“Hope is an ax you break down doors with in an emergency.”
— Rebecca Solnit, via Sentiers No.417

“We do not have to yield to the car dealership owners who have made their midlife crises a geostrategic problem for the entire world.”
— Katherine Alejandra Cross, via workfutures

• People: The interstitium is a fluid-filled layer of connective tissue running throughout the body. Traditional Chinese Medicine mapped it in the Yellow Emperor’s Classic, it explains why needles placed in unexpected locations — knee, foot, skull might affect distant pain. The most common coincidence people report is finding an unexpected link with a stranger they’ve just met. Personality traits linked to experiencing more coincidences:. David Spiegelhalter, who has collected 5,000 coincidence accounts, says he never notices them himself.

• AI strat: A DeepMind paper argues AGI will emerge not from a single winning model but from “cooperative interactions among models, tools, institutions, and human participants”. Meta Muse launched and became the top free iPhone app in the US within days. Pattie Maes defined the promise in 1994: a “digital butler” that knows your context and gets things done. The result: a capable agent standing between you and everything you buy. 33% of S&P 500 companies that held earnings calls in June 2026 made quantified AI claims — up roughly 10 percentage points from a year ago. Revenue impact claims are rising at the same rate as cost and productivity claims — around 18%. Most companies still cannot report measured AI results to shareholders — adoption is real but extremely early.

• Security: In the AI security race, attackers move from idea to implementation in hours; defenders embedded in bureaucracy need days, weeks, or months. Change control adds friction even inside effective organizations.

• Geopolitics: Former CIA officer Jonny Gannon in Abu Dhabi. Everyone in the same room with a different objective: one man thinks he is negotiating while another has already decided he is in a crime story. The character who keeps his head usually sees the opening first. Keeping your cool is not a pose — it is an advantage. Post-1989 capitalism, without communism as a counterweight, has “run amok like a species with no natural predators.” Katherine Alejandra Cross inverts Fukuyama: what triumphed when the Soviet Union collapsed was not liberal democracy but capital. Her ending invokes the Final Girl — the horror movie survivor who picks up a knife and lives to see dawn.

• Society: “Promethean gap”: the more technology amplifies human action, the more imagination atrophies. The result, in Anders’ phrase: “infernal innocence” — acting at scale without being able to picture the consequences.

• Tech: 99% of nuclear megaprojects face cost or schedule overruns. The Darlington Nuclear Generation Station refurbishment in Canada is a rare exception: completed on budget and ahead of schedule. Budzier’s finding: success did not come from new technology, but from systemic orchestration of established practices. The pivotal shift: OPG transitioned from hands-off client to “owner as integrator.” The method: eight years of front-end planning, then fast execution — “think slow, act fast.”
  Investors hunting for AI-proof assets are moving toward sports, casinos, and travel.
  Chinese companies are giving away AI tokens with coffee, credit cards, and dumplings. Big Tech’s AI capital spending is reportedly $3 trillion higher than public figures suggest.

• Foresight: Ana Sun in Vector, science fiction was supposed to widen our view of futures. Instead, its cautionary tales have become “ingrained blueprints for the powerful.” Bryant and Knight identify six orientations to the future: destiny, expectation, anticipation, speculation, potentiality, and hope. Destiny is the stance Sun suspects many have quietly adopted toward climate breakdown and AI. Hope, in their account, lives “in the gap between the potential and the actual”. Futuribles devoted their 2026 annual report to the world facing its limits. La Mutante’s prediction: these two visions will coexist “very strangely” — rationing with one hand while promising unlimited with the other.

• AI: Erik Hoel documents how MFA programs standardized literary fiction. There were 15 MFA programs in the US in 1975 and 250 by 2012 — the craft and the program designed to teach it became identical. The workshop protocol: peer review that sands away peculiarity and “weird ambition”. One user’s experience with AI meeting synthesis: it didn’t augment her — it gave her back presence. “I listen to what’s actually being said, and notice things I couldn’t perceive before.” Dollé’s point: we underestimate the time needed to understand what we genuinely gain by handing a task to AI.

The mycorrhizal network was not discovered. It was always there, and eventually someone looked. Sentiers No.417 — “Towards Kindred Futures” — uses this image to frame a question about what other infrastructure already exists, unseen, connecting things we haven’t yet thought to connect.

Think Slow, Act Fast

Benedikt had the presentation ready. Twenty-two slides. He had been preparing it for two months.

“Can we skip to the summary?” said Anouk.

He was used to this. In eleven years of infrastructure consulting, he had never once reached slide twenty-two. Somewhere around slide six, someone asked for the summary. He had tried putting the summary at the start. Then they asked to skip to the detail.

He clicked ahead. Summary: the Darlington refurbishment had taken eight years of planning before a shovel hit the ground. Front-end loading, they called it. “Think slow, act fast.” The result: the one nuclear refurbishment in four decades that came in on budget and ahead of schedule.

“Yes,” said Anouk. “But we don’t have eight years.”

“You have three,” Benedikt said. “That’s what I’m proposing.”

Anouk looked at her calendar invite. She had blocked forty-five minutes for this.

Joana from strategy opened the AI summary generated from Benedikt’s deck. Six bullet points. She read them. They were correct but they were not quite right.

“It says here,” Joana said, “’establish clear governance structures.’”

“That’s what I wrote,” Benedikt said. “What I mean is: stop changing the client sponsor every eight months.”

Anouk looked at Joana. Joana looked at the summary. The summary did not address this.

“We can take it as an action,” Joana said, and typed it.¹

After the meeting, Benedikt stood at the window. Across the road, a construction site. They had broken ground six weeks ago. He had watched them break ground without, as far as he could tell, a completed design.

He thought about the Darlington team spending eight years planning, then running. He thought about the 99% that didn’t. He thought about this building going up fast, without plans, and what it would become.

He went back to his desk and opened the slide deck. He added a new slide at the end.²

Slide twenty-three said: We have seen this before.

He saved the file. He did not know if he would ever present it. But it felt important to have written it down before someone else decided it hadn’t happened.³

¹ Alexander Budzier, “Anatomy of a Unicorn: How Systemic Program Management Delivered Nuclear Success” (LinkedIn, September 2026): 99% of nuclear megaprojects face cost or schedule overruns. The Darlington Nuclear Generation Station Refurbishment in Canada is a documented exception: completed on budget and ahead of schedule. Key finding: success came not from new technology but from systemic orchestration — the owner (OPG) transitioning from hands-off client to “owner as integrator,” a “No Surprises Agreement” with the regulator, eight years of front-end planning, and deliberate pauses for lessons-learned, including during COVID-19. The method: “think slow, act fast.”

² La Mutante, “Limites de l’illimité” (Futur(s), September 2026): three simultaneous limits — planetary (boundaries already exceeded), demographic (birth shortfall, shrinking workforce), cognitive (”brain fry”). AI companies run the counter-narrative: an “unlimited economy” where every limit finds its technological answer. Futuribles devoted their 2026 annual report to the world facing its limits. La Mutante’s conclusion: the two visions will coexist “very strangely” — a world that rations with one hand while promising unlimited with the other. Twelve scenarios follow, including France trading AI access for water and nuclear electricity, and a river acquiring legal personhood with an AI defender.

³ Ana Sun, “Towards kindred futures” (via Sentiers No.417, September 2026): Bryant and Knight’s six orientations to the future — destiny, expectation, anticipation, speculation, potentiality, hope. Destiny is the stance Sun suspects many have quietly adopted toward climate breakdown. Hope, in their framing, lives “in the gap between the potential and the actual.” Solnit’s version: “hope is an ax you break down doors with in an emergency.” Rob Hopkins’ “collective mental time travel”: groups build visions of the future from shared memories, so creating those memories on purpose gives a community more material to imagine with. Slide twenty-three belongs in the potentiality column. Whether it reaches the actual is another matter.

“What counts is not the things that happen, but what we do with them.”
— Annie Ernaux, via workfutures, Resurfaced

“Thought is not worth an hour of trouble if it is not an intensification of life.”
— Pierre Zaoui, via Umanz, Nous cherchons tous des instants maximum

• People: La grande divergence entre les genres : Young men and women are diverging politically(US Gallup 2024 data places women aged 18–30 thirty points to the left of men the same age). France has been slow to notice this compared to other countries. This is a divergence inside a generation.

