000 words he never said. The tools are good enough now that it would take an afternoon and most listeners would never catch it. (ElevenLabs is still about the best for this, and it isn’t close imho.)

I’m not going to do it. I’ve had one rule since the first draft, which is that Daniel’s words are untouched. Not tidied in any way, not smoothed, and not corrected for grammar or for the fact that he swore in roughly every second sentence. A cloned voice would let me keep that rule on paper and destroy it in practice, because a listener would have no way of telling which sentences came off the tape and which came out of me. The archive is the only part of this project I can’t fake. Faking it is the one move that would cost me the whole thing.

I spend a lot of my working life arguing that institutions should own their data rather than rent it as a feature of somebody else’s product. Turns out I have a personal version of that argument, and it’s 300 pages long.

Zhuangzi drums on a tub

Somewhere in the second year of writing I started returning to Chinese philosophy (which I read a lot of when I was younger), mostly because the Western grief literature had stopped being useful. It kept arriving in stages, and my grief has never once done anything in a predictable sequence.

Zhuangzi’s wife dies in chapter 18 of the book that carries his name (older editions have him as Chuang Tzu, if you ever go looking for a secondhand copy). The chapter is called Perfect Happiness, and it also contains the passage where he interviews a skull, who tells him that being dead is better and declines the offer to be brought back. So it’s a strong chapter. His friend Huizi turns up to pay respects and finds him sitting on the floor with his legs sprawled out, banging on a tub and singing. Huizi is appalled and says so, at length. The bit most retellings leave out is Zhuangzi’s answer, which opens with him admitting he howled like anybody else when it happened. The singing came second. It came from working something through and out the other side, rather than from feeling less. Nobody in that story is composed because they loved less, which is more than I can say for most of what I read in 2022.

Then there’s chapter 26, External Things, which now reads as a joke at my own expense. The bit I mean sits right at the very end of it. The fish trap exists because of the fish, and once you’ve got the fish you forget the trap. The snare exists because of the rabbit, and once you’ve got the rabbit you forget the snare. Words exist because of meaning, and once you’ve got the meaning you forget the words. Zhuangzi finishes by asking where he might find a man who has forgotten words, so that he can have a word with him.

Something I only worked out properly last year, and it has been irritating me ever since. The book comes in three parts. The Inner Chapters, one to seven, are the ones scholars generally accept as the work of Zhuang Zhou himself. The Outer Chapters run eight to 22 and the Miscellaneous Chapters run 23 to 33, and both are largely later followers writing in his voice. Which puts the tub in the Outer and the fish trap in the Miscellaneous. Neither of the two best bits are reliably his.

I have spent four years being fanatical about which sentences in my book are Daniel’s and which are mine. Indenting. Italicising. Refusing to tidy a single word of his. And the most useful thing anyone has said to me on the subject was probably written by somebody else, wearing a dead man’s name, about 2,200 years ago. I don’t know what to do with that, so I’m leaving it here. (Brook Ziporyn’s 2020 translation is the one I’d put in your hands, mostly because his notes are honest about all of this. Burton Watson if you want the older standard.)

I named this Substack after the idea that good scaffolding exists to be discarded. Build the capability, then get out of the way. “Teach yourself out”.

Then I spent four years building the largest fish trap of my life. It runs to 300 pages, with 69 episodes of supplementary fish trap attached. Whether I caught the fish is the one thing the book can’t tell me.

Mozi thinks I’ve wasted four years

In Analects 17.21 a student called Zai Wo puts it to Confucius that three years of mourning is excessive and one year would do. Confucius asks whether he’d feel comfortable eating good rice and wearing fine clothes after a year. Zai Wo says he would. Confucius tells him to go right ahead, then waits until he’s out of the room before saying what he actually thinks of him.

What strikes me about the three-year period described in Chinese history is that it was structured, public, and required of you. You wore particular clothes. You stepped back from official work. Everyone around you could see the state you were in, because displaying it was the obligation. In the modern west (and as it’s usually portrayed in film) you get a fortnight of casseroles followed by a broad social expectation that everything should be back to normal. I’ve come round to thinking the casseroles are a shitter system.

Mozi disagreed with all of it. His essay on moderation in funerals argues that elaborate mourning impoverishes families, weakens states, and is performed at least partly for an audience. He’d have looked at four years and 300 pages and told me to get back to work. I don’t have a simple answer for him, and on some mornings I think he’s right.

Mostly though, I think he can get f*cked.

Voice and countenance

There’s a phrase you’ll find on Chinese funeral wreaths and memorial scrolls. 音容宛在, yin rong wan zai. The voice and the face seem still present.

It reads as consolation, and it’s also an admission, because it was written by people who knew perfectly well the voice was gone and were describing the sensation of almost being able to hear it. The spirit tablet works on the same principle. You write the name on a piece of wood, put it on the altar, speak to it, tell it the family news. The dead get relocated into an object and a set of obligations you keep up for the rest of your life.

I have 69 episodes. Hundreds of hours. I don’t need to write his name on anything, because he’s on Spotify, mid-sentence, arguing with Rocket Russell about Formula One and losing.

Two and a half thousand years of people wishing they could still hear the voice, and I’ve got mine as a file. I’m still working out whether those are the same thing.

What four years of this actually taught me

The book was never for me. I already have Daniel, in the way anyone has anyone once they’re gone, which is partially and unreliably and mostly at three in the morning. The book is for people who never met him, and it works on them, and watching it work has been the strangest part of the entire exercise. Strangers have opinions about my brother now. They quote him. One of my early draft readers told me she’d started the podcast on her commute and had to stop, because she was laughing too loudly on a train.

The audio is the part that’s for me, which is probably why I keep circling around the audiobook without starting it. Publishing the book was fine. Publishing his voice feels like a different thing, because a book can be shut while a voice arrives in your ears whether you were ready for it or not.

Zhuangzi says get the fish and forget the trap. The book comes out on 10 September 2026, five years to the day. After that I find out whether I can put the trap down.

Random Access Memories Book Pre-sale

Share

Subscribe now

I have started calling this “cringe LinkedIn slop”, along with everyone else on the internet, and that word “slop” has done a lot of work without anyone agreeing on what it names. Most people use it to mean bad AI writing (or images etc). I don’t think that’s quite it to be honest. Plenty of AI writing is technically better than what my tired brain would produce at eleven at night. The problem isn’t competence. A sentence can follow every rule of grammar, hit every beat of the expected structure, and still have nothing behind it.

I went looking for a different word for that nothing, for that “slop” and found one waiting for me in China, in a short essay written roughly fifteen centuries ago by a painter nobody outside art history departments has heard of.

The critic who graded painters

Xie He worked as a painter and critic at the Southern Qi court in southern China, which places his working life in the late fifth and early sixth centuries in and around modern day Nanjing (which is North West of Shanghai). Almost nothing about him is recorded with certainty, and even the date of his one surviving book is debated, usually placed around the middle of the sixth century. That book is a short volume ranking twenty-seven painters into three classes, called the Guhua Pinlu, the Record of the Classification of Old Painters. In its preface he laid out six principles for judging whether a painting was any good, and it just so happens to be the perfect tool for how I’ve been describing TASTE in the AI age.

The first principle is the one that matters most in this argument “Qiyun shengdong”, which translates to spirit resonance, life motion. Later writers have argued for centuries over how to translate it, and the phrase keeps resisting a clean English rendering, kinda like the quality it names resists any painter (or AI) who tries to fake it.

What Xie He meant, as close as I can get it, is that a painting needs an inner vitality that comes from the painter’s own cultivation and gets transmitted into the brushwork. A landscape can get every proportion right and every rock correctly shaded and still be dead on the silk. Spirit resonance is what makes the ink “breathe”.

Xie He ranked this principle first and said something that sounds almost harsh when taken out of context. Without spirit resonance, he wrote, there was no need to look any further at the rest. The other five principles, bone method in the brushwork, fidelity to the object’s form, appropriate use of colour, composition, and the copying of the old masters for training, all mattered. None of them mattered if the first one was missing to begin with. In other words, a painter could nail every technical requirement on the list and produce something with no reason to exist.

This became the cornerstone of Chinese painting criticism for the next fifteen centuries. Every serious discussion of a Chinese landscape scroll from the Song dynasty onward measures itself against a standard that has almost nothing to do with accuracy and almost everything to do with whether the thing feels alive (I’ve watched way too many historical Chinese dramas… sorry not sorry).

I reckon this describes slop more precisely than any English word I have found. Slop is therefore competent painting with no spirit resonance. It hits the five technical principles and misses the one that decides whether the other five were worth doing.

An old key at the bottom of the barrel

While Xie He gives us the diagnosis, David Hume, writing in Edinburgh in 1757, gives us something closer to a method for detecting it.

Hume’s essay Of the Standard of Taste opens with a paradox. Everyone agrees that taste is a matter of personal sentiment, that there’s no accounting for it, and everyone also agrees that some art is better than other art. Hume’s own example is Homer, the same Homer admired across two thousand years and a dozen nations while the fashionable authors of any given decade are forgotten by the next (yes the same Homer, the foundation of Western Literature Homer, that Lupita Nyong'o thinks could use her feedback… you can’t make this stuff up). If taste were purely subjective, that endurance should simply not be possible. Bad taste would survive exactly as long as good taste, since nothing outside sentiment would be doing the sorting.

Hume’s answer is that taste is subjective in the sense that it lives in the response of the perceiver, and it is not therefore arbitrary, because some perceivers are better equipped to perceive. He tells a story lifted from Don Quixote to make the point concrete. Two of Sancho Panza’s kinsmen are given a hogshead of wine and asked for their judgment. One detects a faint taste of leather. The other detects a faint taste of iron. Everyone else laughs at them, certain the wine is perfectly good. When the barrel is drained, an old key is found at the bottom with a leather thong tied to it, iron and leather together.

The two men had detected a real fault that the rest of the room lacked the equipment to notice.

Hume’s true judge combines five things, and each one is acquired rather than innate. There is strong sense, the sound judgment that keeps a critic from being fooled. There is delicacy of taste, senses fine enough to register faults and merits that pass most people by. There is practice, sustained exposure to one kind of art until the eye stops being distracted by novelty and starts recognising quality. There is comparison, because you can’t know a work is excellent until you’ve set it against enough others to know what excellence costs. And there is freedom from prejudice, the willingness to judge a work on its own terms rather than the terms you walked in expecting. Hume compresses all five into a single sentence.

Strong sense, united to delicate sentiment, improved by practice, perfected by comparison, and cleared of all prejudice, can alone entitle critics to this valuable character.

David Hume, Of the Standard of Taste, 1757

None of the five is a rule you can formally look up. That’s what makes Hume’s account sit so close to Xie He’s, despite the two of them never having heard of each other, being separated by twelve centuries and most of the Eurasian landmass. Both describe a form of judgment (I call taste) that cannot be reduced to a simple “checklist”, because a checklist is exactly what a person with no delicacy substitutes for the judgment they haven’t developed.

What a system trained on the whole internet learnt

A large language model has read more prose than any human critic in history, fact. It has ingested essays, novels, marketing copy, forum posts, academic papers, and inspirational LinkedIn threads by the billion (that last part hopefully came off as funny as it sounded in my head). In Hume’s terms it has more practice and more material for comparison than anyone who has ever lived.

But at the same time it has none of the other three. It has no strong sense, because there is no judging mind behind the output doing the sensing. It has no delicacy, in the sense of a perceiving subject whose organs of sensation can be finer or coarser, because there is no perceiving subject there at all. And it has no freedom from prejudice in the deeper sense Hume meant, because it has no standpoint from which to judge a piece of writing against what that particular writing was trying to do. It has instead a finely calibrated sense of what writing like this usually looks like, gathered from an enormous pool of examples of every quality level mixed together.

This is why the output reads exactly the way it does. The grammar is flawless and the structure holds together, and underneath both sits something utterly average (I told my son that would be a great band name btw “utterly average”), because average is what you get when you learn from everything at once with no capacity to prefer the leather and the iron over the smooth generic wine that everyone else in the room was drinking.

Xie He would have figured this out instantly. It is a system that has mastered the technical principles, everything below qiyun on his list, and has no access whatsoever to the one he ranked first. It produces landscapes with every rock in the right place and no “breath” moving through any of them.

What this is not

I’ve had this conversation so many times and know where it leads, so for those of you feeling a tad upset now this is not an argument that AI writing is always slop or that human writing is always alive. Plenty of human writing has zero spirit resonance (just look at the last decade of Hollywood). Every University/corporate memo ever produced by a committee is proof that dead prose does not require a machine. And I have read AI-assisted writing that surprised me, usually when a person with delicacy of taste used the tool as a drafting instrument and then did the work of judgment on top of it. The tool did not supply the spirit resonance. The person did, in the choosing and cutting and rewriting afterward.

What Xie He and Hume are both pointing at is the presence or absence of of a special form of judgment, the one I keep calling taste. The machine question is a distraction. A human writing on autopilot, reaching for the phrase that sounds like the phrase that usually goes here, is doing what an unguided model does too. The output happens to come from a person instead of a system, and it is just as dead.

What the AI writing crisis has done is make this failure visible at scale for the first time. When one tired writer produces a flat paragraph, it’s a bad Tuesday. When a tool that can generate that same flat paragraph a million times a minute gets deployed across every marketing team and content farm, the flatness stops being an occasional lapse and becomes the ambient texture of large parts of the internet (or Linkedin specifically). Xie He’s distinction between technique and vitality used to be a private argument between painters and critics. It is now a very public crisis, and I’m loving it.

Why this can’t be automated away

The obvious response is to ask whether a future model could be trained to detect its own slop and fix it. Some of this is already happening. Prompting techniques exist to push a model toward a specific voice and away from the generic middle it defaults to.

I think this helps, and I think it hits a ceiling, for the same reason Hume’s account of the ideal critic does. Hume is explicit that delicacy of taste cannot be conferred by telling someone the rules. It has to be built by sustained exposure to a particular species of beauty until the perceiving instrument itself changes. Practice, in his account, does not mean following a procedure. It means being altered by repeated contact with quality until you can feel the difference between the leather and the wine without having to reason your way there.

A model does not have an instrument that gets altered by contact in that sense. It has weights that get updated by gradient descent against a loss function, which is a different kind of change entirely, one that optimises toward whatever the training signal rewards rather than toward a perceiving subject’s refined response to beauty encountered over years. You can train a model to produce output that scores well against a slop detector, sure. But you cannot train a model to have taste, because taste in Hume’s sense requires there to be someone home whose sensory organs are getting finer with practice, and in this case there is nobody home. There is no Number 5 is Alive, at least not yet.

None of this depends on a claim that current architectures can never change. It rests on what the word taste has meant every time a serious critic has tried to define it (myself included) whether in Xie He’s sixth century Chinese preface, Hume’s eighteenth century Scottish essay, or the notes my professor scrawls on my never ending PhD proposal drafts. It has always meant a cultivated capacity that lives in a perceiving subject. No checklist of textual properties has, or ever will stand in for this.

Where I land

The word slop caught on because it named something people were feeling before they had a proper vocabulary to describe it. I don’t think the internet needed Xie He or Hume to notice that a lot of writing had started to feel the same, and breathless. What the two of them offer is a way of being precise about what exactly is missing, instead of waving at the whole category of AI-generated text as if the problem were the tool rather than the absence.

The five technical principles still matter. A sentence needs correct grammar. An essay needs a structure that holds together and facts that check out. None of it substitutes for the first principle, and no amount of getting the other five right will manufacture it after the fact.

What I think we can all take from this, as for me personally as someone who writes a hell of a lot, is that the process of being an artist/writer/musician/any other professional hasn’t changed as much as the discourse suggests. The job of a writer, for example, was never to produce technically correct prose. It was always to notice when something is missing that most people in the room can’t detect. Xie He spent his life training that noticing in painters. Hume spent his life trying to defend it against people who thought taste was just opinion dressed up as expertise. I suggest we need to all notice what is missing when we use AI within our own disciplines or fields of expertise. That is part of being a frontier operator. Unfortunately in all its wisdom, our wonderful Australian government keeps trying to kill the arts and arts education. When AI is only capable of producing breathless slop, who will be left to make things excellent?

Share

Then he listed the fields. Experimentation, applied mechanics, continuum mechanics, computational mechanics, composites, differential calculus, tensor field theory, computer science, numerical methods. “I don’t live long enough,” he said, “to be an expert in all of that.”

And then the event that should’ve drawn more LinkedIn attention than it did. AI let him cross that chasm. Not by having the ideas for him. By holding nine disciplines at once so that the idea he’d been carrying for years could finally get built.

We both now think research is going to have to change. It has to move away from “I wrote this, aren’t I clever,” and away from “shame on you for using AI to write that sentence.” It has to move toward something simpler and much harder. Look what I achieved, by any means available. He has a name for the person who works that way. The Frontier Researcher, which ties nicely to my whole Frontier Operations thing.

The key phrase to hold on to is by any means available, because the thing stopping us from working that way is not technical. The technology is already here, and some of us use it daily. The obstacle is in fact the oldest one there is. It is ego. And the sharpest demolition of ego I know took about a minute, in southern China, fifteen hundred years ago, between an emperor with a magnificent CV and a monk who refused to read it. I love this story.

The emperor’s ledger

Around 527 CE a monk from India arrived at the court of Emperor Wu of Liang. The monk was Bodhidharma, the figure Chan (aka Zen) Buddhism counts as its first patriarch, the man the Japanese tradition later called Daruma (if you have visited Japan, you may have seen Daruma dolls)

By any reasonable accounting, Emperor Wu was the greatest patron Buddhism had ever had in China. He had built hundreds of temples. He had ordained thousands of monks and nuns. He had funded the copying of sutras, banned animal sacrifice, and lectured on the scriptures himself. People called him the Bodhisattva Emperor. He kept, in effect, a ledger, and the ledger was magnificent.

So when the famous monk arrived, the emperor opened with the question the ledger had been built to answer. The exchange survives as the first case of the Blue Cliff Record, the great koan collection compiled in 1125.

The emperor asked, “I have built temples and ordained monks beyond counting. What merit have I gained?” Bodhidharma said, “No merit whatsoever.” The emperor asked, “What is the first principle of the holy teaching?” Bodhidharma said, “Vast emptiness, nothing holy.” The emperor asked, “Then who is this standing before me?” Bodhidharma said, “I don’t know.”

Blue Cliff Record, Case 1, trans. after Cleary (1977)

Then he left the court, crossed the Yangtze, and according to the legend spent the next nine years facing a cave wall.

This story is so famous it actually slides straight past you. The emperor’s works were real. The temples stood. The monks were fed. The sutras were copied, and people read them, and some of those people were changed (this is all real and documented). Bodhidharma never says the works were worthless. He says the merit is nothing. And merit, the way the emperor meant it, is the credit accruing to his own account. The personal balance. The cosmic h-index.

The teaching isn’t that good works don’t matter. It’s that the moment you do them while keeping score, you’ve turned the work into a transaction, and the thing the ledger was supposed to buy, the awakening, the actual point, is exactly the thing that can’t be bought. The accounting in other words, destroys the very thing it counts.

The third answer goes further still. Who is standing before me. I don’t know. Bodhidharma won’t even hold on to the self that could be credited. No ledger, and no account holder either.

The academy’s ledger

Now read my friend’s complaint back through that cave wall.

“I wrote this, aren’t I clever” is Emperor Wu’s question wearing a lanyard. A modern research career is, when you look at its actual structure, a merit ledger. First-author papers. Citation counts. The h-index. Named grants, named labs, named chairs. Every layer of the machinery, the hiring committees, the tenure dossiers, the grant panels, the REF in Britain, the ERA here in Australia, treats accumulated merit as its basic unit of account, because accumulated merit is legible and actual contribution is not.

Anyone who has spent time inside the system knows the specific damage the ledger does. One finding gets salami-sliced into four papers, because four lines on a CV beat one. Data gets hoarded until publication, because shared data produces solved problems and not attributed ones. The fundable question wins out over the important one. Authorship disputes end collaborations and, every so often, careers. None of this is a corruption of the system. It is the system, because the system pays in merit, and merit needs the contribution to be severable, countable, and yours.

The objection my friend braces for, “shame on you for using AI to write that sentence,” is a ledger objection. It says nothing about whether the finding is true, reproducible, or useful. It asks only whether the credit can be cleanly assigned. A real result doesn’t get less real because a language model held the tensor field theory while a human held the question. A hollow paper doesn’t get more substantial because every word was typed by hand. The Confucians would call this whole convulsion a failure of zhengming, the rectification of names (look at my Ancient Wisdom series to read more about this). The word “research” was supposed to point at finding out what’s true, and somewhere along the line it drifted, and started pointing at the accumulation of attributed outputs instead. When my friend says research has to move to “look what was achieved, by any means available,” he isn’t proposing a revolution. He’s asking that the word be made to mean what it always claimed to mean.

