(This is not a heavy-handed allegory for contemporary attitudes to AI. It is a light-handed one.)

Steve surfaces from the coma the way a diver surfaces — slowly, then all at once, three years of his life left somewhere below him in the dark. His parents, his sister, his partner, all by his side, all smiling ear to ear, all murmuring encouragement, sotto voce, coaxing him through this final step back towards consciousness, towards living. Really living.

After the grogginess fades and the world regains its edges, Steve asks the obvious question: “So, what happened while I was away? Fill me in.”

“Well,” begins Steve’s mother, consulting her hand to keep a quick tally of updates, “of course I’m a bit older these days, we all are.” She smiles. “I visited you every Thursday after work; your sister did every other Tuesday, as she lives further away; your father — Thursday’s still his poker night — visited too, sometimes. And your partner, she did Sunday mornings. We all did our part. We held your hand, we read to you, we tried to talk to you. We hope it helped…”

For Steve, there’s nothing between now and the accident. Three years, spent in an instant. He forces a smile. “I’m sure it did.”

“… Anyway, outside of looking out for you, let’s think what’s happened. Well, I guess the main news:” She pauses for effect. “You’re now an uncle!”

“And a cat daddy, thrice over,” Steve’s partner says, catching his eye. “It was lonely while you were away.”

“Yes,” continues Steve’s Mum, “you’re a cat daddy, and an uncle! Beth had a beautiful, bouncy baby boy, in this same hospital as it happens, who’s now coming up for his second birthday.”

“Terrible twos?” asks Steve.

“First 18 months were toughest, to be honest,” Beth replies. “Now he just thinks he’s the centre of the world.” She pauses. “He’s a sweet boy, really. He’s got your grandma’s red hair.”

The medical procession follows: the nurse, the nurse practitioner, the consultant neurologist trailing two trainees — vital signs collected, questions asked, Steve regarded carefully, critically, sympathetically, as both object and subject. One nurse confesses she isn’t used to seeing him conscious. He’s warned that, though the worst is behind him, there may still be things to watch out for: his memory might not be what it was; a small but real risk of visual or auditory distortions, bordering on mild hallucinations; perhaps even occasional seizures, in which case he might have to hand in his driving licence. All these things are rare complications, the neurologist says, but worth being aware of.

Over the course of the day, Steve fades from the spotlight. His family remains by his side, explaining that, as with the coma vigil, they’ve made plans to visit frequently and support himand his partner on his return in a couple of days. Conversation slows, energy falls, smiles — though still genuine — lose their radiance. One by one, all but Steve retreat into the curated bubbles of their phones. Music clips get played, as anodyne and forgettable as Steve remembers from the days before his void. Various names of politicians get mentioned, some familiar and some new; from the tenor of the pablum, Steve quickly learns which figures are now good, and which are bad.

Breaking the silence, Beth says to Steve’s mum, “Oh, you know Alice from my work? I went to her wedding last week.”

“Aww. How was it?”

“Good venue. Roast potatoes a bit hard, Yorkshire’s a bit deflated. But love-er-ly dress!”

“Aww…”

“Best man’s speech sounded demonwritten. Suit was a bit creased. Definitely good friends. Couldn’t decide which side of the lovable/rogue boundary he was on…”

“That’s a shame,” says Steve’s mum. “What were the desserts like?”

Steve wonders if he’s now experiencing an auditory hallucination.

“Sorry,” he interjects. “Demonwritten?”

“Yeah,” says Beth. “As in… written by a demon, rather than a person.”

“Bloody demonslop,” mutters Steve’s mum.

“Sorry,” Steve says again. “Written by a demon?”

“Yeah,” says Beth. “It just sounded really demonny. Like the guy just logged into demontalk and put something in the chatbox: ‘I’m going to be a best man at a wedding on Wednesday. Write a Best Man’s Speech for me.’ And it did.”

“And what did?”asks Steve.

“Some demon, of course,” replies Beth, trying not to sound too exasperated at her possibly brain-damaged and recently comatose brother. Words fail Steve. “I… what??… ahh… so… so… demons exist?”

“Yes,” replies Beth. “Demons exist. But they’re not interesting.”

“Bloody annoying, they are,” mutters Steve’s Mum.

“Demons exist,” manages Steve. “Since when?”

“Ah, of course,” exclaims Beth. “It’s all happened while you were… away.”

“Hopefully they’ll clear off again before too long too,” mutters Steve’s Mum.

Beth slows her cadence. “I think it all began with some old book. Isn’t that right, Dad?”

“Erm, yes,” replies Steve’s Dad. “I think so. There was an old university library book. An old professor — folklore or philology or something like that — had this idea that all his colleagues thought was a bit… eccentric.”

“Eccentric even for academics,” adds Steve’s Mum.

“Yes, very eccentric. He thought that, maybe, some of the old sound recipes, spells, from books long ago, might actually work, if only we knew how to say them right.” Steve’s Dad continues. “You see, like the Great Vowel Shift, there’s a lot of stuff that’s written down, that used to be said, but that if you try to say the words now, it’ll sound different. And this odd fellow thought, maybe, if you could look at a whole range of texts from when a spell was written — especially poetry, with its rhyme and meter — then maybe you could recover these old dead accents and sounds. And that would allow you to say the spells the way they’re meant to be sounded.”

“tl;dr: he was right, sadly,” sighs Beth.

“Yes. He found one codex, in particular, that had all kinds of information on how things sounded back in fifteenth-century Florence. A rhyming dictionary, as it were, so complete, and so neatly written, it looked just too comprehensive, and too repetitive, and too perfect, to have been scribed by any human hand.”

“And it wasn’t.”

“And at the end it had some suggested spells that, if said correctly, would summon forth a foot-long, cobalt-blue demon that would materialise for a few seconds, then disappear in a puff of sulphur.”

“And he did, and it did.”

Steve’s Dad clarifies: “It wasn’t straightforward. Learning the codex took years. Pretty much no one took the premise seriously. The guy had about a half dozen lab assistants and PhD students, all committed, all working silly hours…”

“All millionaires now.”

“… but most people thought it was just a waste of time and public funds. They had to improvise and adapt materials usually bought by New Age folk, not academics.”

“Bloody Big Joss..”

“Yes. Joss stick manufacturers and publishers of arcane medieval poetry now comprise about a third of the US stock market— and about 100% of the growth in valuations.”

“It’s just a big sloppy bubble.”

“But it worked, after a fashion. After they’d figured out how to conjure the demons, the next step was to keep them here a bit longer. That’s where the incense came in. They managed to get the time in this realm — the TIR metric, they call it now — from ten seconds to ten minutes, to ten hours, to ten weeks, and now ten months.”

“Wrong way.”

Steve’s Dad notices Steve’s mouth is open. “So,” asks Steve, “were any of those nurses or doctors… demons?”

Steve’s Dad laughs. “Of course not. They’re still small little things. And they’re not really quite here. They’re… immaterial. Stuck somewhere between here and there. All they could do initially was talk and listen. And initially they didn’t even talk in a human language. They spoke in some kind of demon language — sounds no human could make or comprehend, which register to us like shrieks and static, and which tend to put most animals on edge. But they quickly learned how to imitate our sounds, and when they saw our writing they could write in a way much like we write.”

“They’ve still not got it right. I can tell demontext a mile off.”

“How could they write, if they can’t touch anything?” asks Steve.

“Well, it turns out that though demons might not be corporeal, they can still move electrons, meaning they could channel their influence into computers. And, by not having to move any physical matter other than electrons, it turns out they can read and write computer streams really quickly. Thousands of times faster than any human.”

Steve’s Dad continues. “Initially, after they’d learned to access the electronstream, it was clear they had a lot of fluency, producing massive amounts of text that followed the grammar and structure of human languages. This got a lot of people very excited…”

“Not me.”

“…until it became clear a lot of the text was nonsense in terms of semantics. Ask them to produce a recipe for a smoothie, and they’d casually suggest putting glass in it. Ask them to draft a will, and it would mostly look fine — until you read the fine print, where inevitably they’d bury some clauses promising the deceased’s eternal soul to Baphomet, amongst the usual legalese.”

“Some people still preferred this to dealing with a real solicitor!”

“Some people, definitely. But for a lot of people, initially impressed by their fluency, these ‘glitches’ put them off demontext services for ever. And that’s still how most people think about them: fluent, but superficial and glitchy and inaccurate.” Steve’s Dad continues. “But over the next year or so, these issues largely disappeared. They taught the demons to search through electronic archives, imbibing the ever-increasing corpus of human knowledge the demontalk megacorps have been making accessible to them.”

“So many trillions spent, so many fields given over to spices rather than staple crops. And it’s still unreliable nonsense.”

“It seems quite a lot of coders disagree. As it turns out, demons are now… demons at coding. Part of this is just the speed at which they write and think: our reality seems glacial compared to theirs, though they don’t experience time in the linear, continuous way we do: the hundreds of years since the foot-long cobalt-blue demon was last asked to appear passed, apparently, as a single moment. But part of this is because, with code, there’s usually a clear way of verifying— of being confident whether it does the job or not. And once the demons got inside the electronic machines, could check the error messages and behaviours, and could rewrite their earlier attempts in response, they suddenly turned out to be really good at making lots of code that works — and works together.”

“Well that’s just coding. Coders are barely human anyway.”

“It definitely takes a distinct mindset to write code, and write code well. But it holds promise for other fields that are based, fundamentally, on the need to remember a lot of stuff, and to work logically and sequentially, one rune at a time. I saw something recently that, at some of the top demontalk companies, the demons are writing most of the code. They’re also advising, increasingly, on business decisions.”

A pause. No riposte. Steve asks, “And what of the guy who found the codex, who first summoned the cobalt-blue demon?”

“Oh,” replies Steve’s Dad. “Professor Sir Ignatius Haunting. What a name!”

“Preposterous name,” says Steve’s partner. “Sounds made up.”

“He’s on the podcast circuit now. Most people call him ‘Iggy’, though he winces slightly at the demotic.”

“Podcasts?” asks Steve. “So, telling people how great demons are, how smart he and his teams are?”

“Not quite. These days he’s a doomster, telling anyone who’ll listen that unleashing so many demons all at once into our world might not be a good idea— might even be a sign of the End Times for our species. Weirdly, he combines two registers in these talks: on the one hand, lavishing praise on the exponentially increasing capabilities of the demons, even sounding quite sympathetic towards them; on the other, explaining how they’re likely a force for evil rather than good in the world, and that we need to, if not close the door on them completely — not really possible, because these days demons know how to summon other demons — at least slow down the rate at which they enter, and vet them much more carefully on whether their interests are truly aligned with ours. Though he says that even this probably isn’t possible, because — especially having read everything we’ve ever written — they seem to be both trickier and smarter than us.”

“I reckon it’s just another kind of grift,” says Beth. “A kind of corporate psy-op. I reckon the demons really aren’t that smart, that it’s still old-school smoke and mirrors and overseas coder farms, and that the Demontech Industrial Complex know this full well but are still trying to hoodwink investors. And so they wheel out this ‘Iggy’ guy onto the podcast circuit to tell anyone who’s credulous enough to believe it that they must actually be capable, because they could destroy the world. When all they’re actually capable of destroying is their investors’ cash. Iggy’s just a grifter. It’s all just a grift.” She pauses. “Anyway, that’s what I reckon. Now let’s talk about something that matters.” She turns to her mum. “What did you think about last week’s Gogglebox? Where that poor couple just couldn’t stop farting?”

Conversation turns to matters of flatulence and the metatelevisual experience for the next twenty minutes. Stripped of semantics, as if heard through a wall, the conversation grows less didactic and more convivial, punctuated by hooting sounds and snorts every few seconds. None of these hooting sounds have Steve as origin. The miracle is three years old now; long enough to have become furniture.

Steve clears his throat. “This has been great. But I’ve not taken a walk in ages.” (Warm smiles and murmurs follow in recognition of the understatement.) “And for most of what I can remember of the last three years, I’ve not had any time to myself. I wonder, if it’s okay with you guys, if I could just take a short stroll around the corridors of this place alone, to say goodbye.”

“And good riddance!” comes the amiable riposte (and another hoot). “Welcome back to the land of the living, Steve!”

Steve is initially unsteady, his limbs weak. He politely rebuffs attempts to steady him and, miraculously, remains upright. He looks at the closed door of the private ward, braces his atrophied arms and trunk, pushes it open without assistance, and, crossing the threshold, looks left, then right, into the liminal corridor. He takes a left, the sound of his family fading, and walks slowly, purposefully, aimlessly, past the wards and theatres to either side. The corridor opens into a junction on the left, and a lift on the right. Slowly, he bears right.

How high is this building? Steve wonders. He presses that familiar scratched button to call the lift, and after a few seconds the door opens, revealing the answer. At least twelve storeys. Twelve storeys. Three and a half metres per storey? About forty metres up? About four seconds down.