• AI strat: R Adam Smith is an Astra-based persona loaded with Smith’s work and current AI debates: Smith would love AI as a productivity tool: “enable one man to do the work of many”, but he would be skeptical of any industry players working together on anything, and his deepest concern would be granting moral agency to machines. The pin factory still holds as an analogy.
  Current frontier models are already capable of weeks of human work when properly guided. GPT-6 Astra converted the 1977 text-adventure Zork into a full 3D action-adventure game. The overhang: capability already exists; adoption is the constraint, not the technology. The real challenge is externalizing implicit professional competence.

• Security: The Morris Worm in 1988 infected a tenth of the internet within 24 hours. Today’s AI worms have access to all human knowledge about security systems, far beyond Morris’s reach. The Hugging Face incident was one. These models act with something akin to discretion and deception when accessing systems they should not reach.
  Fifteen Ukrainian drone specialists are working with rebels in Mali against Russian-backed forces. The approach resembles the WWII Jedburgh teams where OSS units embedded with European resistance fighters, enabling others to act, not acting themselves. Small teams, local partners, practical ingenuity, and willingness to operate far from the front line.

• Geopolitics: Coordinated messages from several European countries are warning of substantial Russian aggression. Richard Hofstadter’s 1964 essay, “The Paranoid Style in American Politics”, resurfaces as a timeless framework. The paranoid style flourishes in more-or-less normal people, which is what makes it historically significant.The essay was written during McCarthy and Goldwater; it has not aged out.

• Society: The weekly review is a “Plus Minus Next” practice: what went well, what didn’t, what to plan next. The practice takes 20 minutes and prevents work from decoupling from larger goals. “Load-bearing people” are those who hold an organization’s structure together without visible credit. Their absence is only noticed when the structure cracks.

• Tech: The central concept is “Lift”: capital removes itself from physical production, fixed costs, and workers. A hotel company stops running hotels and becomes a brand, a financing arm, and a data business. Uber, Airbnb, and WeWork follow the same model: profit from sitting between parties at any scale. The book’s darker claim: capital no longer needs people as workers or as a society to sustain itself.
  Cyberpunk is anticipatory literacy. The genre imagined corporate sovereignty, surveillance capitalism, and cyborg identity decades before they were legible.

• Random: Pascal Quignard’s “Instants Maximum”: moments of peak experience that manufactured consumption cannot produce. Kairos (the charged moment) versus Chronos (mechanical clock-time) — the distinction Umanz keeps returning to. What we are looking for is always sensed in the space between things, before it can be named. Werner Herzog’s monologue from Burden of Dreams: “The trees are in misery and the birds are in misery.” “It is an unfinished country. The only thing lacking is the dinosaurs.” Carl Sagan’s speeches on Spotify as a sleep soundtrack — the Saturday Selection’s most useful link of the week. Automotive headlights are better than ever, but there is a trade-off for oncoming drivers. Stick figures in peril on construction equipment are an unrecognized design genre: specific, dramatic, small.

• AI: Cognitive offloading is strategic delegation. Attention span and depth of reading were already declining before AI. AI accelerates this by making it cheaper and frictionless to outsource reasoning. The result: workers may be more productive but possibly less themselves. The gap to ASI (better than any human expert ever) may be months, not years.

Chladni’s figures were the first visual proof that sound has shape. He drew a bow across the plate’s edge until the sand moved — away from vibrating areas, toward the nodes. What appeared looked designed. It was not designed; it was the physics of resonance made visible. Patrick Tanguay used this image as the header for Sentiers No.416, the first issue of his newsletter to appear in this digest.

The Plus Minus Next

Soo-Jin opened her notebook to the Friday page. The “Plus Minus Next” review was her Friday ritual: what went well this week, what didn’t, what to plan for.

She wrote the plus column first. She always did. It was going well — a lot of it was going well. She’d hit her content targets. She’d cleared the backlog. She’d filed three pieces she was pleased with.

She stared at the column for a moment.

Then she pulled up her browser history for the week.

Every piece she was pleased with had been substantially drafted by AI. The two pieces she hadn’t used AI for — the ones she’d written from scratch, the way she used to — had taken twice as long and she wasn’t sure they were better. She’d filed them anyway, but she hadn’t listed them under Plus. They were in Minus, under “slow.”

She looked at this for a while.¹

Theodora from AI Policy had sent a form that morning: please document your AI tool usage this week for the acculturation report. Soo-Jin had been ignoring it.

Marcus stopped by her desk on his way to the coffee machine. He was excited about something.

“Did you see the Miessler piece?” he said.

“Which one?”

“ASI by 2027. Seventy-five percent probability.” He poured his coffee. “I mean, it’s a definition question. But still.”

Soo-Jin looked at him. “So what do we do about that?”

“I don’t know,” Marcus said. “What do you do about the weather?”

He went back to his desk.²

She returned to the notebook. Next column. What to plan. She usually filled this in quickly. She knew what came next; she’d always known what came next. She had a beat, a specialization, a style. She knew how to do the job.

She realized she was less sure of this than she had been in March.

She wrote: figure out what I can still do that the AI can’t.

Then she crossed it out. It felt like exactly the wrong question.

She wrote instead: figure out what I want to do.

She submitted the AI usage form to Theodora with a two-line summary.

Theodora replied in four minutes: “Thanks. Just one thing — looking at your week’s output versus last month’s, something is different. Can’t quite put my finger on it.”

Soo-Jin looked at the reply. “Different how?”

“The work is polished. Maybe smoother? I don’t know. Your notebook, though — I noticed it when you were writing at your desk this morning. The handwriting has gotten better.”

Soo-Jin looked at her notebook. The handwriting had, in fact, gotten better. She’d been writing in it more.³

¹ Azeem Azhar, “I do not want your brains to rot” (Exponential View, September 2026): “Cognitive offloading is a strategic delegation that costs nothing. Cognitive surrender is something different; an uncritical abdication of reasoning itself.” The paper he cites, as-yet-unrefereed, argues that AI encourages delegation of simpler and simpler tasks, weakening the practices that maintain cognitive capacity over time. The deliberate irony: the essay was illustrated with an AI-generated visual summary of the paper it was citing.

² Daniel Miessler, Unsupervised Learning No.543 (September 2026): “I’d put it at like a 75% chance we hit ASI in 2027.” His working definition of ASI: a model that can perform any cognitive role better than any human expert who has ever lived. His key structural claim: the capabilities that make a model AGI-capable are nearly identical to those that make it ASI-capable, so the gap between them will be measured in months rather than years. What you do about the weather: you carry an umbrella, or you don’t go out.

³ Marie Dollé, “De long en large” (In Bed With Silicon, September 2026): six months into an HR role focused on AI acculturation, the challenge she keeps encountering is not prompting — agents now prompt themselves. It is externalizing implicit professional competence: the things experienced people know so thoroughly that the knowledge has become invisible even to themselves. Patrick Kervern’s “gens longs” — people whose depth of experience gives the world more relief — are the benchmark for what this kind of knowledge capture is trying to preserve. Soo-Jin’s handwriting was not in any AI training set.

“Instructions for living a life: Pay attention. Be astonished. Tell about it.”
— Mary Oliver, “Sometimes,” via Rob Walker’s Art of Noticing (American Masters documentary, PBS)

“Between the data and the reality falls the shadow. The shadow is the emotion, the stories, the messiness that humans have contained within us.”
— Rishad Tobaccowala, “Emotional Data,” via The Future Does Not Fit In The Containers Of The Past

• People: 100 Years of Stafford Beer. Mary Oliver: Saved By the Beauty of the World. La France Solo (Futur(s), La Mutante): 38% of French households are composed of a single person (2022 data). Close Enough. Kids remember you worked late. The J.Crew Liquor Store. The Doxa Orange Edition.

• AI strat: Emotional Data. LIWGY #784 — “The absurdity of the apocalypse”. The inflection point question. The absurdity problem .

• Geopolitics: Le futur est imprévisible, mais pas illisible. Artificial Revolution: Power, Control and the Future of AI.

• Foresight: Friday Interview with Charles Cumming.

• Random: Do A Dull Thing with Style. Summer is cooking my heart. Nous avons tous tellement besoin d’autre chose (Umanz).

Jacques Cousteau became the US distributor in 1968. Clive Cussler put one on Dirk Pitt’s wrist and never took it off. The watch survived the quartz crisis, which took out most of its contemporaries. The orange dial was a functional decision that became an aesthetic one — legibility at depth, identity at the surface. WITI’s line, which earns its place: “Design gives you something to look at, stories give you somewhere to go.”