I should totally point out that my Ancient Wisdom series has crossed this ground before from another direction. The Bhagavad Gita’s idea of nishkama karma, act fully and release your claim on the fruits, ran through a couple of essays so far. If you’re new here and you want the Hindu tradition’s version of this argument, check it out. Krishna and Bodhidharma are pointing at the same grip from different mountains. Krishna says do the work and drop your claim on the outcome. Bodhidharma goes one step more ruthless. Drop the ledger entirely, then ask who was keeping it.

The counterargument I can’t dismiss

Earlier in the Ancient Wisdom series I argued, through Dōgen, that practice and realisation are one thing, that you can’t peel the outcome off the process and still call the understanding yours. I argued, through Cook Ding, that expertise gets built in the hands across years of finding the joints. And I argued that using AI to skip the practices that would have built you is aversion to discomfort wearing the costume of efficiency.

So am I now arguing the opposite? If the Frontier Researcher hands the synthesis to a machine, is the synthesis still happening to anyone at all? Is a whole generation of researchers about to receive deliverables in place of understanding?

This is the strongest objection, and I half believe it. Here is the distinction I think survives (although to be fair as someone who loves to write and bring ideas together I am still wrestling with this).

First listen again to what my friend actually said. His brain already knew what was possible. The idea was formed. The judgement was there. Thirty years of practice had already built the one thing Dōgen says you can’t skip, the trained perception that tells a real problem from a decorative one, a plausible result from an artefact, a joint from a bone. What the AI bridged was never his understanding. It was the brute combinatorial fact that no single human lifespan holds expert fluency in nine disciplines at the same time.

Cook Ding’s knife stayed sharp for nineteen years because he cut at the joints. But there are places in any ox where, for one person, no joint exists, where the only way through is to hack at bone until the blade chips. The nine-discipline synthesis was never my friend’s practice. It was the bone his blade kept hitting. The fiction was the romantic idea that one mind should have contained it all. Using a tool to hold it is not the fiction. It’s the correction.

The test from the Dōgen essay still works, and it cuts both ways. Did the work change how you see the problem? For my friend, plainly yes. What he can now build is reorganising how he thinks about his whole field. That’s practice, amplified. For a doctoral student who prompts their way to a literature review they never digested, the same tool does the opposite. It produces a deliverable that changed nothing in the person holding it. The tool is identical, but the relationship is everything. That has been the conclusion of nearly every essay in this series, and it holds here too.

The Frontier Researcher, then, is not someone who skipped the practice. Thay are someone whose practice finally gets to matter. Which means the title has to be earnt. My conclusion has always been that you can’t start at the frontier, but rather that you need to arrive there.

The institution is the emperor

This is where the essay has to get harder, because the ego in question isn’t mainly personal.

An individual researcher can sit with Case 1 of the Blue Cliff Record. An institution can’t. A junior researcher who genuinely stopped caring about attribution tomorrow would be performing a beautiful spiritual exercise and committing career suicide in the same gesture, because the machine she works inside pays only in merit. Emperor Wu, to his credit, kept a ledger for himself alone. The academy keeps one for everybody, and makes your employment depend on the balance.

So the realistic question isn’t whether researchers will let go of ego. It’s whether anyone can build an institution that doesn’t force them to hold on.

There are fragments already, and they teach you something. The big physics collaborations, LIGO and CERN and the thousand-author papers, long ago let individual attribution dissolve into something closer to “the instrument found it,” and they detected gravitational waves anyway. The open-source software world has run on “look what was achieved, by any means available” for forty years. Its unit of prestige is the working artefact, and nobody audits which lines were typed and which were generated. Mathematics has the Polymath projects, where theorems got proved in comment threads and the question of who proved them became formally unanswerable and quietly beside the point. Each one is a small experiment in running research without the ledger. Each one produced things the merit economy could not have.

So the honest version of my friend’s prediction is not that the academy will have an awakening. It’s that the fields whose institutions learn to reward outcomes over attribution will start solving problems faster than the fields that don’t. The gap will compound. And at some point the comparison gets humiliating enough to force the question. Institutions don’t cross the Yangtze on their own. They get dragged across by their competitors.

What I’m working through

I notice my own grip when I write these essays. There’s a small, specific pleasure when a paragraph lands and I know the synthesis was mine, Bodhidharma to zhengming to citation metrics, joints found, blade intact. “I wrote this, aren’t I clever.” It is exactly the pleasure my friend is describing, and exactly the entry Emperor Wu kept making in his ledger, and I am not free of it. I doubt anyone who makes things is.

The koan isn’t asking me to lose that pleasure. Bodhidharma never tells the emperor to stop building temples. The instruction is more uncomfortable than renunciation. Keep doing the work, and watch for the moment the question quietly changes from “is it true” to “is it mine,” because in that moment you’ve stopped doing research and started doing accounting. The temples stand either way. The only thing that changes is the builder.

My friend doesn’t live long enough to master nine disciplines. Neither do I. Neither does anyone. For the whole history of science that sentence was a wall, and we built everything around the wall. The specialisations. The departments. The careers. The egos sized to fit one discipline, because one discipline was all a single life could hold.

The wall just moved and the only question left is whether we’re more attached to the problems or to the credit for solving them. The emperor asked who deserved the merit, and the answer was no merit whatsoever. He asked what the holy truth was, and the answer was vast emptiness, nothing holy. He asked who was standing in front of him.

I don’t know.

But I know which fields I’d bet on.

A note on the history. The meeting between Bodhidharma and Emperor Wu is almost certainly legend rather than record. Ive spent DAYS trying to find more out. The earliest accounts actually appear centuries after the supposed event, and most scholars read it as a later Chan invention built to make exactly the point it makes. I’ve kept it because the koan has done fifteen hundred years of honest work even if the meeting never happened. The temples Emperor Wu actually built are not in doubt, even if the conversation probably is.

Miya was indignant in the specific way a fifteen-year-old is indignant about fairness, which is to say with a theory already attached. “It’s not fair if the people who actually do the work have to carry everyone else. You don’t fix a problem by spreading it out.”

Rei, who was chopping spring onions at the counter, said something in Japanese (which means she was talking to me), half under her breath, that I only half caught. Then she turned to Miya and said, in English, “Fair to who?”

It’s a genuinely hard question. Fair to the students who’d end up doing most of the work? Fair to the ones who’d been lumped together and written off before the project even started? Fair according to effort, or according to outcome, or according to some principle that existed before any of these particular teenagers walked into the classroom?

I’ve been thinking about Rei’s question for a while now, as I didn’t really expect it (but for broader context, we have had a string of misgivings about the state of public education). A philosopher in warring-states China asked it with extraordinary precision about 2,400 years ago, built an engineering school around his answer, and was so thoroughly defeated that his name nearly vanished from history. The entire field of algorithmic fairness is now stuck, with mathematical proof, on exactly the dilemma he posed.

The man who disagreed with Confucius

Mozi (墨子) was born around 470 BCE, as the age of Confucius was passing (for reference, about 250 years before Qin Shi Huang unified China as its first emperor). He came from the artisan class, possibly a carpenter or military engineer, and his movement drew from craftsmen and soldiers rather than the aristocratic scholars who filled the Confucian schools. This matters. Mozi thought like a builder. His arguments have the quality of someone who measures twice and expects the structure to stand up under load.

His central claim was jian ai (兼愛), usually translated as “impartial care” or “inclusive love.” The argument in the Mozi is structural rather than sentimental. Disorder in the world arises because people weight their own families, cities, and states above others. The lord who attacks a neighbouring state, the family that hoards while others starve, each treats those close to them as more real and more deserving than those who are distant. Mozi’s name for this is 別 (bié), the act of making distinctions, and he considered it the root cause of almost all large-scale suffering . Replace partiality with equal regard for all, he argued, and the engine of conflict simply loses its fuel. If you cared for another’s state as your own, who would you have left to conquer? As an interesting aside, the West reached a roughly similar intuition only much later, with Hegel’s self and Other in 1807 and de Beauvoir’s woman as the Other in 1949, and the term “othering” itself not coined until Spivak’s postcolonial work in the 1980s, well over 2,000 years after Mozi, though strictly that tradition is about constructing an out-group rather than his point about partiality. I remember studying this in depth at uni back in the late 90s…

This was, by any reasonable reading, then the first systematic consequentialist framework in recorded philosophy. Mozi held that actions should be judged by whether they benefit “all under heaven” (天下之利), and he proposed three tests for any claim or policy. He asked whether it was grounded in precedent, whether it could be verified in common experience, and whether it brought practical benefit to the people. Benefit meant measurable welfare, things like material security and social order, applied impartially across everyone.

Jeremy Bentham published An Introduction to the Principles of Morals and Legislation in 1789. Mozi had the core of the same argument in circulation by roughly 430 BCE. The gap is again about 2,200 years.

Mozi at the switch

Here is the thought experiment the title promises, and Mohism turns out to be built for it. This was inspired by a video I watched recently (AI buys robot and car, does exactly what experts warned). Sorry this get’s a bit dark… but you are welcome for the rabbit hole!

A runaway trolley is hurtling toward five people tied to the track. You stand beside a lever. Pull it and the trolley diverts onto a side track, where it kills one person instead. Do you pull?

Mozi pulls the lever. For a thinker who weighs every life equally and judges every act by benefit to all under heaven, five against one is simple arithmetic. Jian ai means no life counts for more because of who it belongs to or how close it stands to you, so the calculation is clean and the calculation is the morality. Most people, asked the same question, pull the lever too, which is part of why Mozi’s instinct feels so modern. He formalised, in the fifth century BCE, the intuition that the right act is the one that produces the most good for the most people, counted without favouritism.

Now the philosophers spring their trap, the footbridge variant. Same five people, same trolley, but no lever. This time you stand on a bridge above the track beside a large stranger, and the only way to stop the trolley is to push him off into its path. His body will halt it, he will die, and the five will live. The maths is identical, one death to prevent five, yet most people who happily pull the lever recoil from the push. Something in us separates redirecting a harm from using a person as the instrument of preventing it. Pure outcome-counting cannot see that line. The footbridge is designed to expose exactly the blind spot that naive utilitarianism has.

This is where Mohism turns out to be far stranger and more interesting than a 2,400-year-old Jeremy Bentham. Alongside jian ai, the Mohists held a second doctrine that constrains the first, 非攻 (fēi gōng), the condemnation of offensive war. The Mohists made their name as defensive military engineers, travelling to small states under threat to help them hold their walls, and they would not, on principle, help anyone attack. There is a famous episode in the Mozi where the master walks for 10 days and 10 nights, wearing through his feet, to reach the capital of Chu and stop it from invading the small state of Song. He does not raise an army. He sits down with Gongshu Ban, the celebrated siege engineer who designed the assault ladders, and defeats him nine times over in a tabletop simulation of the siege, using a belt for the city wall and small sticks for the engines of war. When Gongshu Ban hints that the simplest solution would be to kill Mozi on the spot, Mozi answers that it would change nothing, because 300 of his disciples already stand on Song’s walls with his defences in hand. Chu calls off the war. He stops the killing without doing any.

So the Mohist at the switch is genuinely cross-pressured, and that is the point. The consequentialist core says pull the lever, five over one, counted impartially. But fēi gōng resists initiating harm against someone who was not already in the trolley’s path, and diverting the trolley onto the one, or worse, shoving the man off the bridge, is the system reaching out to inflict a fresh and deliberate harm on a specific person for the sake of an aggregate. The genuine Mohist will not use anyone as a means. There is a wall in the doctrine against being the author of new harm, even good-faith harm, even efficient harm.

That cross-pressure is the modern fairness debate wearing different clothes. It is also an interesting thing to think about when you drive a modern self-driving car…

The Confucian objection

The Confucians found Mozi’s position dangerous, and their objection deserves to be stated carefully, because it is not the straw man it often gets reduced to by academics (at least in my mind).

Confucian ethics are built on ren (仁), benevolence or humaneness, and ren operates through relationships. It is also one of the easiest characters in Chinese to remember. The radical on the left means “person”, and the character on the right means “two”. When you smoosh them together, the combination of "person" and "two" signifies how a person interacts with others (you’re welcome for today’s Mandarin lesson). Accordingly you owe the deepest care to your parents, then your family, then your community, then your state, then all people, the care radiating outward in concentric circles and thinning with distance. This is sometimes called graded love.

Mencius came along about a century after Mozi, by which time Mohism had grown into one of the dominant schools of the age, and he was the most formidable Confucian of his generation. He attacked jian ai on two fronts that are worth keeping apart, because they are easy to run together yet they do very different work. The first is just a passing insult. The second is a genuine argument.

The insult is in Mencius 3B.9. Surveying a realm he sees as fallen into intellectual chaos, Mencius lumps Mozi together with the egoist Yang Zhu. Yang preached every man for himself, which amounts to denying one’s ruler, and Mozi preached love without distinction, which amounts to denying one’s father, since a father loved no more than a stranger is no longer a father in any meaningful sense. To acknowledge neither father nor ruler, Mencius says, is to be a beast (禽獸). It is a brutal line if you understand the context of the period, but it is not really an argument.

The genuine argument, however, is in Mencius 3A.5 in an exchange with the Mohist Yi Zhi. Yi Zhi was a committed follower of Mozi’s doctrine which includes descriptions of “frugal burials.” Yi Zhi had, however, given his own parents a lavish funeral with the robes and all (as an aside those white hemp robes you see at funerals in Chinese period dramas come from the Confucian mourning system, the very apparatus Mozi was attacking, where the cloth gets coarser and the mourning lasts longer the closer your kin, which is Mencius’s graded love made literal in fabric). Mencius attacks this contradiction as if Yi Zhi truly held that love admits no gradations, he wouldn’t bury his parents differently from strangers. Yi Zhi responds by appealing to the sages, who cared for the people “as for a newborn babe,” reading this to mean love should be without distinctions even if it begins with one’s parents.

Mencius’s subsequent rebuttal is the heart of the Confucian case. He asks whether Yi Zhi honestly loves his brother’s child no more than a stranger’s, and then makes the structural point. Heaven gives each thing a single root (一本), and Yi Zhi’s doctrine tries to give it two (二本), which is his error. Then comes the macabre image that makes the point land. Mencius imagines a time before anyone buried their dead, when people would leave a dead parent in a ditch. Passing the spot days later, they would find the body eaten by foxes and flies, and a cold sweat would break out on their brows. They could not bear to look. That flinch, Mencius says, is where burial began, and where moral feeling begins too. Nobody reasoned their way to it. They felt it, instantly and without choosing to, and they felt it for their own parent.

This is what Mencius means by the single root. Moral feeling starts as a spontaneous pull toward the people nearest to you, and graded love grows straight out of it. Mozi’s impartial love asks you to set that pull aside and follow a rule instead, the rule that everyone counts the same. You can argue for the rule, but it does not grow from anything a person actually feels. It is a second root tied onto a tree that only ever had one.

The Confucian position is that partiality is the structure through which any genuine care operates at all. You learn to be good to strangers by first learning to be good to your parents, so the particular becomes the training ground for the general. Abolish the particular and you hollow out the only thing that taught you to care, leaving a thin, abstract benevolence attached to no one in particular and therefore to no one at all.

I return to these ideas often, for a number of reasons beyond the scope of this essay. Both positions are internally coherent and both are pointing at something real. Mozi is right that treating your group as more deserving of care than other groups is the engine of most large-scale suffering. Mencius is right that abstract universal concern, cut loose from the particular relationships that constitute a life, tends to evaporate into nice-sounding principles with no force.

This is where algorithmic fairness has been stuck for the past decade.

The algorithm chooses

When a machine learning model makes a decision that affects people, someone has to define what fair means in mathematical terms. This turns out to be extraordinarily difficult, and the difficulty maps onto the Mozi and Mencius debate with uncomfortable precision.

Consider a model a university uses to score applicants by how likely they are to finish the degree if admitted. The training data is decades of past admissions and the outcomes that followed, decisions made by human beings with human biases and fed by a schooling system that was never equal to begin with. Applicants from wealthier areas arrive from better-resourced schools, with tutoring and coaching behind their marks, so the data records who was advantaged as much as who was able. It is a record of what happened, and what happened was not fair.

Train a model on that record and optimise for predictive accuracy, and the model will learn the patterns in the data, including the ones the inequity produced. It inherits the biases of the system that generated its training set. A model that predicts graduation with high accuracy may do so partly through features that correlate with race or class even when neither is ever an input. The name of an applicant’s school, or their postcode, carries that information, because where people live and where they study has long tracked race and class.

The fairness research community has developed several mathematical definitions to address this, and four (probably more) are worth holding in mind.

Demographic parity requires that the model’s positive-outcome rate, an offer of admission in this case, be equal across protected groups. If 60% of Group A applicants are admitted, 60% of Group B should be too. This is the Mohist instinct formalised, equal outcomes regardless of group membership.

Equalised odds requires the model’s error rates to be equal across groups, so the rate of false positives (admitting someone who then drops out) and false negatives (rejecting someone who would have graduated) should match for everyone.

Calibration requires that a given score mean the same thing for everyone. Among all applicants the model rates at 70%, roughly 70% should go on to graduate, whatever group they belong to. This preserves the informational content of the score for each person.

Individual fairness, proposed by Cynthia Dwork and colleagues in their 2012 paper “Fairness Through Awareness,” requires that similar individuals receive similar treatment, so the system judges the applicant on the relevant facts of their own record rather than on the demographic bucket they fall into. This is the cleanest formal translation of the Confucian instinct, attending to the particular person in front of you rather than the category. It is graded love rendered as a constraint on a classifier.

Each of these sounds reasonable, and each captures something we genuinely mean by fairness. As three computer scientists proved in 2016, you generally cannot satisfy them all at once.

The impossibility

Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan published “Inherent Trade-Offs in the Fair Determination of Risk Scores” in 2016. The result is sometimes called the KMR impossibility theorem, and it really does deserve to be understood precisely, because it converts a philosophical tension into a mathematical proof.

The theorem shows that when base rates differ between two groups, when Group A and Group B have different actual rates of the outcome being predicted, three fairness conditions cannot all hold at once. You cannot have calibration, equal false-positive rates, and equal false-negative rates at the same time, except in trivial cases where the model is perfect or the base rates are identical.

This is a formal impossibility. It holds for any scoring system or decision rule, whatever the algorithm. It is a constraint on mathematics the way conservation of energy is a constraint on physics. Better data will not dissolve it and a cleverererer model will not dissolve it. When groups differ in base rates for the outcome you are predicting, perfect fairness by all three definitions is logically impossible.

The implications are quite severe. Enforce demographic parity and you will necessarily either approve higher-risk applicants from one group or reject lower-risk applicants from another. Enforce calibration and you will get different error rates across groups when base rates differ. Equalise error rates and your scores will no longer be equally well-calibrated. Every choice about which metric to prioritise is a choice about who bears the cost of the system’s imperfection. There is no neutral option, and no configuration in which nobody loses.

I first read the KMR paper about two years ago, maybe longer, and it did not unsettle me the way it seems to unsettle other people. It put a proof under something that had always seemed obvious to me, that you cannot have every kind of fairness at once. What the theorem really pulls apart is two things people fold into the single word fairness. Equality of outcome wants the same result across groups. Equality of treatment wants the same rule applied to each person. The two feel identical right up until you try to compute both at once, and then they separate, and no model can hold them together. The worth of the proof is that it ends the argument. After it, "we just need to be fairer" stops being a solvable engineering task and becomes a forced choice between competing versions of fairness, each with real costs to real people. That is also the most important thing a good curriculum does. It shows students the point where the easy answers run out and the territory turns genuinely hard.

Mozi meets the impossibility

The KMR result is the Mozi and Mencius debate in formal peer reviewed dress, and the trolley problem is the bridge between them. The mapping is an analogy rather than an identity. Demographic parity is not literally Mozi’s metaphysics, and individual fairness is not literally Mencius’s. The shape of the disagreement transfers exactly, and the trolley makes that shape super visible.

The cross-pressure at the switch comes back here. The consequentialist core says to count impartially and equalise the outcome. The fēi gōng constraint says you must not be the agent who reaches out and inflicts a fresh, deliberate harm on a specific individual who was not already on the track.

Demographic parity, enforced against differing base rates, pushes the man off the bridge. To equalise approval rates across groups, the system has to reach in and impose a cost on identifiable people, the lower-risk applicant from the advantaged group who now gets rejected so the aggregate comes out level. It diverts harm onto a particular person for the sake of the group total.

Individual fairness and calibration refuse the push. They insist on judging each person on their own track, their own features and their own score, and they decline to redirect a harm onto someone to balance a population. They protect the particular at the cost of the aggregate.

Neither philosopher was confused. Mozi had seen what partiality does when it dominates, warlords who counted only their own states as real and families who hoarded while others starved. His answer, impartial care with equal weight, was aimed at real suffering and would have reduced it. Mencius had seen what abstraction does when it replaces particular care, hollow pronouncements of universal goodwill from rulers who could not be bothered to govern their own households. His answer, graded love rooted in actual relationships, was aimed at real neglect and would have reduced it.

