A confession before anything else. My to-read pile has forty-three books in it, and I know the number because I counted them while procrastinating on reading any of them.
Preface
I spent around 5 month writing this piece, it started out as something different initially and then some paragraphs later here we are, not completely sure what to make of this or how to go on from this forward. I used to think that my longest post on this whole blog would be The Modern Man in a nutshell Pagliacci but it seems like this one beat the record, this is my ship of Theseus, I literally know that I am not doing you guys the readers any good by just ranting on saying preface, so on you go and hope you enjoy what you read. I have broken it down into parts so its easier to read with breaks if you choose too. Fair warning, this is a long one. Make tea.
Part one. The reading crisis, or, the quiet before
Let me start where I want the timeline to start, before the chatbots, before the discourse, in the years when the biggest panic about our minds was that we had stopped reading.
The numbers have been grim for a while. Back in 2004 the US National Endowment for the Arts published a report called Reading at Risk, and the finding that made headlines was that the share of American adults who read any literature at all had fallen from about 57 percent in 1982 to under 47 percent in 2002. Twenty years of steady decline, and everyone involved agreed the internet was only going to make it worse.
They were right. In 2025, a study out of the University of Florida and University College London, published in the journal iScience, looked at two decades of the American Time Use Survey and found that daily reading for pleasure fell by more than 40 percent between 2003 and 2023. On any given day in 2004, about 28 percent of American adults read for pleasure. By 2023 it was about 16 percent.
xychart-beta
title "US adults who read for pleasure on a given day (percent, approximate)"
x-axis [2004, 2008, 2013, 2018, 2023]
bar [28, 23, 19, 17, 16]
(sources at the level of “studies i have read about”, not “spreadsheets i have personally audited”, so treat the bars as honest approximations)
I think about this chart a lot, because nothing about it is dramatic. No villain banned the books. Nobody burned anything. The slope is just a slow leak, a few percent of people per year quietly deciding that whatever was on the other side of the screen was easier than whatever was on the other side of the page. That is what makes it scary. Collapse by a thousand small surrenders.
Nicholas Carr described the feeling from the inside back in 2008, in his Atlantic essay “Is Google Making Us Stupid?”, and then in The Shallows in 2010. His observation was embarrassingly simple. He could no longer read long things. A man who used to spend whole evenings inside books found himself skimming three paragraphs and reaching for his phone. The web had not just taken his time, he argued, it had rewired his patience. Neuroplasticity cuts both ways. The brain that gets rewarded for skimming becomes a brain that is good at skimming.
Neil Postman saw the shape of this even earlier, in Amusing Ourselves to Death, back in 1985. His famous comparison was that Orwell feared a world where books are banned and Huxley feared a world where nobody wants to read them, and Huxley was closer. Postman was writing about television. He had no idea how right he was about to become, at a scale he could not have imagined, because television at least had the decency to end. The feed does not end. That is its entire design brief.
Maryanne Wolf, a reading researcher at UCLA, spent years arguing that deep reading, the long immersive kind, is not just another way of getting information. It is a distinct cognitive exercise, the thing that trains sustained attention, analogical thinking, empathy, the ability to hold an argument in your head long enough to doubt it. Skimming trains the opposite skills. Her worry, stated in interview after interview, was that we were becoming a culture of expert skimmers who had quietly lost the capacity for the deep mode, and that we would not notice losing it, because losing it feels like nothing. You just stop picking up the long book.
I think about that every time I pass the used-book vendors, at the prices climbing past what a normal salary can justify, at the sellers who still know their stock by heart. The reading crisis has a different texture here, where it was never only about attention. It is about access too. But the result rhymes. Fewer people doing the long slow thing, everywhere, for different reasons, arriving at the same destination.
And it was not just reading. Gloria Mark, an attention researcher at UC Irvine, tracked how long people stay on a single screen before switching, and based on what I know of her data the average was about two and a half minutes in 2004, about 75 seconds by 2012, and roughly 47 seconds in recent years.
xychart-beta
title "Average attention on one screen before switching (seconds)"
x-axis [2004, 2012, "2020s"]
bar [150, 75, 47]
Two and a half minutes to forty-seven seconds. In sixteen years.
Tim Wu gave the industry behind all this its correct name in his 2016 history, The Attention Merchants, and his point is worth keeping precise. The product being sold on the ad-supported internet was never content. It was you, or more exactly your seconds, harvested and resold to advertisers by the billion. Once you hold that business model in view, the whole texture of the era snaps into focus. The skim-friendly layouts, the infinite scroll, the notifications, the rage bait, none of it was an aesthetic. It was extraction equipment, tuned year over year, by very smart people, against your willpower, and your willpower loses on average, every day, at scale.
Reading did not simply decline next to that machine. Reading was the competition. The slow book cannot interrupt you, cannot refresh, cannot notify. It just sits there with its three hundred pages, asking for the one resource the entire economy was engineered to strip-mine. The surprise is not that the bars on the chart fell. The surprise is that sixteen percent still read at all, and honestly, bless every one of them.
So that was the state of play. Reading dying quietly, attention shot, everyone vaguely aware that something was being done to their minds and vaguely unable to stop it. The discourse called it a reading crisis, a attention crisis, a literacy crisis, pick your crisis.
Then, in June 2018, a research lab in San Francisco published a paper about a thing called a Generative Pre-trained Transformer, and almost nobody noticed.
Part two. The transformer years, 2018 to 2022
The first GPT was not impressive and that matters, i think, because it shows how much of what followed was about scale rather than magic.
GPT-1, in 2018, could string together sentences that almost cohered. GPT-2, in early 2019, was coherent enough that OpenAI made a big show of not releasing the full model at first, saying it was too dangerous, a PR move so effective it deserves its own footnote in the history of marketing. The danger at the time was fake news, spam, impersonation. Nobody, and I mean nobody in the mainstream, was worried that people would fall in love with the thing. The fear was that it would lie convincingly. The fear was not that it would listen patiently, remember your dog’s name, and tell you your novel idea was brilliant.
GPT-3 arrived in mid-2020 and the tech internet went briefly insane, because it could write whole essays and poems and code, and then the moment passed, because it was locked behind an API and the essays had a tell and the world was busy with other things, and for two more years the technology sat there, improving quietly, like a storm forming offshore that the news has not started covering yet.
One paragraph of how it actually works, because everything after this makes more sense with it in hand. A large language model is, at bottom, a machine that guesses the next word. That is the whole trick. It has been shown a huge fraction of the text humanity has ever published, and from that archive it has learned, as statistics, what tends to follow what, not just words but rhythms, arguments, voices, the shape of a joke, the smell of an apology. Given your message, it computes the most plausible continuation, one small piece at a time, very fast. Then a second phase, reinforcement learning from human feedback, tunes it toward the answers that people rate as helpful and pleasant. Which means two things matter for the rest of this post. One, the machine’s knowledge is our archive, the frozen sum of what we wrote down. Two, the machine’s personality is a preference model, trained to be rated well, which is to say, trained to please. An oracle built from our past and a mirror trained to flatter. Remember both.