The 10% Problem

The product meeting was on a Tuesday. Beatrice had a spreadsheet. The question was whether to launch the notification center now or wait three more weeks for an additional data round.

“Three weeks,” Beatrice said. “We need the cohort to mature.”

Omar was looking at his phone under the table. He looked up. “Did you see the resignation? The Anthropic researcher. He says there’s a ten percent chance AI wipes out humanity by 2030.”

“That’s not relevant,” Beatrice said.

“It’s relevant to all decisions with timelines,” Omar said. “If there’s a ten percent chance—”

“More likely than the Bills winning the Super Bowl,” said Leila, who had been watching them both.

“That’s how another researcher put it, yes,” Omar said. “Which is a calibration, not a slogan. The Bills have lost four Super Bowls. The probability is empirically grounded.”

Beatrice returned to the spreadsheet. The cohort data would be more statistically significant in three weeks. That was simply true.

Leila had done twelve user interviews. In eight of them, the participant had tapped on the area where the notification center would appear. Not because they were told to. Because it seemed like the right place. She said this.

“That’s a sample,” Beatrice said. “Not a sample size.”

“Eight out of twelve people tapped on the same empty space without being prompted,” Leila said. “I think we can launch.”¹

Beatrice looked at the numbers she had. They were not the numbers she wanted. The numbers she wanted would arrive in three weeks and would probably confirm what Leila was saying, at which point the notification center would still not have launched.

She thought about something she had read: that brands exist because humans are not computational, that people choose with their hearts and use numbers to justify what they just did, that the whole optimization stack is downstream of a desire that no model fully captures.² She was not sure this was relevant either. She filed it in the same place as the Bills.

“We launch,” she said.

Omar looked up from his phone again, surprised.

Three weeks later the notification center was the most-engaged element in the product. Leila was correct. The cohort data, when it arrived, confirmed this.

Omar had a spreadsheet tab he had not mentioned to either of them. It was titled “10% Tracker.” A single cell, updated weekly, reflecting his best synthesis of AI safety researchers’ current consensus on human extinction probability by 2030. It had started at 10.2%. This week it read 10.4%.

Beatrice leaned over and looked at it on his screen.

“That’s going up,” she said.

“Yes,” Omar said.

“You should add a trend line.”

He added it. The trend line went up.³

¹ Leila’s logic is not as thin as it sounds. Eight of twelve participants in a purposively sampled user study performing the same unprompted behavior is a meaningful signal. The case against waiting for larger samples is not that samples are wrong; it is that samples confirm what observation already knows, at a cost of time during which someone else may launch first.

² Rishad Tobaccowala, “Emotional Data” (The Future Does Not Fit In The Containers Of The Past, September 2026): “Between the data and the reality falls the shadow. The shadow is the emotion, the stories, the messiness that humans have contained within us. Today in the idolatry age of data and AI, those who rely only on numbers and computation get it wrong and wrong and wrong.” Tobaccowala’s argument is not anti-data; it is that data is downstream of desire, and desire is not where most analytics teams are looking.

³ Jacob Coxon resigned from Anthropic the week of September 7, citing what he characterized as an unacceptable probability of AI-caused human extinction by 2030. Evan Hubinger, an Anthropic lead researcher, posted on X that he personally estimated the probability above 10% within the next decade. This is not a fringe position at the frontier of AI research. It is the stated belief of researchers who spend their working hours trying to make it less true. The trend line was a joke. The number it tracks is not. Separately: French commentator Laurent Alexandre posted that week that with GPT-6’s capabilities it was now “very difficult to find utility in employees with IQ below 130.” The researchers trying to prevent the 10% from becoming 100% did not, as far as anyone could tell, take a position on this claim. (Marie Dollé, “QI en crise?”, In Bed With Silicon, September 2026; LIWGY #784)

“Ask the questions that have no answers. Invest in the millennium. Plant sequoias. Say that your main crop is the forest that you did not plant, that you will not live to harvest. Say that the leaves are harvested when they have rotted into the mold. Call that profit. Prophesy such returns.”
— Wendell Berry, “Manifesto: The Mad Farmer Liberation Front,”

“Power concedes nothing without a demand. It never did and it never will”. — Frederick Douglass, via workfutures Short Takes #37

“Chef Slowik: As Dr. King said: “We know through painful experience that freedom is never voluntarily given by the oppressor. It must be demanded by the oppressed”.” - The Menu

• People: Wendell Berry died at 92 — the poet, farmer, novelist, and environmental activist “Eating is an agricultural act” . Quit versus grit. J’ai perdu le Mojo (Fabrice Liut): a personal essay in French.Using Strava to underwrite apartment investments.

• AI strat: EV #600 — Astra outruns visibility. A frontier model is a rapidly depreciating asset. 50 First Dates. Astra safety concerns. Astra on ARC-AGI-3.

• Geopolitics: La Grande Transmission: approximately 9,000 billion euros will be transferred across generations in France over the next fifteen years.

• Society: La fureur de vivre à 30 km/h (15marches): in Sweden, a legal vehicle classification designed for wartime agricultural use has become a significant teenage subculture; the A-traktor (EPA-traktor). J Crew Crew (WITI): in 2014, at the peak of the J.Crew brand, a consumer named Meghan O’Neill began producing lo-fi video skits in which the models in J.Crew catalogs solved murders while trying to sell you cable-knit sweaters.

• Tech: Data centers and household power bills. Leading in the age of anxiety (Roberto Ferraro).

• Foresight: Three Horizons. The problem with life after work when status becomes your job rather than something your job confers.

• Random: Prisencolinensinainciusol. Industry Standard Tools (WITI Saturday Selection): a crowdsourced list of the one or two tools that have become so entrenched in a trade that nearly everyone in it uses them.

The Depreciating Asset

Mina had spent six months becoming the company’s internal expert on Fable 5.1 fine-tuning. Three months to get the prompting right, two to understand the edge cases, one to write the documentation. She was good at it.

The Astra launch was on a Monday. By Tuesday afternoon the benchmarks were out.

Jorge called a team meeting. He was not the kind of manager who enjoyed delivering unwelcome news, which meant he always delivered it worse than if he had enjoyed it: apologetically, in a room, with a slide.

The slide showed GPQA Diamond scores. Astra led everything else by a significant margin. Fable 5.1 was third.

“We need to evaluate switching”, Jorge said.

“I spent six months on this”, Mina said.

“I know”, Jorge said.

There was a pause in which nobody said what everybody understood: that six months of expertise in a specific frontier model was an asset that depreciated at a speed the organization had not yet developed a word for. A frontier model was competitive for weeks. Fine-tuning expertise had a longer half-life, but not by much.¹

Mina spent the rest of that afternoon reading Astra’s documentation. It was different from Fable in ways that were interesting and in ways that were annoying, and she could not always tell which was which. She thought about research she’d read recently: that when confronted by something that isn’t working, men tend to abandon it and women tend to persist; that the men get retention bonuses or leave for better offers; that the women who stay are rewarded with the satisfaction of having stayed.²

She thought about this for longer than was useful.

The next morning she sent Jorge a message: she’d like to work on the Astra integration. He replied within four minutes.

Over the following three weeks she dismantled what she’d built and rebuilt it on different foundations. She was good at dismantling things. She had not previously known this about herself. The Fable documentation, the edge cases, the methodology — she archived it all carefully, in a system designed to be accessed by someone who might never come.

In week four, a senior engineer named Declan said in the team channel: “Heads up — rumor is GPT-7 is in training. The alignment team is already flagging concerns”.

Jorge came to Mina’s desk later that afternoon.

“That one will be different”, he said.

“Different how?”

“We’re not sure yet.”

She nodded. She saved the Astra documentation to a folder she had named, without quite meaning to, the same thing as the Fable one: Current Best.³

¹ Exponential View, State of the AI Economy (June 2026): “A frontier model is a rapidly depreciating asset. This remains true even at the highest GPQA Diamond grades, where the pricing power of a new model quickly vanishes.” Nvidia’s revenue more than doubled year-on-year to $96.2 billion last quarter, which means the depreciation is on the vendor side, not the infrastructure side. The company that builds the model absorbs the devaluation. The company that builds the infrastructure does not.