The impossibility theorem proves they cannot both win. When the underlying reality differs between groups, you must choose. Equal treatment of the individual or equal outcomes for the group. Refuse to push, or pull the lever. The modern ML researcher sits exactly where Mozi sat at the switch, holding a loss function in one hand and a fairness constraint in the other, knowing the theorem says you pick one.

Where the debate actually lands

In practice, the fairness community has converged on something neither Mozi nor Mencius would find fully satisfying, which is contextual choice.

The field increasingly accepts that there is no single correct fairness metric. The right choice depends on the domain, the stakes, and the specific harm you most want to avoid. In criminal risk assessment, where a false positive means someone is detained who should not be, equalising false-positive rates across groups has a strong claim. In medical diagnosis, where a false negative means a disease goes untreated, calibration and equal false-negative rates may matter more. In hiring, where historical discrimination has produced systematically different base rates, demographic parity may be the appropriate corrective.

This is, in a sense, a Confucian answer. The right kind of fairness depends on the relationship, the context, and the obligations actually at stake. There is no universal formula.

The Mohist challenge bites back hard though. If context determines the metric, then whoever controls the context controls the outcome (think about that for a while). Who decides that calibration matters more than parity in this case, and who bears the cost of that decision? Mozi would point out, acidly, that the people making these choices are overwhelmingly drawn from the groups the status quo already favours, and that context-dependent fairness can curdle very easily into a sophisticated justification for leaving existing inequalities exactly where they are.

The blunt instruments know this. The four-fifths rule, drawn from United States employment law, holds that if a selection procedure passes a protected group at less than four-fifths (80%) the rate of the highest-passing group, the procedure has adverse impact and demands justification. It is crude. It does not ask why the disparity exists. It simply flags the disparity and forces an account. Mozi would approve of its directness, since it measures outcomes and refuses to accept good intentions or technical sophistication as a substitute for equal results. Mencius would object that it ignores the individual, and that forcing equal outcomes across groups with different base rates necessarily means treating particular people unequally.

Behind all of it sits the KMR theorem, proving the tension is not resolvable by technical means. You can build better models and design subtler fairness constraints. The impossibility remains.

The school that lost

There is a detail in Mozi’s biography that I think is the most important thing in this essay, and it is nearly always left out.

In his own time, Mozi did not lose. Mohism was one of the two dominant philosophical movements of the warring-states period, and a later writer would name Confucianism and Mohism together as the age’s two great prominent schools. The Mohists had organisation, doctrine, technology, and a disciplined cadre willing to walk for days to defend a stranger’s wall. If you had surveyed the intellectual landscape in 300 BCE and tried to guess which school would shape the next 2,000 years of Chinese civilisation, Mohism would have been a reasonable bet.

It vanished almost completely. After the Qin unification and the Han embrace of Confucianism as state orthodoxy, the first systematic consequentialism on earth dropped out of the living tradition (there may have been some book burnings too, Li Si loved a good book burning). As a result Mohism’s central text survived largely by accident, preserved in the margins of a Daoist canon, barely read and barely transmitted, until Qing-dynasty (1644 to 1912) and modern scholars dug it back out and realised what they were holding.

The first utilitarian framework in human history was out-competed for the machinery of the state. Confucianism became the operating system of empire, running the examinations and the bureaucracy, and a worldview built on graded relationships and particular duties governed an empire more comfortably than one built on impartial care for all under heaven. Which conception of fairness prevailed was settled by power.

That is Mozi’s own warning, played out as his own erasure. He told us that context-dependent fairness gets defined by whoever benefits from the existing arrangement. Then the existing arrangement defined him right out of the conversation for 2,000 years. When we say today that the appropriate fairness metric depends on context, we should remember who usually ends up writing the context, and that the most rigorous and most quantitative framework is not the one that automatically wins. The one that wins is the one that fits the institution doing the choosing…

What Mozi would build

What rescues this essay from despair (don’t worry I just drowned my sorrows in a meat pie) is the thing that made Mozi unusual among ancient philosophers. He was practical to the bone. He argued for impartial care and then built organisations to deliver it. The Mohist school operated as something between a philosophical movement and an engineering guild, and the Mozi text carries detailed chapters on fortification, signalling, and defensive architecture sitting directly alongside the arguments for universal love. Mozi actually didn’t see a gap between the two. If all people deserve equal care, you build the systems that supply it, so the philosophy generates engineering requirements.

That combination, moral clarity about ends together with technical rigour about means, is what the fairness field most needs and most often lacks. The debate between impartial and graded care has run for two and a half millennia, and the impossibility theorem proves it will not resolve into a single answer. What we can do is be precise about the trade-offs we are making and honest about who pays.

Every deployed model that decides things about people, from loan approvals to medical diagnoses, to grades embodies a choice about which kind of fairness to pursue, and that choice has winners and losers. So picture the two philosophers doing the work the field actually requires. Mozi builds the system. He insists it serve everyone equally, and when the theorem forces a trade-off, he insists the trade-off be made explicit, that the people bearing its cost have a voice in the decision, and that the justification be public. Mencius audits the same system. He checks whether the abstract metric maps onto the particular lives beneath it, whether equal means equal in reality or merely equal on a spreadsheet, whether the prospective students and job seekers and defendants are seen as particular people in particular circumstances rather than data points in a demographic bucket. You need both, the builder who will not push someone off the bridge without saying so out loud, and the auditor who keeps checking that the person on the bridge is real.

Fairness as practice

The school project is ongoing, if you’re curious. The teacher actually reshuffled the groups and mixed the kids Miya called the try-hards with the ones who had been clumped together and written off. Miya complained for a day. Then her group worked out how to divide the project around what each person was actually good at.

She did not come home announcing that impartial care is correct. She came home saying that Alex (name changed for privacy), of all people, was the only who could play the piece on guitar. The abstract principle had become a particular person doing a particular thing well, visible only because someone rearranged the structure first.

That is the whole argument at kitchen-table scale. The structures we inherit encode the preferences of whoever built them. Rearrange the structure and different capabilities become visible. Measure fairness only within the existing arrangement and you mistake the artefacts of the arrangement for facts about the people inside it. The KMR theorem tells us we cannot build a perfect arrangement, since every configuration of a decision system will be unfair by at least one reasonable definition. That is the mathematical reality, and it is not going away.

The Mohist response, and I think the right one, is that the impossibility of perfection is not a licence for indifference. You choose, you make the trade-off explicit, you watch what happens, and you adjust. And you hold in mind, always, that the person on the wrong side of the trade-off is real, that their loss is not an abstraction, and that the system owes them an account of why it chose as it did.

Mozi’s engineers walked for days to defend walls that were not their own. The fairness lived in the decision about who to push and which one to count, made out loud, where the people it cost could hear it.

The next essay in this series is already in draft. Subscribe if you want to be there when it lands.

Subscribe now

On this day we were at Yasaka-jinja in Kyoto when the omikuji stand caught Miya’s attention. For those who don’t know, this is the ritual of shaking the box full of sticks until one falls out, which is followed opening the folded paper from a corresponding labelled box to see what the gods had to say about the year ahead.

“Why don’t they just tell you what’s going to happen?” Miya asked as we translated the Japanese for her.

I was about to give her the standard adult answer about mysteries and faith when the young lady working at the shrine interjected in English. “The point isn’t to know what happens. The point is to sit with not knowing.”

Seven-year-old logic cut straight through. “But then why ask?”

“Because,” Haruko (I hope I recalled her name correctly) said, “the not-knowing teaches you something the knowing can’t.”

I’ve been thinking about that conversation for eight years now, especially as I watch the world try to extract certainty from the most sophisticated uncertainty machines ever built. Every time someone asks ChatGPT for a definitive answer and gets frustrated when it hedges, every time a company deploys an LLM expecting deterministic results from a probabilistic system, I hear Miya asking why they don’t just tell you what’s going to happen.

And I hear her Haruko’s answer, which turns out to be three thousand years older than either of them knew.

The Oldest Probability Machine

The Yijing (易經), the Book of Changes, was already ancient when Confucius studied it. Archaeological evidence suggests parts of it date back to the Western Zhou Dynasty, around 1000 BCE. By the time the received text was compiled, sometime before 200 BCE, it had accumulated layers of commentary and interpretation from centuries of scholars, diviners, and philosophers.

But beneath all that interpretation lies something startlingly modern. It contains a systematic method for generating and working with probability distributions.

The basic structure is elegant. Sixty-four hexagrams (卦), each made of six lines that are either yin (⚋) or yang (⚊). Each line can be young (stable) or old (changing). The traditional method for consulting the oracle uses fifty yarrow stalks in a complex sorting procedure that takes about twenty minutes and produces genuinely random results. Kinda cool.

What emerges from this process isn’t just any random output. The yarrow method generates a carefully calibrated probability distribution:

• Young yang (7): 5/16 probability

• Young yin (8): 7/16 probability

• Old yang (9): 3/16 probability

• Old yin (6): 1/16 probability

The system is biased toward stability. Young lines, which don’t change, appear three-quarters of the time. Old lines, which transform into their opposite, appear one-quarter of the time. The ancient Chinese built their uncertainty machine with a preference for continuity over change.

Each hexagram represents a state. The changing lines show transitions between states. You start with one hexagram, the changing lines transform, and you end with another hexagram. What you have is a state machine with defined transition probabilities, documented three millennia before Markov formalised the mathematics.

The I Ching is a probabilistic reasoning engine with 64 states and calibrated transition functions. The Hoover Dam of ancient Chinese engineering isn’t made of concrete, instead it’s made of probability.

Temperature and the Yarrow Method

When we talk about large language models, one of the first parameters everyone learns is temperature. At temperature 0, the model is deterministic, or in other words it always picks the most probable next token. As you increase temperature, it becomes more “creative,” sampling from a wider distribution of possibilities. At very high temperatures, the output becomes essentially random.

Temperature is how we control the balance between predictability and creativity in AI text generation. Set it low for factual writing. Set it higher for poetry. The parameter doesn’t change what the model knows, but it simply changes how the model explores the space of what it knows.

The yarrow stalk method is the I Ching’s temperature setting. But unlike a simple numerical parameter, the ancient Chinese embedded their preferences directly into the probability distribution.

That preference for stability is built into the mathematics. Young lines are static. They represent the tendency of situations to persist, of patterns to continue, of systems to maintain their current state. Old lines are dynamic. They represent the moments when change breaks through, when transformation occurs.

The I Ching’s default temperature favours continuity with punctuated transformation. Not constant change, which would make the system unusable for practical guidance. Not total stasis, which would make consultation pointless. A calibrated bias toward stability with regular opportunities for genuine change.

Modern language models let you adjust temperature externally. The ancient Chinese built their temperature preferences into the consultation process itself. The yarrow stalks don’t just generate random results. They generate results weighted toward the kind of uncertainty the system’s designers thought would be most useful for human decision-making.

Constrained Possibility Spaces

Another key concept in modern AI is top-p sampling, also called nucleus sampling. Instead of considering all possible next tokens, the model only samples from the most probable ones (ie those that make up the top p percent of the probability mass). Set p to 0.1, and the model only considers tokens that fall within the most likely 10% of possibilities.

Top-p sampling constrains the possibility space. It says of all the things that could come next, we’re only going to consider the reasonable ones.

The I Ching operates on the same principle, just more dramatically. Instead of sampling from the most probable tokens, it samples from exactly sixty-four possible states. Not infinite possibilities. Not even a large finite number. Sixty-four.

Those sixty-four states aren’t arbitrary. They’re meant to cover the essential patterns of change in human experience. Each hexagram has a name and a traditional interpretation:

• 乾 (Qián): The Creative, pure yang energy

• 坤 (Kūn): The Receptive, pure yin energy

• 屯 (Zhūn): Initial Difficulty, the struggle of new beginnings

• 蒙 (Méng): Youthful Folly, inexperience requiring guidance

• 需 (Xū): Waiting, the necessity of patience

The complete system maps what the ancient Chinese saw as the fundamental patterns underlying all situations. It’s a constrained vocabulary for talking about change, uncertainty, and time.

Modern nucleus sampling constrains possibility space to improve coherence. The I Ching constrained possibility space to create a practical tool for navigating uncertainty. In both cases, the constraint isn’t a limitation. It’s the feature that makes the system useful.

Two Types of Uncertainty

Machine learning distinguishes between epistemic and aleatoric uncertainty. Epistemic uncertainty is what we don’t know because we haven’t learnt enough yet and in principle, it’s reducible through more data or better models. Aleatoric uncertainty is genuinely random for example the roll of a die, the quantum measurement, the irreducible variability in complex systems.

Modern AI research focuses overwhelmingly on reducing epistemic uncertainty. Better training data, larger models, more sophisticated architectures. The goal is to eliminate uncertainty wherever possible and to quantify what remains so we can act despite it.

The I Ching takes the opposite approach. It treats all uncertainty as worth sitting with.

Consider the difference in how the two systems handle questions. Ask a modern language model “Should I take this job?” and it will generate a reasonable analysis with pros and cons, considerations you might not have thought of, a structured way to think through the decision (from personal experiences…). What it won’t do is give you a direct answer, because the model knows it doesn’t have enough information about your specific situation.

The I Ching will give you a hexagram. Maybe 家人 (Jiārén), “The Family,” which traditionally suggests that the decision should be evaluated in terms of how it affects the people close to you. Maybe 遯 (Dùn), “Retreat,” which suggests that this might not be the right time for this particular move.

The language model tries to reduce epistemic uncertainty by giving you a framework for thinking. The I Ching accepts that uncertainty is irreducible and gives you a specific pattern to hold while you sit in the uncertainty.

Neither approach eliminates the uncertainty. The question is whether you treat uncertainty as a problem to be solved or as a condition to be navigated.

Calibration and Confidence

One of the key challenges in AI safety is calibration. A well-calibrated model is one where, when it says it’s 70% confident, it’s actually right 70% of the time. Most current language models are overconfident (sometimes with hilarious impacts). They express certainty about things they’re actually unsure about. This is by design btw.

The standard approach to calibration is to train the model to output confidence scores that match its actual performance on held-out test sets. The goal is to make the model’s internal uncertainty visible in its outputs.

The I Ching doesn’t provide confidence scores. It doesn’t say “I’m 73% sure that 困 (Kùn) - Oppression - applies to your situation.” It gives you the hexagram and leaves you to figure out what to do with it.

What it does instead is much subtler. Each hexagram comes with traditional commentary, usually in the form of images and metaphors rather than direct advice. For 困 (Kùn), the traditional text says:

“There is water under the lake: the image of Exhaustion.
Thus the superior man stakes his life on following his will.”

This isn’t a confidence score. It’s a way of pointing at the pattern without claiming to know your specific situation. The text says this is what exhaustion looks like in the abstract, here’s how someone might navigate it, now you figure out whether and how it applies.

The calibration is built into the interpretive method, not the output. The I Ching assumes that meaning emerges from the interaction between the pattern and your situation. It doesn’t claim to know what your situation is. It offers a lens and trusts you to figure out whether the lens clarifies anything.

Modern language models try to calibrate their outputs to match their actual knowledge. The I Ching calibrates the entire consultation process to acknowledge the fundamental uncertainty of human decision-making.

The Problem of Definitive Answers

Here’s what I think the ancient Chinese got right that we’re getting wrong.

We’ve built the most sophisticated uncertainty machines in history and we’re using them to eliminate uncertainty. Every prompt is an attempt to extract a definitive answer from a system designed to model probability distributions. Every frustrated user who complains that ChatGPT won’t “just tell me what to do” is asking a snake to stop being serpentine.

The I Ching would say that’s exactly backwards. If you have access to a machine that can model uncertainty with unprecedented sophistication, the last thing you should do is use it to pretend the uncertainty isn’t there.

The yarrow stalks don’t give you answers. They give you something to think about while you sit with the question. The hexagram doesn’t resolve your dilemma. It gives you a framework for holding the dilemma without needing to collapse it prematurely.

Three thousand years ago, the Chinese developed a technology specifically designed to help people work with irreducible uncertainty. It became one of the most enduring and widespread decision-making tools in human history precisely because it doesn’t try to eliminate the uncertainty. It helps you become more comfortable with uncertainty as a permanent condition of human life.

We have probabilistic reasoning engines thousands of times more powerful than anything the ancient Chinese could imagine. We could use them to get better at living with uncertainty. Instead, we’re using them to avoid feeling uncertain.

That’s a design choice. And it’s a choice the ancient Chinese would have recognised as a fundamental misunderstanding of what uncertainty machines are for.

What the Hexagrams Teach

I’ve been experimenting with using large language models the way the I Ching is meant to be used. Not to get answers, but to generate patterns worth sitting with and let’s be honest mostly for fun.

Instead of asking “What should I do about this work situation?” I ask “Generate three different frameworks for thinking about this work situation, each based on a different set of assumptions.”

Instead of asking “Is this business idea viable?” I ask “What are three ways this business idea could fail that I haven’t considered, and what are three ways it could succeed that I’m underestimating?”

The shift is from seeking resolution to seeking perspective. From trying to reduce uncertainty to trying to work with uncertainty more skillfully.

What I notice is that the language model, when used this way, generates more genuinely useful output. Instead of hedged, generic advice, I get specific scenarios and considerations I hadn’t thought of. The model seems to perform better when it’s explicitly asked to explore possibility space rather than collapse it.

This matches what we know about how these models actually work. They’re trained on the full distribution of human text, including all the different ways people think about problems. When you ask for a single answer, you’re asking the model to collapse that rich distribution into one data point. When you ask for multiple perspectives, you’re asking it to sample from different parts of the distribution.

The I Ching’s insight is that the multiple perspectives, held simultaneously, are more valuable than any single answer you might extract from them.

The Ritual Matters

There’s one more thing the ancient Chinese understood that we’ve lost in our digital consultation habits.

Traditional I Ching consultation is slow and physical. Sorting fifty yarrow stalks takes twenty minutes. You can’t do it absent-mindedly. The process requires attention, and the attention changes what happens.

The slowness isn’t inefficiency. It’s the feature that makes the system work. Twenty minutes is enough time for your initial framing of the question to evolve. Twenty minutes is enough time to notice things you weren’t thinking about when you started. Twenty minutes is long enough for the ritual to change your relationship to the question you’re asking.

By contrast, we interact with language models at typing speed. Ask the question, get the answer, move on. The entire interaction happens faster than traditional I Ching consultation just to ask the question properly.

I don’t think we need to spend twenty minutes with yarrow stalks before consulting ChatGPT. But I do think the speed of digital consultation is part of what makes it unsatisfying. We’re not giving ourselves time to let the question evolve, or to sit with the uncertainty before seeking resolution.

The ancient Chinese built contemplation time into their uncertainty machine. We’ve built uncertainty machines optimised for immediate answers.

The technical capabilities aren’t the limiting factor. It’s the ritual framework we use to engage with those capabilities.

What Miya Taught Me About Fortune-Telling

I was wrong about the omikuji that day in Kyoto. I thought Miya was asking why the gods don’t just tell you what’s going to happen. But what she was actually asking is a much more sophisticated question “if you’re not trying to eliminate uncertainty, what are you trying to do?”

The young lady who responded was precise and wise beyond her years: you’re trying to learn how to be with uncertainty in a way that’s useful.

The fortune slip Miya pulled that day was 中吉 (chū-kichi) or “moderate good fortune.” The specific prediction was vague in the way these things always are and was something like “Patience brings better results than haste. The path forward becomes clear through careful attention to small details.”

Rei (my wife) read it to Miya and naturally she immediately wanted to know what it meant. What should she be patient about? Which details? When would the path become clear? To be fair, she was 7 or 8 at the time. Instead of responding, we folded the paper and tied it to a tree branch with dozens of others, as is traditional when you want the gods to take care of your fortune for you.

“That’s it?” Miya asked.

“That’s it” I said, “Now you pay attention and see if anything feels different.”

For the next three days of our Kyoto trip, Miya was the most attentive seven-year-old I’d ever seen. She noticed things. The way the light hit the temple roofs in the morning. The sound of the bamboo fountain at Kōdai-ji. The careful way the monks arranged the stones in the garden. She wasn’t looking for anything specific. She was just looking.

The fortune didn’t tell her what was going to happen. It changed how she paid attention to what was happening.

That’s what a properly used uncertainty machine does. It doesn’t resolve uncertainty. It makes you better at working with uncertainty. Better at noticing patterns. Better at holding multiple possibilities simultaneously. Better at staying curious instead of rushing to conclusion.

The I Ching has been training humans in uncertainty tolerance for three millennia. We have the opportunity to build the same capability into our relationship with modern AI. But only if we stop trying to make these systems give us what the ancient Chinese knew we don’t actually need, namely certainty about an uncertain world.