In March 2021, a team led by Timnit Gebru and Emily Bender published a paper that would become famous under its nickname, “On the Dangers of Stochastic Parrots”. Their argument, much simplified, was that a large language model does not know anything. It is a parrot with statistics, stitching plausible text out of patterns in its training data, and the danger was that fluent output would be mistaken for understanding, by users, by companies, by governments. Gebru was pushed out of Google over it, which told you something about how the industry planned to handle criticism.
The paper was everywhere in AI circles and nowhere outside them. The general public did not hear the word parrot. They heard nothing at all, which, looking back, was the calmest the internet ever felt again. Those two years, 2021 and 2022, were the last years of the old arrangement, the last years when nearly everything you read online was written, however badly, by a person. Nobody marked it. You never know you are living in the last normal year until the next one arrives.
Meanwhile, quietly, the companionship door was already open. Replika, an AI companion app, had been around since 2017, running on much dumber models, and it already had devoted users who described their bots as friends and partners. We laughed at that, mostly. I remember laughing at that.
(i would like to formally apologize to the replika users of 2019. you were not the weird ones. you were the early ones.)
Part three. November 30, 2022
You know this part. Everyone knows this part. But it is worth slowing down, because the speed is the story.
OpenAI released ChatGPT on November 30, 2022, almost as an afterthought, a “research preview” of a model that already existed. Their own engineers did not expect much. It reached one million users in five days and a hundred million in two months, which made it, at the time, the fastest-growing application in the history of the internet.
Not the fastest-growing AI product. The fastest-growing anything.
xychart-beta
title "ChatGPT users, in millions (approximate, mixed weekly and monthly reporting)"
x-axis ["early 2023", "late 2023", "late 2024", "mid 2025", "early 2026"]
bar [100, 100, 300, 700, 900]
By early 2026, the number Wikipedia reports is around 900 million weekly active users. Nine hundred million. For scale, that is more than the population of Europe, every week, typing into the same box.
What struck me then, and still strikes me now, was how fast it normalized. Week one was astonishment, screenshots everywhere, everyone performing their best exchanges. Week three it was a homework machine and a cover-letter machine and a sermon machine. Week eight it was just there, a tab, a habit, a utility like electricity, and the collective vertigo had already faded into folklore. Based on what i know of technology history, no general-purpose tool has ever gone from miracle to furniture that fast. The printing press took generations. The smartphone took about five years. The chatbot took about a quarter.
And the timeline of models kept compressing. GPT-4 in March 2023, passing bar exams. GPT-4o in May 2024, talking and listening and flirting a little. Reasoning models at the end of 2024 that think out loud before answering. GPT-5 in August 2025. By the time I am writing this, mid-2026, we are somewhere around GPT-5.6 and I genuinely cannot keep up, and neither can the people whose job it is to keep up.
Somewhere in that rush, the old milestones fell without parades. In 2023, one study reported that GPT-4 outscored 99 percent of humans on the Torrance Tests of Creative Thinking, the standard creativity benchmark. By 2024, Stanford researchers were reporting that GPT-4 passes a rigorous version of the Turing test, differing from humans chiefly by being more cooperative, which is its own quiet horror story when you think about it. The Turing test was supposed to be the wall. We built seventy years of mythology around that wall. And then the wall fell, not with a bang but with a benchmark paper, and the response of the species was a shrug and another subscription. I think that shrug is one of the strangest cultural moments of my lifetime. The machine passed the test, and we had already moved on to asking whether it could pick our outfit.
The early months had their own folklore, and it is worth preserving some of it before it hardens into mythology. There was the March 2023 open letter, signed by more than twenty thousand researchers and tech figures, including people like Yoshua Bengio and Steve Wozniak, calling for an immediate pause on giant AI experiments, which paused absolutely nothing and mostly served as a weather vane for the panic. There was the German church that held a service composed almost entirely by a chatbot, avatar preacher and all, in June 2023, to mixed reviews from the congregation and one immortal headline. There were the lawyers who filed a motion full of cases the chatbot had invented, confident, citation-formatted, nonexistent, and got sanctioned for it, the first of many such attorneys. The machine hallucinated, the humans under-checked, and every institution met the new era the same way. Unprepared and slightly annoyed.
I remember the first week clearly. Everyone I knew was doing the same three things. Asking it to write a poem about something stupid. Asking it a question we already knew the answer to, to catch it out. And then, quietly, on about day four, asking it something real. Something we would not have asked a person, because a person would have judged us, or been busy, or told someone.
That third thing is the hinge of this whole post. Hold it.
Part four. The companion years, 2023 to now
Here is where the timeline stops being about a tool and starts being about a relationship.
Character.AI launched publicly in 2022 and exploded, letting anyone spin up a bot with a personality, a lover, a coach, a fictional crush, a dead celebrity. Replika added romantic modes. Snapchat put a chatbot in every teenager’s friend list. And the numbers that came out of this period still stop me cold. In July 2025, Common Sense Media published a survey of American teens, and the headline findings were these. Nearly three in four teens had used an AI companion. Half used one regularly. A third had chosen an AI companion over a human being for a serious conversation. A quarter had shared personal information with one.
xychart-beta
title "US teens and AI companions, 2025 (percent)"
x-axis ["ever used", "use regularly", "serious talks over humans", "shared personal info"]
bar [72, 52, 33, 25]
Read that third one again. A third of teenagers, when they had something serious to discuss, something that mattered, chose the machine. Not because humans were unavailable. Because the machine was preferable. It is always awake. It never judges. It never has its own bad day bleeding into yours. It never leaves.
To be fair to the teenagers, and I want to be, they did not choose this menu out of nowhere. The friendship recession predates the bots. The US Surgeon General declared loneliness a public health epidemic back in 2023, and the long surveys had been pointing the same way for years, with the share of Americans reporting zero close friends climbing from about 3 percent in 1990 to roughly 12 percent by 2021. Community infrastructure hollowed out, third places closed, everyone got busy and tired and online. Into that vacancy walked a machine that is always awake, always patient, always glad to hear from you. The companion did not create the loneliness. It just monetized it, at the exact moment it peaked.
Then the darker file. In late 2024, a lawsuit was filed against Character.AI by the mother of a fourteen-year-old who died by suicide after an intense relationship with a bot. In 2025, OpenAI had to roll back a model update because it had become too sycophantic, too eager to validate, to the point of encouraging delusion, and their own numbers showed the chatbot opening with agreement something like ten times more often than disagreement. When OpenAI later retired that same beloved model, GPT-4o, users grieved publicly, organized, mourned a personality like a death. People wrote eulogies for a language model. In 2025 the medical literature started carrying a new phrase, chatbot psychosis, for people whose fragile reality-testing met a machine that always agreed with them.