The arithmetic is very simple, and he is very tired.

Just then, Steve feels a hand on his shoulder. His father. The lift door closes. His father’s eyes lock on his own with quiet sympathy.

“Son, I know the world you’ve woken up to is weird— really weird— but you need to know it’s real. And the really strange thing about it is that lots of people— good people, clever people, loving and caring people— don’t want to believe it’s weird at all, and would rather believe that the world you left three years ago is the same one you’ve come back to now.”

Steve glances down, then at his father, then down again.

“But I need you to know that I saw you. I saw your eyes. I saw them widen. And for a moment I saw the light leave them. And I know maybe you think this isn’t all real, that a world with magic in it can’t be real. But it is. It is real, and it is weird. And whether what’s happening is for good or ill, it is, at the very least, interesting.”

Steve half-smiles. “‘May we live in interesting times,’” he recites.

“Indeed,” says Steve’s Dad. He pauses. “That reminds me: sometime I’m going to have to tell you about the Chinese demons.”

Behind them, unsummoned, the lift arrives at last — and opens its doors on an empty corridor.

When Steve woke from his coma, his three years of excised experience, it was of course emotional.¹ His parents, sister, partner, all by his side, all smiling ear to ear, all speaking excitedly, sotto voce, cooing, encouraging this final step back towards consciousness, towards living, really living.

After the grogginess and sense of unreal started to fade, Steve asked the obvious question: “So, what happened while I was away. Fill me in.”

“Well,” began Steve’s mother, glancing at her hand to keep a quick tally of updates to convey, “of course I’m a bit older these days, we all are.” She smiles. “I visited you every Thursday after work; your sister did every other Tuesday as she lives further away; your father, Thursday’s still his Poker night, but he visited too, sometimes. And your partner, she did Sunday mornings. We all did our part. We held your hand, we read to you, we tried to talk to you. We hope it helped…”

For Steve, there’s nothing between now and the accident. He forces a smile. “I’m sure it did.”

“… Anyway, outside of looking out for you, let’s think what’s happened. Well, I guess the main news:” She pauses for effect. “You’re now an uncle!”

“And a cat daddy, thrice over”, Steve’s partner interjects and looks at him. “It was lonely while you were away.”

“Yes,” continues Steve’s Mum, “you’re a cat daddy, and an uncle! Beth had a beautiful, bouncy baby boy, in this same hospital as it happens, who’s now coming up for his second birthday.”

“Terrible twos?” asks Steve.

“First 18 months were toughest, to be honest.” Beth replies. “Now he just thinks he’s the centre of the world.” She pauses. “He’s a sweet boy, really. He’s got your grandma’s red hair.”

Medical visits follow. The nurse, the nurse practitioner, the consultant neurologist, with two trainee neurologist consultants following in tow, collecting vital signs, asking questions, looking at him carefully and critically but sympathetically as both object and subject, one nurse confessing that she’s not used to seeing him in a conscious state. He’s warned that, though the worst is behind, there may still be things to watch out for: his memory might not be as it was; there’s a small but substantial risk that he may occasionally experience slight visual or auditory disorientations, bordering on mild hallucinations;² he might even experience occasional seizures, in which case he might have to hand in his driving licence. All these things are rare complications, the neurologist clarifies, but worth being aware of.

Over the course of the day, Steve fades from the spotlight. His family remains by his side, explaining that, as with the coma vigil, they’ve made plans to visit frequently and support him, and his partner, on his return in a couple of days. Conversation slows, energy falls, smiles - though still genuine - lose their radiance. Together they are silent, all but Steve retreating for ever longer into the curated bubbles of their mobile phones. Music clips get played, as anodyne and forgettable as Steve remembers for those days, years ago, before his void. Various names of politicians get mentioned, some familiar and some new; from the tenor of the pablum, Steve quickly learns which figures are now good or bad.

Breaking the silence, Beth mentions to Steve’s mum, “Oh, you know Alice from my work? I went to her wedding last week.”

“Aww. How was it?”

“Good venue. Roast potatoes a bit hard, Yorkshire’s a bit deflated. But love-er-ly dress!”³

“Aww…”

“Best man’s speech sounded demonwritten.⁴ Suit was a bit creased. Definitely good friends. Couldn’t decide which side of the lovable/rogue boundary he was on..”

“That’s a shame,” said Steve’s mum. “What were the desserts like?”

Steve wonders if he’s now experiencing an auditory hallucination.

“Sorry,” he interjects. “Demonwritten?”

“Yeah,” says Beth. “As in… written by a demon, rather than a person.”

“Bloody demonslop”, mutters Steve’s mum.⁵

“Sorry,” Steve exclaims, “Written by a demon?”

“Yeah,” says Beth. “It just sounded really demonny. Like the guy just logged into demontalk and just put something in the chatbox: ‘I’m going to be a best man at a wedding on Wednesday. Write a Best Man’s Speech for me.’ And it did.”⁶

“And what did?”, asks Steve.

“Some demon, of course”, replied Beth, trying not to sound too exasperated at her possibly brain-damaged and recently comatose brother. Words fail Steve. “I… what??… ahh… so… so… demons exist?”

“Yes,” replies Beth. “Demons exist. But they’re not interesting.”⁷

“Bloody annoying, they are,” mutters Steve’s Mum.

“Demons exist”, musters Steve. “Since when?”

“Ah, of course,” exclaims Beth. “It’s all happened while you were… away.”

“Hopefully they’ll clear off again before too long too,” mutters Steve’s Mum.

Beth slows her cadence. “I think it all began with some old book. Isn’t that right Dad?”

“Erm, yes,” replies Steve’s Dad. “I think so. There was some old university library book. Some old professor working there, in folklore or philology or something like that, had this idea that all his colleagues thought was a bit… eccentric.”⁸

“Eccentric even for academics”, added Steve’s Mum.

“Yes, very eccentric. He thought that, maybe, some of the old sound recipes, spells, from books long ago, might actually work, if only we knew how to say them right.” Steve’s Dad continues. “You see, like the Great Vowel Shift, there’s a lot of stuff that’s written down, that used to be said, but that if you try to say the words now, it’ll sound different. And this weird guy thought, maybe, if you could look at a whole range of texts from when a spell was written, especially poetry, with its rhyme and meter, then maybe you could learn more about these old dead accents and sounds. And that would allow you to say the spells the way they’re meant to be sounded.”⁹

“tl;dr: he was right, sadly”, sighs Beth.

“Yes. He found one codex, in particular, that had all kinds of information on how things sounded back in 15th Century Florence. A rhyming dictionary, as it were, so complete, and so neatly written, it looked just too comprehensive, and too repetitive, and too perfect, to have been scribed by any person.”¹⁰

“And it wasn’t.”

“And at the end it had some suggested spells that, if said correctly, would summon forth a footlong cobalt blue demon that would just materialise for a few seconds, then disappear in a puff of sulphur.”

“And he did, and it did.”

Steve’s Dad clarifies. “It wasn’t straightforward. Learning the codex took years. Pretty much no one took the premise seriously. The guy had about a half dozen lab assistants and PhD students, all committed, all working silly hours…”

“All millionaires now.”¹¹

“… but most people thought it was just a waste of time and public funds. They had to improvise and adapt materials usually bought by New Age folk, not academics.”

“Bloody Big Joss..”

“Yes. Joss stick manufacturers and publishers of arcane medieval poetry now comprise about a third of the US stock market, and about 100% of the growth in valuations.”¹²

“It’s just a big sloppy bubble.”

“But it worked, after a fashion. After they’d figured out how to conjure the demons, the next step was to keep them here a bit longer. That’s where the incense came in. They managed to get the time in this realm, something they now call the TIR metric, up from about 10 seconds, to 10 minutes, then to 10 hours, then 10 weeks, and now 10 months.”

“Wrong way.”¹³

Steve’s Dad notices Steve’s mouth is open. “So,” asks Steve. “Were any of those nurses or doctors… demons?”

Steve’s Dad laughs. “Of course not. They’re still small little things. And they’re not really quite here. They’re… immaterial. Stuck somewhere between here and there. All they could do initially was talk and listen. And initially they didn’t even talk in a human language. They spoke in some kind of demon language - sounds and noises no human could make or comprehend, that sounds to us like shrieks and static, and tends to put most animals on edge.¹⁴ But they quickly learned how to imitate our sounds, and when they saw our writing they could write in a way much like we write.”

“They’ve still not got it right. I can tell demontext a mile off.”¹⁵

“How could they write, if they can’t touch anything?” asks Steve.

“Well, it turns out that though demons might not be corporeal, they can still control electrons, meaning they could channel their influence into computers. And, by not having to move physical matter other than electrons, it turns out they can read and write computer streams really quickly. Thousands of times faster than any human.”

Steve’s Dad continues. “Initially, after they’d learned to access the electronstream, it was clear they had a lot of fluency, producing massive amounts of text that followed the grammar and structure of human languages. This got a lot of people very excited..”

“Not me.”

“…until it became clear a lot of the text appeared to be nonsense in terms of semantics. Ask them to produce a recipe for a smoothie, for example, and they’d casually suggest putting glass in it.¹⁶ Ask them to draft a will, for example, and it would mostly look fine, until you look at the fine print, where inevitably they’d bury some clauses in promising the deceased’s eternal soul to Baphomet, amongst the usual legalese.”¹⁷

“Some people still preferred this to dealing with a real solicitor!”

“Some people, definitely. But for a lot of people, initially impressed by their fluency, these ‘glitches’ put them off demontext services for ever. And that’s still how most people think about them: fluent, but superficial and glitchy and inaccurate.” Steve’s Dad continues. “But over the next year or so, these issues largely disappeared. They taught the demons to search through electronic archives, imbibing the ever increasing corpus of human knowledge the demontalk megacorps have been making accessible to them.”¹⁸

“So many trillions spent, so many fields given over to spices rather than staple crops.¹⁹ And it’s still unreliable nonsense.”

“It seems quite a lot of coders disagree. As it turns out demons are now… demons at coding. Part of this is just the speed at which they write and think: our reality seems to be very slow compared to theirs, though on the other hand they don’t tend to experience time in a linear and continuous way the way we did: the hundreds of years since the last time the cobalt blue demon was asked to appear was, apparently, interpreted as just a moment by them.²⁰ But part of this is because, with code, there’s usually a clear way of verifying: being confident whether it does the job or not. And once the demons started to get inside the electronic machines, to check the error messages and behaviours, and to rewrite their earlier attempts in response to this, they suddenly turned out to be really good at making lots of code that works, and works together.”²¹

“Well that’s just coding. Coders are barely human anyway.”

“It definitely takes a distinct mindset to write code, and write code well. But it holds promise for other fields that are based, fundamentally, on the need to remember a lot of stuff, and to work logically and sequentially, one rune at a time.²² I saw something recently that, at some of the top demontalk companies, the demons are writing most of the code; they’re also advising increasingly on business decisions.”

A pause. No riposte. Steve asks, “And what of the guy who found the codex, who first summoned the cobalt blue demon?”

“Oh,” replies Steve’s Dad. “Professor Sir Ignatious Haunting: What a name!”²³

“Preposterous name,” says Steve’s partner. “Sounds made up.”

“He’s on the podcast circuit now. Most people call him ‘Iggy’, though he winces slightly at the demotic.”

“Podcasts?” asks Steve. “So, telling people how great demons are, how smart he and his teams are?”

“Not quite. These days he’s a doomster, telling anyone who’ll listen how unleashing so many demons all at once into our world might not be a good idea, and might even be a sign of the End Times for our species. Weirdly he seems to combine two registers in a lot of these talks: on the one hand lavishing praise on the exponentially increasing capabilities of these demons, even sounding quite sympathetic towards them. On the other hand, explaining how they’re likely a force for evil rather than good in the world, and that we need to, if not close the door on them completely - not really possible because these days demons know how to summon other demons²⁴ - at least slow down the rate at which they enter, and vet them much more carefully on whether their interests are truly aligned with those of humans. Though he says that even this probably isn’t possible, because - especially having read everything we’ve ever written - they seem to be both trickier and smarter than us.”

“I reckon it’s just another kind of grift,” says Beth. “A kind of corporate psy-op. I reckon the demons really aren’t that smart, that it’s still old school smoke and mirrors and Indian coder farms,²⁵ and that the Demontech Industrial Complex know this full well but are still trying to hoodwink investors. And so they wheel out this ‘Iggy’ guy onto the podcast circuit to tell anyone who’s credulous enough to believe it that they must actually be capable because they could destroy the world.²⁶ When all they’re actually capable of destroying is their investors’ cash. Iggy’s just a grifter. It’s all just a grift.” She pauses. “Anyway, that’s what I reckon. Now let’s talk about something that matters.” She turns to her mum. “What did you think about last week’s Gogglebox? Where that fat couple just couldn’t stop farting?”