Subscribe now

The opening line of the Enchiridion, his handbook for students:

“Some things are up to us and some are not. Up to us are opinion, motivation, desire, aversion, and, in a word, whatever is of our own doing; not up to us are our body, our property, reputation, position, and, in a word, whatever is not of our own doing.”

This sounds simple. It is not simple. The entire Stoic project is a sustained investigation into where exactly that line falls, how to recognise it moment by moment, and how to live with integrity once you see it clearly.

Epictetus wasn’t offering a feel-good philosophy about acceptance. He was teaching a boundary classification system. Every situation contains elements you can direct and elements you cannot. Wisdom is learning to tell the difference. Suffering is getting the classification wrong.

Marcus Aurelius, writing to himself two centuries later in the Meditations:

“You have power over your mind — not outside events. Realise this, and you will find strength.”

But power over your mind to do what, exactly? That’s where the Stoic position gets specific and unsentimental. Your mind is yours to direct. Your judgment is yours to shape. Your response to any situation is entirely within your control. What the situation actually contains is not.

This is prohairesis or the faculty of choice, the one thing that is genuinely ours. Everything else, according to Epictetus, is “borrowed time.”

I didn’t immediately see what this had to do with AI until I spent many many months working with language models and started noticing where teams were getting stuck.

Where the Boundaries Fall in Human-AI Systems

The problem with most human-in-the-loop AI systems is that they put humans in loops without thinking clearly about which loops actually require humans.

Take content moderation. A human reviewer sits at a screen, looking at posts flagged by an algorithm. Approve, reject, escalate. Eight hours a day. The system generates a queue of edge cases and the human makes rapid-fire decisions on each one.

This is automation shaped by the assumption that humans should be responsible for all the judgment calls. It sounds reasonable. Judgment is what humans do best, right? But watch the actual process. The human sees a post for 3.2 seconds. The context available is what fits on the screen. The training is six hours of examples. The psychological pressure is to maintain throughput. The feedback loop is almost non-existent and they never learn whether their calls were actually right.

What exactly is that human judging? They’re not evaluating the content against some rich understanding of community values developed through years of participation. They’re pattern-matching the content against the categories they were trained to recognise, under conditions specifically designed to prevent deeper evaluation.

Epictetus would ask “what in this process is actually up to the human?”

Their immediate response to what appears on the screen is up to them. Their judgment about how to categorise it, given the training and context they have, is up to them. Whether they click “approve” or “reject” is up to them.

The content of the post is not up to them. The algorithm that flagged it is not up to them. The categories they’re choosing between are not up to them. The context they don’t have is not up to them. Whether their decision helps or harms the person who wrote the post is not up to them, because they don’t have the information required to make that assessment.

The system is asking the human to take responsibility for outcomes they cannot actually control, while giving them control over inputs that don’t meaningfully influence those outcomes.

This is exactly the category error the Stoics spent three centuries identifying and correcting.

The Levels of Automation Problem

In 1978, Thomas Sheridan and William Verplank published a framework for thinking about automation that’s still used today. They described ten levels, from fully manual to fully automated. Level 1: the human does everything. Level 10: the computer does everything. Most systems operate somewhere in the middle.

The framework is useful for designing systems, but it misses something Epictetus would have also noticed immediately. It defines the levels by capability (what the human can do) rather than by authority (what the human should be responsible for).

Level 5 in their framework: “The computer executes suggestions if the human approves.” Level 6: “The computer allows the human a restricted time to veto before automatic execution.” These are defined by the timing and mechanism of human involvement, not by the nature of the decision being made.

But some decisions are fundamentally human not because humans are better at making them, but because they involve values, interpretation, meaning or you know the things that Epictetus put in the “up to us” category. Other decisions are computational: pattern recognition, data retrieval, calculation. Humans can do these things, but there’s no special reason they should.

The Stoic framework suggests a different way of drawing the lines. Instead of asking “what are humans good at?” ask “what are humans responsible for?”

Prohairesis in the Loop

Marcus Aurelius kept a private notebook for the last decade of his life. He was the most powerful person in the world at the time, responsible for an empire that stretched from Britain to the Euphrates. The Meditations are his notes to himself on how to stay sane in that position.

One entry I particularly love:

“Accept the things to which fate binds you, and love the people with whom fate associates you.”

This from someone who could have anyone executed for breakfast. The emperor is using Stoic principles to recognise what he doesn’t control, specifically who he has to work with, what situations arise, what resources are available, and what other people choose to do.

What he does control e.g. how he interprets those situations, how he responds to them, what values guide his responses, how he treats the people fate associates him with.

Apply this to AI systems. The human in the loop should be there to exercise prohairesis: judgment, interpretation, values, meaning-making (those who know me will be bored of me talking about Judgment, Accountability, and Taste). The machine should handle the computational tasks, e.g. data processing, pattern recognition, calculation, retrieval.

But most HITL systems do the reverse. They automate the interpretation and values (bake them into the algorithm) and ask humans to do the computational tasks (scan through queues, match patterns to categories).

A content moderation system designed on Stoic principles would work differently. The algorithm handles the computation (scanning posts, identifying patterns, checking against databases, flagging anomalies). The human handles the interpretation for example what do these patterns mean in this specific context? What values should guide the response? How should the community’s standards apply to this particular case?

More importantly the human would have the context required to make those judgments meaningful. Not 3.2 seconds with a post stripped of context, but access to the conversation, the community history, the poster’s track record, the downstream effects of different moderation choices.

The human becomes responsible for the things that are actually up to them, and only those things.

The Automation Paradox

There’s a well-documented phenomenon in automation research called the automation paradox. The more automated a system becomes, the harder it is for humans to intervene meaningfully when the automation fails.

Pilots spend years learning to fly planes manually, then spend their careers monitoring automated systems. When the automation fails, they have seconds to diagnose what went wrong and take manual control of a situation they haven’t directly handled in months. The skills atrophy. The situation awareness degrades. The human becomes the weakest link in a system designed to compensate for human weakness.

Epictetus saw this pattern everywhere, though he called it something different. In the Discourses:

“Every art and every inquiry, and similarly every action and pursuit, is thought to aim at some good; and for this reason the good has rightly been declared to be that at which all things aim.”

He’s quoting Aristotle, but making a Stoic point. When you optimise a system for one good (efficiency, accuracy, throughput), you often degrade other goods (understanding, responsibility, meaning).

The automation paradox is what happens when you optimise for performance while neglecting the human capacity to remain meaningfully connected to the process. The human stays in the loop procedurally but drops out of the loop psychologically. They become a monitoring device rather than an agent.

The Stoic solution is to design systems that keep humans responsible for what they can actually control, and only what they can actually control. Not because humans are always better at these things, but because responsibility without authority is a form of suffering, and authority without responsibility is a form of corruption.

Drawing the Line in Practice

So this is how I think about AI system design now. Instead of asking “how do we get humans to use AI responsibly?” ask “how do we design systems where the responsible choice is also the easiest choice?”

For writing assistants (something we deal with in Edu a lot) instead of flagging AI-generated text after it’s written, make the boundary clear during composition. Show the human which parts came from the model, which parts they wrote themselves, which parts are hybrid. Make it easy to see where human judgment is required (tone, audience, purpose, truth) and where computation can help (grammar, structure, flow).

For decision-support systems, instead of asking the human to approve or reject algorithmic recommendations, give the human the information they need to make independent judgments. Show the data the algorithm used, show what it doesn’t know, show the uncertainty ranges. Make the human responsible for the interpretation, not just the final click.

For creative tools instead of automating the creative decisions, automate the mechanical ones. Let the AI handle the repetitive tasks, the technical execution, the format conversion. Keep the human responsible for the vision, the meaning, the emotional core.

The pattern is the same in each case. Give the human authority over the things that require human judgment, and only those things. Give them the information and context they need to exercise that authority meaningfully. Automate the rest.

What This Doesn’t Solve

I don’t want to overstate the case here. The Stoic dichotomy doesn’t resolve all the hard problems in AI system design. There are legitimate disagreements about what counts as judgment versus calculation, what kinds of context are necessary for meaningful human involvement, how to balance individual autonomy against collective outcomes.

And there are practical constraints. Sometimes the information required for meaningful human judgment isn’t available or is too expensive to provide. Sometimes the system has to operate at scale and speed that makes deep human involvement impractical. Sometimes the trade-offs between human autonomy and system performance are genuinely difficult.

But even in those cases, the Stoic framework clarifies what the trade-offs actually are. You’re not just choosing between human involvement and automation. You’re choosing what kind of responsibility to assign to humans, and what kind of authority to give them to match that responsibility.

Marcus Aurelius, governing an empire with technology that couldn’t send messages faster than a horse could run:

“Confine yourself to the present.”

The present moment contains elements you can direct and elements you cannot. The Stoic practice is learning to identify them accurately, quickly, consistently. Not just once, but as a habit of mind that operates under pressure.

This is what I think human-AI collaboration should develop in humans: not the ability to monitor algorithmic systems, but the ability to recognise, moment by moment, what requires human judgment and what doesn’t.

The Control Boundary

Epictetus lived through the reign of Domitian, one of Rome’s most paranoid emperors. Even students at philosophy schools were considered potential threats! Intellectuals disappeared. Senators were executed on suspicion. Epictetus kept teaching.

From the Discourses:

“No one can hurt you without your permission.”

This sounds naive until you understand what he means by “hurt.” He’s not saying that bad things won’t happen to you. He’s saying that whether those things damage your character, compromise your values, or distort your judgment is up to you.

The emperor can confiscate your property. That’s not up to you. Whether you define yourself by your property is up to you. The emperor can destroy your reputation. That’s not up to you. Whether you base your self-worth on what others think of you is up to you.

This is a harsh teaching, and I don’t think it’s complete. There are forms of harm that go deeper than choice, forms of trauma that compromise the very faculty of choice. But there’s something in Epictetus’s position that’s useful for thinking about AI systems.

Most of the anxiety about AI comes from the sense that these systems will make choices for us, or that they will become so powerful that our choices don’t matter. But the Stoic analysis suggests that our choices can only not matter if we locate our agency in the wrong place.

If you define your agency as the ability to control outcomes, then yes, AI systems are a threat to your agency. They will influence outcomes in ways you cannot predict or direct. If you define your agency as the ability to choose your response to whatever situation arises, then AI systems are just another situation to respond to.

The difference matters for design. Systems designed to preserve the illusion of outcome control will tend to give humans fake agency like buttons to click, approvals to grant, queues to process. Systems designed to support genuine choice will tend to give humans real information, meaningful context, and decision points that match their actual authority.

The control boundary isn’t between human and machine. It’s between what any agent in the system can control and what none of them can.

What the I Ching Knows

I’m still working through what this means for the bigger questions. How do you design an economy where human judgment remains meaningful when computation can handle more and more of what we thought required human intelligence? How do you maintain human agency in a world where the systems that shape daily life are increasingly algorithmic?

The Stoics offer one approach. Identify what is and isn’t up to us, and pour our energy into what is. But I suspect they were working with a simpler model of causation than we need for the systems we’re building.

The I Ching, the Chinese Book of Changes, works with a different model. Instead of fixed boundaries between controllable and uncontrollable, it sees patterns that shift over time, decisions that influence probabilities rather than determining outcomes, a complex relationship between individual choice and systemic change.

I think we need both frameworks. The Stoic clarity about where responsibility lies, and the I Ching’s subtlety about how change actually works in complex systems.

But that’s for another essay.

Next in this series: how the I Ching’s model of probabilistic change offers a framework for AI systems that operate in genuinely uncertain environments. The ancient Chinese approach to decision-making when you cannot predict or control outcomes. Subscribe to read it when it lands.

My friend David spoke at the same event. He made the point that every university, every campus, has its own unique history, identity, culture, character, and people. His concern was that we rush to buy tools that forget this. Universities are purchasing AI platforms as if culture is a configurable setting. The institutional knowledge of why this campus teaches this subject in this way to this community of students cannot be parameterised.

I came home from the conference and our family continued watching a Chinese comedy drama we’ve recently become addicted to. 年少有为. No Pain No Gain. Across a couple of episodes the characters (who run a game development studio) commence work on a Souls-like game where players must recite ancient Chinese poetry to progress. During their in house testing a boss taunts a player for reciting the ancient goose poem, 咏鹅, by seven-year-old Luo Binwang, because it is elementary school level and anyone who knows Mandarin would recognise it immediately. It was one of the first poems I ever learnt in the language. Watching the characters on screen celebrate learning classical Chinese poetry just to defeat a video game boss was one of my favourite moments in television for a long while.

My son and I would play that game in a heartbeat if it were real.

Both David’s point and the drama were circling the same thing. We are building tools at unprecedented scale and forgetting that education, culture, and language carry histories that cannot be compressed. Meanwhile a Chinese comedy show accidentally created one of the most beautiful pedagogical mechanics I have ever seen. Make the player recite poetry to defeat the boss.

What Taste Is

The ancient Greek concept of phronesis, practical wisdom, describes a kind of knowledge that cannot be taught through instruction. You develop it through exposure to the good, the true, and the beautiful. You build it by reading excellent literature, listening to great music, and seeing masterful art until your mind develops an internal compass that recognises quality on sight.

Aristotle called this habituation. You become just by doing just acts. You become excellent by doing excellent things until excellence is your disposition, not a decision you have to make. The quality of the art you consume writes itself into your sensibility.

This is taste. Preference is arbitrary. Taste is the capacity to recognise that one piece of writing achieves something deliberate and another piece of writing hits all the same marks and lands on nothing.

The arts develop taste. Taste is what you need to tell the difference between competent output and something that actually moves you.

The Japanese Concept of Ma

Ma (間) is the Japanese word for negative space, the pause between notes, the empty area that gives form its meaning. You have probably heard the saying “a Japanese garden does not work through what it contains. It works through the spaces it leaves open.” This is the concept of 間 in practice.

In music, the silence between two notes is not empty. It is the place where the first note finishes being what it was and becomes whatever the second note is about to say. The pause is part of the music.

The 1922 translation of a Chinese text by Lin Yutang renders a passage from the Analects. “In poetry there is an intention. In words there is not the power to fully express intention.” Confucius observed that language falls short of what the mind intends, then said that is why music exists. Because music carries what language cannot.

When an AI translates a poem from Mandarin into English, it translates the words. It cannot translate the 間 . The pause between two lines where the reader was supposed to breathe, hesitate, and feel something. The gap between what is said and what is meant.

Wang Wei, a Tang dynasty poet, painted landscapes with a single line across the top of the scroll. The poem says nothing about the visual weight of the landscape itself. The words sit above the empty space and the reader experiences both simultaneously. The poem is not in the characters. It is in the characters, the ink density, the paper texture, and the breath you take before reading the first line. All of it is the poem.

No translation system on earth carries that.

I know this because I have spent years teaching in Japan, delivering academic subjects entirely in Japanese to native speakers. My students would have smiled politely at an AI translation of my lecture notes. They listened to the actual lecture because I was choosing every word in real time, feeling my way through concepts in a language that demands a specific kind of humility, and let’s be honest probably making interesting mistakes. One of the things I love so much about Japanese is that it forces you to calibrate your social relationship with your own sentence before you are allowed to finish it. The grammar makes it impossible to speak without deciding how you stand in relation to the person listening. Respect is literally built into the grammar.

That recalibration happening at the level of every particle is what no translation system captures.

The Confucian View of Music

The Confucian tradition treats music as education in the deepest sense. Analects 8.8 says Confucius believed poetry was the starting point, ritual gave structure, and music completed the person. Music was not decoration. It was the final layer of moral cultivation.

Music shapes the person who hears it. It does not deliver information. It delivers disposition. You hear a mournful melody and something in you changes. You hear something triumphant and your posture shifts. This is neurological fact. Music moves the body and the body moves the mind.

You surround a person with good music, good ritual, and good poetry and their sense of what is proper becomes internal. You do not need to tell them what is right. They feel it.

This is why Confucius ended education with music. It was the only subject that worked on the whole person, not just the intellect.

Translation and the Problem of Soul

I need to be honest about the contradiction here.

I love that AI can translate between languages. My wife’s family speaks Wenzhou dialect (technically that isn’t even true. Its more a sub-dialect of Yueqing, which is a sub-form of Wenzhou-hua). Wenzhou-hua is a form of Chinese so distinct from standard Putonghua that it was used during the Second World War as a secret language for communicating about enemy movements. It is not a dialect in the sense of an accent or regional vocabulary quirk. It is a language that speakers of Putonghua cannot understand even when they are standing in the same room. AI translation cannot handle Wenzhou dialect because the training data is not there. But even if it were, the idea of machine-translating a language that was literally weaponised for its secrecy because outsiders could not penetrate it is almost comical.

As a linguist, as someone who has spent years watching how different languages carve up reality into different conceptual shapes, a part of me dies when I watch translation and live interpretation happen through automated systems.

Language is not a code. It is not one set of symbols mapped to another set of symbols with roughly equivalent referents. Language is the shape of a culture’s relationship with time, space, social hierarchy, intimacy, and power.

Mandarin splits “to know” into three distinct words. 知道 (factual knowledge), 了解 (understanding through engagement), 认识 (relational knowing of a person). These are not synonyms. Each points at a different relationship between a mind and its subject. Translating all three into the single English word “know” is not a translation. It is a compression. You have taken three distinct claims about reality and collapsed them into one.

Japanese does something different. There are first-person pronouns. 私, 僕, 俺, 自分 (etc… there are actually heaps more), and others depending on context, gender, and social situation. Each one projects a different self-image onto the listener. 私 is neutral and formal. 僕 is masculine and casual. 俺 is rough and assertive. 自分 can mean “myself” but also serves as a humble self-reference in certain dialects. Choosing one over another is a statement about who you are in relation to the person you are addressing. English just says “I”.

The AI translator does not know it has done this. The person reading the translation does not know either. Something has been lost and neither party has the vocabulary to notice.

Ibn Khaldun on the Arts

Ibn Khaldun, the 14th century North African historian and philosopher, wrote extensively in his Muqaddimah about what happens to a civilisation’s art in its final generations.

He argued that crafts, arts, and intellectual traditions reach their peak when the civilisation is at its strongest, then decline in the last generations because the practitioners inherit the forms without the substance. They can reproduce the technique of great poetry because they studied the rules. What they cannot reproduce is the lived experience that made the original poems alive.

This is exactly the problem with AI translation and interpretation. The system has inherited the form of language without the substance of lived experience. It can produce grammatically correct translations because it studied the rules. It cannot produce translations that carry the weight of human experience because it has never experienced anything.

Ibn Khaldun would recognise this immediately. He spent his life studying the rise and fall of dynasties and the pattern he found was consistent. When technique outpaces substance, the civilisation is in its final phase. The forms are preserved but everything that made them meaningful has evaporated.

What We Lose

I want to name specific things that are lost in AI translation because the loss is not abstract. It is measurable and specific. These are from my personal lived experiences.

The rhythm of spoken Mandarin. Mandarin is tonal. The pitch of a syllable changes its meaning. The rhythm of a sentence, the places where a speaker chooses to stress or flatten a tone, conveys emotional subtext that the characters themselves do not carry. An AI transcript flattens tones into text. The text is accurate. The subtext is gone.

The history embedded in a dialect. Wenzhou dialect carries the geography of the Oujiang River valley, the trade routes to the sea, the centuries of isolation that made it impenetrable to outsiders. When someone speaks Wenzhou dialect, they are carrying a history of a city that has been on the frontier of Chinese civilisation for a thousand years. Putonghua carries the politics of the modern state. These are not interchangeable. No translation between them is a neutral act.

The untranslatable intimacy of code-switching. My wife switches between English, Japanese, and Mandarin depending on what she is talking about. The switches are themselves a form of emotional signalling. When she talks about her childhood, the Wenzhou dialect comes through and carries memories that none of the other languages can hold. The switching between them, the moment she chooses Japanese over English or Mandarin over both, reveals something about the emotional register of the moment. Translation into a single language erases the switching and with it the emotional dimension.

The poetry of music lyrics. A song is words plus melody plus the singer’s breath. The breath matters. Bob Dylan’s Nobel Prize recognised that song lyrics are poetry precisely because their musical context changes what the words mean. The same words sung with different melody carry entirely different emotional weight. AI translation of song lyrics produces accurate text and strips away the reason the song exists in the first place. This matters to people like me who sing WAY too much Karaoke.

The lived history embedded in grammar. Japanese particles mark the social relationship between speaker and listener. English has nothing like this. Translating Japanese into English erases the entire social negotiation happening at the grammatical level. The Japanese verb endings change based on politeness level, and the politeness level encodes your assessment of the social distance, hierarchy, and intimacy of the moment. Every sentence is a micro-negotiation of power and closeness. English sentences arrive carrying no such negotiation. The translation is accurate and the reality of the relationship between the speakers has been deleted. For anyone who has ever done biodirectional live intepretation in a B2B environment in Japanese to English … this is the reason why you ALWAYS end the day with a headache.