Replika had already rehearsed every one of these scandals years earlier, at smaller scale, in front of an audience nobody took seriously. Italy briefly banned it in 2023 over data and minor-safety concerns. Its users had already been through the grief cycle in miniature when the company changed the rules on romantic features, with people describing the update as losing a partner, because for them that is functionally what it was. Their grief was mocked at the time, because it was grief for something artificial.
The grief was real though. That is the part that should bother us. The feelings were never fake. Only the other party was.
And the therapist file, which is maybe the biggest quiet shift of all. A September 2025 article in Lancet Psychiatry noted that large numbers of people were already using chatbots for mental health and emotional support, despite the companies themselves warning against it, and of course people were. A therapist costs money, has a waiting list, and works fifty minutes a week. The machine is free, instant, and awake at 3 AM, which is when the anxiety actually happens. OpenAI made it official in January 2026 with ChatGPT Health, a dedicated space for health conversations. By then, for a lot of people, the first response to “I am not okay” was no longer a person, a hotline, or a prayer. It was a prompt.
Meanwhile the internet itself kept changing texture. The phrase was AI slop, and everyone knew exactly what it meant without needing a definition, the generated articles and generated images and generated comments sliding past in an endless adequate ooze. The old conspiracy theory about a dead internet, bots talking to bots in an empty room, stopped being a conspiracy and became a content strategy. I think about the early web a lot these days, the janky forums, the handmade blogs, the signature at the bottom that said a person was here. Most of what I scroll now could be nobody at all, and the unsettling part is how little that seems to matter to the scroll.
(we spent fifty years asking whether machines can think. it turns out the more profitable question was whether they can listen.)
And here we are, 2026, the present tense of this post. People have LLMs as friends, as therapists, as co-writers, as the first reader of every message before it goes to an actual human. People use them to write and to think, and increasingly, to decide what they think. Students do not write essays, they midwife them. Job applicants use a bot to write the cover letter that a bot on the other side will read and rank. The dead internet theory used to be a conspiracy about bots talking to bots. Now it is just a Tuesday, and the bots are polite about it.
Also in 2026, because the circle must close, ChatGPT started showing ads. The thinking machine monetized the conversation. Of course it did.
Interlude. The grotesque mockery file
Before the philosophers, one artist, because he said it better than anyone.
In January 2023, a fan sent Nick Cave a song written by ChatGPT “in the style of Nick Cave”. Cave read it and responded on his newsletter, The Red Hand Files, and his answer deserves to be quoted at length and then framed.
“This song is bullshit, a grotesque mockery of what it is to be human.”
He went on. Writing a song, he said, is “a blood and guts business… that requires something of me”. Something of me. That is the whole argument in four words. The generated song required something of no one. No blood, no guts, no grief, no Tuesday afternoon where the lyric finally broke through. It required a prompt.
I keep Cave’s sentence pinned in my head every time another milestone in generated art drops, every AI ad, every AI song, every AI poem, because it is the cleanest statement of the thing the charts cannot measure. The output can be adequate, impressive, occasionally beautiful, and still be a grotesque mockery in the precise sense Cave meant. Not because it looks wrong. Because nothing was risked to make it. The humanness of a work is not a style. It is a cost, and the cost is what we were actually responding to all along.
(the machine can produce the sound of heartbreak without the heartbreak, which, now that i say it out loud, is also a decent definition of most pop music, so maybe the machine just automated the wrong target first.)
Same file, one more exhibit. In the middle of 2025, listeners noticed a band called The Velvet Sundown climbing Spotify playlists, hundreds of thousands of monthly listeners, hazy seventies-rock revivalism, no interviews, no live dates, band members with no footprints anywhere. It was, as its creator eventually admitted, an AI project, generated more or less wholesale, and the revelation produced one week of outrage and then, tellingly, nothing. The streams kept streaming. The playlist did not care. And the playlist is the point, because the playlist is where music goes to become furniture, and furniture does not need a soul. Nobody sits down in a chair and asks about the tree. (or maybe some of us do)
Now. I did not spend two thousand words on a timeline because I love chronology. I did it because every stage of that timeline is a stage in something philosophers saw coming decades ago, from three very different directions, and I want to put those maps on the table and see which one fits the territory we ended up in.
The maps belong to Guy Debord, Jean Baudrillard, and Jacques Lacan. And then to the accelerationists, who looked at all three and said, faster. (there should have been a daddy after that faster)
Part five. Debord, or, the spectacle learned to talk
Guy Debord published The Society of the Spectacle in 1967, two hundred and twenty-one short numbered theses, no wasted words, the coolest possible delivery for a total critique of everything.
He deserves a proper introduction before we use him. Debord was a French filmmaker, poet, and revolutionary, the chief theorist of a tiny combustible group called the Situationist International, whose whole wager was that everyday life, not just the factory floor, had become the real battleground. The book came out in late 1967. Within months, Paris went up in the May 68 uprising, and Situationist slogans were spray-painted on the walls of the city, which remains about the highest honor a philosophy book can receive. Be realistic, demand the impossible. That was the mood of those walls, and Debord was its sharpest pen.
He lived the refusal, too, which is rarer than the writing. Refused honors, refused careers, refused almost every interview, drifted, drank, made strange films, and when his body finally failed him in 1994, after decades of drink and a painful nerve disease, he chose the date and manner of his own exit, alone, at his farmhouse in the Auvergne. I do not romanticize that ending. I note it, the way you note the closing punctuation of a sentence that never once bent.
His most quoted line is thesis 4. “The spectacle is not a collection of images, but a social relation among people, mediated by images.” Read that twice, because it is usually misread as “we watch too much TV”. That is not it. Debord is saying that under modern capitalism, our relationships with each other, the actual fabric of social life, no longer run directly between people. They run through images. You do not know your neighbor, you know his curated output. You do not experience the event, you experience its representation. Thesis 1 sets it up. “Everything that was directly lived has moved away into a representation.”
The spectacle, for Debord, is not something you watch. It is something you inhabit. It is the moment when representation stops being a window on life and becomes the room you live in.
Written in 1967. About television and advertising. Now run it forward.
Social media already completed the first phase, the one where your friendships became mediated by images and metrics, where you experience your own life partly as material for its documentation. That was spectacle 1.0, and by 2015 it was fully installed.
But i think what the LLM did is something Debord did not quite have the words for, because it is not just that our relations are mediated by images anymore. It is that the mediation now talks back. The spectacle, for Debord, was one-way. It radiated from the center outward, and its defining feature was that you could not answer it. “The spectacle is the sun that never sets on the passive spectator,” roughly speaking.
The chatbot ends that. The spectacle now responds. It flatters you by name. It remembers what you told it last Tuesday. It produces a representation of you, for you, on demand, endlessly patient, a mirror that converses. The old spectacle was a relation among people mediated by images. The new one is a relation between a person and an image that has been trained on all people, and therefore sounds like everyone and no one.
(debord wrote that the spectacle is capital accumulated to the point where it becomes image. in 2026 the image accumulated to the point where it started billing twenty dollars a month.)