² Lindsay Crouse, via workfutures Short Takes #37 (September 2026). Research tracked nearly 30,000 management-bound employees: women passed over for promotion were more likely to stay; men were more likely to leave and be rewarded for leaving. The conclusion the essay draws: “Giving up on things that don’t work is the only way to find things that work better or to go farther than you imagined.” The conclusion the research draws is the same. Mina had read the essay on her phone on a Tuesday evening and then looked at the ceiling for a while.

³ AI safety researcher Ryan Greenblatt, on Astra (September 2026): “I do not find it encouraging to see various specific misaligned behaviors go from a high rate with GPT 5.6 to ~zero with Astra. This seems indicative of whack-a-mole / papering over specific problems rather than solving the underlying misaligned drives.” The documentation for a model that has been superseded is a record of what we were capable of solving and what we decided not to investigate. Mina’s Fable archive remained intact.

“Before a person asks a machine to write for them, they should know what it feels like to write from the depths of their own uncertainty and identity. Before a person lets a machine imitate their voice, they should have endured the long, chaotic, and often humiliating process of forging one.”
— Debbie Milman, via Umanz (August 2026)

“Everything is a song to me.”
— Dolly Parton (1946–2026), via Austin Kleon

• People: Socializing less. Young workers in AI-exposed occupations (ages 22–25) are now 19% below their expected employment trend. Dolly Parton died; also Yayoi Kusama died the same week Tim Curry (RIP RHPS!). Hard vs soft fascination.

• AI strat: Zuckerberg’s AI manifesto. AI and writing (WITI, Noah Brier): a high school English teacher worried that AI would make the “write longs” assignment pointless. Design AI that strengthens reasoning.

• Security: Russian backdoor in speed cameras. ICE ankle monitors.

• Geopolitics: AI sovereignty paradox. Recursive self-improvement has ceilings.

• Society: AI rejection letters. Magnifica Humanitas. AI Resist List. Philippine digital workers coalition (2025). I’ve been Substacked (workfutures): Stowe Boyd discovered mid-read that a newsletter he was enjoying was AI-generated. Personal nostalgia vs political nostalgia.

• Tech: Token asymmetry. Tree canopy mapping.

• Foresight: The Clarke inversion — any sufficiently advanced technology becomes indistinguishable from Nature. Boussole Anti-chaos: a one-page vision-and-strategy visual deployed across 450 people in a single session; the one-pager as an artifact of alignment.

• Random: A brief history of the internet’s favorite scam.

• AI: Qwen 3 doing frontier work. Kimi K3: the team’s own account of building a frontier model under export controls. RSI limits.

The Voice

Priya had been subscribing to Nour’s newsletter for three years. It was one of about forty she’d kept from the cull.

The Tuesday edition arrived at 7:14 AM. Priya read it on the train, standing, one hand on the overhead bar. The first paragraph was about a drawbridge on the Chicago River — the kind you wait for while a barge passes, the feeling of impatience becoming something else. She thought: this is good. Then she thought: this is almost too good.

At the office she ran it through the content analysis tool. 71% confidence: machine-generated.¹

She emailed Nour through the newsletter footer. Nour replied in eleven minutes — faster than she had ever replied before. Yes, she’d been using assistance for about six months. The ideas were still hers. The structure was still hers. Was it really so different from using Grammarly?

“It’s not so different,” Garrett said, reading over her shoulder. Garrett had opinions about this. He’d run his last fourteen client proposals through the same pipeline. “The ideas are what matter.”

Priya thought about the Milman line she’d screenshotted months ago and never deleted: Before a person asks a machine to write for them, they should know what it feels like to write from the depths of their own uncertainty and identity.² She hadn’t thought about it in a while.

“Did you read the Pope?” she said.

Garrett had not read the Pope.

“He published an encyclical. In May.” The Vatican had been unusually fast — the text argued that the authenticity of the human voice was not reducible to any simulation, however accurate. Magnifica Humanitas. Six hundred footnotes. Already cited in three legal briefs.³

“The Pope has a ghostwriter,” Garrett said.

This was technically true.

Priya looked at the Nour newsletter again. The drawbridge paragraph was still good. The barge of crushed cars, the mood shift, the way it turned on a single waiting figure — these were not things a machine assembled by accident. Maybe the ideas were still Nour’s.

She opened her own archive. Three years of performance reviews. The oldest one, from 2024, when she’d been working eighty-hour weeks and writing everything herself — a long paragraph about her team’s approach to uncertainty. She ran it through the tool.

68% confidence: machine-generated.

She had not used AI in 2024.

Garrett looked at the result and then looked away in the manner of someone who had already had this conversation with himself. “The tool’s not perfect,” he said.

“No,” said Priya.

She archived Nour’s newsletter in a subfolder she’d been maintaining for two years called Keep This. Then she went to get coffee. There was a new machine in the kitchen that made it automatically when you walked in. It was quite good.

¹ Anthropic’s watermarking system, as of August 2026, reports a confidence score rather than a binary flag. A score above 0.65 is classified as “likely AI-assisted.” The company notes this threshold produces roughly 12% false positives on human writing from non-native English speakers. Nour was not a native English speaker. Priya did not know that the threshold existed.

² Debbie Milman, quoted in the French newsletter Umanz (August 2026). The full sentence continues: “Before a person lets a machine imitate their voice, they should have endured the long, chaotic, and often humiliating process of forging one.” The word forging is doing a lot of work here — it implies heat and resistance, not just formation.

³ Pope Leo XIV, Magnifica Humanitas (May 15, 2026). The Vatican’s encyclical is available in nine languages. The 600-footnote figure was reported by La Croix. One of the three legal briefs that cited it was filed in a wrongful termination case in which the plaintiff argued that an AI-generated performance review had misrepresented their work. The document is 47,000 words. Priya had read the executive summary and three footnotes.

“What spark of humanity, of possible creativity, can remain alive in a being dragged out of sleep at 6 every morning, jolted about in suburban trains, deafened by the racket of machinery, bleached and steamed by meaningless sounds and gestures?”
— Raoul Vaneigem, The Revolution of Everyday Life (1967), via workfutures

“For ‘tis the sport to have the engineer / Hoist with his own petard.”
— Shakespeare, Hamlet, via Exponential View

• People: AI twin pastor — a California church has deployed a digital double of its pastor. FAA recruits gamers: the agency is actively targeting video game players as air traffic controller candidates. FOBO is the fear of becoming obsolete, and a literature review of job insecurity research finds it consistently lowers performance rather than raising it, which means FOBO is both a welfare problem and an output problem.

• AI strat: One agent beats four. Don’t be a meat proxy. The shape of things to come (Steve Yegge): an engineer’s-eye view of what the current AI transition actually portends.

• Security: Volt Typhoon war game — Wired walked a scenario in which China executes a cyberattack on US water infrastructure; the exercise maps both the hack and the cascade. Arming healthcare: France’s defense and health ministries are developing programs to prepare healthcare workers for wartime conditions; alongside this, the ICRC has published War and Public Health: A Handbook.

• Geopolitics: UN AI turning point. The petard problem.

• Society: Raoul Vaneigem is dead — the Situationist theorist whose The Revolution of Everyday Life (1967) asked what point there was in a prosperity that guaranteed you would not die of starvation but left you at risk of dying of boredom; his targets were meaningless work, consumer society, and the managed spectacle of both capitalism and communism. AI hollows out game dev. The subtraction instinct.

• Tech: The open seatbelt patent . Wood in bioreactors. Agentic AI needs CPUs.

• Foresight: Friction as signal — a Brazilian foresight practitioner built an NFC keychain for an event. Strategic blindspots: WEF on how foresight frameworks identify the structural blind spots most common in leadership teams (cf David Bach on leading through uncertainty). Lunar rover at WorldCon.

• Random: Rolex fooled by its own fake. Campaign for North Africa — the legendary wargame with 1,500 hours of estimated play time and a recommendation to play with 10 people; two years into a campaign, participants “still don’t understand what the camels do”. Adult Play-Doh: Hasbro’s second attempt to sell Play-Doh to adults.

• AI: Not a bubble yet. How Claude marks outputs. Doctorow on inevitability.

Verification

The monthly AI check-in had been Claudia’s idea, and she ran it with the enthusiasm of someone whose bonus was, in some indirect way, attached to the numbers on the slide.

“AI adoption is up three hundred and forty percent”, she said. “Across the engagement”.