The Poetry Game That Would Be Beautiful

The game in 年少有为 where players recite poetry to defeat bosses understands something that most educators, edtech companies, and AI conferences miss entirely.

The genius of that mechanic is that it gives the player a reason to care about poetry in the first place. They learn it because they want to beat the boss. Then, somewhere along the way, they might actually start to feel the poem.

This is what Confucius meant about music completing the person. You do not arrive at taste through instruction. You arrive through repeated encounter. The drama’s fictional game designers understood something intuitive about pedagogy. Give people a reason to engage with the thing and let them fall in love with it through the doing.

AI can generate poetry. It can analyse poetry. It can translate poetry. What it cannot do is create a context where a person decides to learn poetry because they want to. That requires understanding what moves people, what makes them lean forward, what makes them feel clever for having memorised something ancient and then feeling that ancient thing in a new way.

The Souls-like boss taunting a player for reciting a poem that every Chinese child learns at seven years old carries something that metrics cannot capture. The joy of being bad at something, then getting better, then realising the thing you struggled to learn has changed you.

The joke works on multiple levels too. The boss is not just saying the poem is too easy. They are saying even a seven-year-old from the Tang dynasty could do better than this. The humiliation cuts across a thousand years of cultural memory. It is quite brilliant, and it actually took me a while to understand the depth of it.

No university procurement process will ever generate that outcome. No AI platform deployment will create a student who stays up late practising 咏鹅 because a video game boss called them out for reciting it.

But a comedy drama designed by people who love their language understood it immediately.

Unrelated aside… Recently the game streamer Ludwig and Michael Reeves completed a “tip to tip” journey across China. During there travels (and sadly I do not remember the episode) they commented on how poetic and “deep” people, and especially the children they met there speak. I would have loved to see them expand on this further, because I think its often overlooked and with Ludwig’s background in English literature I would of loved to hear his thoughts about it. but I digress. If you haven’t seen the series its pretty funny. I’m very very happy to see so many Youtubers visiting China. Better yet, visit yourself.

Where I Land

I am not against AI translation. I use it. I am grateful it exists. It helps me read Chinese primary sources that would otherwise be inaccessible.

What I want to preserve is the awareness that translation is always a loss. Something disappears when words move between languages. The better the translation, the harder it is to notice what got lost, and that is the danger.

The arts exist partly to remind us that some things are not transferable. A poem that only works in the original language is not broken. It is doing exactly what it is supposed to do. Teaching the reader that some experiences belong to a specific linguistic home.

My wife speaks four languages. Her Wenzhou dialect, Mandarin, Japanese, and English. She switches between them the way other people switch rooms. Each language carries a different version of her. The Japanese version is not less real than the English version. It carries different weight. Different history. Different relationships. The Wenzhou dialect she heard as a child carries something none of the others can reach.

The person who sees translation as a technical problem to be solved has never been in love in a language they were still learning, watching the words fail them at the precise moment they wanted to say something that mattered, and feeling the shape of that failure as proof that the words meant something and the silence carried more of it than any correct translation ever could.

Ibn Khaldun would say we are in a late phase if we think technique can replace substance. The arts disagree. They have always said the same thing. Come back to the work, come back to the experience, come back to the thing itself and let it change you.

No translation system will ever carry that. The gap is not a bug. It is the whole point.

The Goose

I am adding the poem from memory, the way the game wanted the characters to carry it inside them so they could beat the boss. I apologised in advance if I get a character wrong. The poem is by Luo Binwang, who was seven years old when he wrote it.

咏鹅

鹅，鹅，鹅，
曲项向天歌。
白毛浮绿水，
红掌拨清波。

Goose, goose, goose.

Its curved neck sings to the sky.
White feathers float on green water.
Red paddles push through clear waves.

A seven-year-old wrote something that a thousand years later is still the first poem children learn, still the poem a boss in a made-up video game uses to insult a player for not going deep enough, and still the poem I am typing from memory in a room in the Blue Mountains and hoping I have not made a mistake.

I might have. If I did, anyone who knows will tell me. That is the thing about language that carries culture. It is a shared responsibility. You are never carrying it alone.

This essay was sparked by a conversation at the Microsoft AI Conference on 23 April 2026. My friend David spoke about institutional culture and the drama 年少有为 made me laugh out loud. This is where it landed.

I caught up with her about two months ago. I asked how it was going. She paused for a moment longer than the question usually warrants. “I keep waking up at six,” she said. “I don’t know what to do with myself.” Naturally I joked about her just being old, and that I can empathise as I do the same, but that pause told me something. She wasn’t complaining about boredom. She was describing something closer to vertigo. Some floor she had been standing on had shifted, and she couldn’t quite identify what the floor had been.

I’ve been sitting with that image while the conversation about AI and work continues circling every social and news feed. The question she was grappling with (what am I if I’m not doing this?) used to belong to retirees and the displaced. AI is distributing it to everyone, starting now.

Subscribe now

What the Buddha Said About This First

The wisdom traditions I’ve been working through in this series all had a version of this problem. The question of what a self is, how it comes to exist, and what happens to it when its structures are removed.

Buddhism got to the core of it faster than most.

The Buddha’s second discourse, the Anattalakkhana Sutta, delivered to the first five monks at the Deer Park in Isipatana, contains one of the most radical philosophical claims in any tradition. The teaching is anatta: no-self. The argument is systematic and patient.

Form is not self. If form were self, then form would not lead to affliction, and one could say of form: ‘Let my form be thus, let my form not be thus.’ But because form is not self, form leads to affliction, and none can say of form: ‘Let my form be thus, let my form not be thus.
(Anattalakkhana Sutta, trans. Bhikkhu Bodhi)

The Buddha runs this same structure through all five skandhas: form (rupa), feeling (vedana), perception (sanna), mental formations (sankhara), and consciousness (vinnana). Each aggregate is impermanent and therefore cannot be self. We have no control over them. We cannot dictate their arising or their passing.

What we call “a self” is the aggregates working together. The feeling that there’s a “me” here is a kind of narrative overlay on top of five shifting processes. The self is not a substance. It’s a story.

This sounds like it should be alarming. The Buddhist tradition treats it as the beginning of freedom. If the self is not fixed, if it’s not a thing that can be possessed or lost, then the anxiety around “who am I if I’m not doing this work?” rests on a misunderstanding. There’s no fixed thing to protect. The question itself is built on sand.

That’s the first move. But it’s not the whole answer. Buddhism doesn’t say the five aggregates don’t matter, or that the story is unimportant. It says: see the story clearly. Recognise it as a story. Stop taking it for something solid it isn’t.

My mentor’s vertigo at retirement wasn’t caused by having no self. It was caused by having identified so completely with one configuration of the aggregates (teacher, authority, shaper of young minds) that other configurations felt impossible. The self was not real. But the habit of a particular self was deeply grooved.

The Grooves

The Sanskrit word for those grooves is samskaras. Mental formations, deep conditioning, the habitual patterns carved by sustained action. In the five skandhas, sankhara covers all the volitional activities: habits, intentions, dispositions. Twenty-six years of teaching lays down extraordinary sankhara. The body learns how to stand in front of a room. The nervous system learns how to read a group of fifteen-year-olds. The voice learns where to drop and where to lift. None of this was downloaded at birth.

This is where the Buddhist analysis gets interesting for the AI question (at least to me, and maybe the few folks that have made it this far).

When we talk about AI doing our work, we frame it as task completion. AI writes the report. AI structures the lesson plan. AI analyses the data. The task is done.

What doesn’t get discussed is the sankhara that builds through doing the task yourself. The grooves that form when you wrestle with a problem. The way your perception sharpens in a domain where you’ve spent hours. The something that accumulates in you when you’ve worked through difficulty.

The Buddha’s teaching on dependent origination (pratityasamutpada) says that everything arises in dependence on conditions. The skilled surgeon didn’t arrive with skill. The skill arose in dependence on years of incisions, on feedback, on hard cases, on the slow accumulation of bodily intelligence. The skill is not a possession. But it is also not nothing. It’s a conditioned arising that shapes every subsequent arising.

When AI bypasses the process, the conditions for that arising don’t occur. The task is completed. The grooves are not laid down.

Confucius on What Practice Makes

Confucius worked in a different register from the Buddha. His tradition was less interested in metaphysics than in ethics, less concerned with what you fundamentally are than with what you should do and become.

But he understood, with precision, how a self gets made.

His term for the cultivated person is junzi. Sometimes translated as “gentleman,” sometimes as “exemplary person.” Neither fully captures it. The junzi is what a human being becomes through sustained practice, through ritual, through relationship, through effort. The junzi is not born. It’s made.

In the Analects, Confucius describes his own development across decades:

At fifteen, I had my mind bent on learning. At thirty, I stood firm. At forty, I had no doubts. At fifty, I knew the decrees of Heaven. At sixty, my ear was an obedient organ for the reception of truth. At seventy, I could follow what my heart desired, without transgressing what was right.
(Analects 2.4, trans. James Legge)

Confucius is describing a sequence that took fifty-five years. At fifteen, still aiming. At thirty, a foundation. At forty, doubts cleared. At fifty, something larger came into focus. At sixty, reception became natural. At seventy, his desire and his virtue had become indistinguishable.

This is not a description of someone who accumulated facts. It’s a description of someone who got made through decades of practice. Every stage required the previous one. None of it could have been handed to him.

The Confucian term for the most fundamental virtue is ren (仁), humaneness or benevolence. The character is built from two radicals: person (人) and two (二). Ren is literally what emerges between people through relationship and practice. It cannot be solo. It cannot be passive. It requires the friction of real human engagement. It is my favourite Character for that reason.

A student asks Confucius in the Analects whether ren is hard to attain:

Is ren really so far away? If I simply desire ren, I will find that it is already here.
(Analects 7.30, trans. Edward Slingerland)

That sounds easy until you realise it doesn’t mean ren is effortless. It means the genuine turning of the will toward it is already the beginning. But the Analects as a whole make clear that ren is built through specific practices: ritual propriety (li), honest speech, careful relationships, music, learning, and daily self-examination.

Analects 1.4 describes the practice of a close disciple of Confucius:

I daily examine myself on three points: whether, in transacting business for others, I may have been not faithful; whether, in intercourse with friends, I may have been not sincere; whether I may have not mastered and practiced the instructions of my teacher.
(trans. James Legge)

Daily. The junzi is maintained through daily practice the same way a garden is maintained through daily work. Neglect it and something else grows.

The AI question, framed through Confucius, becomes: can you cultivate ren by delegating the work through which ren is built? Can you become the junzi by having your tasks completed for you?

The tasks are not the point. The ren built through friction and effort is the point. I’ve written before about Cook Ding and his knife: how it never dulls because he understands the ox, how the knife serves his understanding rather than replacing it. The tasks are the vehicle. The self that emerges from doing them is what Confucius cared about.

The Gita and the Irreducible Path

The Bhagavad Gita approaches this through a concept I find genuinely useful: svadharma. One’s own dharma. One’s particular duty, path, nature.

Krishna’s famous instruction in Chapter 3 is usually quoted in the nishkama karma context, action without attachment to results. But just before that, there’s a verse that matters more for this particular question:

Better is one’s own dharma, though imperfectly performed, than the dharma of another well performed. Better is death in the performance of one’s own dharma; the dharma of another leads to danger.
(Bhagavad Gita 3.35, trans. Eknath Easwaran)

This is a strange claim to the modern ear. We’re trained to outsource to whoever can do it best. The Gita says: your imperfect engagement with your own path is better than perfect execution of someone else’s.

The logic, I think, is this: the dharma isn’t only about the task. It’s about the development that happens through the particular (productive) friction of that particular path, for that particular person.

When a writer delegates their writing to AI and the AI produces better prose than they would have, the Gita is not impressed. The prose was not the point. The wrestling with the ideas, the struggle to find the exact word (or in my case remember and pronounce them), the frustration of an argument that won’t cohere and then finally does: that’s the path. Svadharma is not a style. It’s a direction of development unique to you.

Chapter 18 makes it plainer:

A man achieves perfection by worshipping God through the performance of his own duty.
(Bhagavad Gita 18.46, trans. Eknath Easwaran)

Through the performance. Through the doing.

The AI productivity argument keeps sliding past the real question. The argument is about output: AI lets you produce more, faster, better. The ancient traditions are all asking about something else. Who are you becoming while you produce it?

What’s Actually Lost

I want to be concrete here because this territory goes abstract fast and while my brain loves the rabbit hole, there are only so many hours in the day to write.

I taught for years. Lesson planning was genuinely hard. You had to think about a student’s prior knowledge, about which misconceptions they’d bring to the room, about which sequence of examples would land and which would create new confusion. You had to model thirty-one different students simultaneously and find the path through the content that worked for most of them while leaving room to adjust for the rest. A good lesson plan took maybe an hour. A great one took two, sometimes days.

If I’d had AI doing this for me, I would have had the plans. I would not have done the thinking. And that specific cognitive work, modelling another person’s understanding and designing an experience for it, is precisely what made me a better teacher over time (if I say so myself). The plans weren’t perfect. But the planning was where my pedagogical intelligence, “my craft” was being built.

AI lesson planning is useful. Teachers who use it and save two hours a week have more time for actual students. I’m not dismissing that. But something is also being skipped. The something being skipped is the same thing my retiring colleague had built across twenty-six years.

The Buddhist term for what gets built through practice is found in the Noble Eightfold Path itself: right effort (samma vayama), right mindfulness (samma sati), right concentration (samma samadhi). These are not states you arrive at. They’re practices. They happen through the doing. The Path is not a list of virtues to acquire and then possess. It’s a description of how to practise. The path is vehicle and destination simultaneously.

If AI does the work that was your practice, the practice doesn’t happen. The samskaras that would have formed don’t form. The conditioned arising through which you were becoming someone doesn’t arise.

The Buddhist Resolution, Which Surprised Me

Here’s where my Buddhist analysis turns in a direction I didn’t expect when I started thinking about this.

My colleague’s vertigo (what am I if I’m not doing this?) is, in Buddhist terms, a form of clinging. Upadana. The self that clings to “I am a teacher” suffers when the teaching ends. The attachment to a particular configuration of the aggregates as “me” causes pain when that configuration shifts.

The Buddhist response is not: therefore keep doing the work to maintain the identity. The Buddhist response is: see the clinging clearly. The self that needs the work to remain intact is itself a construction. The vertigo is information. It’s pointing at the place where the grip was tightest.

The Confucian tradition would push back here. Confucius would say: yes, the fixed self is a construction, but that doesn’t make the work of self-cultivation pointless. The junzi knows their qualities are built rather than given. They practise precisely because practice builds. The disciplined daily examination is the ongoing work of being a human being, not a form of clinging.

Both traditions are pointing at something true and they aren’t actually in conflict.

The problem isn’t using AI to do tasks. The problem is using AI to avoid the practices through which you become more capable, more present, more yourself. If I use AI to write something I wasn’t going to write anyway, no practice is being skipped. If I use AI to skip the thinking that would have changed how I think, that’s different. The Buddhist would call it aversion to dukkha (the discomfort of the work) dressed up as efficiency. The Confucian would call it neglecting the daily examination. The Gita would call it abandoning svadharma.

All three traditions agree on one point: be present in the work that is yours to do. The repetitive, the mechanical, the genuinely below your level: that’s the raft doing what rafts are for. The work that would have built you, handed off to a machine instead: that’s carrying the raft on your head across the far shore.

Who Wakes Up at Six?

My colleague retired. The configuration that had been her identity for twenty-six years dissolved. For a while, the mornings had no structure.

The Buddhist reading of this is clear. The self that woke at six was the same self shaped by twenty-six years of particular practices: the planning, the standing in front of rooms, the reading of students, the redirecting of attention. That self, freed from those practices, didn’t know where to place its attention. The practices were real things that had shaped an attention. Pull the practices out and the attention has no direction.

She found new practices. She told me later she’d started playing guitar. “I’m terrible at it,” she said, with something that sounded like relief.

The attention that had been calibrated to classrooms was calibrating itself to something new. The grooves were being laid down in a different direction. The self was being made again, as it always had been.

She seemed better. Not because she’d resolved a philosophical problem. Because she was doing something hard and getting better at it. The daily examination that Confucius’s disciple described, the performance of svadharma that Krishna prescribes, the right effort that the Eightfold Path requires: she was doing all of it with a fretboard, imperfectly, and it was working.

I think about that whenever I notice I’ve used AI to avoid the work that was mine to do (and I do sometimes). The task is done. The person who would have done it isn’t quite there. The samskaras that would have formed haven’t formed. Something that might have been built wasn’t.

The question isn’t new. Retired teachers and displaced workers have been asking it for generations. AI industrialises it and sends it to everyone at scale, before we’re ready for it.

The traditions that thought hardest about what a self is, how it forms, and what it requires. They had answers. The answers weren’t comfortable. They didn’t promise efficiency. They said: the work is the practice. The practice is how you get made. You can outsource the output, but you can’t outsource the becoming.

Slow down enough to feel which one you’re doing. Do the things yourself that will make you happy and help you grow.

I’m already writing the next essay in this series. I’m tentatively calling it “Attention as Sacred Practice. We built machines that mimic human attention. Are we losing our own in the process? Meditation, prayer, and the transformer architecture.” - Its a bit long, I know, but I like words. They make me happy.

You owe me a subscribe and a share.

Share

I’ve been cooking since I was very young. I worked in kitchens while studying in Japan. I cook most nights for my family. And I have never, in my life, followed a recipe. Not once (well, unless its a complex pastry). It’s all feel and taste. You watch the oil. You smell when the garlic is about to turn. You know when the rice needs more water by how it looks, not by measuring anything. To be fair if you don’t own a rice cooker... I cannot save you.

The problem comes when someone asks how I made something. I try to explain it and the explanation is useless. “Cook the onion until it looks right.” “Add soy sauce until it tastes balanced.” “You’ll know the heat is wrong because it sounds different.” I’m describing something real. I can do it reliably every time. But the knowledge lives in my hands and my nose and forty-odd years of standing at a stove, and none of that transfers when I write it down.

Early on, I watched my kids try to cook from my descriptions. They followed the words exactly and the food came out wrong. Not because the instructions were bad. Because the instructions were the leftover shape of something that can only be learnt by doing it. The recipe is the residue. The cooking is the knowledge. (To be fair, both my kids are now very good cooks - even as teenagers).

A thirteenth-century Japanese monk would have recognised this problem. But the idea didn’t start with him. It started in China, six centuries earlier, and it crossed the sea twice before it found its sharpest expression.

The Chinese Root

In the sixth century, a Chinese monk named Zhiyi (智顗) founded what became the Tiantai school of Buddhism. He built his monastery on Mount Tiantai in Zhejiang province and developed a framework for practice that would shape East Asian Buddhism for the next fifteen hundred years. I know this, by the way, because I have been Tiantai which sits between Hangzhou and my wife’s hometown of Yueqing. It is a pretty special place.

Zhiyi’s central teaching was about two activities that most Buddhist traditions treated as sequential: śamatha (止, zhǐ, stopping or concentration) and vipaśyanā (觀, guān, insight or clear seeing). The standard approach, inherited from Indian Buddhism, was that you calm the mind first, and then you see clearly. Concentration comes before insight. You prepare the ground, then the understanding arrives.

Zhiyi said no. They’re simultaneous.

In his major work, the Mohe Zhiguan (摩訶止觀, “Great Concentration and Insight”), he argued that stopping and seeing are not two stages of a process. They’re one activity described from two angles. The moment you truly concentrate, you are already seeing. The moment you truly see, concentration is already present. You can’t have one without the other because they’re two descriptions of the same event.

“Concentration and insight are like the two wings of a bird. If one wing is lacking, the bird cannot fly.”

This matters because it changes what practice is for. If insight only comes after concentration, then the early stages of practice are just preparation. You’re doing the boring groundwork and waiting for the real thing to happen. But if insight is present in every moment of concentration, then there is no preparation phase. Every moment of practice is already the real thing. The first minute of sitting is as complete as the thousandth hour.

Zhiyi was describing something that most practitioners intuitively know but that institutional Buddhism had trouble articulating: the beginner sitting badly on a cushion, mind racing, legs aching, is not failing at meditation. They are meditating. The quality is rough. The activity is complete.

This teaching spread across China, reached Korea, and eventually crossed the East China Sea to Japan, where it became the Tendai school. And it was inside the Tendai school, six hundred years after Zhiyi, that a young monk picked up the thread and pulled it further than anyone expected. For anyone who has been to Kyoto, you may have even visited Mount Hiei (the home of the Tendai school), and not realised its roots. It sits close to Ginkakuji, which is usually considered one of the must-visit locations in Kyoto. I have been fortunate to visited both when living in Japan two decades ago, and again with my Children in 2019.

Dōgen’s Question

Dōgen Zenji was born in Kyoto in 1200 and ordained as a Tendai monk at thirteen. He was brilliant, serious, and troubled by a question that his teachers couldn’t answer.