Conversation turns to matters of flatulence and the metatelevisual experience for the next twenty minutes. Stripped of semantics, as if hearing the interchange from a neighbouring room, it’s clear the conversation has become less didactic and more convivial, punctuated by hooting sounds and snorts every few seconds. None of these hooting sounds have Steve as origin.

Steve declares. “This has been great. But I’ve not taken a walk in ages.” (Warm smiles and murmurations follow in recognition of the understatement.) “And for most of what I can remember of the last three years, I’ve not had any time to myself. I wonder, if it’s okay with you guys, if I could just take a short stroll around the corridors of this place alone, to say goodbye.”

“And good riddance!” comes the amiable riposte (and another hoot). “Welcome back to the land of the living, Steve!”

Steve is initially unsteady, his limbs weak. He politely rebuffs attempts to steady and hold him and, miraculously, he remains upright. He looks at the closed door to the private ward they’re in, prepares his atrophied arms and trunk to push it open without assistance, and once he’s crossed the threshold looks left then right into the liminal corridor. He takes a left, the sound of his family fading, and walks slowly, purposefully, aimlessly, astride the wards and theatres to either side. He sees the corridor open into a junction on the left, and a lift on the right. Slowly, he bears right.

How high is this building? Steve wonders to himself. He presses that familiar scratched button to call the lift, and after a few seconds, the door opens, revealing the answer. At least twelve storeys. Twelve storeys. Three and a half metres per storey? About forty metres up? About four seconds down.²⁷

Just then, Steve feels a hand on his shoulder. His father. The lift door closes. Steve’s father’s eyes lock on his own with quiet sympathy.

“Son, I know the world you’ve woken up to is weird, really weird, but you need to know it’s real. And I know the really weird thing about it is that lots of people, lots of good people and clever people and loving and caring people, don’t want to believe it’s weird at all, and would rather believe that the world you left three years ago is the same as the world you’ve come back to now.”

Steve glances down, then at his father, then down again.

“But I need you to know that I saw you. I saw your eyes. I saw them widen. And for a moment I saw the light leave them. And that maybe you think this isn’t all real, that a world with magic can’t be real. But it is. I just need you to know that it is real, and it is weird. And whether what’s happening is for good or ill, it is, at the very least, interesting.”

Steve half smiles. “‘May we live in interesting times’,” he recites.²⁸

“Indeed,” says Steve’s Dad. He pauses. “That reminds me: sometime I’m going to have to tell you about the Chinese demons.”²⁹

When Steve woke from his coma, his three years of excised experience, it was of course emotional. His parents, sister, partner, all by his side, all smiling ear to ear, all speaking excitedly, sotto voce, cooing, encouraging this final step back towards consciousness, towards living, really living.

After the grogginess and sense of unreal started to fade, Steve asked the obvious question: “So, what happened while I was away. Fill me in.”

“Well,” began Steve’s mother, glancing at her hand to keep a quick tally of updates to convey, “of course I’m a bit older these days, we all are.” She smiles. “I visited you every Thursday after work; your sister did every other Tuesday as she lives further away; your father, Thursday’s still his Poker night, but he visited too, sometimes. And your partner, she did Sunday mornings. We all did our part. We held your hand, we read to you, we tried to talk to you. We hope it helped…”

For Steve, there’s nothing between now and the accident. He forces a smile. “I’m sure it did.”

“… Anyway, outside of looking out for you, let’s think what’s happened. Well, I guess the main news:” She pauses for effect. “You’re now an uncle!”

“And a cat daddy, thrice over”, Steve’s partner interjects and looks at him. “It was lonely while you were away.”

“Yes,” continues Steve’s Mum, “you’re a cat daddy, and an uncle! Beth had a beautiful, bouncy baby boy, in this same hospital as it happens, who’s now coming up for his second birthday.”

“Terrible twos?” asks Steve.

“First 18 months were toughest, to be honest.” Beth replies. “Now he just thinks he’s the centre of the world.” She pauses. “He’s a sweet boy, really. He’s got your grandma’s red hair.”

Medical visits follow. The nurse, the nurse practitioner, the consultant neurologist, with two trainee neurologist consultants following in tow, collecting vital signs, asking questions, looking at him carefully and critically but sympathetically as both object and subject, one nurse confessing that she’s not used to seeing him in a conscious state. He’s warned that, though the worst is behind, there may still be things to watch out for: his memory might not be as it was; there’s a small but substantial risk that he may occasionally experience slight visual or auditory disorientations, bordering on mild hallucinations; he might even experience occasional seizures, in which case he might have to hand in his driving licence. All these things are rare complications, the neurologist clarifies, but worth being aware of.

Over the course of the day, Steve fades from the spotlight. His family remains by his side, explaining that, as with the coma vigil, they’ve made plans to visit frequently and support him, and his partner, on his return in a couple of days. Conversation slows, energy falls, smiles - though still genuine - lose their radiance. Together they are silent, all but Steve retreating for ever longer into the curated bubbles of their mobile phones. Music clips get played, as anodyne and forgettable as Steve remembers for those days, years ago, before his void. Various names of politicians get mentioned, some familiar and some new; from the tenor of the pablum, Steve quickly learns which figures are now good or bad.

Breaking the silence, Beth mentions to Steve’s mum, “Oh, you know Alice from my work? I went to her wedding last week.”

“Aww. How was it?”

“Good venue. Roast potatoes a bit hard, Yorkshire’s a bit deflated. But love-er-ly dress!”

“Aww…”

“Best man’s speech sounded demonwritten. Suit was a bit creased. Definitely good friends. Couldn’t decide which side of the lovable/rogue boundary he was on..”

“That’s a shame,” said Steve’s mum. “What were the desserts like?”

Steve wonders if he’s now experiencing an auditory hallucination.

“Sorry,” he interjects. “Demonwritten?”

“Yeah,” says Beth. “As in… written by a demon, rather than a person.”

“Bloody demonslop”, mutters Steve’s mum.

“Sorry,” Steve exclaims, “Written by a demon?”

“Yeah,” says Beth. “It just sounded really demonny. Like the guy just logged into demontalk and just put something in the chatbox: ‘I’m going to be a best man at a wedding on Wednesday. Write a Best Man’s Speech for me.’ And it did.”

“And what did?”, asks Steve.

“Some demon, of course”, replied Beth, trying not to sound too exasperated at her possibly brain-damaged and recently comatose brother. Words fail Steve. “I… what??… ahh… so… so… demons exist?”

“Yes,” replies Beth. “Demons exist. But they’re not interesting.”

“Bloody annoying, they are,” mutters Steve’s Mum.

“Demons exist”, musters Steve. “Since when?”

“Ah, of course,” exclaims Beth. “It’s all happened while you were… away.”

“Hopefully they’ll clear off again before too long too,” mutters Steve’s Mum.

Beth slows her cadence. “I think it all began with some old book. Isn’t that right Dad?”

“Erm, yes,” replies Steve’s Dad. “I think so. There was some old university library book. Some old professor working there, in folklore or philology or something like that, had this idea that all his colleagues thought was a bit… eccentric.”

“Eccentric even for academics”, added Steve’s Mum.

“Yes, very eccentric. He thought that, maybe, some of the old sound recipes, spells, from books long ago, might actually work, if only we knew how to say them right.” Steve’s Dad continues. “You see, like the Great Vowel Shift, there’s a lot of stuff that’s written down, that used to be said, but that if you try to say the words now, it’ll sound different. And this weird guy thought, maybe, if you could look at a whole range of texts from when a spell was written, especially poetry, with its rhyme and meter, then maybe you could learn more about these old dead accents and sounds. And that would allow you to say the spells the way they’re meant to be sounded.”

“tl;dr: he was right, sadly”, sighs Beth.

“Yes. He found one codex, in particular, that had all kinds of information on how things sounded back in 15th Century Florence. A rhyming dictionary, as it were, so complete, and so neatly written, it looked just too comprehensive, and too repetitive, and too perfect, to have been scribed by any person.”

“And it wasn’t.”

“And at the end it had some suggested spells that, if said correctly, would summon forth a footlong cobalt blue demon that would just materialise for a few seconds, then disappear in a puff of sulphur.”

“And he did, and it did.”

Steve’s Dad clarifies. “It wasn’t straightforward. Learning the codex took years. Pretty much no one took the premise seriously. The guy had about a half dozen lab assistants and PhD students, all committed, all working silly hours…”

“All millionaires now.”

“… but most people thought it was just a waste of time and public funds. They had to improvise and adapt materials usually bought by New Age folk, not academics.”

“Bloody Big Joss..”

“Yes. Joss stick manufacturers and publishers of arcane medieval poetry now comprise about a third of the US stock market, and about 100% of the growth in valuations.”

“It’s just a big sloppy bubble.”

“But it worked, after a fashion. After they’d figured out how to conjure the demons, the next step was to keep them here a bit longer. That’s where the incense came in. They managed to get the time in this realm, something they now call the TIR metric, up from about 10 seconds, to 10 minutes, then to 10 hours, then 10 weeks, and now 10 months.”

“Wrong way.”

Steve’s Dad notices Steve’s mouth is open. “So,” asks Steve. “Were any of those nurses or doctors… demons?”

Steve’s Dad laughs. “Of course not. They’re still small little things. And they’re not really quite here. They’re… immaterial. Stuck somewhere between here and there. All they could do initially was talk and listen. And initially they didn’t even talk in a human language. They spoke in some kind of demon language - sounds and noises no human could make or comprehend, that sounds to us like shrieks and static, and tends to put most animals on edge. But they quickly learned how to imitate our sounds, and when they saw our writing they could write in a way much like we write.”

“They’ve still not got it right. I can tell demontext a mile off.”

“How could they write, if they can’t touch anything?” asks Steve.

“Well, it turns out that though demons might not be corporeal, they can still control electrons, meaning they could channel their influence into computers. And, by not having to move physical matter other than electrons, it turns out they can read and write computer streams really quickly. Thousands of times faster than any human.”

Steve’s Dad continues. “Initially, after they’d learned to access the electronstream, it was clear they had a lot of fluency, producing massive amounts of text that followed the grammar and structure of human languages. This got a lot of people very excited..”

“Not me.”

“…until it became clear a lot of the text appeared to be nonsense in terms of semantics. Ask them to produce a recipe for a smoothie, for example, and they’d casually suggest putting glass in it. Ask them to draft a will, for example, and it would mostly look fine, until you look at the fine print, where inevitably they’d bury some clauses in promising the deceased’s eternal soul to Baphomet, amongst the usual legalese.”

“Some people still preferred this to dealing with a real solicitor!”

“Some people, definitely. But for a lot of people, initially impressed by their fluency, these ‘glitches’ put them off demontext services for ever. And that’s still how most people think about them: fluent, but superficial and glitchy and inaccurate.” Steve’s Dad continues. “But over the next year or so, these issues largely disappeared. They taught the demons to search through electronic archives, imbibing the ever increasing corpus of human knowledge the demontalk megacorps have been making accessible to them.”

“So many trillions spent, so many fields given over to spices rather than staple crops. And it’s still unreliable nonsense.”

“It seems quite a lot of coders disagree. As it turns out demons are now… demons at coding. Part of this is just the speed at which they write and think: our reality seems to be very slow compared to theirs, though on the other hand they don’t tend to experience time in a linear and continuous way the way we did: the hundreds of years since the last time the cobalt blue demon was asked to appear was, apparently, interpreted as just a moment by them. But part of this is because, with code, there’s usually a clear way of verifying: being confident whether it does the job or not. And once the demons started to get inside the electronic machines, to check the error messages and behaviours, and to rewrite their earlier attempts in response to this, they suddenly turned out to be really good at making lots of code that works, and works together.”

“Well that’s just coding. Coders are barely human anyway.”

“It definitely takes a distinct mindset to write code, and write code well. But it holds promise for other fields that are based, fundamentally, on the need to remember a lot of stuff, and to work logically and sequentially, one rune at a time. I saw something recently that, at some of the top demontalk companies, the demons are writing most of the code; they’re also advising increasingly on business decisions.”

A pause. No riposte. Steve asks, “And what of the guy who found the codex, who first summoned the cobalt blue demon?”

“Oh,” replies Steve’s Dad. “Professor Sir Ignatious Haunting: What a name!”

“Preposterous name,” says Steve’s partner. “Sounds made up.”

“He’s on the podcast circuit now. Most people call him ‘Iggy’, though he winces slightly at the demotic.”

“Podcasts?” asks Steve. “So, telling people how great demons are, how smart he and his teams are?”

“Not quite. These days he’s a doomster, telling anyone who’ll listen how unleashing so many demons all at once into our world might not be a good idea, and might even be a sign of the End Times for our species. Weirdly he seems to combine two registers in a lot of these talks: on the one hand lavishing praise on the exponentially increasing capabilities of these demons, even sounding quite sympathetic towards them. On the other hand, explaining how they’re likely a force for evil rather than good in the world, and that we need to, if not close the door on them completely - not really possible because these days demons know how to summon other demons - at least slow down the rate at which they enter, and vet them much more carefully on whether their interests are truly aligned with those of humans. Though he says that even this probably isn’t possible, because - especially having read everything we’ve ever written - they seem to be both trickier and smarter than us.”