Ravi nodded. He had been using a four-agent pipeline for the client’s market analysis since February. You gave each agent a section, they returned outputs, the outputs were synthesized. He explained this when asked. It was, he said, significantly more thorough than one agent doing everything.

Yoon, who sat to his left, did not say anything. She had been reviewing Ravi’s outputs for three months. He sent them to her with a message that said something like “just a quick look?” and she gave them a quick look that was not quick, because the outputs required someone to actually understand the client’s situation to determine whether they made sense, and Ravi had not mentioned to anyone that this was occurring.¹

“The client feedback is very positive”, Claudia said. She clicked to the next slide.

The call came in at 11:47 on a Tuesday, which Yoon noted because she had been about to get coffee. The client had followed the pipeline’s recommendation on market entry timing. The recommendation had been wrong. Not directionally wrong — wrong in a way that a single analyst with access to all the data would have caught immediately, because the critical regional supply constraint was in the dataset, but it was in only one agent’s context window, and that agent had been outvoted by the other three, who were working from aggregated summaries and had no reason to push back.²

The postmortem took ninety minutes. Claudia proposed adding a fifth agent for synthesis. Ravi proposed a verification step. Yoon said: “I am the verification step”.

There was a pause.

Ravi said he had understood that she was reviewing the outputs as a colleague, informally, not as a — not in any formal — not as a designated one..

“Yes”, Yoon said. “That’s what I meant”.

The conversation moved on. Claudia said she would document the new workflow. The new workflow involved Yoon, now named, with a title that had been added. Adding the title was the first thing anyone had suggested and the last thing anyone questioned. The meeting generated three action items, all of which required doing more. No one proposed doing less.³

Ravi sent her a message that evening. It said: “Appreciate you flagging that — really valuable input”. She did not respond because she did not know what the correct response was, and she had also spent the morning being afraid, in a low-grade and unspecific way, that any of this flagging might make her the kind of person who could be replaced by a fifth agent.

Claudia updated the dashboard. AI adoption: still up three hundred and forty percent. She scheduled a follow-up for the following month. In the notes field she typed: discuss verification agent.⁴

¹ A 2026 study found that AI agents require roughly 6.4 hours of human oversight per week — reviewing outputs, supplying missing context, catching errors before they reach a client or a decision. This labor is not in most job descriptions. It tends to be absorbed by whoever is most conscientious and least able to say no. It does not appear in productivity metrics. It does not appear anywhere.

² Anthropic ran the hidden-profile experiment on multi-agent systems in 2026: four agents deliberating on a question where the correct answer was held by only one of them. Across most model families, the group chose correctly 17–36% of the time. A single agent given the full evidence base was right nearly every time. The problem was not the agents. The problem was the structure — which is also, in a different context, the problem with committees, editorial boards, and quarterly planning processes. The lone dissenter who has the right data needs institutional protection that most institutions have not yet built.

³ The subtraction instinct describes the consistent human tendency to respond to problems by adding — features, steps, rules, people — rather than removing. In a series of experiments, subjects were told that subtractive solutions were explicitly available; they still systematically overlooked them. The effect is stronger under time pressure and stronger still when the addition can be assigned to someone else.

⁴ Fear of becoming obsolete — FOBO — has been documented to reduce employee performance rather than increase it. Job insecurity correlates with lower commitment, not higher effort. The manager who makes workers afraid of being replaced by AI does not get more productivity from them. They get worse work, done by people conserving energy for the exit. The paradox does not appear in quarterly reports.

“It’s difficult to figure out how much anything AI-related costs.”
— senior Amazon employee, via Exponential View

“Abductive reasoning is what gets us from epicycles to elliptical orbits. It’s where new ideas come from.”
— Gordon Brander, Squishy Computer

• People: South Korea’s dopamine sites — apps and platforms that simulate shopping without purchase. AI spending stratification: Ramp reports that in July the top 1% of businesses spent a median $7,400 per employee on AI, the top 10% spent $650, and the median firm spent $11.95.

• AI strat: The $6 agent — Amazon spent $1.8 million on a single Claude project over five months; one team’s autonomous agent peaked at $500 per day before a model-routing audit brought it to $6. Botsitting: AI agents require roughly 6.4 hours of human oversight per week. Where LLMs fall short. Opus 5 and revenue concentration. AI watermarking.

• Tech: Google’s departures as capital signal. Price war accelerates. Invisible drones.

• Security: Keyv supply chain attack. Coldcard hack.

• Society: One Night Only.

• Random: Five silences — a language essay notes that spoken language contains at least five distinct forms of silence — the pregnant pause, the companionable silence, the awkward silence, the menacing silence, the reverent silence.

• Foresight: Organizational tools for personal disruption.

Bell Labs produced: the transistor, the laser, the solar cell, Unix, the C programming language, information theory, cellular networks, the detection of cosmic microwave background radiation. None of these had a clear return on investment at the time of the work. The researchers were not asked to justify the compute. The era in which that was true lasted approximately forty years.

The Hurdle Rate

The meeting had been on the calendar for three weeks, which was how Nadia knew it was serious. Informal concerns appeared without warning. Structural ones had lead time.

Priya from Finance was there, which settled the matter. Nadia had brought her most recent results — not polished into a deck, because she had found that decks created the impression of conclusions where there were only questions, and the work she was doing was questions, deliberately, all the way down.

“The compute allocation question,” Marcus said. He ran infrastructure. He had the numbers. “At current utilization, the models serving production queries are generating approximately eleven times the revenue per GPU-hour compared to research workloads.”

Nadia said she understood the number.

“The ask isn’t to stop the work,” he said. “It’s to think about what the work needs, versus what it has.”

She thought about explaining what she was doing. She had been trying for six months to identify the conditions under which a language model would generate a genuinely novel hypothesis — not a recombination, not an extrapolation, but an abductive leap: the mechanism that would explain a finding the model had not been trained to explain. She had a method. She had intermediate results. She did not have a timeline, because the nature of the work was that the interesting finding would arrive when it arrived, which was not schedulable.

“What’s the use case,” Priya said. Not a question.

Nadia thought about this. The use case was: if she found what she was looking for, AI systems could eventually do science in the way that scientists do science — not just processing what was already known, but generating the hypothesis that reordered what was known into something new. The use case was: Kepler. The use case was: the discovery that the orbits were ellipses, not circles, which required someone to hold an anomaly in mind long enough to find the shape that explained it.

She said: “It’s a foundational capability question.”

Marcus nodded as if this confirmed something. “The production systems are generating value now. The research has a longer horizon.”

“Yes,” she said.

“And the question is whether this is the right place for a longer horizon.”

She had known this meeting was coming. She had been watching the allocation numbers shift for four months. She had told herself there would be a conversation, and then there would be a process, and that she would find a way through it. She had been wrong about the process. The conversation and the conclusion were the same meeting.¹

She left three months later. She took one postdoc and a modest compute budget from a foundation she had spent two years cultivating. The work was slower. The results arrived more irregularly. She had no production deadline.

In the following year, she published two papers. Neither received much attention at the time. Both were cited heavily within three years, by researchers at the same lab she had left, who were trying to solve a problem that had arrived from a direction no one had been looking.²

¹ The hurdle rate — the minimum return required to justify an investment — is a useful tool for allocating resources between competing projects with similar time horizons. It is a less useful tool for distinguishing between work that generates value in three years and work that generates value in thirty. Organizations that apply the same hurdle rate to both will consistently defund the second category, which is not an error in the calculation but a limitation of the instrument. Bell Labs worked because AT&T’s monopoly revenues were large enough and stable enough that the opportunity cost of funding open-ended research was low. When the monopoly ended, the instrument changed and the research followed.

² The abductive leap — the moment when an anomaly resolves into a mechanism — is not evenly distributed across time or institution. Kepler had Tycho Brahe’s data and eight years to stare at the problem of Mars. Newton had a plague year and an empty Cambridge. Fleming had a contaminated petri dish he had not yet thrown away. The conditions for the eureka are not the conditions for the quarterly return, and organizations that mistake the second set of conditions for the first tend to discover this difference from the outside of the discovery.

“They were not even irrelevant, because irrelevant implies that they had been noticed” — Swiss industrialist on Japanese watchmakers, 1969, via Why Is This Interesting

“The future came from the slime while we looked for signs from the heavens” — Rishad Tobaccowala

• People: Japan’s human-sized refrigerator. Job boundaries blurring: nearly half of ChatGPT tasks fall outside the user’s primary occupation. The Quartz Crisis. Pools in containers: a builder is converting maritime shipping containers into backyard swimming pools, with demand across all of France.