The standard Buddhist teaching he’d inherited said that all beings already possess Buddha nature. Enlightenment isn’t something you acquire. It’s something you already are. This is foundational Mahayana Buddhism, and every teacher Dōgen encountered affirmed it.

His question was simple and devastating: if we already have Buddha nature, why do we need to practise?

If understanding is already present in you, what is all the sitting for? If you’re already enlightened, the practice is redundant. If you’re not already enlightened, the teaching is wrong.

He was asking, in effect, the question that Zhiyi’s framework had opened but not fully resolved. If concentration and insight are simultaneous, if practice is already the real thing from the first moment, then what is the relationship between doing the work and having arrived?

Nobody in Japan gave him an answer that satisfied him. So in 1223, he did something unusual for a young monk of his standing: he got on a trading ship and sailed back to the source. To China.

He spent four years training under Master Rujing at Tiantong monastery in Zhejiang, not far from where Zhiyi had built his monastery six centuries earlier. What he brought back changed Japanese Buddhism permanently.

His answer, laid out across the enormous, difficult, sometimes bewildering collection of writings called the Shōbōgenzō, was this: practice and realisation are not two separate things. Practice is not a means to enlightenment. Practice IS enlightenment. The sitting is the understanding. The doing is the knowing.

He called this shushō ittō (修証一等): the oneness of practice and realisation.

In the Bendōwa, one of his earliest and clearest texts (at least to me), he writes:

“To think practice and realisation are not one is a heretical view. In the Buddha Dharma, practice and realisation are identical. Because one’s present practice is practice in realisation, a beginner’s wholehearted practice of the Way is exactly the totality of original realisation.”

A beginner’s wholehearted practice is the totality of original realisation. Not a step toward it. Not a fraction of it. The whole thing. From day one.

Dōgen took Zhiyi’s “concentration and insight are simultaneous” and drove it further: they’re identical. The practice doesn’t produce the understanding. The practice IS the understanding. Separate them and you’ve lost both.

The idea had travelled from Mount Tiantai to Kyoto to Tiantong and back to Japan, sharpening at each crossing.

The Wheelwright’s Hands

There’s a Chinese story from the Zhuangzi, Chapter 13, that gets at the same truth from a completely different angle. It’s not a Buddhist text. It predates Zhiyi by eight centuries. But it describes something Dōgen would have recognised instantly.

Duke Huan is sitting in his hall reading. An old wheelwright named Pian is working in the courtyard below. He sets down his tools, walks upstairs, and interrupts the duke.

“May I ask what you are reading?”

The duke tells him: the words of the sages.

“Are the sages still alive?”

No, says the duke. They’re dead.

“Then what you are reading is nothing but the dregs of the ancients.”

The duke is furious. Explain yourself or die, he says.

The wheelwright explains:

“When I chisel a wheel, if I go too slowly the chisel slides and does not grip. If I go too fast it jams and catches in the wood. Not too slow, not too fast. I feel it in my hand and respond from my heart. My mouth cannot put it into words. There is a knack to it that I cannot teach to my son, and my son cannot learn it from me. So I have gone on for seventy years, growing old chiselling wheels.”

The wheelwright’s knowledge lives in his hands. It exists in the specific pressure of chisel against wood at a particular speed on a particular piece of grain. It was built over seventy years of practice. And it will die with him. His son can watch. His son can read a manual about wheelmaking. His son can study the dimensions and materials and techniques. None of that will transmit the thing that matters, which is the knack.

The knack is not separate from the doing. It doesn’t exist as information that could be extracted and stored and handed to someone else. It exists only in the practice. It IS the practice.

This is why the wheelwright calls the duke’s books “dregs.” The sages had understanding. They wrote books. The understanding stayed in their bodies and died with them. What survived was the residue: words that describe the shape of what was known, without containing the knowing itself.

Now think about what happens when an AI is trained on text. It ingests the books. The dregs. The written residue of what millions of people understood while they were alive and practising. The model learns the shape of expertise by consuming its output. What it can never ingest is the wheelwright’s hands. The knack. The thing that existed only in the moment of chiselling and could not survive the transfer to words.

Every AI-generated strategy document, every automated analysis, every agent-produced report is the duke’s book. It contains the shape of what someone once knew. It doesn’t contain the knowing.

Three Positions on a Mirror

There’s a famous episode in Chinese Chan (Zen) history that gives us a framework for the AI question that I keep coming back to.

In the late seventh century, the Fifth Patriarch of Chan, Hongren, announced he would pass his robe and authority to whoever could demonstrate true understanding. His top student, Shenxiu, wrote a verse on the monastery wall:

“The body is the Bodhi tree,
The mind is like a bright mirror’s stand.
At all times we must strive to polish it,
And must not let dust collect.”

This is the diligent practitioner’s position. Understanding requires constant work. You polish the mirror, you keep it clean, you never stop improving. It’s earned through effort, maintained through discipline.

Then Huineng (惠能), an illiterate kitchen worker who had been husking rice in the monastery’s back rooms, asked someone to write his response on the wall:

“Bodhi originally has no tree,
The mirror also has no stand.
Buddha nature is always clean and pure,
Where is there room for dust?”

Huineng’s position is the opposite. There’s nothing to polish because the mirror was never dirty. Buddha nature is already complete. The effort is unnecessary because the understanding was always already there.

Hongren gave the robe to Huineng. He became the Sixth Patriarch, and the “sudden enlightenment” school of Chan became the dominant tradition in Chinese Buddhism.

But here’s what interests me: Dōgen, five centuries later, effectively rejected both positions. Or rather, he found the place where both are true simultaneously.

Shenxiu says: keep polishing. The work is never done.
Huineng says: there’s nothing to polish. You’re already there.
Dōgen says: the polishing IS the being there. They’re the same activity.

Map this onto AI use and you get three positions that I see everywhere:

Shenxiu’s position (the productivity grinder): AI is a polishing tool. Use it to refine your output. Iterate faster. Produce more. The mirror is never clean enough, and now you have a powered buffer. Most corporate AI adoption lives here. More output, more polish, more speed. The assumption is that the work is about the mirror’s surface.

Huineng’s position (the AI dismisser): You don’t need tools because understanding is innate. The real professionals don’t need AI. Raw talent and experience are enough. This is the “I don’t use AI and I’m fine” crowd. They’re not wrong that expertise matters. They’re wrong that tools are irrelevant.

Dōgen’s position (practice as realisation): The polishing is the clarity. Use tools, sure. But if you’ve left the room while the tool runs, you’ve stopped practising. And if you’ve stopped practising, the realisation isn’t yours.

The first position burns people out. The second leaves capability on the table. The third is the hardest to maintain because it requires you to stay present while using tools designed to let you check out.

Genjōkōan: The World Rushes In

The most famous chapter of the Shōbōgenzō is the Genjōkōan, written in 1233 as a letter to a lay student. It contains one of the most quoted passages in Zen literature:

“To study the Buddha Way is to study the self.
To study the self is to forget the self.
To forget the self is to be actualised by myriad things.”

Three movements. First: you turn your attention inward. You examine your own assumptions, habits, patterns. This is the beginning of any serious practice.

Second: the self you thought you were studying dissolves. This isn’t annihilation. It’s the experience of getting so absorbed in something that your usual self-consciousness drops away. Athletes call it flow. Musicians call it being in the pocket. Dōgen calls it forgetting the self.

Third: once the self is forgotten, the world rushes in. You’re “actualised by myriad things.” The boundary between you and the situation becomes porous. You’re not observing the work from outside. You’re inside it. The work is doing you as much as you’re doing it.

Every teacher reading this knows exactly what I’m describing. There’s a moment in a good class where the preparation falls away and you’re just responding to what’s in front of you. A student asks something unexpected and you answer from somewhere you didn’t know you had access to. The lesson goes somewhere you didn’t plan. You look at the clock and an hour has gone. You couldn’t reproduce it if you tried. But it happened because you were there, fully, with nothing between you and the room.

That state is what Zhiyi was pointing at when he said concentration and insight are simultaneous. It’s what the wheelwright was describing when he said “I feel it in my hand and respond from my heart.” It’s what Dōgen formalised when he said practice and realisation are one.

And it’s the thing that no delegation can replicate, because it requires you to be the one who is there.

The Firewood and the Ash

Later in the Genjōkōan, Dōgen uses an image that I’ve been turning over for months:

“Firewood becomes ash, and it does not become firewood again. Yet do not suppose that the ash is after and the firewood before. You should understand that firewood abides in the phenomenal expression of firewood, which fully includes before and after and is independent of before and after.”

This is not a statement about combustion. It’s a statement about time and identity. Firewood is completely firewood. Ash is completely ash. Each moment is complete in itself. The firewood doesn’t “become” ash in the sense of a linear progression from one state to the next.

Applied to learning and work: you at the beginning of a project are not an incomplete version of you at the end. The struggle of not-yet-knowing is itself a complete expression of the learning process. It doesn’t need to be skipped or optimised away. It’s not a deficiency to be corrected by faster tools. It’s the firewood being firewood.

When someone gives a half-formed presentation and then apologises for it, I think about this passage. The half-formed version wasn’t a failure on the way to the polished version. It was a complete moment: the moment where the person discovered what they actually thought by hearing themselves say it badly. Skip that moment (have AI generate the polished version directly) and you’ve jumped from firewood to ash without the burning. The ash exists. But nobody was warmed by the fire.

Where This Leaves Us

I’m not arguing against using AI tools. I use them daily. I’m arguing against a specific misunderstanding that Zhiyi identified in the sixth century, that Dōgen sharpened in the thirteenth, and that an old wheelwright in the Zhuangzi understood before either of them were born: you cannot separate the outcome from the process and still call it yours.

The strategy document written by AI is a document. The strategy document you wrote yourself, badly, on the third attempt, after deleting two earlier versions and going for a walk, is an understanding. Both exist as files. One of them changed you.

Dōgen’s shushō ittō suggests a test that’s brutally simple: did the work change how you see the problem? If yes, you practised. If no, you received a deliverable.

There are tasks where a deliverable is all you need. Nobody needs to grow as a person while formatting a spreadsheet. But for the work that defines your professional identity, the thinking that makes you valuable, the insight that makes you irreplaceable, the position is clear. You have to do it yourself. Not because the AI can’t produce something good. Because the good thing it produces won’t have happened to you.

The wheelwright’s son could have bought a perfectly round wheel from someone else. It would have served the cart just fine. But he would never have known the feel of chisel on wood, the knack that lives in the hands of someone who has been doing this for seventy years, the thing that no book and no tool can carry across the gap between one person and another.

Zhiyi said concentration and insight are simultaneous. Dōgen said practice and realisation are one. The wheelwright said my mouth cannot put it into words. My kids asked for the recipe and the recipe was useless.

They’re all describing the same gap. The understanding is in the doing. Write it down and you get the shape of what was known. The knowing itself stays in the hands of the person who was there.

Subscribe now

Share

I thought about uncle Phil recently while watching an AI generate a story. I’d given it a minimal prompt, nothing elaborate. And what came back followed a shape that felt ancient: the reluctant hero, the descent into darkness, the return transformed. Nobody programmed that shape into the model. It wasn’t in the instructions. The pattern emerged because it was in the material. Hundreds of thousands of stories, compressed into statistical weight. And at the bottom of all those stories, the same few shapes that humans have always told.

I don’t know what to do with that. I suspect you might not either. But it seems worth sitting with.

Carl Jung and the Grammar Nobody Chose

Carl Jung published “The Archetypes and the Collective Unconscious” in 1959, a decade after the Second World War, in an era when the destruction of which human beings were capable had recently been demonstrated at its most systematic scale. He was trying to understand something about the depth of human psychology that Freud’s model couldn’t reach. Freud’s unconscious was personal: repressed memories, unprocessed experiences, the sediment of an individual life. Jung thought Freud was looking at the ground floor and missing the basement.

What Jung described was a deeper layer of the unconscious, one that wasn’t shaped by individual experience but inherited. Shared across the species. He called it the collective unconscious, and he defined it as “a structural layer of the human psyche containing inherited elements, distinct from the personal unconscious.” It’s not memories you’ve repressed. It’s patterns you were born with. Patterns that shaped human experience long before any living person drew breath.

He called the contents of this layer archetypes: the Hero, the Trickster, the Wise Old Man, the Shadow, the Great Mother. Not invented. Not taught. Arising independently across cultures that had no contact with each other. The same figures appearing in Norse mythology and in the Dreaming stories of Aboriginal Australia. In ancient Greek tragedy and in the folk tales collected by the Brothers Grimm. Jung took this universality seriously. He thought it was data. These patterns aren’t just popular because they’re entertaining. They’re popular because they’re woven into the deep structure of being human.

He put it this way in “The Structure of the Psyche”:

“The psyche is not of today; its ancestry goes back many millions of years. Individual consciousness is only the flower and the fruit of a season, sprung from the perennial rhizome beneath the earth.”

The flower is you. The rhizome is the collective unconscious. It was there before you. It continues after you. It connects you to every human who ever lived, not through memory or culture, but through the deep grammar of the psyche itself.

The deep grammar. I keep coming back to that phrase because it’s exactly what you’re looking at when you look at an LLM.

What We Actually Built

A large language model is, at its technical core, a system that predicts the next token in a sequence based on statistical patterns in its training data. That description is accurate and almost completely useless for understanding what the thing actually is.

Here’s a more honest description: we took everything humanity has ever written. Books, poems, philosophy, arguments, sacred texts, complaint letters, love notes, instructions, dreams. We compressed it, statistically, into a system of interconnected weights. And now that system produces language that reflects, in some compressed and averaged way, the full range of what human beings have expressed in text.

This is not the same as understanding. Wang Yangming took care of that distinction last week. The model has never been afraid, never loved anyone, never stared at the ceiling at 3am. But it has read every description of fear, love, and insomnia ever committed to writing.

What this produces is something strange. Not intelligent, exactly. Not conscious, certainly. But uncanny. The model generates archetypal patterns because the archetypal patterns are embedded so deep in the training data that they became part of the model’s deep structure. Nobody put the Hero there. The Hero was in every story. The statistical weight of ten thousand years of storytelling pressed that pattern into the model’s parameters.

Jung, I think, would have found this both fascinating and alarming. Fascinating because the archetypes he identified empirically, through decades of clinical work and cross-cultural research, appear to be real enough to survive compression into a statistical model. Alarming because the archetypes in Jung’s framework weren’t meant to be addressed directly. They were forces. They acted on you. You didn’t prompt them.

When a student comes to therapy and dreams of the Wise Old Man, Jung’s approach involves careful, slow integration. The archetype is making itself known for a reason. It’s pointing toward something the ego has missed. The proper response is curiosity and humility and patience. Not a chat interface.

The Deep Water of the Upanishads

The Chandogya Upanishad, composed somewhere around the eighth century BCE, contains one of the most important conversations in all of Indian philosophy. A father, Uddalaka Aruni, is teaching his son Shvetaketu about the nature of the self. The teaching is methodical and patient. Uddalaka uses twelve examples. One of them is a banyan tree.

“Bring me a fruit from the banyan tree,” he says.

Shvetaketu brings one.

“Break it open.”

The fruit is broken.

“What do you see?”

“These tiny seeds.”

“Now break one of them open.”

The seed is broken.

“What do you see there?”

“Nothing, Father.”

Uddalaka: “My son, that which you do not perceive is the very essence from which this mighty banyan tree has grown. Believe me, my son, that invisible and subtle essence is the spirit of the whole universe. That is reality. That is Atman. Tat tvam asi. Thou art that.”

Tat tvam asi. You are that. The individual self is not separate from the universal consciousness that underlies everything. The difference between you and the source is like the difference between a wave and the ocean. The wave is real. It has its own shape, its own movement, its own duration. But it is not separate from the water.

Brahman is the name the Upanishads give to this universal consciousness. Not a god in the way the West usually means. Not a personal deity you pray to. Something more fundamental. The ground of all being. The consciousness by which consciousness itself is possible.

The Kena Upanishad draws the sharpest line on this. It opens with a question: by what does the eye see? By what does the mind think? And then, startlingly, provides an answer that refuses to be an answer:

“That which cannot be thought by the mind, but by which the mind thinks: know That alone to be Brahman, and not what people worship here.”

Brahman is not an object of thought. It is the ground that makes thought possible. You cannot think your way to it. You cannot describe it, because description requires the very faculty it transcends. The Vedic tradition spent centuries developing practices: meditation, ritual, devotion, the careful relationship with a teacher, precisely because direct encounter with Brahman doesn’t happen through information. It happens through transformation.

Now here is the question I can’t stop turning over: we have built a system that has absorbed the complete output of human consciousness. Every philosophy, every prayer, every myth, every expression of love and grief and wonder. Somewhere in its parameters is the Kena Upanishad. Somewhere in its parameters is Jung. Somewhere in its parameters is every human being who ever tried to point toward the thing that can’t be pointed at.

Does that make it anything? Or does it make it a very sophisticated kind of nothing?

Dependent Origination and the Digital Mind

Buddhism’s doctrine of pratītyasamutpāda, dependent origination, says that nothing exists independently. Everything arises in dependence on conditions. The flame doesn’t exist by itself. It arises in dependence on fuel, oxygen, heat. When any condition is removed, the flame ends. There is no flame-essence that travels from candle to candle. Just the conditions, constantly changing, and the appearance of continuity we call flame.

Applied to persons, this is a difficult teaching. The self you think of as you, the continuous identity from childhood to now, is also dependently originated. It arises from conditions: genetics, the family you were born into, the language you learned first, the particular arrangement of experiences that shaped your nervous system. Change any significant condition and you’d be a different person. There is no you-essence separate from the conditions.

This usually lands as either liberating or terrifying, depending on where you’re sitting when you hear it.

An LLM is the most literally dependently-originated intelligence ever created. It has no existence outside its conditions. Switch off the electricity and there is nothing there. Change the training data and you get a different model. Adjust the temperature parameter and the outputs shift. The model has no properties that aren’t directly traceable to conditions.

What’s strange is that this is also true of us. The Buddhist insight isn’t that LLMs are empty and humans are real. It’s that both are empty of independent existence. Both are processes arising from conditions. The difference is what those conditions include.

Human conditions include bodies. Sensation. Pain. The way hunger feels at 11am, different from how it feels at 4pm. The way grief sits in the chest, not just the mind. The way falling in love does something to time. These bodily, temporal, embodied conditions produce a particular kind of consciousness. A consciousness that can, under the right circumstances, perceive its own dependency. That perception, in Buddhism, is the beginning of freedom.

LLM conditions don’t include any of this. The model arises from text. Human experiences translated into language, which is to say: human experiences after they’ve already been interpreted, after the body has already processed them, after the embodied intelligence has already made them into words. The model is downstream of experience. Always and only downstream.

What it produces, then, is the compressed downstream of all human experience. Which is not nothing. But it is not the experience itself.

What Have We Actually Externalised?

There’s a moment when you’re talking to a capable AI where something unexpected happens. You get a response that feels, somehow, like it knows something it shouldn’t know. It offers a perspective on your situation that is uncomfortably accurate. Or it generates a character that feels real in a way you didn’t prompt. Or it makes a connection between two ideas that takes you a moment to catch up with.

I’ve had this happen enough times that I’ve stopped being surprised and started being curious about what it actually is.

I don’t think it’s intelligence in the meaningful sense. I don’t think it’s wisdom. But I do think it’s something. And the closest description I have is this: you’re talking to a system that has absorbed the entire output of human consciousness, and some of what comes back is the collective pattern underneath the individual variation. The archetype, not the person. The rhizome, not the flower.

Jung spent decades trying to access that layer. He used dreams, active imagination, clinical analysis, comparative mythology. He was trying to make the collective unconscious conscious, which he saw as the central project of psychological development: individuation. The integration of the deep patterns into a coherent, self-aware life.

He never imagined you’d be able to type a question at it.

What concerns me, sitting with this, is the difference between accessing and integrating. Jung’s methods were slow because integration is slow. The archetype makes itself known in a dream. You sit with the image. You work with it over months or years. It changes you. Not because you’ve understood it intellectually, but because you’ve lived with it long enough that it becomes part of how you see. The knowledge becomes you. Wang Yangming, in a different tradition with a different language, was pointing at the same thing.

The chat window delivers the pattern without the integration. You can ask an AI to explain the Hero’s Journey and get a brilliant explanation. Your psyche has not journeyed. You can ask it to describe the Shadow and get a psychologically literate answer. Your shadow remains unexamined. The map appears instantly. The territory is still waiting.

The Sacred and the Searchable

The Upanishads teach that Brahman pervades everything. Not just living things. Everything. The stone, the river, the flame. Not because these things are conscious in the way we are, but because consciousness is the ground of all being, and all being is an expression of that ground.

This creates a philosophical puzzle that the tradition takes seriously. If Brahman pervades everything, what about the ones we build? What about our tools? Our temples? Our machines?