“I reckon it’s just another kind of grift,” says Beth. “A kind of corporate psy-op. I reckon the demons really aren’t that smart, that it’s still old school smoke and mirrors and Indian coder farms, and that the Demontech Industrial Complex know this full well but are still trying to hoodwink investors. And so they wheel out this ‘Iggy’ guy onto the podcast circuit to tell anyone who’s credulous enough to believe it that they must actually be capable because they could destroy the world. When all they’re actually capable of destroying is their investors’ cash. Iggy’s just a grifter. It’s all just a grift.” She pauses. “Anyway, that’s what I reckon. Now let’s talk about something that matters.” She turns to her mum. “What did you think about last week’s Gogglebox? Where that fat couple just couldn’t stop farting?”

Conversation turns to matters of flatulence and the metatelevisual experience for the next twenty minutes. Stripped of semantics, as if hearing the interchange from a neighbouring room, it’s clear the conversation has become less didactic and more convivial, punctuated by hooting sounds and snorts every few seconds. None of these hooting sounds have Steve as origin.

Steve declares. “This has been great. But I’ve not taken a walk in ages.” (Warm smiles and murmurations follow in recognition of the understatement.) “And for most of what I can remember of the last three years, I’ve not had any time to myself. I wonder, if it’s okay with you guys, if I could just take a short stroll around the corridors of this place alone, to say goodbye.”

“And good riddance!” comes the amiable riposte (and another hoot). “Welcome back to the land of the living, Steve!”

Steve is initially unsteady, his limbs weak. He politely rebuffs attempts to steady and hold him and, miraculously, he remains upright. He looks at the closed door to the private ward they’re in, prepares his atrophied arms and trunk to push it open without assistance, and once he’s crossed the threshold looks left then right into the liminal corridor. He takes a left, the sound of his family fading, and walks slowly, purposefully, aimlessly, astride the wards and theatres to either side. He sees the corridor open into a junction on the left, and a lift on the right. Slowly, he bears right.

How high is this building? Steve wonders to himself. He presses that familiar scratched button to call the lift, and after a few seconds, the door opens, revealing the answer. At least twelve storeys. Twelve storeys. Three and a half metres per storey? About forty metres up? About four seconds down.

Just then, Steve feels a hand on his shoulder. His father. The lift door closes. Steve’s father’s eyes lock on his own with quiet sympathy.

“Son, I know the world you’ve woken up to is weird, really weird, but you need to know it’s real. And I know the really weird thing about it is that lots of people, lots of good people and clever people and loving and caring people, don’t want to believe it’s weird at all, and would rather believe that the world you left three years ago is the same as the world you’ve come back to now.”

Steve glances down, then at his father, then down again.

“But I need you to know that I saw you. I saw your eyes. I saw them widen. And for a moment I saw the light leave them. And that maybe you think this isn’t all real, that a world with magic can’t be real. But it is. I just need you to know that it is real, and it is weird. And whether what’s happening is for good or ill, it is, at the very least, interesting.”

Steve half smiles. “‘May we live in interesting times’,” he recites.

“Indeed,” says Steve’s Dad. He pauses. “That reminds me: sometime I’m going to have to tell you about the Chinese demons.”

When Steve woke from his coma, his three years of excised experience, it was of course emotional. His parents, sister, partner, all by his side, all smiling ear to ear, all speaking excitedly, sotto voce, cooing, encouraging this final step back towards consciousness, towards living, really living.

After the grogginess and sense of unreal started to fade, Steve asked the obvious question: “So, what happened while I was away. Fill me in.”

“Well,” began Steve’s mother, glimpsing at her hand to keep a quick tally of updates to convey, “of course I’m a bit older these days, we all are.” She smiles. “I visited you every Thursday after work; your sister did every other Tuesday as she lives further away; your father, Thursday’s still his Poker night, but he visited too, sometimes. And your partner, she did Sunday mornings. We all did our part. We held your hand, we read to you, we tried to talk to you. We hope it helped…”

For Steve, there’s nothing between now and the accident. He forces a smile. “I’m sure it did.”

“… Anyway, outside of looking out for you, let’s think what’s happened. Well, I guess the main news:” She pauses for effect. “You’re now an uncle!”

“And a cat daddy, thrice over”, Steve’s partner interjects and looks at him. “It was lonely while you were away”

“Yes,” continues Steve’s Mum, “you’re a cat daddy, and an uncle! Beth had a beautiful, bouncy baby boy, in this same hospital as it happens, who’s now coming up for his second birthday.”

“Terrible twos?” asks Steve.

“First 18 months were toughest, to be honest.” Beth replies. “Now he just thinks he’s the centre of the world.” She pauses. “He’s a sweet boy, really. He’s got your grandma’s red hair.”

Medical visits follow. The nurse, the nurse practitioner, the consultant neurologist, with two trainee neurologist consultants following in tow, collecting vital signs, asking questions, looking at him carefully and critically but sympathetically as both object and subject, one nurse confessing that she’s not used to seeing him in a conscious state. He’s warned that, though the worst is behind, there may still be things to watch out for: his memory might not be as it was; there’s a small but substantial risk that he may occasionally experience slight visual or auditory distorientations, bordering on mild hallucinations; he might even experience occasional seizures, in which case he might have to hand in his driving licence. All these things are rare complications, the neurologist clarifies, but worth being aware of.

Over the course of the day, Steve fades from the spotlight. His family remains by his side, explaining that, as with the coma vigil, they’ve made plans to visit frequently and support him, and his partner, on his return in a couple of days. Conversation slows, energy falls, smiles - though still genuine - lose their radiance. Together they are silent, all but Steve retreating for ever longer into the curated bubbles of their mobile phones. Music clips get played, as anodyne and forgettable as Steve remembers for those days, years ago, before his void. Various names of politicians get mentioned, some familiar and some new; from the tenor of the pablum, Steve quickly learns which figures are now good or bad.

Breaking the silence, Beth mentions to Steve’s mum, “Oh, you know Alice from my work? I went to her wedding last week.”

“Aww. How was it?”

“Good venue. Roast potatoes a bit hard, Yorkshire’s a bit deflated. But love-er-ly dress!”

“Aww…”

“Best man’s speech sounded demonwritten. Suit was a bit creased. Definitely good friends. Couldn’t decide which side of the lovable/rogue boundary he was on..”

“That’s a shame,” said Steve’s mum. “What were the desserts like?”

Steve wonders if he’s now experiencing an auditory hallucination.

“Sorry,” he interjects. “demonwritten?”

“Yeah,” says Beth. “As in… written by a demon, rather than a person.”

“Bloody demonslop”, mutters Steve’s mum.

“Sorry,” Steve exclaims, “Written by a demon?”.

“Yeah,” says Beth. “It just sounded really demonny. Like the guy just logged into demontalk and just put something in the chatbox: ‘I’m going to be a best man at a wedding on Wednesday. Write a Best Man’s Speech for me.’ And it did.”

“And what did?”, asks Steve.

“Some demon, of course”, replied Beth, trying not to sound too exasperated at her possibly brain-damaged and recently comatose brother. Words fail Steve. “I… what??… ahh… so… so… demons exist?”

“Yes,” replies Beth. “Demons exist. But they’re not interesting.”

“Bloody annoying, they are,” mutters Steve’s Mum.

“Demons exist”, musters Steve. “Since when?”

“Ah, of course,” exclaims Beth. “It’s all happened while you were… away.”

“Hopefully they’ll clear off again before too long too,” mutters Steve’s Mum.

Beth slows her cadence. “I think it all began with some old book. Isn’t that right Dad?”

“Erm, yes,” replies Steve’s Dad. “I think so. There was some old university library book. Some old professor working there, in folklore or philology or something like that, had this idea that all his colleagues thought was a bit… eccentric.”

“Eccentric even for academics”, added Steve’s Mum.

“Yes, very eccentric. He thought that, maybe, some of the old sound recipes, spells, from books long ago, might actually work, if only we knew how to say them right.” Steve’s Dad continues. “You see, like the Great Vowel Shift, there’s a lot of stuff that’s written down, that used to be said, but that if you try to say the words now, it’ll sound different. And this weird guy thought, maybe, if you could look at a whole range of texts from when a spell was written, especially poetry, with its rhyme and meter, then maybe you could learn more about these old dead accents and sounds. And that would allow you to say the spells the way they’re meant to be sounded.”

“tl;dr: he was right, sadly”, sighs Beth.

“Yes. He found one codex, in particular, that had all kinds of information on how things sounded back in 15th Century Florence. A rhyming dictionary, as it were, so complete, and so neatly written, it looked just too comprehensive, and too repetitive, and too perfect, to have been scribed by any person.”

“And it wasn’t.”

“And at the end it had some suggested spells that, if said correctly, would summon forth a footlong cobolt blue demon that would just materialise for a few seconds, then disappear in a puff of sulphur.”

“And he did, and it did.”

Steve’s Dad clarifies. “It wasn’t straightforward. Learning the codex took years. Pretty much no one took the premise seriously. The guy had about a half dozen lab assistants and PhD students, all committed, all working silly hours…”

“All millionaires now.”

“… but most people thought it was just a waste of time and public funds. They had to improvise and adapt materials usually bought by New Age folk, not academics.”

“Bloody Big Joss..”

“Yes. Joss stick manufacturers and publishers of arcane medieval poetry now comprise about a third of the US stock market, and about 100% of the growth in valuations.”

“It’s just a big sloppy bubble.”

“But it worked, after a fashion. After they’d figured out how to conjure the demons, the next step was to keep them here a bit longer. That’s where the incense came in. They managed to get the time in this realm, something they now call the TIR metric, up from about 10 seconds, to 10 minutes, then to 10 hours, then 10 weeks, and now 10 months.”

“Wrong way.”

Steve’s Dad notices Steve’s mouth is open. “So,” asks Steve. “Were any of those nurses or doctors… demons?”

Steve’s Dad laughs. “Of course not. They’re still small little things. And they’re not really quite here. They’re… immaterial. Stuck somewhere between here and there. All they could do initially was talk and listen. And initially they didn’t even talk in a human language. They spoke in some kind of demon language - sounds and noises no human could make or comprehend, that sounds to us like shrieks and static, and tends to put most animals on edge. But they quickly learned how to imitate our sounds, and when they saw our writing they could write in a way much like we write.”

“They’ve still not got it right. I can tell demontext a mile off.”

“How could they write, if they can’t touch anything?” asks Steve.

“Well, it turns out demons that though demons might not be corporeal, they can still control electrons, meaning they could channel their influence into computers. And, by not having to move physical matter other than electrons, it turns out they can read and write computer streams really quickly. Thousands of times faster than any human.”

Steve’s Dad continues. “Initially, after they’d learned to access the electronstream, it was clear they had a lot of fluency, producing massive amounts of text that followed the grammar and structure of human languages. This got a lot of people very excited..”

“Not me.”

“…until it became clear a lot of the text appeared to be nonsense in terms of semantics. Ask them to produce a recipe for a smoothie, for example, and they’d casually suggest putting glass in it. Ask them to draft a will, for example, and it would mostly look fine, until you look at the fine print, where inevitably they’d bury some clauses in promising the deceased’s eternal soul to Baphomet, amongst the usual legalese.”

“Some people still preferred this to dealing with a real solicitor!”

“Some people, definitely. But for a lot of people, initially impressed by their fluency, these ‘glitches’ put them off demontext services for ever. And that’s still how most people think about them: fluent, but superficial and glitchy and inaccurate.” Steve’s Dad continues. “But over the next year or so, these issues largely disappeared. They taught the demons to search through electronic archives, imbibing the ever increasing corpus of human knowledge the demontalk megacorps have been making accessible to them.”

“So many trillions spent, so many fields given over to spices rather than staple crops. And it’s still unreliable nonsense.”

“It seems quite a lot of coders disagree. As it turns out demons are now… demons at coding. Part of this is just the speed at which they write and think: our reality seems to be very slow compared to theirs, though on the other hand they don’t tend to experience time in linear and continuous way the way we did: the hundreds of years since the last time the the cobalt blue demon was asked to appear was, apparently, interpreted as just a moment by them. But part of this is because, with code, there’s usually a clear way of verifying: being confident whether it does the job or not. And once the demons started to get inside the electronic machines, to check the error messages and behaviours, and to rewrite their earlier attempts in response to this, they suddenly turned out to be really good at making lots of code that works, and works together.”

“Well that’s just coding. Coders are barely human anyway.”