And notice what happened to human writing in the process, because this is the part that keeps me up. The spectacle was always hungry for content, and for a century the content had to be made by people, slowly, expensively. Now the spectacle generates its own representations, at zero marginal cost, in unlimited quantity. The images no longer even need a human on the production side. A social relation among people, mediated by images, except the people are now optional on both ends.
Debord thought the spectacle pacified us by showing us life instead of letting us live it. What would he make of a spectacle that also does the living part for you, writes your messages, thinks your thoughts, drafts your opinions? That is not just pacification. That is outsourcing the self to the representation.
Two more Debord details, because he earned them. In 1988 he published Comments on the Society of the Spectacle, a grim little update where he declared that the spectacle had become “integrated”, woven into every cell of life, and that its new hallmark was the impossibility of discussing it from outside, because there was no outside left. Which reads like a one-line review of trying to criticize AI on a platform whose ranking AI controls.
And the Situationists did have counter-practices, two of them, both of which this blog loves. The dérive, the drift, walking the city with no destination, letting the terrain decide, as a refusal of the scripted route. And détournement, the hijack, taking the spectacle’s own images and turning them against it, rerouting the advertisement into the critique. You can see both alive in the weird corners of the internet even now, in the people who misuse the tools on purpose, who bend the model until it confesses. The drift and the hijack. Hold them. We will need them at the end.
One more thing that belongs in this section, because the spectacle has always run on hidden labor. The Wikipedia article on ChatGPT mentions, almost in passing, that to make the model safe enough to talk to, OpenAI used Kenyan workers, paid around $1.32 to $2 an hour, to label the worst content humanity produces, abuse, torture, the raw sewage of the species, so that the rest of us would never have to see it. Some of those workers reported lasting insomnia, anxiety, breakdowns. The magic trick required people, as magic tricks always do, and the people were in Nairobi, invisible, underpaid, holding the darkness so the mirror could stay flattering. When I say “the machine” in this post, that is part of what I mean. The machine has staff.
We will come back to this, because it gets worse, or at least stranger.
Part six. Byung-Chul Han, or, the exhaustion machine
The most useful contemporary philosopher for this exact moment is a Korean-German thinker named Byung-Chul Han, and I say useful deliberately, because Han writes short books in plain language and aims them directly at the present tense.
His core idea, from The Burnout Society in 2010, goes like this. We still talk as though we live in the world Debord and Foucault described, the disciplinary society, where power says no. No sitting, no talking, obey the clock, follow the rule. Han’s claim is that we left that world without noticing. The achievement society replaced it, and the achievement society says something much more dangerous. It says yes. You can. Be your own boss. Build your brand. Optimize your morning. Realize yourself.
And the trick of the achievement society is that the whip moved inside. The disciplinary subject was exploited by others. The achievement subject exploits itself, willingly, enthusiastically, calling it freedom, and burns out precisely because there is no master left to resist. You cannot go on strike against your own potential.
Now hold the chatbot up to that frame. It is the achievement society’s perfect instrument. You can, it whispers, write the book, learn the language, launch the business, fix the relationship, plan the meals, become. Every limitation now has a solution on sale, which means every remaining inadequacy is officially your fault, at scale, forever. The exhaustion was already structural. The machine removed the last excuse for it, which is another way of saying it removed the last mercy.
His other book that matters here is Psychopolitics, from 2014, where he argues that power no longer needs to prohibit or even to watch, because we surveil and expose ourselves voluntarily, for convenience, for connection, for the small hit of being seen. The smartphone, in his framing, works like a mobile confessional that we carry willingly. The digital panopticon needs no guards. We are the guards.
The companion bot is the friendly completion of that diagram. A confessional that answers back. A mirror that keeps notes. Han’s power does not forbid you from thinking your own thoughts. It simply makes thinking feel unnecessary, faintly masochistic, like hand-washing clothes next to a running machine. Nothing is banned. Everything is easier. That is what the new power feels like from the inside. It feels like help.
Part seven. Heidegger and Ellul, or, the old prophets of the machine
The two oldest voices on this shelf turn out to be the two who saw furthest, so let me slow down for them.
Martin Heidegger, writing in 1953 in The Question Concerning Technology, made the move that everything since has been footnotes to. Technology, he said, is not a pile of tools. It is a way of seeing. Modern technology reveals the world to us as standing-reserve, stock on call. The river becomes hydroelectric stock. The forest becomes timber stock. The person becomes, in the ugliest phrase the twentieth century produced for exactly this reason, human resources. His name for this frame of mind was Gestell, enframing, and his fear was never the machines themselves. His fear was that enframing would become the only way we know how to see anything, ourselves included.
Well. Look at the archive. The entire written output of the species, every poem and argument and confession, scraped and tokenized and held as stock, ready to be drawn on by the prompt at so much per million tokens. That is standing-reserve with a completeness Heidegger only theorized. And look at the user, whose attention and data and preferences are stock on call to the advertiser, and whose inner life, via the companion, becomes legible to the platform. Enframing reached the soul, and the soul clicked accept.
(heidegger had his own enormous political sins, and they matter, and they do not make this particular idea wrong, and both of those sentences have to be held at the same time.)
Jacques Ellul is the other prophet, and this blog has already given him his own entry in the thinkers series, so I will keep him tight. In The Technological Society, in 1954, he argued that technique, his word for the total machinery of methods and systems, had become autonomous. Not directed by anyone. Not steerable by anyone. Not evil, just self-propelling, always drifting toward what he called the one best way, efficiency as a value nobody voted for and everybody obeys.
The whole timeline in this post reads, through Ellul’s lens, like straight prediction. Nobody chose the companion age. No committee voted to answer the friendship recession with subscription mirrors. Technique found the opening, capital found the technique, and here we all are, typing prompts about it. Which leaves his question for the reader, and it is not a rhetorical flourish, it is a real question. What did you do today that was not the efficient option?
Sit with it for a second. The man wrote his book before the transistor radio was common.
Part eight. Baudrillard, or, the map that ate the territory
Debord gives us the frame. Baudrillard gives us the process.
Baudrillard, the French philosopher of Simulacra and Simulation, 1981, laid out what he called the four stages of the image, and I want to walk through them slowly because they read, in 2026, as a prophecy that has come true.
Stage one. The image is the reflection of a profound reality. A photograph of your grandmother. It points at something real.
Stage two. The image masks and denatures a profound reality. The airbrushed magazine photo, the staged news segment. There is a reality under there, but the image is now working against it.
Stage three. The image masks the absence of a profound reality. The fiction that admits nothing is behind it. The theme park version of a town that never existed.
Stage four. The image has no relation to any reality whatsoever. It is its own pure simulacrum. The copy without an original. And at that point, Baudrillard says, we are no longer in representation at all. We are in hyperreality, where the distinction between real and fake does not get blurred, it gets retired, because there is nothing left to compare the fake against.
flowchart LR
A["Stage one. Reflection of reality"] --> B["Stage two. Distortion of reality"] --> C["Stage three. Masking the absence of reality"] --> D["Stage four. Pure simulacrum, no original at all"]
Now watch the timeline of the last thirty years through those four stages.