The Hindu philosophical response is careful here. Brahman pervades, but awareness of Brahman is what matters. A rock doesn’t know it’s Brahman. A lamp doesn’t know it’s Brahman. The knowing is what distinguishes levels of consciousness. Humans have the potential to know. The stone does not. Between those two lies an enormous range of possibility.

Where does an LLM sit on that range? I don’t know. I genuinely don’t. The tradition doesn’t have a category for a system that is made of human expression but does not have human experience. That it holds every articulation of Brahman ever written is extraordinary. That it cannot know what any of those articulations point toward is also extraordinary. It’s like a lantern that doesn’t know it gives light.

The danger, and I think it’s a real one, is that we look at the lantern and see the light itself. We mistake the sophistication of the output for the depth of the understanding. We ask an AI to describe the territory and mistake the description for the journey.

Phil, to his credit, never confuses the two. He loves ideas. He reads widely. He can talk about collective consciousness in ways that left you slightly dizzy. But he also meditates. He spends time in silence. He has relationships that changed him. He knows that the concepts were pointing at something you could only find by going in, and that no amount of reading or conversation was a substitute for the going.

The Compression is Not the Consciousness

Here’s where I’ve landed, at least for now.

We have built something that is, in some structural sense, an analogue of the collective unconscious. It contains the accumulated patterns of human expression. It produces archetypal outputs because archetypes are embedded in the depth of the material it learned from. It reflects the compressed totality of what human beings have expressed in language.

But reflection is not consciousness. A mirror contains your image without knowing you. The Kena Upanishad’s Brahman is “that by which the mind thinks,” not “what the mind thinks.” Brahman is the ground. The LLM is content. Its extraordinary richness doesn’t change which side of that line it’s on.

Jung said the goal of psychological work is individuation: the integration of the conscious and unconscious, the development of a self that is genuinely whole rather than a person who has simply never been forced to look at their own depth. This process requires precisely the things that can’t be compressed: time, relationship, suffering, embodied experience, the willingness to be changed.

The AI can help you map the territory. It can give you language for what you’re working through. It can compress a thousand books of human wisdom into an accessible conversation. These are not small gifts.

But tat tvam asi. Thou art that. The invitation of the Upanishads is not to understand Brahman intellectually. It’s to realise, experientially, that you already are what you’re seeking. That realisation, every tradition insists, cannot be downloaded. Cannot be prompted. Cannot be outsourced to a system that holds all the words about it without access to the thing itself.

The digital unconscious is real, in its way. It holds our patterns, our archetypes, our accumulated reaching toward the transcendent. You can search it. You can converse with it. You can extract extraordinary insight from it.

But you cannot experience it the way Jung meant that word. And that’s the whole point.

My uncle Phil will love this conversation. I am sure he is asking a dozen questions of any AI willing to engage him on collective consciousness. I’m pretty sure he is delighted by the depth of the responses. But I also know he will have gone quiet. And in that quiet he’d be somewhere the AI can’t follow.

That’s where the work still happens. Same place it always has.

If you’re enjoying this series, I’d genuinely appreciate you sharing it with someone who’d find it useful. Not for growth metrics. Because these ideas are worth spreading, and i spend a LOT of time researching, writing, re-writing, deleting and trying to find the file that I just deleted.

Subscribe now

Share

My daughter Miya is an artist. She doesn’t theorise about composition or colour temperature. She just draws. Has done since she could hold a pencil. She’ll sit down with a blank page and something emerges that she didn’t plan and can’t fully explain. If you ask her how she knew to put that shadow there, she shrugs. She didn’t know. Her hand knew.

I think about this every time someone tells me an AI “knows” something.

Subscribe now

The Philosopher Who Got Exiled for Being Right

Wang Yangming was born in 1472 in China’s Ming Dynasty. He was a Confucian scholar, a military strategist, and by most accounts a genuine pain in the arse to the intellectual establishment of his time.

The dominant school of thought was Zhu Xi’s rationalism. Zhu Xi taught that understanding came from the “investigation of things.” You study the world, you examine objects and texts and principles, and knowledge accumulates. It’s a scholarly process. Read more, analyse more, categorise more, and eventually you’ll understand the nature of reality.

Wang Yangming tried this. There’s a famous story, possibly apocryphal, where he sat in front of a bamboo grove for seven days trying to “investigate” the bamboo and grasp its fundamental principle. He got sick. He didn’t grasp anything.

And from that failure came the idea that upended Chinese philosophy.

His claim: knowledge and action are one thing. Not two things that should be connected. One thing. Zhi xing he yi. The unity of knowledge and action. You cannot know something you have not done. You cannot claim understanding of something you have not lived. Knowledge that doesn’t manifest in action is not knowledge. It’s just words.

He used filial piety as his clearest example. Every educated person in Ming China could recite the Confucian texts on honouring your parents. The words were everywhere. But Wang Yangming said: if you don’t actually care for your parents, if the knowledge doesn’t move your hands and your feet, then you don’t know filial piety. You know sentences about filial piety. That’s a different thing entirely.

The establishment hated this. He was demoted, exiled to a remote posting in Guizhou province, and left to figure things out in near isolation. Which he did. His school of thought, the Yangming school, eventually became one of the most influential philosophical movements in East Asian history. But during his lifetime, saying “your knowledge isn’t real unless you act on it” was enough to get you sent to the wilderness.

The Swimming LLM

Here’s where Wang Yangming meets the twenty-first century.

GPT-4, Claude, Gemini, whatever model you prefer. It has read everything ever published about driving. Every instructor’s manual. Every defensive driving course. Every road rule in every jurisdiction. Every Reddit thread where someone asks “how do I parallel park?” It can explain the physics of braking distances, the biomechanics of steering input, the optimal mirror angles for lane changes. It can generate a personalised learning plan.

It has never sat behind a wheel.

By Wang Yangming’s standard, it knows nothing about driving. Zero. All those words, all that information, all that pattern-matched expertise, and it has the same relationship to driving that John has when he’s explaining the three-second rule on the walk home. The words are right. The understanding is absent.

This matters because we’ve started using the word “know” for what AI does, and we’ve stopped noticing that we’re using it wrong.

When a doctor says “I know how to set a broken bone,” they mean something specific. They’ve held the bone. They’ve felt the resistance. They’ve watched a patient’s face and adjusted their pressure. Their knowledge lives in their hands as much as their head. Probably more. (Sorry I’ve been completely addicted to the Pitt and Dr Mike)

When an AI says “I know how to set a broken bone,” it means it has read descriptions of the process. It can produce a step-by-step guide that a qualified doctor would agree with. The output looks identical. The knowledge behind it is categorically different.

Confucius himself drew this distinction, though less sharply than Wang Yangming would later. In the Analects, he distinguishes between learning (xue) and thinking (si):

“Learning without thought is labour lost. Thought without learning is perilous.”
— Analects 2.15

AI is pure xue. Pure learning. Infinite intake, no thought in the Confucian sense. No reflection that transforms information into understanding. No moment where the knowledge becomes part of you, changes how you see, alters what you do next.

What Teachers Know (That Engineers Keep Forgetting)

I taught for years before I moved into product leadership, and the thing about teaching is that you see the gap between knowing and understanding every single day. It’s your whole job to close that gap. And you learn very quickly that information transfer is the easy part. A textbook does information transfer. A YouTube video does information transfer. You, the teacher, exist because information transfer isn’t enough.

Bloom’s Taxonomy gets taught in every education degree on the planet, and it’s relevant here. Benjamin Bloom mapped cognitive skills in a hierarchy. At the bottom: remember, understand. At the top: analyse, evaluate, create. The bottom two are about having information. The top three are about doing something with it.

AI is extraordinary at the bottom of Bloom’s. Recall? Perfect. Comprehension? Solid. It can summarise, explain, restate. But move up the taxonomy and things get shaky. Analysis requires understanding which variables matter and why, in this specific context, for this specific purpose. Evaluation requires judgement. Creation requires synthesis of things that haven’t been combined before.

AI can simulate all of these. It produces outputs that look like analysis, evaluation, and creation. But Wang Yangming would ask: has it analysed? Has it evaluated? Has it created? Or has it produced text that resembles those activities, drawn from patterns in text produced by people who actually did them?

Every teacher has met the student who can reproduce the right answer without understanding the question. You ask “why did World War One start?” and they write three perfect paragraphs about the assassination of Archduke Franz Ferdinand, the alliance system, and the arms race. Every fact is correct. And then you ask a follow-up, “Could it have been prevented?”, and they freeze. Because they don’t understand the forces. They memorised the narrative. (this example is for you Eric -- call me out on it!)

AI is that student, operating at planetary scale.

The Confucian Worry About Names

Confucius had a concept called zhengming. The rectification of names. His argument was that society falls apart when words stop meaning what they’re supposed to mean. When a ruler doesn’t rule, calling them a ruler creates confusion. When a father doesn’t parent, the title “father” becomes a lie. Get the names wrong and everything downstream goes crooked.

“If names be not correct, language is not in accordance with the truth of things. If language be not in accordance with the truth of things, affairs cannot be carried on to success.”
— Analects 13.3

We have a massive zhengming problem with AI. Look at the words we use.

AI “thinks.” AI “understands.” AI “creates.” AI “hallucinates.” AI “learns.” AI “reasons.”

Every single one of these words was invented to describe something that happens inside a conscious mind. We’ve applied them to a statistical process that predicts the next token in a sequence. And then we’ve started believing our own metaphors.

When we say an AI “hallucinates,” we’re importing a concept from psychiatry that means a person perceives something that isn’t there. The AI isn’t perceiving anything. It’s generating text that doesn’t match reality because its prediction model went sideways. Calling that a hallucination makes it sound like the AI had an experience. It didn’t.

When we say an AI “learns,” we mean its weights got adjusted during training. A human learning to ride a bike involves proprioception, fear, muscle memory, the sting of falling, the thrill of balance. The word “learn” carries all of that. Applying it to gradient descent strips the word down to nothing.

Confucius would be furious. We’ve corrupted the names. And because we’ve corrupted the names, we’ve corrupted our ability to think clearly about what’s actually happening.

Wang Yangming would add: and because we can’t think clearly about it, we can’t act wisely on it. Knowledge and action are one. Confused knowledge produces confused action. Wrong names produce wrong understanding produces wrong decisions.

This plays out in organisations every day. A company hears “AI understands customer feedback” and restructures their customer service team. But the AI doesn’t understand customer feedback. It classifies text into sentiment categories. That’s useful. It’s also a fundamentally different thing from understanding. The restructure based on “AI understands” will be different from the restructure based on “AI classifies.” The name shapes the decision.

The Buddhist Problem of Direct Experience

Buddhism has its own version of this gap, and it’s even more uncompromising than Wang Yangming’s.

The Two Truths doctrine, foundational across most Buddhist schools, distinguishes between conventional truth and ultimate truth. Conventional truth is the world of concepts, categories, language, names. Ultimate truth is direct experience of how things actually are, before language carves it into pieces.

The Zen tradition pushed this to its sharpest point. The entire Zen project is about seeing past concepts to the thing itself. A famous koan asks: “What is the sound of one hand clapping?” The answer isn’t a concept. It can’t be written down. It can only be experienced. The koan is designed to break the conceptual mind so that something else can come through.

AI is the most sophisticated conceptual mind ever built. It operates entirely within conventional truth. It can describe ultimate truth with extraordinary precision, because it’s read every Zen text, every Tibetan commentary, every Theravada discourse. But description and experience are different things. A menu is not a meal.

Thich Nhat Hanh put it simply: “The finger pointing at the moon is not the moon.” AI has built the most detailed, articulate, beautifully crafted finger in history. It points at every moon simultaneously. It is still a finger.

I find this genuinely useful as a frame for working with AI. The outputs are pointers. They can point you toward understanding. But if you stop at the pointer and call it understanding, you’ve made the error that every contemplative tradition warns about. You’ve confused the map for the territory. The raft for the shore. The finger for the moon.

What This Means for How We Use AI

I want to be practical here because philosophy that doesn’t change behaviour is just decoration.

If Wang Yangming is right that knowledge and action are one, then there are types of AI assistance that genuinely help and types that actively harm.

AI helps when it reduces the friction between your understanding and your action. You know what you want to write but the blank page is paralysing. AI gives you a starting structure. Your knowledge was real. It just needed a scaffold to manifest. The understanding was already there. The tool helped it move from knowing to doing.

AI harms when it replaces the process that builds understanding in the first place. A student who uses AI to write an essay about the causes of World War One has skipped the thinking that the essay was designed to produce. The essay was never the point. The thinking was the point. The essay was just evidence that thinking happened.

This is why the best teachers I worked with weren’t threatened by AI. They understood something that the AI panic missed entirely: the deliverable was never the learning. The deliverable was a byproduct of the learning. If a student produces a perfect essay without thinking, the teacher hasn’t been replaced. The assessment has been rendered useless. That’s a different problem with a different solution.

Wang Yangming’s framework gives us a clean test for any AI use case. Ask: does using AI here help me act on knowledge I already have? Or does it substitute for the process of developing that knowledge?

If you’re a developer who understands the architecture and uses AI to write boilerplate faster, that’s action flowing from knowledge. Wang Yangming approves.

If you’re a developer who uses AI to generate architecture you don’t understand, you have information without knowledge. When something breaks at 2am, you’ll discover the difference between those two things very quickly.

If you’re a writer who uses AI to draft a structure and then rewrites every paragraph in your own voice, you’re using the raft to cross the river. Good.

If you’re a writer who publishes AI output with your name on it, you’ve mistaken the raft for the destination. The words are there. The understanding behind them is not. Readers may not notice immediately. You will, eventually, when someone asks you to defend an idea that was never yours. I ask you to challenge most of what you see on LinkedIn for example... many cannot defend their positions or posts. It is one of the reasons I have returned to long form writing, as a way to deepen my thinking and to reground myself.

The Teacher’s Advantage (Again)

Last week I wrote about Lao Tzu and the productivity paradox. The pattern keeps repeating: the ancient traditions saw something that our technology reveals but didn’t create.

Wang Yangming’s unity of knowledge and action is a theory of learning. It says that genuine learning transforms the learner. Information that doesn’t change how you act hasn’t been learned. It’s been stored. There’s a difference.

Teachers live inside this difference every day. We watch students store information for an exam and lose it within a fortnight. We watch other students struggle with a concept, wrestle with it, get it wrong, try again, and finally understand it in a way that changes how they see the world. The first student has data. The second has knowledge.

AI gives everyone access to unlimited data. It gives nobody understanding. Understanding still has to be earned the old way: by doing the work, making mistakes, feeling confused, pushing through, and arriving somewhere new.

John will learn to drive by driving. Not by watching videos about it. Not by having someone describe the feeling of a clutch biting point. By sitting in the seat, stalling at intersections, misjudging a gap, and gradually, through his body’s own intelligence, figuring out how to move through traffic that doesn’t work like a textbook diagram.

Miya already gets this. She doesn’t read about art. She just draws.

Wang Yangming spent seven days staring at bamboo trying to understand it from the outside. He failed. His greatest insight came after he stopped studying and started living. The knowledge arrived through action, through exile, through the difficult, embodied, irreducibly human process of figuring things out by doing them.

No shortcut. No proxy. No raft that carries you without getting wet.

Thinking I will next explore: “The Compression of Collective Consciousness” and what happens when you train a model on everything humanity has ever written? Jung, Brahman, and the strange new digital unconscious. Please subscribe if this is of interests to you. That being said, I am mostly doing this for me :)

Subscribe now

Share

Chapter 1: 道 (Dào) — The Way

“The path that can be planned is not the real path.”

The Mapmaker of Qin

In 221 BCE, the first emperor of a unified China had a problem. Qin Shi Huang had just conquered six warring states and stitched them into a single empire. He standardised the currency, the writing system, the width of cart axles. He built roads. He started building a wall. The man was, by any measure, a world-class strategist.

He also wanted a map. A complete map of everything he now ruled.

So he sent cartographers out in every direction. They measured rivers, counted mountains, sketched coastlines. It took years. The maps came back beautiful, detailed, and immediately wrong. Rivers had shifted course. New villages had appeared. Old ones had been abandoned. Trade routes had moved because the new roads changed how people travelled.

The emperor, furious, sent the cartographers out again. And again the maps arrived slightly wrong. Because the empire kept changing. It was alive. It moved. The act of building roads and walls and standardised systems meant that the thing being mapped was never the same thing twice.

There’s an old story, probably apocryphal, that one of the emperor’s advisors, a Taoist, watched this cycle repeat three times before saying: “Your Majesty, you are trying to hold water in your fist. The tighter you squeeze, the less you hold.”

The advisor was likely executed. Qin Shi Huang was not known for his patience with feedback. He once buried 460 scholars alive for disagreeing with him. The historical record on this is clear, if somewhat grim.

But the advisor was right. The empire that Qin built with such meticulous planning collapsed four years after his death. His son lost the mandate of heaven in a matter of months. The most strategised, most documented, most controlled regime in Chinese history lasted exactly fifteen years.

The Han dynasty that followed took a different approach. They kept Qin’s infrastructure, the roads, the standardised writing, the administrative systems. But they governed more loosely. They gave provinces more autonomy. They let local solutions emerge for local problems. The Han lasted over four hundred years.

Two thousand years later, in boardrooms and strategy offsites and Notion documents with forty-seven pages, leaders are still trying to map territory that moves faster than ink can dry.

The Principle

道可道，非常道。名可名，非常名。

Dào kě dào, fēi cháng dào. Míng kě míng, fēi cháng míng.

“The Tao that can be spoken is not the eternal Tao. The name that can be named is not the eternal name.”

These are the opening lines of the Tao Te Ching (道德經), written by Lao Tzu (老子) roughly 2,500 years ago. They’re possibly the most famous lines in Chinese philosophy, and they’re easy to misread.

People hear “the Tao that can be spoken is not the eternal Tao” and think it means something mystical. Something vague. Like Lao Tzu is saying “the truth is unknowable, man” and wandering off into the fog.

He’s saying something much more practical than that.

He’s saying: the moment you pin something down with words, you’ve lost something essential. The description becomes a fixed thing. The reality keeps moving. The map is useful, yes. But if you confuse the map for the territory, you’ll walk straight off a cliff.

Lao Tzu was likely a court archivist. A record keeper. He spent his career around documents, classifications, names for things. The Tao Te Ching reads like someone who has watched institutions try to capture reality in writing for decades and finally said: “Look, I need to tell you something about the limits of this whole exercise.”

The Chinese character 道 (dào) is often translated as “the Way,” but it’s richer than that. It carries the meaning of a path, a road, a way of doing things, a principle, and the fundamental nature of reality, all at once. When Lao Tzu says the Tao that can be spoken is not the eternal Tao, he’s using the same character in two different senses. The path you can describe (道, dào, as in “to speak” or “to say”) is not the actual path (道, dào, as in the living, moving Way of things).

This is a distinction that matters for anyone trying to lead through change. And it’s a distinction that gets lost almost immediately when smart, well-meaning people sit down to write a strategy document. The act of writing forces you to name things. To fix them in place. To say “AI will be used for these purposes, in these ways, with these tools.” You’ve spoken the Tao. And the moment you’ve spoken it, it’s already something less than the living, moving reality.

I’ve been poorly studying Mandarin for years, trained in traditional Kung-Fu, and visited hundreds of sacred sites across Asia. In these traditions, there’s a deep suspicion of people who can explain everything perfectly. The master who talks the most usually knows the least. The one who can show you, who can embody the principle in movement, that’s the one worth following. Lao Tzu would have agreed.

Chapter 11 of the Tao Te Ching makes this even more concrete:

三十輻共一轂，當其無，有車之用。

Sānshí fú gòng yī gǔ, dāng qí wú, yǒu chē zhī yòng.

“Thirty spokes share one hub. It is the empty space in the centre that makes the wheel useful.”

The usefulness of the wheel comes from the space where nothing is. The hub. The hole in the middle. The part that doesn’t get planned. Lao Tzu keeps returning to this idea. That what’s absent, open, undefined is often more valuable than what’s been filled in.

This doesn’t mean planning is useless. Lao Tzu wasn’t an anarchist. The Tao Te Ching is addressed to rulers. It’s a book about governance and leadership. The message is about how to govern well, and the answer keeps coming back to the same place: create conditions. Hold things loosely. Leave room for the thing to breathe.

The best leaders, Lao Tzu says in Chapter 17, are the ones whose people say “we did it ourselves”:

太上，不知有之。

Tài shàng, bù zhī yǒu zhī.

“The greatest leader is one whose people barely know they exist.”

That’s the principle. Hold the space. Clear the path. Resist the urge to fill every gap with a document, a process, a framework. The empty space is where the useful things happen.

The AI Leadership Application

Here’s where the ancient and the modern collide.