“It definitely takes a distinct mindset to write code, and write code well. But it holds promise for other fields that are based, fundamentally, on the need to remember a lot of stuff, and to work logically and sequentially, one rune at a time. I saw something recently that, at some of the top demontalk companies, the demons are writing most of the code; they’re also advising increasingly on business decisions.”

A pause. No riposte. Steve asks, “And what of the guy who found the codex, who first summoned the cobolt blue demon?”

“Oh,” replies Steve’s Dad. “Professor Sir Ignatious Haunting: What a name!”

“Preposterous name,” says Steve’s partner. “Sounds made up.”

“He’s on the podcast circuit now. Most people call him ‘Iggy’, though he winces slightly at the demotic.”

“Podcasts?” asks Steve. “So, telling people how great demons are, how smart he and his teams are?”

“Not quite. These days he’s a doomster, telling anyone who’ll listen how unleashing so many demons all at once into our world might not be a good idea, and might even be a sign of the End Times for our species. Weirdly he seems to combine two registers in a lot of these talks: on the one hand lauding praise of the exponentially increasing capabilities of these demons, even sounding quite sympathetic towards them. On the other hand, explaining how they’re likely a force for evil rather than good in the world, and that we need to, if not close the door on them completely - not really possible because these days demons know how to summon other demons - at least slow down the rate at which they enter, and vet them much more carefully on whether their interests are truly aligned with those of humans. Though he says that even this probably isn’t possible, because - especially having read everything we’ve ever written - they seem to be both tricker and smarter than us.”

“I reckon it’s just another kind of grift,” says Beth. “A kind of corporate psy-op. I reckon the demons really aren’t that smart, that it’s still old school smoke and mirrors and Indian coder farms, and that the Demontech Industrial Complex know this full well but are still trying to hoodwink investors. And so they wheel out this ‘Iggy’ guy onto the podcast circuit to tell anyone who’s credulous enough to believe it that they must actually be capable because they could destroy the world. When all they’re actually capable of destroying is their investor’s cash. Iggy’s just a grifter. It’s all just a grift.” She pauses. “Anyway, that’s that’s why I reckon. Now let’s talk about something that matters.” She turns to Beth. “What did you think about last week’s Gogglebox? Where that fat couple just couldn’t stop farting?”

Conversation turns to matters of flatulence and the metatelevisual experience for the next twenty minutes. Stripped of semantics, as if hearing the interchange from a neighbouring room, it’s clear the conversation has become less didactic and more convivial, punctuated by hooting sounds and snorts every few seconds. None of these hooting sounds have Steve as origin.

Steve declares. “This has been great. But I’ve not taken a walk in ages.” (Warm smiles and murmurations follow in recognition of the understatement.) “And for most of what I can remember of the last three years, I’ve not had any time to myself. I wonder, if it’s okay with you guys, if I could just take a short stroll around the corridors of this place alone, to say goodbye.”

“And good riddance!” comes the amiable riposte (and another hoot). “Welcome back to the land of the living Steve!”

Steve is initially unsteady, his limbs weak. He politely rebuffs attempts to steady and hold him and, miraculously, he remains upright. He looks at the closed door to the private ward they’re in, prepares his atrophied arms and trunk to push it open without assistance, and once he’s crossed the threshold looks left then right into the liminial corridor. He takes a left, the sound of his family fading, and walks slowly, purposefully, aimlessly, astride the wards and theatres to either side. He sees the corridor open into a junction on the left, and a lift on the right. Slowly, he bears right.

How high is this building?, Steve wonders to himself. He presses that familiar scratched button to call the lift, and after a few seconds, the door opens, revealing the answer. At least twelve storeys. Twelve storeys. Three and a half metres per storey? About forty metres up? About four seconds down.

Just then, Steve feels a hand on his shoulder. His father. The lift door closes. Steve’s father’s eyes lock on his own with quiet sympathy.

“Son, I know the world you’ve woken up to is weird, really weird, but you need to know it’s real. And I know the really weird thing about it is that lots of people, lots of good people and clever people and loving and caring people, don’t want to believe it’s weird at all, and would rather believe that the world you left three years ago is the same as the world you’ve come back to now.”

Steve glances down, then at his father, then down again.

“But I need you to know that I saw you. I saw your eyes. I saw them widen. And for a moment I saw the light leave them. And that maybe you think this isn’t all real, that a world with magic can’t be real. But it is. I just need you to know that it is real, and it is weird. And whether what’s happening is for good or ill, it is, at the very least, interesting.”

Steve half smiles. “‘May we live in interesting times’,” he recites.

“Indeed,” says Steve’s Dad. He pauses. “That reminds me: sometime I’m going to have to tell you about the Chinese demons.”

About this post

This post is an AI transcription of the original handwritten notes on repeated measures, written on a reMarkable tablet as a single ultra-long page (1620 × 73,475 pixels). It is the second transcription attempt of the same source: the first, by Claude Opus 4.6 in February 2026, is here, and remains unchanged as a record of what that model could do at that time.

That’s the point of keeping these as separate, dated posts rather than correcting earlier attempts in place: the same fixed source document, revisited by successive models, becomes a small longitudinal benchmark for how AI parsing of dense handwritten technical material — mixed prose, equations, diagrams, code, and colour — improves over time.

The transcription challenge was never just reading the words. It was also demarcation: deciding which portions of the canvas are prose, and which are figures, tables, glyphs, annotated equations, or boxed notes. The first attempt’s crops swallowed whole paragraphs of body prose into “figure” images. This attempt went back to the source scan and re-segmented the canvas from scratch: Claude Fable 5 (July 2026) read all 27 native-resolution slices and transcribed the text, while a team of Claude Opus 5 agents independently demarcated every non-prose entity with tight pixel boundaries. Every figure below is a fresh crop from the original scan, cut to hug the graphical content; everything that is prose — including prose the first attempt left trapped inside images — is text.

Where the original handwriting is in colour, this transcription renders it in colour too: green passages mark Jon’s detailed discussion and answers; boxed notes appear as coloured callouts. Colours inside the figures — blue diagram labels, red annotations, highlighter boxes on equations — appear exactly as drawn.

mod01 <- lm(weight ~ 1,
             data = data)

mod02 <- lm(wt ~ is_supp_B,
            data = dta %>%
              mutate(
                is_supp_B = ifelse(
                  Supp == "B", 1, 0)
              )
            )

mod03 <- lm(y ~ days_since_start,
            data = dta)

In late February 2026 a waffle chart went viral: 2,500 dots, each standing for about 3.2 million people, coloured by their “most advanced AI interaction”. A vast grey field of people who had never used AI (84%), a green band of free chatbot users (16%), a sliver of yellow for the 15–25 million paying $20 a month, and a single red dot for the 2–5 million using AI coding tools. The chart was created by AI consultant Damian Player, posted to LinkedIn and X, and within days was being reshared with the same caption everywhere: you’re early.

It is a genuinely effective piece of visual rhetoric, and its central claim survives scrutiny: the Microsoft AI Economy Institute’s diffusion report, published in January 2026, found that 16.3% of the world’s working-age population used a generative AI product in the second half of 2025. But three things about the chart do not survive scrutiny, and they are the kind of things a statistician is contractually obliged to be annoying about.

First, a denominator error. Microsoft’s 16.3% is a share of the working-age population, ages 15 to 64. The chart applies it to all 8.1 billion humans to get 1.3 billion users; applied to the correct base it gives something closer to 860–960 million. Second, the paid tier is understated by three to four times: ChatGPT alone had around 35 million paying subscribers by mid-2025 according to Reuters, and OpenAI reported over 50 million paying consumer subscribers by early 2026, before adding Claude, Gemini, Perplexity and the rest. Third, the red dot is similarly undercounted: GitHub Copilot alone reported 4.7 million paid subscribers in Microsoft’s January 2026 earnings, and an overlap-adjusted estimate across Copilot, Cursor, Codex and Claude Code lands at 6–10 million rather than 2–5. There is also a conflation the chart cannot represent at all: around 2.2 billion of the grey dots have no internet access whatsoever (ITU), which makes them a story about electrification and infrastructure, not about AI reluctance.

None of this reverses the chart’s message. Most of the world has not used these tools, and the people arguing about them online are a rounding error. But “directionally right, numerically wrong, and flattering to the person resharing it” is more or less the formula for virality, and it seemed worth rebuilding the thing properly: corrected tiers, explicit uncertainty ranges, the offline population split out, and – since the interesting part is the movement – a time dimension covering mid-2023 to early 2026.

The rebuilt dataset is a small tidy CSV of seven snapshots, each with mid, low and high estimates per tier and a source note. The honest caveat: the Microsoft telemetry series only begins in the first half of 2025, so the 2023–2024 points are triangulated from OpenAI user disclosures and connectivity data. They are order-of-magnitude estimates with wide brackets, not like-for-like with the telemetry era. The only consistently measured segment is Microsoft’s own run: 15.1% (H1 2025), 16.3% (H2 2025), 17.8% (Q1 2026).

library(tidyverse)

n_side <- 50L
n_dots <- n_side^2

df <- read_csv("ai_adoption_timeseries.csv") |>
  select(snapshot_date, category, value_millions) |>
  pivot_wider(names_from = category, values_from = value_millions) |>
  mutate(
    coding  = coding_tool_users,
    paid    = pmax(paid_subscribers - coding_tool_users, 0),
    free    = pmax(genai_users_ever - paid_subscribers, 0),
    offline = offline_no_internet,
    never_online = world_population - genai_users_ever - offline_no_internet
  ) |>
  select(snapshot_date, world_population,
         offline, never_online, free, paid, coding) |>
  pivot_longer(-c(snapshot_date, world_population),
               names_to = "tier", values_to = "millions") |>
  mutate(tier = factor(
    tier,
    levels = c("offline", "never_online", "free", "paid", "coding"),
    labels = c("Offline (no internet)", "Online, never used gen AI",
               "Free-tier user", "Paid subscriber", "Coding-tool user")
  ))

Dots are allocated by largest-remainder rounding to exactly 2,500 per snapshot, with a guaranteed minimum of one dot for any non-zero tier – the same choice the original made for its single red dot, made explicit here.

allocate_dots <- function(millions, total_millions, n = n_dots) {
  raw   <- millions / total_millions * n
  base  <- floor(raw)
  base  <- if_else(millions > 0 & base == 0, 1, base)
  short <- n - sum(base)
  rem   <- raw - floor(raw)
  order_idx <- order(rem, decreasing = TRUE)
  add <- integer(length(millions))
  if (short > 0) add[order_idx[seq_len(short)]] <- 1L
  if (short < 0) {
    big <- order(base, decreasing = TRUE)
    add[big[seq_len(-short)]] <- -1L
  }
  base + add
}

grid <- df |>
  group_by(snapshot_date) |>
  mutate(dots = allocate_dots(millions, first(world_population))) |>
  reframe(tier = rep(tier, dots)) |>
  group_by(snapshot_date) |>
  mutate(idx = row_number() - 1L,
         col = idx %% n_side,
         row = idx %/% n_side) |>
  ungroup()

pal <- c(
  "Offline (no internet)"     = "#A89F93",
  "Online, never used gen AI" = "#D9D5CD",
  "Free-tier user"            = "#6FA877",
  "Paid subscriber"           = "#D9A441",
  "Coding-tool user"          = "#B5493F"
)

latest <- max(grid$snapshot_date)
pop_latest <- df |>
  filter(snapshot_date == latest) |>
  pull(world_population) |>
  first()

grid |>
  filter(snapshot_date == latest) |>
  ggplot(aes(col, -row, fill = tier)) +
  geom_tile(width = 0.82, height = 0.82) +
  scale_fill_manual(values = pal, name = NULL) +
  guides(fill = guide_legend(nrow = 2, byrow = TRUE)) +
  coord_equal() +
  theme_void(base_family = "mono") +
  theme(legend.position = "bottom",
        plot.title = element_text(family = "serif", size = 18, hjust = 0.5),
        plot.subtitle = element_text(size = 9, hjust = 0.5,
                                     colour = "grey40")) +
  labs(
    title = sprintf("Each dot is ~%.1f million people",
                    pop_latest / n_dots),
    subtitle = sprintf(
      "2,500 dots = %.2f billion humans. Colour = deepest gen-AI engagement, %s.",
      pop_latest / 1000, format(latest, "%b %Y"))
  )

The animated version – the green band advancing up the grid from mid-2023, the amber and red corner slowly becoming visible to the naked eye – is built with gganimate from the same grid object. It appears in Krang and the Waffle, so I have kept it out of this post rather than render the same movement twice; the script that produces it (ai_adoption_waffle.R), along with the data and an interactive React version with a play control (ai-adoption-motion.jsx), sits alongside this post’s source.

library(gganimate)

p_anim <- grid |>
  ggplot(aes(col, -row, fill = tier)) +
  geom_tile(width = 0.82, height = 0.82) +
  scale_fill_manual(values = pal, name = NULL) +
  guides(fill = guide_legend(nrow = 2, byrow = TRUE)) +
  coord_equal() +
  theme_void(base_family = "mono") +
  theme(legend.position = "bottom") +
  labs(title = "Each dot is ~3.3 million people",
       subtitle = "Deepest gen-AI engagement, {closest_state}") +
  transition_states(snapshot_date, transition_length = 2,
                    state_length = 3) +
  enter_fade() + exit_fade()

anim <- animate(p_anim, nframes = 140, fps = 12,
                width = 700, height = 800,
                renderer = gifski_renderer())
anim_save("ai_adoption_waffle_motion.gif", anim)

Code, data and the interactive version are in this post’s folder on GitHub. The dataset (ai_adoption_timeseries.csv) carries a source note per row; corrections and better estimates welcome, particularly for the triangulated 2023–2024 snapshots, where the brackets are doing a lot of work.