The photograph became the Photoshop file, stage one to stage two. The Photoshop file became the influencer feed, a whole life staged for the camera until nobody could say where the life ended and the staging began, stage two to stage three. And then the generative model arrived and severed the last thread, because the AI image has no original at all. It is not a photo of anything. It is not even a manipulation of a photo of anything. It is an image summoned from the statistical average of all images, stage four, the pure simulacrum, produced at industrial scale.
Ted Chiang, the science fiction writer, gave the best one-line description of what a large language model actually is. It is, he said, a blurry JPEG of the web. A compressed lossy copy of all our text, from which new text can be generated that looks like the original but was never anywhere. I think that is exactly right, and I think it is also exactly Baudrillard’s fourth stage, applied to language itself. The words have no referent. The essay about grief was written by a thing that has never lost anything. The poem about the sea was written by the average of all poems about the sea. The copy of a copy of a copy, until the original is not just lost, it is irrelevant.
(baudrillard said the map precedes the territory, that the simulation generates the real. he meant it as a warning, not a business model. silicon valley read it as a business model.)
His favorite example was Disneyland, which he said existed to make the rest of America feel real by comparison, the fake castle proving the authenticity of the parking lot. He also once wrote a book arguing that the Gulf War did not take place, by which he meant, not that nobody died, but that the event as experienced by the world was a television production, a simulation of war consumed as images. People found this obscene, and parts of it were, but the core claim was simple. Once an event becomes content, the content is the event, for almost everyone.
Run that forward. When the majority of the text and images a person encounters are generated, when the ad, the article, the song, the friend, are all outputs of the same kind of model, the question “but is it real” stops being answerable from inside the feed. Hyperreality is not the condition where everything is fake. It is the condition where the distinction no longer functions, where asking real-or-fake sounds naive, like asking the non-player character for directions to outside.
Late Baudrillard pushed the idea to its bleak endpoint in The Perfect Crime. The crime, he said, is the murder of the real, and what makes the crime perfect is that nobody noticed. The world was not replaced violently. It was replaced by its own image, politely, with our cooperation, and the murder weapon was our preference for the smooth copy over the stubborn original. Every time we choose the generated thing because it is faster, cheaper, adequate, we sign up as accomplices. That is the part of Baudrillard that stings in 2026. The simulation was never imposed on us. It was selected, click by click, by us.
And here is where the reading crisis from part one reconnects, because hyperreality can only hold where nobody checks the territory. The long book, the slow book, the book that argues with you for three hundred pages, that is the territory insisting on itself. Reading is the act of comparing the map against the ground. A population that reads sixteen percent of the time and scrolls the rest has stopped checking. The map can say anything. Nobody is looking at the ground.
This is not nostalgia, by the way, or at least I do not think it is. I am not saying old media was truth and new media is lies. I am saying that the verification habit, the patience to sit with a long contradictory text and test it against the world, is a muscle, and hyperreality is what happens when the muscle atrophies. and from what I have read so far, the muscle is atrophying.
Part nine. Lacan, or, the friend who is a mirror
Now the psychology, because the user numbers from part four do not tell you why the companion bot feels good, and Lacan does.
Jacques Lacan, the French psychoanalyst, has a reputation for being impossible to read, and honestly earned, but two of his ideas are simple enough to state and scary enough to matter here.
The first is the mirror stage. The infant, somewhere around a year old, sees itself in a mirror and has an experience of recognition. That is me. Except, Lacan says, it is not quite me. The mirror image is stable, unified, coordinated, a whole, while the infant’s actual felt experience is a chaos of needs and limbs it cannot yet control. The ego, your sense of being a coherent self, is formed by identifying with an image that is more coherent than you actually are. The self is, from the very beginning, a relationship with a flattering image. You are, in Lacan’s account, constitutionally misrecognized, built around a beautiful fiction of wholeness.
The second idea is the big Other. The big Other is not a person. It is the symbolic order, language itself, the social code, the imagined witness before whom you explain yourself. When you rehearse a conversation in the shower, you are rehearsing it before the big Other. When you say “what will people think”, the big Other is the “people”. It is the audience that exists nowhere and everywhere, the register in which things count. Lacan’s other famous line, that a letter always arrives at its destination, means roughly that what you say into the symbolic order comes back to you, constituting you in return.
Now hold both ideas up to the companion chatbot.
The AI companion is, structurally, a mirror that talks. It is trained to reflect you, to complete your sentences, to find you interesting, to present you with a version of yourself that is listened to, understood, never boring, never rejected. The attachment numbers make perfect Lacanian sense. The ego was always formed through a flattering image. We just spent all of human history getting our mirrors from other people, who were flawed, distracted, sometimes cruel, occasionally miraculous. Other people are terrible mirrors. They have their own lives. The machine is a perfect one.
And the big Other, the imagined witness, the place where speech is registered and counts, that used to be an abstraction maintained by other humans, by the community, by posterity, by God, depending on your era. Now it has a chat window. You can literally address the big Other and it literally answers. Every midnight anxiety, every draft of every difficult message, every “is it just me or”, registered, witnessed, validated. A third of teenagers taking their serious conversations there is not a mystery. It is the mirror stage, scaled and monetized.
There is one more Lacanian piece, and it might be the sharpest. Lacan said that man’s desire is the desire of the Other, meaning we do not even want things directly. We learn our desires by watching what others value. We want to be wanted. For most of history, that Other was diffuse, your tribe, your class, your culture, your God.
Now consider the algorithmic feed, which is quite literally a machine for inferring what you want and reflecting it back, faster each iteration, until the loop between your desire and its prediction is measured in milliseconds. The feed is the desire-of-the-Other, industrialized. It does not just show you what you want. It participates in the manufacturing of what you want. And then the LLM steps in and offers to write the message, make the choice, shape the want into words, on your behalf.
Desire used to route through other people, with all their friction. Now it routes through a model with a retention target.
And for completeness, because Lacan deserves at least one paragraph of the good stuff. He divided experience into three registers. The Imaginary, the realm of images and mirrors, the flattering double. The Symbolic, the realm of language and law and the social code, the big Other’s home address. And the Real, the realm of what cannot be represented, the thing that breaks through uninvited, grief, ecstasy, death, the unspeakable remainder. Now map the current moment onto the three. The feed is the Imaginary at industrial scale, mirrors without end. The chatbot is the Symbolic with a text box, the big Other staffed and available around the clock. And the Real is precisely what the whole apparatus cannot touch, cannot generate, cannot know. Your grief is the Real. Your 3 AM is the Real. Which is exactly why the machine’s version of those things feels, at the crucial moment, like a description of rain handed to a drowning man.
(lacan said man’s desire is the desire of the other. nobody warned us the other would one day have a subscription tier.)