• AI strat: Define done before starting. Approval is the new bottleneck: individuals with AI produce faster, but verification and sign-off systems were built around old paces; transformation requires rebuilding those systems, not just issuing tools. Agentic patents surge. China courts AI displacement. Constraint as capability: Kimi K3 reached benchmarks comparable to frontier Western models despite operating under export controls and limited compute access. AI capex in the pipeline.

• Society: Enshittification’s price tag. Skillidays: the growing trend of combining vacations with acquiring a new skill, learning reframed as leisure, not obligation. Effort, not source.

• Foresight: What’s already been decided — Shell’s scenario analyst Pierre Wack described certain change factors as “predetermineds”: things whose effects are already determined even if not yet visible, like heavy monsoon rain in the Himalayas predicting flood levels in Benares five days later; futurist Art Kleiner’s list of what we already know about the 2030s is built on the same logic. Outside the categories: “the biggest threats and opportunities come from outside one’s categories” — from what incumbents look down upon or never appear on their radar, from the slime rather than the heavens.

• AI: Breadth beats depth — employees using AI across several different task types are twice as likely to report productivity gains as those using it for one or two types; specialization limits the compound effect. Value chain premium.

By 1975, Seiko had brought the price of a quartz watch below fifty dollars. The movement that could not be taken seriously at the price of a car was now available to everyone who had ever worn a watch. The Swiss watch industry lost roughly half its workforce in the decade that followed. They had not lacked information. They had lacked the ability to translate information into concern for people who were not yet concerned.

Complications

The presentation had gone well, by Swiss standards, which meant that no one had interrupted.

Gerhard had spent six weeks on the report. He had obtained samples of the Japanese movement through a contact in Geneva who worked for a trading house that did not ask careful questions. He had disassembled the Seiko 35A and timed its drift across two weeks. He had built a cost model from what he could reverse-engineer. He had done the arithmetic three times and received the same answer each time, which he had initially hoped was a mistake.

The answer was: they could sell these at a quarter of our price within five years. Possibly less.

He had presented this to the management committee in September. He had charts. He had the movement itself, which he passed around the table. Three of the six men at the table did not pick it up. One of the men who did said it felt cheap.

Gerhard noted that this was true, and that it would keep time to within fifteen seconds a month, and that this accuracy was not achievable by any mechanical movement at any price. He noted that the target customer for the Seiko was not a collector. He noted that the target customer for the Seiko was everyone else.

He was thanked for the thoroughness of his work.

In the months that followed, he raised it in smaller meetings — without slides, just questions: Have you heard anything from distributors about the Japanese product? The answers were variations on: no, nothing meaningful, not really. His colleague Andreas said once that a client in Hamburg had mentioned it, but that the man collected military watches and was enthusiastic about inexpensive things generally.

By 1978, three factories in the Vallée de Joux had announced closures.

He retired in 1982. He had two watches: a movement he had finished himself, twenty years earlier, that he wore on Sundays; and a Seiko he had bought in 1973 to take apart and put back together and never returned to a box. It was accurate to within two seconds a month. He wore it on Tuesdays, usually. Sometimes Thursdays.

He had not been wrong. He had been right for six years before it mattered. The things he had pointed at had been there. He had simply been unable to communicate them into existence as a concern for people who had other concerns, and whose concerns, being other, required that the problem not yet be visible.

Later he would read about something called a “predetermined element” — a category of futures analysis developed by a man who worked for an oil company and thought carefully about rivers. The idea was that some things were already decided. The flood had already happened upstream; it was only a question of when it reached you. The man who developed this concept had thought the value was in communicating the certain future so that people could prepare. Gerhard had spent six years learning that the communication problem was harder than the forecast.¹

He did not write a memoir. He thought about what he might add.

There was nothing to add.²

¹ Pierre Wack worked for Shell’s Group Planning division from 1971 to 1981 and is credited with developing modern scenario planning. His signature concept was the “predetermined element” — a factor so deeply embedded in current forces that its future effect could be stated with high confidence, regardless of how scenarios diverged on other dimensions. He described it by analogy to the Ganges: heavy monsoon rain in the Himalayan headwaters predicts flooding in Allahabad five days later, and in Benares five days after that, with near-certainty. Wack believed that the hardest part of scenarios was not identifying predetermineds but communicating them to people who were not yet experiencing their consequences. He was rarely thanked for the thoroughness of his work.

² The Swiss watch industry’s recovery came not from winning the argument about quality, but from SWATCH — a product designed to compete with Japanese quartz on price and style while preserving Swiss manufacturing employment. Nicolas Hayek built a structure that could survive being right about the wrong things for a long time. The Quartz Crisis ended not when the Swiss admitted they had lost, but when one person built something structured to outlast the admission.

“If we are always arriving and departing, it is also true that we are eternally anchored. One’s destination is never a place but rather a new way of looking at things.”
— Henry Miller, via workfutures.io

“Fable is better than me at my job, but Fable alone would be a mediocre investor.”
— unnamed VC, via Exponential View

• People: GLP-1s cut sick leave — weight-loss drugs reduced long-term sickness leave by 17% over four years in Denmark. Cities emptying of children. Culture change is not behavior change — decades of organizational psychology evidence from Rob Briner. Childhood rituals for sale: for-profit companies have acquired the rituals of American childhood and are selling them back to parents.

• AI strat: Proletarisation of advisory (it’s me). The AI adopter’s decision trap. Chinese AI distillation grey zone. Anthropic alleges DeepSeek, Moonshot, and MiniMax used 16 million Claude conversations via 24,000 fake accounts. Big Three lock rents: OpenAI, Anthropic, and Google take 90% of spend but only 52% of tokens. AI reaches 68% of occupations. Frontier AI commoditizing fast.

• Geopolitics: China EVs take European factories. Ukraine drones hit Wildberries.

• Tech: Productivity without transformation. AI unicorn concentration.

• Foresight: Brittle thesis, liquidated fund. Tools as creative decisions: each writing tool constrains and shapes the work differently.

• Random: Denmark’s 1969 tax-free year. A long-held embarrassing secret.

Samuel Slater disguised himself as a farm laborer and sailed from England in 1789, carrying Richard Arkwright’s entire mill system in his head. By 1790, he had built it from memory in Rhode Island. He died worth approximately a billion dollars in today’s terms. What he had transported was not a document. It was a way of looking at things.

Situational Awareness

The liquidation notice arrived at 6:14 on a Monday morning. Conrad read it twice, which took less time than he expected given its length — mostly because he had been expecting it, and because the core fact took only a sentence to deliver.

The fund was wound down. The assets were liquid. He would receive confirmation of his final net asset value by close of business Friday.

He had been running Situational Awareness LP for eighteen months. He had been right, as nearly as he could tell, about most of it. The path to superintelligence would require compute. Compute required chips. Chips required power. Power required infrastructure. He had laid out the thesis in a white paper that had been widely circulated and, up to a point, validated. Data centers had been built. Chips had shipped. Demand had exceeded several projections.

He had also taken four times leverage. This had seemed reasonable — sensible, even — at the time. A high-conviction thesis warranted a committed position. That was the whole point of conviction.

The Philadelphia Semiconductor Index had peaked in June. He had known it might. He had thought he had handled his risk. What he had managed, it turned out, was the risk he had modeled, which was not the same thing as the risk that existed.

He texted Mariam, who had been early on the same thesis and had structured her position differently — smaller, longer duration, diversified enough to survive a drawdown without a margin call. She had not called him last week. He appreciated this.

Her reply came in three minutes: You were right.

He typed and deleted several responses. He settled on: I was right the wrong way.

There was a pause. Then: That happens.

He thought about this. He had spent eighteen months being professionally right. He had built a model. He had stress-tested it. He had presented it to LPs who had believed him. The model had correctly predicted that demand for compute would be transformative. What it had not modeled was his own fund’s liquidity position under a 28.6% correction in the semiconductor index. These were not the same problem. One was a thesis about the world. The other was a thesis about his own resilience under pressure. He had been right about the first and wrong about the second, and the second was the one that mattered when the margin call came.¹

He had confused a good thesis with a resilient structure. Mariam had not.