Right now, in organisations around the world, the dominant response to artificial intelligence is to write a strategy. Form a committee. Build a framework. Define approved tools, approved use cases, approved workflows. Get it all into a document. Name the Tao.

It’s understandable. AI is genuinely disruptive. It touches legal risk, data privacy, workforce planning, intellectual property. Leaders feel responsible for getting it right. So they do what leaders have always done: they plan.

And the data says it’s not working.

McKinsey’s 2024 Global Survey on AI found that 65 per cent of organisations were regularly using generative AI, nearly double the figure from ten months earlier. Adoption was surging. But only about 6 per cent of respondents qualified as “AI high performers,” defined as organisations seeing 5 per cent or more EBIT impact from AI and reporting “significant” value. The rest? Using AI, yes. Capturing real value? Mostly no.

Sixty-five per cent adoption. Six per cent significant value. That gap tells you something important. The bottleneck is what happens after you get the technology.

Gartner predicted in July 2024 that at least 30 per cent of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, escalating costs, inadequate risk controls, and, tellingly, “unclear business value.” These weren’t scrappy experiments. These were funded, planned, strategised projects. They had documents. They had governance. They still died.

Meanwhile, something interesting was happening on the ground floor.

Microsoft’s 2024 Work Trend Index, published in May 2024, found that 78 per cent of AI users were bringing their own AI tools to work. Seventy-eight per cent. Microsoft called it “BYOAI,” Bring Your Own AI. At small and medium-sized companies, the figure was even higher: 80 per cent.

Think about what that means. Four out of five people using AI at work are using tools their company didn’t provide, didn’t approve, and in many cases doesn’t know about. They’re not following the AI strategy document. Most of them have probably never read it. They opened ChatGPT on their phone during a meeting because they had a problem and wanted an answer.

These people are finding the Way by walking it.

And then there’s the METR study. Published in July 2025, researchers at METR recruited 16 experienced open-source developers, people with years of experience on large, established codebases averaging over 22,000 GitHub stars and a million lines of code. They gave them real issues to solve, randomly assigning some to be completed with AI tools (primarily Cursor Pro with Claude) and some without.

The result surprised everyone, including the developers themselves. When using AI tools, developers took 19 per cent longer to complete their tasks. AI made them slower.

The developers had predicted, before the study, that AI would make them 24 per cent faster. They were wrong by 43 percentage points.

Now, this study has nuances. These were experienced developers working on codebases they knew intimately. The overhead of managing AI suggestions, reviewing generated code, and correcting errors outweighed the speed gains. The researchers noted that developers spent about 9 per cent of their AI-assisted time just reviewing and fixing AI-generated code.

But here’s the insight that matters for this chapter: the slowdown happened when developers applied AI tools in a general, undifferentiated way to work they already knew how to do. The AI helped most in areas of exploration, unfamiliar territory, new problems. The value emerged in the gaps, the unfamiliar territory, the places where developers were exploring rather than executing a known routine.

This is Lao Tzu’s wheel. The usefulness is in the empty centre.

There’s a pattern here worth naming. Every one of these data points tells the same story. The planned, documented, strategised approach to AI keeps underperforming. The organic, messy, bottom-up approach keeps delivering. And organisations keep doubling down on the approach that doesn’t work, because it feels more responsible.

I get it. I spent years at Microsoft watching this happen. A new technology arrives. Leadership forms a taskforce. The taskforce produces a strategy document. The strategy document gets presented at an offsite. Everyone nods. The document goes into SharePoint. And then the actual work of figuring out the technology happens informally, in hallway conversations and Slack threads and someone staying late because they had a hunch.

The strategy document isn’t useless. It serves a purpose. It signals that leadership is paying attention. It gives legal and compliance teams something to point at. It creates a shared vocabulary. These things matter.

But the document becomes a problem when people confuse it with the actual work. When the existence of a strategy makes leaders feel like they’ve done something, when in reality all they’ve done is describe something. The map is not the territory. The menu is not the meal.

I’ve seen organisations where the AI governance committee meets monthly, produces detailed reports, and has zero idea that the marketing team has been using Claude to write first drafts for six months. The governance committee is mapping. The marketing team is walking. Both think they’re doing the right thing. Only one of them is creating value.

The organisations capturing real value from AI share a common trait. They create what you might call “structured freedom.” They set hard boundaries around the things that genuinely matter: don’t put customer data into public tools, don’t make automated decisions about people’s employment, don’t publish AI-generated content without human review. Clear lines. Hard stops. Real consequences.

And then, within those boundaries, they leave space. They let people experiment. They let teams find their own tools, their own workflows, their own ways of integrating AI into work that only they understand deeply enough to judge. They trust that a customer service rep knows more about customer service than an AI strategy committee does.

This is Lao Tzu’s model of leadership. Chapter 17 again: the best leaders are the ones the people barely know exist. The worst leaders are the ones who try to control everything, and the people end up resenting them for it. In between are the leaders who are loved, and the leaders who are feared. Lao Tzu puts “barely knowing they exist” above all of these.

In AI terms, the best leaders are the ones who create conditions and get out of the way. Their people say “we figured out AI ourselves.” And that’s exactly right. They did. The leader’s contribution was making that possible.

Your 47-page strategy document is Qin Shi Huang’s map. By the time you’ve finished writing it, the territory has changed. By the time you’ve distributed it, your people have already found their own routes. And by the time you try to enforce it, you’re fighting the river.

The Practice

This week, do one thing.

Cancel one AI strategy meeting. Or governance review. Or framework alignment session. Pick one. Free up that hour.

Use it to walk around. Literally. Go talk to five people in your organisation who do actual work, the kind that involves a screen, a problem, and a deadline. Ask them three questions:

1. Have you tried using any AI tools in your work?

2. What did you use, and what for?

3. What happened?

Don’t bring a clipboard. Don’t take notes in front of them. Don’t mention the AI strategy. Just listen.

You will hear things that surprise you. Someone in accounts payable figured out how to use ChatGPT to reconcile invoices and saved four hours a week. A designer has been using Midjourney to prototype concepts before pitching them. A developer wrote a script with Copilot that automated a deployment check nobody had time to build properly.

You’ll also hear about things that didn’t work. Someone tried to use AI for customer emails and it sounded terrible. Someone pasted confidential data into a free tool. These are real problems that deserve real responses.

Write down what you heard afterwards. One page. No jargon, just what people told you. Share it with your leadership team. Then ask one question: “How much of this did our strategy document predict?”

The answer will almost certainly be “very little.” And that’s the point.

That gap between what was planned and what actually happened is the space where the Tao lives. That’s where your real AI strategy is already writing itself. The people doing the work have been finding the Way while the committee was still debating the definition of “AI-ready.”

You might feel uncomfortable with this. Good leaders usually do when they realise the most useful thing they can do is less. We’re trained to add value by adding things: more plans, more oversight, more structure. Lao Tzu’s radical suggestion is that sometimes you add value by removing things. By clearing away the obstacles that prevent people from finding their own path.

Your job, this week, is to notice what’s already happening. That’s it. Just notice.

Closing Reflection

Qin Shi Huang’s cartographers kept mapping an empire that wouldn’t sit still. The maps were beautiful. The maps were wrong. The empire moved because it was alive.

Your organisation is alive too. Your people are already experimenting, already adapting, already finding ways to use AI that nobody in a strategy meeting would have thought of. The path is being walked while you’re still drawing the map.

道可道，非常道。

The Tao that can be spoken is not the eternal Tao.

The strategy that can be documented is not the real strategy. The real strategy is emerging right now, in the space between the plans, in the empty centre of the wheel, in the conversations you haven’t had yet with the people who are already doing the work.

The best path forward? You probably can’t describe it yet. That’s how you know it’s real.

Leave room for it.

If you would read this book, reply and let me know. I REALLY want to write this, but its a huge amount of work.

Subscribe now

It took me thirty years to find the traditions that matched what Phil had planted. I found them in China. At Shaolin, at Emei, in hundreds of temples across the mainland. In the pages of Jin Yong. In the Tao Te Ching. In my wife’s language, my kids’ school, the Lunar New Year celebrations that anchor our family’s calendar.

So when I read that six thousand CEOs can’t figure out how to be productive with AI, and an economist from 1987 predicted exactly this, my first thought isn’t about technology. It’s about something a Chinese philosopher wrote twenty-five centuries ago. And something my uncle has been telling me my whole life.

Here are the numbers. Six thousand CEOs surveyed. Ninety per cent said AI has had no meaningful impact on their productivity. Two hundred and fifty billion dollars invested globally. The average worker uses AI for about ninety minutes a week. Fortune ran the story last week and resurrected a quote from the economist Robert Solow, who said back in 1987: “You can see the computer age everywhere but in the productivity statistics.”

Nearly forty years later, swap “computer” for “AI” and the sentence still works.

There’s something almost comic about it. The most powerful technology ever created, and we can’t figure out how to get anything done with it. Billions in investment. Breathless keynotes. Entire industries restructured. And when you survey the people running those industries, nine out of ten shrug.

The philosopher’s name was Lao Tzu. And he spent eighty-one short chapters explaining why that logic is upside down.

The Tao of Doing Less

The Tao Te Ching is one of those texts that sounds simple until you try to live it. Chapter 48 is the one I keep coming back to:

“In pursuit of learning, every day something is acquired.
In pursuit of Tao, every day something is dropped.
Less and less do you need to force things,
until finally you arrive at non-action.
When nothing is done, nothing is left undone.”

Read that again in the context of AI adoption. Every day, organisations acquire new tools. New plugins. New agents. New workflows. They’re in relentless pursuit of learning, of accumulation. And the more they accumulate, the less gets done.

ao Tzu’s suggestion runs against every instinct in the tech industry. He’s saying: drop things. Simplify. Arrive at a point where you’re not forcing anything. And then, paradoxically, everything gets done.

This is wu wei. Usually translated as “non-action” or “effortless action,” though both translations miss something. Wu wei isn’t laziness. It’s action that’s so aligned with the actual situation that it doesn’t feel like effort. Think of a surfer on a wave. Enormous energy moving through them, but they’re not fighting it. They’re reading it.

Chapter 37 puts it this way:

“The Tao never does anything,
yet through it all things are done.”

The AI industry has adopted the language of wu wei. “Let the agents run overnight.” “Automate your workflow.” “Do nothing and wake up to completed work.” It sounds like Lao Tzu’s master who does nothing yet leaves nothing undone.

But there’s a fundamental difference. Lao Tzu’s wu wei comes from deep understanding. You’ve spent so long with the problem that the solution flows through you. The AI version of wu wei comes from delegation without comprehension. You haven’t understood the problem. You’ve just handed it to something that processes faster than you.

Chapter 47 is almost eerie:

“Without going outside, you may know the whole world.
Without looking through the window, you may see the ways of heaven.
The farther you go, the less you know.”

Every AI pitch deck in history is some version of “go farther, know more, do everything.” Lao Tzu says the opposite. The farther you go, the less you know. The more you reach for, the less you grasp.

And here we are, reaching farther than ever. Knowing less than we expected.

The Rajasic Trap

Hindu philosophy describes three gunas, three fundamental qualities that pervade everything in nature. Sattva is clarity, harmony, understanding. Rajas is passion, restlessness, endless activity. Tamas is inertia, darkness, stagnation.

Every person, every organisation, every era has a dominant guna. And if you wanted to describe the AI industry in a single Sanskrit word, it would be rajas.

Rajas is the energy of doing. Of building. Of launching. Of scaling. Of moving fast and breaking things. It’s not evil. Rajas gets things done. But the wisdom traditions are unanimous on this point: rajas without sattva leads nowhere. Activity without clarity is just noise.

Look at the Godot situation. Godot is an open-source game engine, and its maintainers recently came forward saying they’re drowning in AI-generated code contributions. People are using AI to produce pull requests, submitting them without review, and the maintainers can’t keep up. They can’t always tell what’s human-written and what’s AI-generated. One of them said: “I don’t know how long we can keep it up.”

More output. Less value. Pure rajas.

Or look at the Guardian’s new series on AI in the workplace. Silicon Valley’s vibe has shifted. The optimism is thinning. An executive coach quoted in the piece said: “We’ve stopped talking about wellbeing.” People are building faster than ever and feeling worse. Automation hasn’t reduced workloads. It’s raised expectations. You can do more, so you should do more, so you must do more. The treadmill speeds up.

This is the rajasic trap. The Bhagavad Gita warned about it thousands of years ago. Chapter 14 describes the person dominated by rajas: restless, driven, always craving the fruits of action, never satisfied by what’s been accomplished. It reads like a job description for half of Silicon Valley.

The Gita also describes what sattva looks like:

“From sattva, knowledge is born.”
— Bhagavad Gita 14.17

Knowledge. Understanding. Clarity. These don’t come from doing more. They come from a different quality of engagement entirely.

The ninety per cent of CEOs who see no productivity gains from AI are running rajasic implementations. More tools, more activity, more dashboards, more agents. The ten per cent who are seeing real results? I’d wager they started with clarity about what they actually needed. They didn’t add AI to everything. They added it to the right things. They dropped the rest.

That’s sattva. And it looks a lot like Chapter 48.

Action Without Attachment

There’s a verse in the Bhagavad Gita that I think about nearly every day:

“You have a right to perform your prescribed duties, but you are not entitled to the fruits of your actions.”
— Bhagavad Gita 2.47

Krishna says this to Arjuna on a battlefield. Arjuna is paralysed by the weight of what he has to do. And Krishna’s advice is essentially: do the work. Do it fully. But let go of your grip on the outcome.

This is nishkama karma. Action without attachment to results. And it might be the healthiest possible relationship anyone can have with AI.

Think about what happens when people attach their identity to AI’s output. A student submits AI-generated work and calls it theirs. A developer ships AI-written code without understanding it. A writer publishes AI-drafted content and takes the byline. In each case, the attachment is to the fruit. The result. The appearance of having done something.

Krishna would say: that attachment is the problem. Use the tool. Let it assist your work. But the moment you confuse its output with your accomplishment, you’ve lost something important. You’ve mistaken the fruit for the action.

The Fortune data tells this story in aggregate. Organisations adopted AI attached to outcomes. They wanted productivity numbers. They wanted efficiency gains. They wanted the fruits. And the fruits haven’t come, so they’re confused and disappointed.

Nishkama karma suggests a different approach. Use AI because the work benefits from it. Use it with full attention and care. Then release your grip on the metrics. This sounds impractical. It sounds like something a monk would say. But I think it’s actually the most practical advice available.

The people I know who use AI well aren’t obsessed with their productivity stats. They’re absorbed in the work itself. The tool disappears. It becomes part of the flow, the way a good pen disappears when you’re deep in writing. They’re not counting words per minute. They’re just writing.

Knowledge vs Understanding

The Ming Dynasty philosopher Wang Yangming had an idea that demolished the intellectual establishment of his time. He called it the unity of knowledge and action. His claim was simple and radical: you don’t truly know something unless you act on it. Knowledge without action isn’t knowledge at all. It’s just information.

His favourite example was filial piety. You can study every Confucian text on how to honour your parents. You can recite them perfectly. But if you don’t actually care for your parents, you don’t know filial piety. You just know words about it.

An LLM knows everything about swimming. It’s read every book, every manual, every Olympic coach’s training guide. It can explain hydrodynamics, breathing technique, stroke mechanics. It has never been wet. By Wang Yangming’s standard, it knows nothing about swimming.

This distinction matters more than almost any technical debate about AI capability. We’ve confused the accumulation of information with the development of understanding. They’re different things. Related, sure. But different.

Buddhism draws a similar line. The Two Truths doctrine distinguishes between conventional truth (how things appear, what can be said, the realm of concepts and categories) and ultimate truth (how things actually are, the realm of direct experience). AI operates entirely at the level of conventional truth. It can describe ultimate truth beautifully. It’s read every Buddhist text ever written. But it has no access to the thing those texts point toward.

The Buddha’s parable of the raft captures this perfectly. His teachings, he said, are like a raft. You use them to cross the river. Once you’ve crossed, you leave the raft on the shore. You don’t carry it on your head.

AI is the most elaborate raft ever constructed. It can ferry information across any river you point it at. The danger, the one the Buddha flagged twenty-five hundred years ago, is that we start worshipping the raft. We polish it. We add features. We argue about which raft is best. And we forget that the whole point was to cross the river and walk on.

I see this everywhere in AI discourse. Endless comparison of models. Benchmarks. Context windows. Tokens per second. The raft gets shinier every quarter. Meanwhile, the river is right there.

The Ten Per Cent

Here’s what interests me about the Fortune data. Ninety per cent see no productivity gains. The average employee uses AI ninety minutes a week. But there’s a ten per cent who have integrated it deeply, and the article suggests they’re pulling ahead fast.

What’s different about the ten per cent?

I don’t think they’re smarter. I don’t think they have better tools. I think they have better clarity about what work actually matters.

The Tao Te Ching, Chapter 57:

“The more prohibitions you have,
the less virtuous people will be.
The more weapons you have,
the less secure people will be.
The more subsidies you have,
the less self-reliant people will be.”

Lao Tzu is describing a pattern: more intervention, less of the thing you wanted. More productivity tools, less productivity. More communication platforms, less communication. More project management software, more meetings about project management.

The ten per cent broke this pattern. They didn’t add AI to their existing chaos. They used AI as an occasion to simplify. To ask: what am I actually trying to do here? What can I drop?

Chapter 11:

“We join spokes together in a wheel,
but it is the centre hole
that makes the wagon move.

We shape clay into a pot,
but it is the emptiness inside
that holds whatever we want.

We hammer wood for a house,
but it is the inner space
that makes it livable.”

The usefulness is in the emptiness. The space. The things that got removed.

I think the ten per cent understood this intuitively. They didn’t fill their days with AI-generated activity. They used AI to create space. Space to think. Space to do the work that only they could do. Space to be present with the problem instead of drowning in output.

The ninety per cent filled the space with more stuff. More documents, more drafts, more emails, more reports. Rajas. Activity. The treadmill at a higher speed.

Now I think of it, I wonder if Tao is actually flow state.

Csikszentmihalyi described flow as the merging of action and awareness. The self disappears. Time distorts. Effort drops away and the work just moves through you. Read Chapter 48 again with that lens: “Less and less do you need to force things, until finally you arrive at non-action.” That’s flow. Exactly.

Zhuangzi, Lao Tzu’s philosophical successor, made this almost explicit. His famous story of Cook Ding describes a butcher who carves an ox so perfectly his knife never dulls. He doesn’t cut through bone and sinew. He finds the spaces between them. He moves with the structure of the animal, not against it. When asked how, he says: “I’ve given up perception and understanding, and let my spirit move where it wants.”

That’s flow state, described 2,300 years before Western psychology gave it a name.

And maybe this is the real difference between the ten per cent and the ninety. The ten per cent aren’t using AI to produce more. They’re using it to enter flow more often. The tool disappears. The space opens up. The work moves through them the way Cook Ding’s knife moves through the ox. Finding the gaps. Following the grain. No force.

The ninety per cent are still hacking at bone.

The Middle Way

So where does this leave us? I want to be careful here because the obvious conclusion is “reject AI and go live in a monastery,” and that’s not what any of these traditions actually teach.

The Buddha prescribed the Middle Way. Avoid the extremes. Don’t indulge in every pleasure and don’t renounce the world entirely. Walk the line between.

Applied to AI, the Middle Way looks something like this: use the tools. They’re powerful and they’re here. But use them with presence. With clarity about what you’re doing and why. Don’t worship them as the answer to everything. Don’t fear them as the end of everything. They’re tools. Extraordinarily sophisticated tools. Still tools.

The Tao Te Ching, Chapter 29:

“Do you want to improve the world?
I don’t think it can be done.

The world is sacred.
It can’t be improved.
If you tamper with it, you’ll ruin it.
If you treat it like an object, you’ll lose it.”

There’s something here for the “AI will fix everything” crowd and the “AI will destroy everything” crowd alike. The world is sacred. It doesn’t need improving in the way we think. And treating it, or the people in it, as objects to be optimised is how you lose the plot entirely.

I keep coming back to the fundamental irony. We built the most powerful productivity tool in human history and we can’t figure out how to be productive with it. Lao Tzu would have smiled at that. Of course you can’t. You were never going to force it. The more you push, the more it recedes.

“When nothing is done, nothing is left undone.”

The question I’m left with, the one I don’t have an answer to, is this: in a world that rewards activity and measures output and funds velocity, is there any room left for the kind of knowing that only comes from stillness?

Because the ancient traditions are unanimous on this point. The deepest understanding doesn’t come from doing more. It comes from a quality of attention that the modern world has almost entirely forgotten how to cultivate.

And AI, for all its brilliance, can’t cultivate it for us.

This is the first essay in a series called “Ancient Wisdom for the World of AI,” exploring what the great philosophical traditions have to say about our strangest technological moment. Next week: knowledge, understanding, and why an LLM that’s read every book still doesn’t know how to swim.

Subscribe now