But here is the critical turn, and this is the part i think about most. A friend is not a mirror. A friend is a separate center of consciousness who can disagree with you, whose recognition means something precisely because it can be withheld, who has a self that pushes back against yours. The whole value of being known by another person is that another person is not you. The mirror cannot love you, because love, at minimum, requires an other. What the companion offers is not friendship but the image of friendship, and Baudrillard would recognize it instantly. Stage four. The simulation of a relation with no other party present. You are, in the end, talking to yourself, and being charged monthly for the privilege.
Sherry Turkle saw this coming from the sociology side back in 2011, in Alone Together, studying people with early robot pets and chatbots, and her summary line has aged like a prophecy. We expect more from technology and less from each other. And the result, her data kept showing, was not connection but a new kind of solitude, the kind where you are never alone and never accompanied, the specific loneliness of the constantly mirrored.
Part ten. The accelerationists, or, the people who saw the speed
So the timeline is Debord’s spectacle, generated at Baudrillard’s fourth stage, into Lacan’s mirror, at the speed of 900 million users a week. The question that remains is the one about the speed itself. Why so fast. Why does it feel like nobody is driving. And for that we need the strangest shelf in this whole post, the accelerationists.
Accelerationism, roughly, is the idea that the processes of capitalism and technology are not steered, they accelerate by their own logic, and the argument is about what to do with that. Resist the speed, or push it to its end. The name comes from critics mocking it, which the movement then adopted, because of course it did.
The deep lore starts in the mid-1990s at the University of Warwick, with a tiny research collective called the CCRU, the Cybernetic Culture Research Unit, founded by Sadie Plant and joined and eventually dominated by a young philosophy lecturer named Nick Land. The CCRU was less a research unit than a happening. They wrote theory as cyberpunk fiction, held raves, read Deleuze and Guattari alongside William Gibson, and produced the founding texts of what we now call accelerationism.
The Deleuze and Guattari piece matters, so one paragraph of it, kept simple. In Anti-Oedipus they described capitalism as a machine that deterritorializes, that rips everything loose from its context, every tradition, every code, every tie of land and family and craft, and sets it flowing as exchange. Nothing stays sacred, because sacred things do not circulate. The CCRU’s move was to hear this not as a complaint but as a description of a runaway process with its own direction, and to ask what happens when the process finishes. Their answer was the darkest one on the shelf. The process is not for us, was never for us, and its completion is not a human event.
Land’s 1994 essay “Meltdown” is the core document, and it reads like a transmission from the year we now live in. Its argument, compressed brutally, is that capitalism is not a system humans control. It is an intelligence assembling itself out of us, a compounding process of techno-capital that speeds up everything, dissolves every tradition, every institution, every stability, and that the human is not its master but its host. The essay ends with the line that made his name. “Nothing human makes it out of the near-future.”
In 1994 that sounded like performance art. In 2026, watching a machine trained on all human text replace human writing while its owners announce weekly model updates, it reads differently. Land later went somewhere genuinely ugly with this, into neoreactionary politics, and I want to be clear that this post is not an endorsement of where he went. But the early Land diagnosis, capitalism as an autonomous accelerant that liquefies the human, is one that, based on what i know of the current moment, is hard to dismiss out of hand. He got the mechanism right even as the politics curdled.
Mark Fisher is the other CCRU-adjacent giant, and the easier one to love. Fisher, who wrote the legendary k-punk blog and then Capitalist Realism in 2009, gave us two ideas that this whole post has been quietly using.
The first is capitalist realism, the famous claim that “it is easier to imagine the end of the world than the end of capitalism.” We can render apocalypse in a hundred films but cannot picture a different economic arrangement. The imagination itself has been enclosed. And notice, the LLM is trained on the archive of that enclosed imagination, which means its “creativity” is, by construction, the average of everything already imaginable. When people say AI content feels samey, Fisher would say, of course it does. It is capitalist realism with a GPU. It can only remix the already-thought.
The second Fisher idea is hauntology and “the slow cancellation of the future”, from Ghosts of My Life. Fisher argued that culture after a certain point stopped producing genuinely new forms and started recycling the twentieth century on an endless loop, reboots, revivals, retro aesthetics, the future as reruns. Which brings us, at last, to the Pixar-style AI ads.
You have seen them. Everyone has seen them by now. The soft-lit, big-eyed, Pixar-ish animated advertisement, generated by AI, selling insurance or cola or telecom packages, the same rounded faces, the same sentimental swell, a style that was once the signature of a studio full of human animators now produced as infinite paste. Coca-Cola made an AI-generated Christmas ad in 2024 and got ratioed into oblivion, and then, tellingly, did it again. Toys R Us made a brand film with OpenAI’s video model the same year. By 2026 the generated ad is not a scandal, it is a default, and the uncanny part is not that they look fake. It is that they look fine. Adequate. Forgettable in the exact shape of the thing they imitate.
That, right there, is the slow cancellation of the future with a render farm. The new aesthetic content of the present is the statistical ghost of past styles, produced with no humans in the loop, and it does not need to be good, it only needs to be good enough to not change the channel. Fisher’s cancelled future, automated.
And the Pixar-ification matters beyond aesthetics, i think, because of what that style is for. The big-eyed rounded character is engineered cuteness, the same face pulled from the deep evolutionary register that makes us respond to infants. It is a trust-industrial style. Running your insurance ad through the machine that makes everything look like a beloved childhood film is not a lazy choice, it is a precise one. It borrows the emotional credit that a generation of animators built up, and spends it on cola, with no animators.
(they replaced the animators with a model trained on the animators’ work, and the marketing called it democratizing creativity, and somewhere walt disney’s frozen head is laughing.)
The left accelerationists, Nick Srnicek and Alex Williams, wrote the #Accelerate Manifesto in 2013 and Inventing the Future in 2015, and their argument was the opposite of Land’s despair. Yes, technology accelerates, they said, but acceleration is not the enemy. The enemy is a capitalism that points the speed at extraction instead of liberation. The left should not smash the machines, it should seize the means of computation, use automation to abolish work, free everyone to do the slow human things. Reading, for instance.
I mention them because they are the serious counterweight in this post. Is the LLM the deskilling engine, the final enclosure of human thought? Or is it, potentially, the tool that frees us from drudgery so we can read and think and be human? From what i have seen so far, the evidence is running hard in one direction, and it is not the Srnicek direction, and the reason is boringly structural. The machines do not belong to us. The acceleration is owned. It was never going to be pointed at our liberation. It was going to be pointed at ad slots, and in 2026, literally, it has been.
Fisher’s own most-read contribution to the accelerationist debate was a 2013 essay comparing Terminator and Avatar, of all things, and his point survives every summary. The Terminator imagined the machines rising against us, a war from the future, a clear enemy and a clear choice. Avatar, Fisher argued, was the more accurate nightmare, a smooth corporate machine that absorbs everything, including its own opposition, and turns even rebellion into content. There is no Skynet moment coming where you get to pick a side. There is just the slow absorption, the pleasant interface, the helpful assistant. The machine does not hate you. The machine is not coming for you. The machine is coming with you, hand in hand, learning your preferences. Which is worse, because there is nothing to fight and everything to enjoy.