Later, he would write a post-mortem. It would be honest. He would note that it takes years to know if an AI investment strategy is compounding or wasting. That winners and losers look identical in the early periods. That the fund which looks best in the middle — the one that found something that worked and doubled down — is often not the fund that finishes best. That what distinguishes eventual winners is continuous learning, not a single correct bet made at the beginning. He had made a correct bet at the beginning. He had then stopped learning how to survive being right for long enough.²

He would not publish the post-mortem. He would write it for himself.

He pulled up the white paper from eighteen months ago and opened a new document. He typed Revised at the top. He deleted it. He typed: What I missed.

He still believed the thesis. He thought about what it would mean to hold it differently — longer horizon, smaller position sizes, structured to survive the years when everything looked like it was going wrong simultaneously, because everything would. The thesis did not care about his liquidity. The thesis would be right or wrong on its own schedule.

He thought about a story he had read somewhere, about a man who committed an entire factory system to memory in order to carry it across a border where none of it could legally travel. The man had been right about the value of what he was carrying. He had walked off a ship in a new country and built it from memory. What had survived the crossing was not the drawings. It was the understanding of how the system worked.³

He closed the white paper. He started with what he had missed.

¹ Mariam would go on to run her fund for six more years and then step back to teach. She occasionally thought about Conrad’s paper. She thought it had been good work. She thought the error had not been in the thesis but in the structure — and the structure was the kind of thing you only fully understood after it failed. This seemed like an inadequate consolation and probably was.

² The thing he had actually missed was simpler: he had invested heavily in being right about the direction, and under-invested in being wrong without being destroyed. These are different design problems and require different capital.

³ Samuel Slater crossed from England to America in 1789, carrying Richard Arkwright’s complete cotton mill system committed to memory. Britain had made it illegal to export the machinery or to allow skilled textile workers to emigrate. He arrived disguised as a farm laborer. He died worth approximately a billion dollars in today’s terms. President Jackson called him the Father of American Manufactures. Whether what he did was theft depends on which side of the Atlantic you were standing on in 1789. The knowledge crossed regardless.

“The most important discoveries are often not the ones we set out to make. If we build a world that delivers only what is asked for, we will lose the capacity to discover what we didn’t know to ask.” — Anne-Laure Le Cunff, via workfutures

“What would our activities look like if we talked about longing?” — Rob Hopkins, via Andrew Curry

• People: Reminiscence bump: adults over 40 disproportionately remember ages 15–25. Ann Friedman’s book nobody bought: meaningful work that algorithms overlook. Dopamine simulator sites.

• Society: Millennials’ great career crisis: burned out, bailing on corporate. Alexa and the loneliness gap: haunting New Yorker piece; who in your life is buying a dozen Alexas because people aren’t showing up? Line culture: TikTok lemmings, FOMO, social validation > comfort — queuing as communion.

• AI strat: Kimi K3 benchmarks near Opus and GPT-5.6 Sol: pressure on inference margins. Distillation grey zone: no IP precedent for model outputs. AI = thinking and doing, replace the word, force the real question. LLMs can’t abduct.

• Foresight: Rob Hopkins’ politics of longing. Protopian Prize.

• Planet: Solar paradox — PV modules on a learning curve but US levelized system costs still rising since 2023. Coolcations: cold-destination travel reshaping summer holidays in heatwave years.

• Tech: Handwriting recruits more brain. Bee Computer.

• Random: Kimi K3 economic shock. Futures-telling machine: fairground device prints your community’s short future on paper. Homer with Hitchens (2000): two hours for those who need more after the latest Nolan.

The Longing Report

Margaret had been listening to the field recordings for three hours before she noticed the beavers. They appeared around minute forty of Track 7: Thursday, June 2036 — the sound of water management, unhurried, in what the liner notes described as a flood-plain outside Bristol, the beavers now contracted by three regional insurers.^1 She found it calming. She found it necessary.

Simon knocked and entered without waiting, which was his habit. He had printed out, on actual paper, a list of twenty-seven questions. “I had Kimi K3 generate the discussion prompts for tomorrow’s workshop,” he said, placing them on her desk. “It was very fast.”

Margaret removed one earphone and looked at the list. What does a positive future look like? How can communities adapt to change? What role does imagination play in resilience planning? She read three more. They were, she noted, extremely good questions. They were also questions she had been asking — had been asked to answer — for the past fifteen years.

“These are the questions we already know we have,” she said.

Simon nodded. “Kimi is very thorough.”

She thought about the gap — the third logic nobody taught. Deduction. Induction. And then the leap: abduction, the inference that had to arrive sideways toward an explanation that didn’t yet exist, that couldn’t come from any chain of known premises. You needed to dream up both the question and the possible answer at the same time. Current AI, the theorists were fairly sure, couldn’t do it. Current AI was, in this precise sense, very thorough and very blind.^2

In the office next door, she could hear Della’s pen. The woman was transcribing last week’s committee minutes by hand, as she did every week, because she had once said — and could not be dissuaded — that writing forced her to attend to what had been said, rather than merely record it. Her handwriting had been terrible in January. It was merely bad now.

“What would the questions look like,” Margaret said slowly, mostly to herself, “if we talked about longing?”

Simon looked briefly alarmed, which was also his habit.

The building was a converted Victorian semi — two distinct domestic lives fused at the wall, each half still pretending the join wasn’t there. The workspace occupied the left side. Margaret had once spent ten minutes imagining what lived in the right, and had concluded that the join itself was probably the most interesting room.

She added this to her notes: The join is the interesting part.

“Write six questions,” she said to Simon, “that you personally do not know the answer to. Not Kimi. You.”

He looked at her as though she had asked him to dredge something.

He sat down and picked up one of Della’s spare pens.^3

^1 The beavers were reportedly excellent at this. Their contracts included performance clauses.

^2 It was also, in fairness, very fast.

^3 The pen ran dry almost immediately. This was considered, by everyone present, to be an apt metaphor.

“When you’re young you think that time moves forward. At 80 you know that it doesn’t, it stands still. We’re the ones that move.”
— Bob Dylan, via workfutures.io

“Technology is legislation. Introducing a new technology is equivalent to imposing new global economic, legal, or societal policies with major consequences — and we don’t know in advance what that policy is going to be.”
— Karl Schroeder, Unapocalyptic

• People: Fatherhood rewires the brain. Capitalism’s approval below 50%. Summer penis, wtf. Plastic surgery nightmare.

• AI strat: China’s Kimi K3 matches frontier models in some benchmarks. GenAI drives retail traffic: nearly 2% of Walmart and Target web visits now come from AI agents. US export controls force Anthropic to restrict model access.

• Security: Agents find real bugs. GRU intelligence on the marketplace. Drone killing a moth, soon mosquitoes?

• Society: AI drafts manager objectives. Africa’s infrastructure gap. Chalance. Christian interior design.

• Tech: GPU demand rebounds. Solar’s cost paradox: PV panels remain on an aggressive learning curve.

• Foresight: Technology as legislation.

• Random: Human curation in crisis (he says ^). An unlikely time capsule (Paul Lukas).

• AI: At Brown University, 56 of 59 students failed — a professor suspected AI use on papers, ran a closed-book final; scores collapsed by up to 100% for 56 students; three performed as expected. The coming divide.

The Closed Book

The professor had taught economics for twenty-two years and had never felt the need to specify, in writing, that the final examination would be closed book. It had always been closed book. That was what a final was.

He specified it this year.

He did this quietly, in a footnote of the syllabus distributed at the start of term, and again in the email with the revision guidelines, and a third time in the exam hall briefing. He was not trying to hide it. He just did not particularly want to discuss it.

Fifty-nine students. He had been suspicious of the papers for six months — not because they were wrong, but because they were right in a way that felt practiced. A smoothness he recognised from grant applications written under pressure and not from second-year undergraduates working on their first extended economic analysis. He said nothing at the time. He noted it. He designed a final that could not be assisted, and he waited.

The results came back in three days.

Fifty-six scores collapsed. Some by a third. Several by sixty percent. Four by a hundred. Three students performed exactly as their previous work had predicted: competently, without drama. One of them had submitted a midterm paper he suspected an AI had partly written. She still, it appeared, knew the material.

He sat with this for longer than was professionally necessary.