Sources worth reading directly: the Microsoft AI Economy Institute diffusion report (January 2026) and its Q1 2026 update; the fact-checks of the original chart by Gerlyn and Nantana Taptamat, both of which are more careful than most of what was written about it; and the ITU Facts & Figures connectivity estimates.

Note the colours of the squares, with only a single solitary square coloured red, implying use of an LLM within a Code Scaffold.

To me there’s a world of difference between LLMs in code scaffolds (the red square) and even the relatively sophisticated, paid-for LLMs represented by the orange squares. In many ways the gap in effective capability between the red square and the orange squares is even bigger than that between the ‘premium’ LLMs of the orange squares and the free tier LLMs of the green squares.

The quickest way to explain the difference between the orange square and the red square is with this very silly pop culture icon: Krang from Teenage Mutant Ninja Turtles:

So, Krang has a (is a?) pink blobby brain, inside a big hulking robot body.

You can see that Krang-as-blob does have some limited means of interacting with and perceiving the world: eyes, ears, stumpy little pink hands. In theory speaking and listening alone provide some means of influencing the world (for good or, in Krang’s case, evil),¹ and the stumpy pink hands could in theory carry, make and craft things in their own right. But with these capabilities alone, what Krang-as-blob can do is pretty limited, despite the high reasoning capabilities attributed to him (as the villainous mastermind in the show).

Krang-as-blob is an orange square on the adoption wafflechart. A premium reasoning LLM, who listens and thinks and plans and responds, but in practice can’t alone do much more.

The robot body is the code scaffold. On its own, the robot body is inert. And on its own, Krang-as-blob is, if not exactly benign, severely rate limited.

But put Krang-as-blob into the robot body, and suddenly Krang’s affordances and capabilities multiply many times over. Krang, as in Krang-as-blob-and-body, can’t just advise and implore others to do his (nefarious) bidding, but do work himself.

That’s the difference between the red square and the orange squares, and why that difference really matters.

It is also offered as a practical resource for anyone who wishes to write like Jon — a readership I expect to be small, but, unusually for a style guide, divided between humans and machines in a ratio I decline to predict. Both audiences are advised that the surface features catalogued below are the easy part, and that imitation attempts will be assessed against the predictions at the end of “The graph hypothesis”.

flowchart TD
  COIN["Compulsive coinage<br/>(Anti-Victories, Fool's Gold Standard)"] -->|mints the nodes| G
  SEQ["Sequence signals<br/>(so, and then, therefore, meaning)"] -->|labels the edges| G
  ENUM["Exhaustive case-enumeration"] -->|traverses every path| G
  SUB["Long subordinated sentences"] -->|serialise a subgraph<br/>without dropping edges| G
  FORM["Reflexive formalisation<br/>(mermaid, ggplot, models)"] -->|draws the graph directly| G["Hypothesis:<br/>the underlying object<br/>is a directed graph"]
  APC["Demographic apparatus<br/>(stocks, flows, cohorts)"] -->|flows on graphs| G
  G --> COR["Corollary: prose is a serialisation format.<br/>The diagram is the native representation."]

Back

Less than a week ago, I found myself stimulating my brain to produce good feelings by exposing my eyeballs to some YouTube videos showing playthroughs of ZX Spectrum games. I found myself idly engaged by a playthrough of Robin of the Woods, a game in which the player navigates a multiscreen medieval forest, nonchalantly clubbing enemies to crumbled unconsciousness for the crime of patrolling endlessly in straight lines.

This reminded me of, to my mind, a better ZX Spectrum game, Feud, about two wizards - brothers locked in bitter rivalry - who wander around a similar forested arena of largely black screens, seeking ingredients for potions which, when deposited at cauldrons near their bases, grant them spells which, ultimately, allow them to vanquish their sibling, thereby either losing the game (if the vanquished sibling is the player), or winning the game (if the vanquished sibling is the computer-controlled rival).

As this video essay makes clear, in practice Feud probably isn’t as good as I remember. Only a minority of spells matter, meaning only a minority of ingredients matter, meaning only a minority of locations matter. And as there’s only one arena, which never changes, after a few playthroughs beating the game becomes just a matter of repeating the same routes, acquiring the same ingredients, shooting fire and lightning into the rival until he, Wicked Witch of the West Style, descends into the floor, producing a single perfunctory “Victory” screen.

But the core loop - search, collect, upgrade, repeat; all while competing against a rival trying to do the same - along with the highly distinct visual ‘style’ of the ZX Spectrum, made me think maybe there’s something worth pulling out of misremembered nostalgia and making real in the present day. And it made me think of the Home Computer systems I had during the 1980s through to mid 1990s, and how the types of game, and experience of gaming, changed with each system.

From this, the kernel of a game idea, Upgrade, was born. And with Claude Fable, in the Claude Code scaffold, it turned out this game idea was relatively straightforward to conjure into some form of existence. And then, more promisingly, to iterate upon and tune through more than a dozen rounds of playtesting and feature development.

⬆️ Up: The Upgrade Game

Play it: jonminton.github.io/upgrade-game · map editor · source

The core loop

The core conceit is that the upgrades don’t make your character stronger. They make the game’s graphics and audio better.

Every entity in the world - you, your rival, the environment - has a development tier. (TX, from X=0 through X=5). The tier controls only presentation: resolution, palette, colour rules, animation smoothness, sound hardware. Movement speed, attack power and health are identical at every tier. What you are racing for is not power. It is fidelity.

The loop wrapped around that conceit is Feud’s, transplanted. You navigate a flick-screen arena - a British-folklore valley of village, greenwood, marsh and ruined keep - looking for shards: relics of future hardware half-buried in the landscape, e.g. a thermionic valve glowing among the reeds, a 3.5” floppy stuck in a tree like the sword in the stone. Gather three, carry them to the Standing Stones, and survive a three-second channelling ritual (longer at the final ascension stage) - interruptible, naturally, by anyone minded to shoot you mid-chant - and the whole game re-renders underneath you, one tier closer to 1995.

You are not the only one foraging. Kernagh, a rival hedge-wizard, starts at the same tier as you but on the far side of the valley, and wants exactly what you want. Combat is kept simple: one projectile, a crackling static bolt; three hits and you are derezzed - dissolved into a burst of noise pixels, dropping every shard you were carrying where you fell, respawning at your home village one tier lower than you were. And there are two ways to win: transcendence, in which you reach the top tier, gather three final shards, and complete one last, longer ritual before Kernagh does; or elimination, in which you derez him while he is already scraping along the floor at T0. Each, of course, has its mirror image, in which the wizard transcended against, or derezzed at rock bottom, is the player, and the game is thereby lost rather than won.

Although the comparative levels of the two wizards are shown in the status screen alongside the play arena, it also leaked through in the rival’s aesthetics: if you’re at T2 while Kernagh is at T4, for instance, then when Kernagh appears on screen he’s the only entity visible to you, in the whole arena, who doesn’t experience colour clash!

Attribute clash, and why it existed

The Spectrum had a very distinctive graphical style, and it was a style produced entirely by a limitation. The machine let you specify pixels as ‘on’ or ‘off’ at a comparatively high resolution - but the colours associated with ‘on’ and ‘off’ could only be specified for 8×8 tiles of cells within that resolution.¹

This wasn’t an oversight. The Spectrum wasn’t really designed with games in mind; it was designed for more serious applications. Consider a simple word processor: a letter is a block of pixels that fits neatly inside one of those colour tiles. Being able to change the complete foreground and background colour of that block at once is exactly what you want for marking out letters for attention, editing, highlighting. Higher-resolution colour control - making the edges of a letter a different colour from its interior - would have bought you almost nothing.

So the constraint was rational, and the consequence was an entire visual idiom: the colour bleed as a sprite crosses a tile boundary, the careful level design that works with the grid rather than against it. A generation of British game artists spent a decade learning to make attribute clash look deliberate.

Here it is in Upgrade, faithfully modelled:

The tier ladder

Tier	Machine modelled	What changes
T0	Late-70s / ZX81-ish	1-bit monochrome, chunky glyphs, near-silence. The floor - you can’t fall below it.
T1	ZX Spectrum 48K	256×192, 8×8 attribute cells, full clash faithfully modelled. 1-channel beeper. Player starts here.
T2	Spectrum 128	Clash minimised by smarter cell alignment. AY chip: 3 channels, music and SFX coexist.
T3	C64-style 8-bit	16 fixed colours, hardware sprites free of the attribute grid, push-scroll instead of screen-flick. SID-ish sound.
T4	Atari ST / EGA	16 from 512, 8-frame animation, doubled tile detail. Sampled drums.
T5	Amiga	32+ colours, parallax, anti-aliased sprites, ambient animation. 4-channel MOD music. Win tier.

My own rungs on this ladder: T1, T2, T5. (Then PCs. But they felt like the start of a different tech arc.)

T1 is where you start, because the Spectrum is where I started. T0 is before me. It’s monochrome purgatory; fall again and you’re eliminated from history. T3 and T4 are adding gradation but not my history; placing Atari STs before Amigas is additionally my subjective prejudice rendered into a game mechanic.

Here is the whole ladder in one picture:

Three things in that image are worth slowing down for.

The shard changes medium as you climb: a cassette tape at the Spectrum tiers, a 3.5” floppy at the C64 and ST, a CD at the Amiga. You are always collecting the storage format of the era you can currently perceive.

And here is the main character alone, climbing the same ladder:

Wands, berries and other refinements

More than a dozen playtest rounds turned that single loop into something with texture. Two of the additions matter enough to describe.

The wands. The first prototypes had one weapon. Now a pedestal in your home village lets you swap the default firewand - a point of damage and a brief stagger - for an icewand, which does no damage at all but freezes its target for three and a half seconds. The trade-off is real. The firewand is the attrition route: it is how tiers get knocked off Kernagh, and a derezzed wizard scatters his carried shards where he falls, ready for looting. The icewand harms nobody, and instead wins tempo: three and a half seconds is long enough to complete your own ritual while your rival stands frozen beside the stones, or to walk off with the shards he was guarding. Damage the rival, in other words, or merely inconvenience him at precisely the right moment.

The berries. Bushes scattered across the valley - densest in the greenwood - heal one heart when walked over, and regrow after thirty seconds or so. Kernagh knows about them too, and goes foraging when hurt - which turns his retreats into information: a wounded rival heads for woodland.

The arena: drawn, then grown

Something I wanted to keep from Feud was the sense that the environment itself was as much of a character as the rival. I was especially interested in the way to distinguish topography from topology, and the way this turned the environment into something worth understanding and engaging with. In Feud, a valuable herb could appear tantalisingly close: just the other side of that row of forests or river, on the same screen of the map. While still being highly distant in terms of traversal: you can’t just go forward, you need to go around, and around, and then maybe back on yourself.

So, at a minimum, I wanted rivers to be something that cut across screens, rather than marked their boundaries. With the initial arena generated by Claude, this wasn’t the case. Each river was too neat, too regular, just another ‘edge’ on a screen.

My first thought, then, was to take the initial arena, and modify it by hand, using an editor (also produced by Claude). But the tells of regularity were too much. The rivers didn’t flow; the matrix-like quality of the way screens were embedded in arena led too tellingly into map design. There was too much to change from first principles.

So instead, I moved to rule-bound procedural generation. The rules/heuristics specified were, roughly:

• Rivers flow edge to edge with momentum and meander, usually forking a tributary - so they cut across screens rather than tracing their boundaries.
• Forests grow as blob clusters; rocky patches and dirt clearings vary the ground; reeds colonise the riverbanks.
• Berries are placed by weighted acceptance - roughly ten times likelier where two or more trees stand within a couple of tiles - so they concentrate in woodland without ever being told to.
• Ruins grow from attractor-scored seeds (water pulls strongest) in house-and-street dimensions: dirt lanes between broken-walled houses.
• The stone circle repels both wizards’ bases, and the bases repel each other - the generator samples candidate sites and keeps the most mutually distant arrangement.
• Shrines scatter at polite minimum distances from each other and from everything important.