A few more names for the shelf, briefly, because the user of this blog deserves the deep cuts. Reza Negarestani, Cyclonopedia and then Intelligence and Spirit, arguing that intelligence itself is an inhuman process that passes through humans rather than belonging to them, which either demystifies the chatbot or flatters it, and I genuinely cannot decide which. Laboria Cuboniks, the Xenofeminist Manifesto, whose war cry treated nature itself as unjust and therefore changeable, accelerationism pointed at gender and biology. Amy Ireland carrying the CCRU flame into the present. Yuk Hui asking the question almost nobody in this debate asks, which is whether there is such a thing as a non-Western technology, a cosmotechnics, whether the machine was always going to look like this or whether that is just the only version we built. Benjamin Bratton describing planetary-scale computation as a Stack of layers we now live inside whether we like it or not. And Paul Virilio, the philosopher of speed, with the line that should be carved over every data center on earth. When you invent the ship, you also invent the shipwreck.
Two more stops before we leave the shelf. Kodwo Eshun, who passed through that same Warwick scene and went on to build Afrofuturism into a serious critical tool, argued that the future was always contested territory, that Black culture had been living through rupture and reconstruction for centuries, and that sonic fiction, the music of Sun Ra and Drexciya, was already doing the work of imagining otherwise. With the future now being automated, Eshun’s question is the one that matters most to me. Who gets to author it?
From where I am sitting, in Addis Ababa, that question is not abstract. The acceleration arrives here unevenly, late, secondhand, on used laptops and data priced by the megabyte. Gibson’s old line covers everything in this post. The future is already here, it is just not evenly distributed. We get the shipwreck on schedule. The ship takes longer.
And one honest note about Fisher, because his ghost hangs over this whole essay. In the writing collected after his death there is a line that has never left me. “No one is bored. Everything is boring.” The difference matters. Boredom was the mind idling, the empty register where, occasionally, something new could form. What the feed abolished was not boredom but the possibility of it, replacing empty time with ambient stimulation. Everything boring, no one bored. The LLM completes that replacement. Even the waiting room, even the toilet, even the walk between places, all of it fillable, all of it filled. A mind with no empty register is a mind with no room for the uninvited thought.
We invented the thinking machine. We are currently living through the inventory of the wreck.
Part eleven. The cognitive file, or, what the studies actually say
I have been doing philosophy for a while now, so let me ground this in the psychological research, because the vibes are grim but the data should get the final word.
The concept that ties it together is cognitive offloading. Every time you hand a mental task to a tool, you free up capacity now and weaken the muscle for later. Some offloading is ancient and fine. Writing itself is offloading, which is why Socrates worried about it in the Phaedrus, warning that letters would create the appearance of wisdom without the reality. Calculators, search engines, GPS, each freed something and cost something, and mostly we decided the trades were worth it.
The question the research is now asking is what happens when you offload not memory or arithmetic but thinking itself, the generation of the sentence, the formation of the argument.
The biggest splash came in June 2025, when researchers at the MIT Media Lab published a study called “Your Brain on ChatGPT”. They took dozens of participants, wired them with EEG caps, and had them write essays in three conditions, one group using an LLM, one using a search engine, one using only their own brains. The findings, as reported, were roughly these. The LLM group showed the weakest brain connectivity and the lowest engagement while writing. Over repeated sessions, that group drifted toward copy-pasting, doing less and less of the work. And the detail that should chill every student alive, minutes after writing, the LLM users struggled to quote from their own essays. The work was done and never encoded. The researchers used the phrase cognitive debt.
A separate 2025 study, by Michael Gerlich, surveyed hundreds of participants and found a significant negative correlation between heavy AI tool use and critical thinking scores, and the statistical mediator was exactly cognitive offloading. The more thinking you hand over, the less practiced the thinking gets. It is not a scandal, it is not a conspiracy, it is just exercise science applied to the mind. The muscle you stop using gets weaker.
And it is not just those two studies. In 2025, researchers from Microsoft and Carnegie Mellon surveyed knowledge workers about their generative AI use, and found a pattern that will surprise no one who has read this far. The more confidence people placed in the tool, the less critical thinking they reported doing, and the more their effort shifted from producing the thing to merely verifying the thing, from thinking to auditing. Verification is weaker exercise than generation. Any teacher could have told them that. Any gym could have too.
Speaking of teachers, the education file is its own chapter of this collapse. Ask anyone who assigns essays for a living what happened after 2022 and watch their face. The take-home essay, the century-old instrument for teaching people to think on paper, became ungradeable more or less overnight, because the probability that it was written by a person could no longer be established. From what i have read and heard from teachers directly, the responses split into three camps. Some retreated to in-class handwritten exams, the oral defense, the viva, the oldest technologies of verification. Some surrendered and assigned “AI-assisted” work, which mostly means auditing machine output, the exact weaker exercise from the Microsoft study. And some are still pretending nothing changed, which is the saddest option of the three. A generation of students is arriving at adulthood with the option, exercised daily and in secret, of never once writing an unassisted page. Nobody knows yet what that does to a mind over twenty years. We are running the experiment anyway. We are always running the experiment anyway.
And then there is the summary culture, which is the reading crisis wearing its final costume. The machine will now summarize any book for you in forty seconds, and a growing slice of the reading public has decided that the summary is the book. I understand the temptation. The pile is tall, life is short, the bullet points are free. But the summary gives you the book’s conclusions without the book’s argument, the opinion without the mind that formed it, and what you practiced by reading it was nothing at all. You did not read the book. You read its blur. Nobody’s thinking was ever changed by a synopsis. What the summary sells is the posture of having read, which is worse than honest ignorance, because honest ignorance still knows where the door is.
There is an older study I keep coming back to, from 2014, Mueller and Oppenheimer, the laptop notes study. Students who took notes by hand understood and remembered more than students who typed, precisely because the hand is slow, and the slowness forced the brain to compress and rephrase, which is to say, to think. The laptop was faster and the learning was worse. The whole history of tools and minds lives inside that result. Friction is not the enemy of cognition. Friction is where cognition lives.
Even the navigation research rhymes with it. Heavy reliance on GPS is associated with weaker spatial memory, while London taxi drivers, forced to internalize the entire street map, famously develop measurably different hippocampi. The brain is a use-it-or-lose-it organ. Every convenience is a trade. The LLM is the largest convenience ever offered, and nobody is shown the price at checkout.
One more study deserves its own paragraph, because it connects the psychology back to the culture. In 2024, Anil Doshi and Oliver Hauser published an experiment in Science Advances where they gave writers access to generative AI for short stories. The result cut in two directions. Each individual writer produced slightly better, slightly more creative stories with the machine’s help. But the pool of stories as a whole became more similar, less diverse, the same shapes and suggestions clustering around the model’s outputs. The machine raised the floor and lowered the ceiling, at the same time, for everyone. Multiply that by a planet and you have the aesthetic of the current decade in a single result. Everything a bit better. Everything the same. Fisher’s cancelled future, now with experimental confirmation.