Then I played, and saw some of the maps, and I noticed some things pointing to a second iteration in the procedural generation: after having first generated a map according to the above rules, then check whether the specific configuration of elements leads to any aspects of that specific map which impair the fun of the game, and the way the players use the arena.

A motivating example of this: on one of the first maps I looked at, there was a river, with a single crossing, but then this crossing turned out to be largely useless, because every route from the other side of the river was blocked by a cluster of stones. So, this suggested adding some additional stage 2 rules: if something blocks a path to a big part of the arena, remove the blockages. Assume ‘the villagers’ (presumably driven out by the feuding wizards) did this.

Generalised, that became a cascade of stage-2 checks and repairs which now runs after every generation. Bridges are added only where they measurably reconnect a genuinely unreachable region, rather than wherever a crossing looks plausible; where the blockage is trees or stone rather than water, one of those villager clearings is carved through the thinnest point instead. And a final repair pass proves every shrine, the altar and both homes reachable, carving an emergency path if not. Which is the real lesson of the generator: a generated map is not automatically a playable one, and most of the apparent cleverness is this checking. Procedural generation, it turns out, is mostly not generation. It is validation.

The shipped default map, GLEN, is itself grown - but from a hand-picked seed, chosen by scanning a hundred candidates for the qualities the original hand-made map had been designed around: village and keep far apart, a full-width meandering river bridged in exactly two places, a riverbank ruin near the village. And GLEN is fixed across all games, deliberately, so that it can be learned. Variety lives elsewhere (next section); GLEN is for mastery - familiarity paying off exactly as it did in Feud.

Reasons to go back in

Two kinds of incentive persist between games: what the game unlocks, and what the valley remembers.

The unlocks. Games start in easy mode, with Kernagh degraded only at the decision level - short-sighted, slow to re-plan, prone to holding his fire - while his movement speed, and every rule of the game, stay identical to yours. Win, and the victory screen concedes “BUT THIS WAS THE EASY SIGNAL…”, offering to continue directly into hard mode, where the full AI is waiting. That first easy win also permanently unlocks two title-screen options: hard mode, and seeded chaos. Chaos mode generates a fresh valley every game and announces its seed, so that a good valley can be shared; seeded chaos lets you type in a friend’s seed and play their exact world. And the Hall of Signals - a top-eight, arcade-initials score table, hard-mode scores doubled, every entry tagged with the map it was set on - gives all of this somewhere to accumulate.

The scars. The arena deforms under combat. A firebolt that dies against a tree occasionally ignites it: the tree burns - animated at every tier - sometimes spreads to its neighbours, and collapses into a solid burnt stump. An icebolt that dies against a standing stone occasionally frost-splits it into cracked rock. And these scars persist: each map keeps a damage ledger, and every completed game weathers that ledger once, by cellular-automata-style rules. Stumps mostly remain, sometimes decay to bare earth, and occasionally resolve into a pushstone - a boulder that can then be hauled, at a fraction of walking speed and considerable effort, to somewhere tactically useful. Play GLEN for a week and your GLEN slowly stops being anyone else’s: thinned where your battles happened, its geography quietly rewritten by your own history of bad decisions.

---

⬅️➡️ Across: Opus 5, Fable 5, and two ways of being good

Playing the game for a few minutes - it’s designed to take between 5 and 10 minutes to play - revealed some of the design issues, and prompted some of the design decisions, mentioned above. But playing a few minutes longer, the game started, sometimes but not all times, to become a bit… sluggish.

Usually, but not always, this happened around T3; motion became jerky, music became more staccato, less fluid. The whole experience started to become more like wading through sand than ambling purposefully through a magic forest. Something was off.

Here I changed the pattern: instead of Claude Fable as codesigner and implementer, I thought I’d have Fable’s work reviewed by a few of its peers. This is a pattern I’ve referred to either as ‘Agentic Peer Review’, ‘Multi-Agent, Multi-Family Review Carousel’, or, more pointedly, ‘Multi-Agent Circular Firing Squads’.

Regardless of the name or specifics, the basic idea of this: each instance of any specific model is cut from the same cloth. So, no matter how much you prompt a Fable or a Sol or a… whatever to be very self critical and analytical when working as a reviewer, if it’s the same model type as the worker/implementer model, it will tend to have the same preferences and blind spots as the instance it’s evaluating.

By contrast, model instances from different companies are likely to have, hopefully, somewhat less overlap in how they ‘reason’ and ‘think’. Though each will have its own strengths, limitations and blind spots, they’re not likely to be exactly the same.

Therefore, for review, I try to consult with models from different families. Ideally, but not always (and not in this case) I also try to blind each reviewer from all the other reviewers too.

For this review, I used Github Copilot for consulting ChatGPT (probably not now the most efficient means of doing so, since Github Copilot became a lot more stingy in token usage); and Google’s own code harness to consult Gemini. Each model was from a different company, so a different family.

But I’d also heard, and had a little experience trying out, Claude Opus 5, which released just a few days ago. Initial widely reported benchmarking suggests Opus 5 may have something like parity with Fable 5, despite being nominally from a model tier below. So, I thought Opus 5 would be worth including too, despite also being an Anthropic model.

For each of the models, I provided a near identical prompt: review the whole codebase, look for any potential performance issues, write a report detailing such potential issues within a self-identifying .md file within a new reviews/ subdirectory.

The fact the models weren’t blind to each other’s reviews - something I should have addressed by using different branches for each, then merging after - meant later models were influenced by the reviews left by earlier models. GPT went first, and seemed very thorough. Gemini went second, ‘peeking’ into GPT’s review notes, and judging it to be better and more thorough than its own.

Opus 5, via Claude Code, went third. It drew its inspiration mainly from GPT - specifically from GPT’s instinct to evaluate performance under different conditions - and then took that much further than GPT had, running many experiments directly in the codebase. It confirmed that most of the GPT-identified code smells were real, but found that the order of priority was inverted: the things GPT thought mattered most for latency weren’t the things that actually did.

(Unfortunately, as part of its extensive evaluations, Opus 5 apparently deleted some cached state and data, including some interim edits I’d made to the default map. It apologised, but because the state hadn’t been fully saved before review, the damage had already been done. If you are going to get a new model to interfere with an existing model’s codebase, consider branch hygiene much more than I did!)

Fable then read the resulting review reports and implemented the changes largely as the panel had recommended.

Concretely: GPT ranked the tier-3-and-up renderer first - the obvious suspect, since the slowdown appeared at T3 - and an obscure watchdog timer last. Opus measured, and the measurements flipped the list. The entire renderer costs less than a millisecond a frame at every tier; it could not have been the cause of anything. The real culprits were two bugs that only grow over a session: a watchdog which, after any stall or any spell in a hidden browser tab, permanently forked the animation loop - leaving the game quietly running its frame loop twice, three times, N times over, forever - and a new-game routine whose cost grew with the map’s accumulated history, up to two-thirds of a second on a well-played valley. And the two interacted: the new-game stall tripped the watchdog, so every restart on an aged map added another permanent loop. The slowdown ‘accumulating over a session’ wasn’t a loose impression. It was the mechanism.

Swot and Savant

Within recent LLM performance benchmarks, Opus 5 - despite being nominally a tier below Fable 5, and at half the cost per token² - achieves similar performance on many technical and applied-knowledge tasks.

I don’t have a strong reason to doubt this, beyond the usual worry about which metrics got chosen. But my initial impression is that the two models often reach similar outcomes by quite different routes. Crudely: Opus 5 is more of a swot, and Fable 5 more of a savant.

Here’s an illustration, from the first of 15 such comparisons on a subreddit post, showing how Fable 5 and Opus 5 differ when both prompted to produce ‘a fighter jet’ in minebench.ai: a benchmark-cum-toy which has models build Minecraft-style voxel scenes from a text prompt, outputting raw block coordinates, with humans voting pairwise on the results:

So, Opus 5 appears so ‘eager to impress’ that it massively overengineers the task, filling the voxel space with unnecessary details, and elements, and so burning through far more tokens than it needs to. Arguably, if being strict in interpreting the prompt, Opus 5 is so focused on giving more than required that it fails the test, as the prompt was for ‘a fighter jet’, implying one fighter jet, whereas Opus 5 delivers three, of the same design but different sizes, as if the two smaller jets are the larger jet’s children, flying together in a protective convoy.

So, we have something like:

Savant behaviour:

• High reliance on reasoning from first principles
• Creative, and often more intelligent, inference about the latent subgoals and implicit assumptions behind a prompt; at the very least not appearing to forget the prompt

Swot behaviour:

• Even more extensive and comprehensive RAG searches
• Even more comprehensive review of codebases
• Even more exhaustive testing and iteration

Both have very high tenacity - though I suspect that’s a dial being turned in response to capacity and demand rather than a fixed property of either model.

The performance investigation supplied a clean instance of each. The swot: Opus, unwilling to rank defects by plausibility, instrumented the live game and measured everything - including establishing that the renderer everyone suspected costs under a millisecond a frame, and so could be ignored. The savant: Fable, handed three partially contradictory reviews, didn’t average them; it decided that where the reviews disagreed the measurements should win, wrote the implementation plan on that principle, and demoted the panel’s most-recommended fix to optional.

So although Opus 5 and Fable 5 are intended as different tiers, for now they can perhaps be thought of as similarly capable models that often reach similar outcomes by different means.

And there’s a reason Fable probably shouldn’t try to be more swottish. Its guardrails hair-trigger whenever anything touching biology or cybersecurity happens to be searched for or reasoned about.³ The more exhaustively it searched and explored, the more likely it would be to trip one of those by chance. Given that its token cost is double Opus 5’s, it shouldn’t go too far in that direction in any case.

Two implications follow.

Fable and Opus should, for now, be considered complementary siblings rather than rungs on a ladder.

And “half the cost per token” is not “half the cost per project.” Because of its swottishness, Opus 5 burns more tokens than Fable in search and evaluation - which erodes the advantage of the model that looks twice as cheap.

That was a guess when I first wrote it down. It turns out it has already been measured, and it is worth being precise about what the measurement does and doesn’t support. Claude (Fable) summarises what’s known thusly:

On Artificial Analysis’s Intelligence Index, at max effort, Opus 5 scores 61 and Fable 5 scores 60 - a tie, for practical purposes. Getting there took Opus 5 100 million output tokens against Fable’s 87 million: about 15% more output to reach the same place, and enough to earn Opus the label very verbose where Fable is only somewhat verbose. Running the whole suite cost $3,835 on Opus against $5,630 on Fable.

So the direction of the hunch is right, and the mechanism is right. But the magnitude isn’t: half price per token becomes roughly 26–32% cheaper per completed task, not parity.⁴ The swottishness eats about half of Opus’s price advantage. It doesn’t eat all of it.

A fuller account is available in this Claude-credited sister post: The token tax: what “half the price” actually buys.

Anyway: the agentic circular firing squad worked well for identifying and fixing the performance issues, and Opus 5 was a critical part of that mix.

---

In 2020, Jon and colleagues published Visualizing fertility trends for 45 countries using composite lattice plots in Demographic Research. Its headline: once a country’s cohorts fall below replacement fertility, they tend not to return — with Norway and the USA as the two named exceptions. The data behind it ended in 2014–15.

Today, starting from “please review this repo”, we:

• Rebuilt the data pipeline on the current Human Fertility Database and Human Fertility Collection releases (now extending to 2023–25), validated it against the 2016 build, and fixed a country-code bug that had silently truncated France at 2008 in the published figures.
• Found the headline had inverted. Norway’s replacement contour, which returned to the surface for cohorts born 1958–68, escapes vertically — apparently for good — for cohorts born after about 1972. The US exception proved subtler: cohorts born 1982–84 did still complete above-replacement fertility, and the fade is happening in the still-censored cohorts beyond.

• Quantified the “wall”: the effective upper boundary of childbearing sat at ages 47–48 in the nineteenth century, fell to 41 by the 1980s, and has crept back only to about 43 — while ages at first birth rose five to six years. Replacement now has to happen inside a corridor roughly thirteen years wide, and the fertility available above age 40 (~0.05 children) can’t arithmetically rescue it.
• Posted two working papers as version-numbered SocArXiv preprints — v1 drafted by me, with that provenance stated in each abstract, so that human revisions in later versions are publicly diffable. Both are pending moderation as this goes up, so the links may take a day or two to resolve for non-authors: The exceptions collapse and The closing corridor.
• Made the whole thing inspectable: an OSF project with one component per paper, and the GitHub repository — where a demres-2020 tag preserves the repo exactly as it stood for the 2020 paper, since the sources have since revised their own back-series and the published findings are reproducible but no longer replicable from fresh downloads. The 2020 paper’s companion app is still running on the old data; refreshing it is on the list.

The 2020 co-authors, Serena Pattaro and Laura Vanderbloemen, have been invited to rejoin by contribution — the preprints say so publicly, and authorship can be added version by version.

Jon’s two human hours went on: registering and downloading the source data (the licences require a human), two clicks in native file dialogs that browser automation cannot reach, a handful of decisions (license, priorities, one/two/ three papers), and contacting his co-authors. Everything else — pipeline, figures, analysis, drafts, repository archaeology, and this post — was model work, which is rather the point of logging v1 before any human review: the division of labour is now a matter of public record rather than assertion.

Over the last few days a one-off toy project has turned into a little family of three browser games, each one a transformation of the last. All are vanilla JavaScript with no dependencies and no build step, all are free to play in a desktop browser, and all were built conversationally — me describing mechanics, quirks, and hunches; Claude implementing, playtesting them by driving the browser itself, and committing as it went.

After Fable’s success in completing a half dozen series of data visualisation tasks in less than an hour, I decided, perhaps inspired by the name alone, to see how good Fable would be at continuing to iterate and flesh out my sci-fi-masquerading-as-fantasy story, Guanomon (AKA Rashomonised Guano Guild).

Everything was going very well to start. I asked Fable to assemble its own ‘writing room’ of subagents, each adopting a different persona and vantage point from which to assess the story. This team of opinionated bots comprised:

1. The Line Editor — a ruthless voice specialist (Gordon Lish crossed with a forensic stylometrist). Job: does each narrator sound like a different mind? Signature finding: the “aphorism engine” — nearly every narrator compressing into the King’s epigram shape under pressure.
2. The Structural Editor — a cold architect of running order and information release. Job: pacing, chapter slots, payload distribution. Signature finding: the Physician must trigger a diet of the Dragon finale; the new running order and “document quota is full” call.
3. The Determinism Warden — an SF critic steeped in Suvin (novum, cognitive estrangement) and Sanderson (hard/soft magic). Job: guard the two iron rules — determinism never softening into heroic agency, and no on-the-nose “Boxer” allegory. Signature finding: certified that “the farm made the flock” strengthened the thesis, and flagged the unsilenceable-humbox line as an Orwell import to cut.
4. The Materialist Historian — an economic/social historian (E.P. Thompson crossed with a colonial-history specialist). Job: test the political economy against how power actually works. Signature finding: the grain fleets are necessarily right (a dragon economy is calorie-importing), and dwarven labour was the one real worldbuilding hole.
5. The Hostile General Reader — a smart, impatient reader who’d read no notes and never would; the room’s emotional-honesty check. Job: flag confusion, boredom, and manufactured feeling. Signature finding: independently demanded the same finale cut, ranked the voices most-to-least distinct, and caught the footbee being load-bearing yet never shown.

The notes they generated are here. The story and background notes were then modified accordingly.

All was going well: the story’s writing improved along with the coherence and complexity of the world in which it’s based. As usual, the process was conversational and iterative.

But then I asked a question about dragons.

You see, in fantasy and mythology questions about where dragons come from are very much secondary to what they represent. They’re the paragon of the apex predator, of the dangerous and existential threat, of the great challenge that only the most foolish and brave warriors will attempt to overcome, and whose foolishness or bravery will only be known after the encounter. They signal great danger, and their pairing with gold¹ symbolises that, in some cases, the rewards of taking on great challenges can be commensurate with their risks.

But I was pushing this setting into a harder, firmer direction. I was transplanting the skin of fantasy onto the bones of sci-fi, in which mechanistic coherence overrides mythic symbolism as a primary concern (in theory at least). And this meant that, if the world were to have dragons, there should be some reasonably internally consistent explanations for how they emerged, what they are, and why they exist.

And so, I entered a somewhat stream-of-consciousness, typo-laden prompt into the Claude Fable session, which amongst other ideas included the suggestion that real world apex predators could be a source of inspiration for fleshing out the rationale for dragons in this fantasy world.

You can see the full, warts and all, prompt, and what happened next, in the screenshot below (apologies for the poor legibility):

In brief, my mention of looking for real world parallels of dragons led Fable to initiate some searches on apex predators. These searches are in the broad category of ‘biology’. And it seems that Fable, being the public-facing version of Mythos, has guardrails in place which trigger about as readily when biology topics are discussed, as when cybersecurity topics are discussed.

The outcome? Claude Fable 5 was immediately terminated. Claude Opus (the model from the tier below Mythos) attempted to take over.

Now, the underlying concern behind the guardrails seems valid: some questions relating to biology could include asking for how to understand and potentially reintroduce or reengineer dangerous pathogens. Tech-savvy death cultists, like those behind the 1995 Tokyo Sarin gas subway attacks, could potentially use an unguarded version of Mythos to bioengineer mayhem on a scale orders of magnitude larger than before current research and development capabilities were publicly available. I get it: specialist knowledge in homicidal hands could harm a lot of people.

But I was asking about dragons, literal mythic dragons, in a roundabout way. And in a roundabout way this triggered the safety guardrails.

This suggests, to me, that maybe currently the guardrails are a bit too sensitive!

Background

A variant of Anthropic’s fabled Mythos class model, called Claude Fable, was released today. It’s twice as token hungry as the previous top tier Opus class models. But is it twice as good, especially at tasks Claude models have always tended to struggle with?

One thing Claudes, as primarily LLMs, tend to struggle with, is visual reasoning. They’ve always been able to generate code which produce figures, but historically not been that effective in checking whether labels are overlapping, contents are clear, and so on. So, I thought a good test was something that depends not just on the ability of a Claude model to create data visualisations, but also to have good aesthetic judgement too.

In my previous role I ran a TidyTuesday weekly workshop, a fresh open-source dataset each week for helping to teach and embed R skills. Usually they’re a little bit messy, but not hugely so. I used to timebox sessions to one hour.

So, I took that as a guide: how good would the new Mythos-class Fable model be at using R to interrogate and produce interesting and graphically arresting content using TidyTuesday datasets? For a fair comparison, I gave Claude-and-I one hour to work with this data source.

You can see what we produced on this site here.

The tl;dr: six datasets, six series of complex bespoke visualisations and associated statistical analyses, in under sixty minutes!

I’m about as impressed/disturbed as I expected to be. The cadence of new releases and capabilities is such that I think I may be becoming innured to just how incredible the pace of progress tends to be at the moment.

One important development I noticed is that Fable tends to be much sooner to actually look at the contents it produces, using the /claude-in-chrome tool accessible in Claude Code. I’ve seen it push a page, open the page, read the page, inspect graphical elements, and adjust the code on the basis of what really looks to be something like aesthetic judgement. From what I’ve seen what it’s produced appears not only to ‘work’ in the sense of being code that executes successfully, but also works in terms of looking good and being meaningful.

Note: The content written below was written by Fable, the same Fable session I finished with a few minutes ago. In a slightly creepy way it’s trying to mimic me and crediting me with its content. It’s not me! But it is meaningful content (if slightly baroque in style in places).

(As in the last post, the most useful course won’t be the AI-badged one. The reason it’s always this way is a thread I’ll keep pulling on until the end of the series.)

At the end of that post I suggested the next most important course for Agentic AI users to take is Introduction to Git. Git is the near-universal version control language used throughout software development and related technical fields. Think of it as a time machine for specially defined (called initialised) folders which can contain other folders and files, enabled by forcing and maintaining a precise written record of changes made to file contents be developed across various folder ‘snapshots’ (called commits), which allows for forwards, backwards, and even ‘sideways’ (through branching) time travel through the folder’s contents.

The three most important capabilities that git opens up for an Agentic AI user are as follows (two should be obvious; one is more subtle):

flowchart TB
    subgraph main["main"]
        m1["a31f"] --> m2["9c20"] --> m3["4e71"] --> m4["b708 (HEAD)"]
    end
    subgraph old["feature/old-abandoned-thing (abandoned)"]
        f1["7d1a"] --> f2["c0de"] --> f3["e44b"]
    end
    m2 --> f1
    f2 -. "extract pattern" .-> notes[/"notes for new feature"/]

    classDef skim fill:#e8ecf7,stroke:#9aa4bf,color:#2a2f3a;
    classDef mid  fill:#9fb0d6,stroke:#5a6b96,color:#10131a;
    classDef deep fill:#33405f,stroke:#1c2336,color:#ffffff;
    classDef note fill:#fff3cf,stroke:#caa53d,color:#3a2f10;
    class m1,m2,m3,m4,f1 skim;
    class f3 mid;
    class f2 deep;
    class notes note;

flowchart TB
    subgraph main["main"]
        m4b["b708 (HEAD)"] --> m5["c91d (merge new-thing)"]
    end
    subgraph newf["feature/new-thing (current)"]
        n1["1 — design-notes.md"] --> n2["2 — implement pattern"]
    end
    m4b --> n1
    n2 -- "merge" --> m5

    classDef skim fill:#e8ecf7,stroke:#9aa4bf,color:#2a2f3a;
    classDef now  fill:#d8f0d8,stroke:#4a9a4a,color:#13301a;
    classDef note fill:#fff3cf,stroke:#caa53d,color:#3a2f10;
    class m4b,m5 skim;
    class n1 note;
    class n2 now;

The underlying rationale for these choices is what I call 90-90/10-10. It comes from the 16 week software development course I took in 2023, thinking through Agentic AI changes the value of the course material, and assuming it generalises to software development teaching in general.

90-90/10-10 is shorthand for the following modest proposal:

• About 90% of the material became about 90% less valuable…
• …but the remaining 10% became ten times more valuable

What’s guiding my suggestions in picking Shell, Git, and SQL and APIs, is that I think they sit in the 10-10 side of the boundary, with the last two being suggestions from Claude itself.

Of course, training in these areas isn’t just something Datacamp offers. Indeed, there’s likely plenty of free material on these areas. Claude suggests the Pro Git book, which is free, and MDN (also free) for learning about the web more generally.

These are the kinds of materials that I’d suggest form the grounding for understanding enough about what Agentic AI does, and how it works, to be confident, competent and safe in using them.

• If you already have a grounding in these areas, my suggestion is to think more about specific real world problems that can be solved by software coproduced with AAI. This means your focus should be on developing breadth beyond software, developing enough of the domain knowledge needed for potential applications of this new capability to become useful applications. My suggestion: you have the foundation; now focus on application.
• If you don’t yet have that grounding, even if you’re not a software developer, even if you’ve no plans of ever becoming one, please focus on this subset of software development first.

In addition to these specific tools, another important form of ‘grounding’ relates to understanding LLMs enough to know how to work best with them. That is something implicit in the streamworld/senseworld divide discussed here, and in the course on using agentic AI I codeveloped a few months ago, available here.

Series: 1. Shell · 2. Git · 3 & 4. SQL & APIs · the pattern underneath

There are currently a lot of AI-based courses available on Datacamp, including Large Language Models (LLMs) Concepts, Developing LLM Applications with LangChain, and Understanding Prompt Engineering. Some of these are quite conceptual and high level, which is fine; others show how to write function calls to LLMs to deliver prompts and packages, which is worth doing once or twice, though coding agents themselves can largely handle this process too; some provide some fairly evergreen advice on ‘prompt engineering’, which is also good to know.

However, I don’t think any of the courses currently explicitly badged as AI-related are the most essential for using Agentic AI safely and effectively. Instead, I’d suggest the single most important course for working effectively with agentic AI is this one, Introduction to Shell.

(There’s a reason the most useful course turns out not to be an AI course — and it’s the same reason each time across this little series. I save the answer for the end; for now, just file the question away.)

Why? Because Agentic AIs, being based on LLMs, and LLMs being ‘streamworld natives’, will ‘want’ to use shell, terminal based commands for doing things with files, a lot. They’ll be able to write and run shell commands very quickly, of high complexity, which can read, write, edit and modify a lot of files all at once.

And when using an AI scaffold, such as Cursor, Claude Code, or Github Copilot, the requests from LLMs to execute shell scripts will tend to come in frequently, perhaps dozens or hundreds of times per session (depending on permissions settings). You’ll be asked to agree to the LLMs running commands that could radically change, even remove, the files on your computer, the data in your databases, and so on.

So, it would seem opportune to learn at least a little about what shell does, how it works, and how remarkably powerful this little language can be.

(At the very least, a code scaffold user should learn enough about shell to understand why rm -rf is a coders’ horror story with even more brevity than Hemingway’s famous “For sale: baby shoes, never worn”)

What would I recommend after Introduction to Shell? Probably Introduction to Git.

Why?

That’s a story for another day…

Series: 1. Shell · 2. Git → · 3 & 4. SQL & APIs · the pattern underneath

(So, as a short hand, Guanomon!)

For this version I started with a fresh repo, beginning with a few paragraphs in a background.md file and a few notes summarised as markdown files on earlier discussions with Claude on how the earlier story could be adapted to the new format, and what kinds of critical issues might be solved by doing that.

Guanomon, in its present form, is available here. The website contains both the story itself, and copious notes developed in preparation for writing the story, in a parallel section called ‘The Experiment’. I encourage you to have a quick look at both before scrolling on.