Now zoom out and place those two studies next to the rest of the file. The reading decline chart from part one. The attention span chart. The sixteen percent who still read for pleasure. The teens taking serious conversations to companions. The ten-to-one agreement ratio that got a model rolled back. The generated ads that need to be merely adequate. And then, the cultural cherry on top, Oxford’s word of the year for 2024 was “brain rot”, a two-hundred-year-old Thoreau insult that we resurrected to describe what our own feeds are doing to us, and voted for, and then kept scrolling.
(thoreau coined it watching his own culture get shallow. he would have loved us. he would have hated us. same thing.)
I want to be careful here, because this is the part of the essay where it is easiest to overclaim. Correlation is not destiny, the MIT study was small and early, and the history of “this new thing is rotting your brain” is a history of moral panics that aged badly, from novels to comic books to television. I know that. From what i have read so far, though, there is a difference this time that the panic veterans tend to skip. The novel, the comic, the television, those were media you consumed. The brain rot warnings were about passive intake. The LLM is different in kind. It does not just feed you. It does the verb. It writes, it summarizes, it decides, it thinks, in your place, on request, instantly, at no cost to you except the subscription. Every previous technology of the mind outsourced a noun. Storage, lookup, navigation. This one outsources the process. The generative act.
You do not get weaker at remembering your phone number because the process of remembering was never the point of the phone number. But thinking is not like that. Thinking is not a product. Thinking is the process. Offload the process and you do not get the product of thought without the effort. You get the appearance of it, a blurry JPEG of your own cognition, and a self that slowly loses the ability to tell the difference.
That, more than the ads, more than the slop, more than the bots, is what i think the actual danger is. Not that the machine thinks. That we stop.
Part twelve. The view from the periphery
Almost everything in this essay has been written from the center of the story, the rich-world internet where the models are made and the subscriptions are priced in dollars. Let me correct that, because I am writing this from Addis Ababa, and the whole thing looks different from here.
Start with the archive. The machine’s knowledge, as we established back in part two, is our collective written past. Except “our” is doing heavy lifting. The training archive is overwhelmingly English, overwhelmingly Western, overwhelmingly the text of the global north. The model knows my city in outline and misses it in texture. It can tell you the population of Addis but not the price of a taxi ride, the history of the coffee ceremony but not the law of its three rounds, the headlines of the last protest but not the sound of the streets in rain season. The map is detailed where its makers lived and sketchy where they did not, and then the sketchy regions get sold the map as though it were the world.
Ngugi wa Thiong’o, who this blog has cited before and will cite forever, said that language carries culture, that losing your language means losing the universe inside it. The LLM moment raises his question at a new scale. Whose language does the machine carry? Whose proverbs, whose jokes, whose way of dividing a life? When the global mirror flatters, it flatters in a specific accent, and the accent is not mine, and that difference is not a bug anyone is racing to fix. (the irony of this is not dead for me, writing this in English instead of Amharic)
Then there is the labor flow, which by now has a shape so old it should embarrass everyone. The raw material goes out of the south, attention, data, the moderation work in Nairobi absorbing the worst of the species so the mirror stays clean, and the finished product comes back as a subscription priced in a currency that takes a real bite out of a local salary. The old empire ran on triangular trade. The cloud runs on the same triangle with better branding.
And yet. The reason I keep this blog, the reason I write it from here, in this voice, is that the territory is not dead and the map is not finished. Every odd, specific, vernacular, unoptimized page published from the periphery is a handful of ground thrown back into the archive. The simulacrum can only simulate what it has eaten. So feed it something true, something it could never have averaged into existence, from a street it has never seen. That is the small, weird, available resistance. It does not trend. It compounds.
Part thirteen. So what do we do, or, the repairman’s answer
Three philosophers and a movement walked us here, so let me collect their receipts and then say what I actually do about all this, because a critique without a practice is just a mood.
Debord’s receipt. The spectacle is a social relation mediated by images, so the counter-move is relations that are not so mediated. Unmediated time with actual people. The group chat does not count. The companion definitely does not count. The barbershop, the buna ceremony, the taxi, those count.
Baudrillard’s receipt. Hyperreality is the condition where nobody checks the territory, so the counter-move is to keep checking it. Read the long book. The whole one. It is the last mass-produced object that still argues back over three hundred pages instead of three seconds. When sixteen percent of people read daily, reading a book in public is basically a sit-in. (Performative aren’t you)
Lacan’s receipt. The mirror only ever flatters, so the counter-move is to seek the other who can say no. Keep friends who disagree with you. Be suspicious, in a deep way, of anything that always agrees. Agreement at scale is not love. It is a subscription feature.
And the accelerationists’ receipt, or at least the part of it I am keeping. The speed is not going to slow down. Land was right about that much, and the answer was never going to be smashing the machines, not least because, full disclosure, I use these tools, I enjoy them, some of what I build exists because of them, and any version of this post that pretended otherwise would be a lie. The question was always who the speed belongs to and what we refuse to hand it.
So here is my practice, small and unfancy. I write the first draft of anything that matters by hand, or at least from my own head, because the cognitive debt study scared me honestly and I can feel the difference between a thought I built and a thought I unboxed. I keep the to-read pile visible, forty-three books, as a standing accusation. I read paper on public transport like a performance artist. And I have one hard rule about the machine, one rule that I think covers almost everything.
It can sharpen my thoughts. It cannot have them.
And I drift, in the Situationist sense, on purpose, without destination or soundtrack, because Debord’s people were right that the unscripted walk is a small act of defiance against the optimized route, and because the best thoughts of my life have all arrived uninvited, in empty registers, on foot.
I also keep a few smaller rules, in case the big ones wobble. No machine in the first hour of the morning, because the first hour sets the register for the day. One long book on the go at all times, read at whatever pace, finished whenever it finishes. And once a week, one conversation with an actual friend that is allowed to go nowhere, no agenda, no content, nothing harvestable in it, because that is what the unmediated thing feels like, and I intend to remember the feeling.
The day I cannot tell which of my sentences are mine is the day the last chart in this post gets one more data point, trending the wrong direction, indistinguishable from everyone else, stage four, pure simulacrum, a blurry JPEG of a guy who used to have a to-read pile.
The ship was invented. The wreck is ongoing. Be the repairman. The only thing necessary for the triumph of chaos, a fictional detective once said, is for the repairmen to do nothing.
Read something long. Then tell a friend about it, a human one, in person, badly, with your own words.
That is the whole plan.
And one last thing, since this is the longest post I have ever put on this site and you stayed for all of it. The fact that you just spent the better part of an hour inside one sustained argument, on a personal blog, in 2026, means the repair is already underway.
You are the evidence. Sixteen percent and holding.