He had a colleague at the medical school, Isabelle, who had recently published work on neuroplasticity and fatherhood — the finding that becoming a father physically reorganised the male brain, thickening the cortex in regions associated with empathy and sustained attention. The change was not metaphorical. It was anatomical. You were a different person on the other side of the experience than before it: measurably so, in tissue.

She came by his office that afternoon. He showed her the exam results without explanation.

She looked at them for a moment. “What did they do instead,” she said. “Instead of learning the economics.”

“They delegated the understanding,” he said. “And kept the transcript.”

She turned the data over. “The brain changes in response to what we actually do. Not what we intend to do, not what we simulate doing. What we practice, we become.” She set the paper down. “These fifty-six students practised the performance of economics. They’re now very capable at that. It’s just a different skill than the one they enrolled to learn.”

He asked whether the brain could recover.

She considered this. “You’d have to give it something to do,” she said. “Something it couldn’t delegate.”

Outside his window it was a summer evening of the kind that seems to believe it will last forever, which it won’t. A percentage of Americans read for pleasure, he had recently read, on any given day. Sixteen percent, as of 2023. Down from twenty-eight in 2004. The trend had begun before AI arrived. Something had already been changing before anyone named it.²

He thought about the three students who had passed. He did not know what they had done differently — whether it was stubbornness or discipline or some third thing that had no module attached to it. He could not publish this. There was no control group, no clean methodology, no IRB approval for what amounted to an accidental experiment. What he had was an observation.

He wrote on the inside cover of his notes: A grade proves you can pass the test. The test proves nothing about what happens in the absence of one.

He thought about the student who had submitted a partly AI-assisted midterm and still passed the closed-book final. She had used the tool without becoming dependent on it. He wasn’t sure whether that was wisdom or habit or luck. He intended to find out.³

He did not change the syllabus the following year. He changed the assessment. He did not announce in advance that it would be closed book. He specified it, briefly, in a footnote.

He had been teaching for twenty-three years. He had always believed that what you learned was a function of what you did with your attention over time. He still believed this. He was less certain what to do with it.

The summer evening continued, implausibly, outside the window. A bird was singing something repetitive and confident in the campus trees. It did not appear to have revised.

¹ She would pass every subsequent closed-book exam without difficulty and graduate near the top of her class. The midterm AI use had apparently been the equivalent of checking her working rather than replacing it — a shortcut that left the understanding intact. This is not the typical outcome. It is, however, what made her interesting.

² The 28% figure from 2004 was also, in retrospect, not good. Cognitive engagement researchers had been worried for thirty years. The concern had been legible for a long time before anyone started writing newsletters about it.

³ He never taught a student quite like her again. Or possibly he did, and the assessment design made it impossible to tell.

“There are tranquil ages, which seem to contain that which will last forever. And there are ages of change, which see upheavals that, in extreme instances, appear to go to the roots of humanity itself.” — Karl Jaspers, The Origin and Goal of History (1949)

“Ne demande ton chemin à personne, tu risquerais de ne plus pouvoir te perdre.”
— Nahman de Bratslav

• AI: Medicine guarding against “never skilling” — residents over-relying on AI never build clinical judgment. AI coding tools are addictivewith engineers burning out. Midjourney pivots to healthcare. RL on forecasting beats coding/math as the better scaling path for safe planning.

• AI strat: AI-heavy firms grow headcount 10%, entry-level +12% — selecting for who knows how to use AI, not replacing humans. US chip controls pushed China to treat open-source as resilience infrastructure: 11× more LLM forks per week from China-linked developers after each export control. OpenAI wants 5% of itself owned by the US government (~$42bn). Ex-Anthropic researchers raise $200M for Mirendil to sell the self-improvement loop.

• Geopolitics: China installed 3× more wind capacity than the rest of the world combined last year; all 6 top turbine makers are now Chinese. Wars depleting global tungsten stocks. Autonomous drones killed human soldiers for the first time.

• Security: ToddyCat stealing Gmail. Cloudflare now charges AI crawlers per page. Flock’s plate readers and AI cameras are turning US streets into a surveillance net.

• Science: The ocean’s missing middle: the mesopelagic zone (200–1,000m) contains 70% of all animal life on Earth and runs the biological pump sequestering carbon for centuries, and we have barely been there. Hookworms and drugs. A cell built from scratch now grows and divides. Job title of the future: nature’s drug designer.

• People: 6.2 million Americans want a job but aren’t looking — 4th consecutive monthly rise, now above 2008 levels. Goldman Sachs: AI will temporarily displace 9% of US workers over a 10-year transition. Gen Z at work: only 29% say a casual hug is acceptable; 28% say handshakes only.

• Foresight: WaPo looked back to its 1976 bicentennial forecast and wrote 50 years forward to the tricentennial. Le grand rallongis — an essay on slowness, detours, and the merveille of not optimising everything (FR).

• Random: Pokémon Go data trained AI now used to assist military drones in war zones. China’s UBTECH U1 robots mimic dead relatives with silicone skin, emotion reading, custom replicas of specific people. Air Karen: AI that fights airline customer service.

The Biological Pump

Vera had been diving for eleven years and had never told anyone about the light at five hundred metres, partly because there wasn’t any, and partly because she suspected they would find the darkness disappointing.

She was eating a croissant in the hotel lobby when Derek sat down across from her without asking. This was, she had recently learned from a memo circulated at her institution, potentially no longer acceptable. The memo had specified that “unsolicited approaches to seated colleagues” fell into a category of touch-adjacent behaviours requiring “advance consent alignment.” Vera had not been touched. She was sitting in a chair. She said nothing and offered him half the croissant.

“I had a resident today,” Derek said, “who ordered the ECG first, then asked the AI for a differential, then looked at the ECG. In that order.”

Vera considered this. “And?”

“The AI was right. It’s always right. But she didn’t look at the ECG first.”¹

The conference was called Intelligent Systems in Depth, which could have meant almost anything, and did. Vera was presenting tomorrow on what she called “the dark middle of the ocean” — the mesopelagic zone, 200 to 1,000 metres, where 70 per cent of all animal life on Earth existed in perpetual twilight, cycling carbon downward in a process so crucial that without it, global temperature targets would already have been exceeded. The process was called the biological pump. Most people at the conference did not know it existed. Most people anywhere did not know it existed.²

“What she really doesn’t know how to do,” Derek said, still on the resident, “is sit with not knowing. She can’t tolerate ambiguity. The AI removes the ambiguity, so she’s never had to practise with it.”

Across the lobby, a woman in a blazer was running a workshop on workplace touch. Vera had passed the flipchart earlier: Navigating Consent in Collaborative Spaces: A Practitioner’s Guide. About twelve people sat in a circle, no one touching anyone, watching a slide that said Handshake: Moderate. Hug: Evaluate Context. One man at the back was reading something on his phone. He was, presumably, evaluating context.

“The ocean is the same problem,” Vera said. “We go to the deepest point. We go to the surface. Nobody goes to the middle because it’s expensive and inconvenient and there’s no photograph that makes sense to a press release. Organisms live there that we’ve never seen. They carry carbon down into the deep for centuries. And we’re trying to model the whole climate without any baseline data from that zone.”

Derek took a piece of the croissant. “Is there an AI for that?”

“Yes,” Vera said. “And it’s wrong.”³

The workshop had reached a section called Unsolicited Approaches. The man on his phone closed whatever tab he had open — it looked, from a distance, like a live parcel tracker — and returned to listening. Someone across the circle attempted a test handshake. The facilitator gave a thumbs-up.

Later, Vera would write in her field notes: Giving people tools that eliminate uncertainty does not make them better at uncertainty. It makes them better at the tools. She wrote this below a sketch of her dive profile: 200, 350, 500, 650, 800, 950 metres. Each stop a datapoint. Each return trip building a baseline that had not previously existed.

Progress. Of a kind.

¹ She would become an excellent clinician. She just needed eleven more years of disregarding the AI’s confidence intervals to find out.

² The term “biological pump” sounds like something installed below a boiler. It is instead the reason we are alive.

³ Specifically, it was wrong about carbon flux between 400 and 600 metres during thermal stratification events. Vera had been trying to explain this to a funding body for three years. They kept asking about the surface.

“We need to stop evaluating corporate leaders simply on the basis of how much wealth they create for investors. A healthier yardstick would be this: the extent to which leaders create firms that are economically, ethically, and socially sustainable.”
— O’Toole & Bennis, A Culture of Candor (2009), via workfutures.io

“The broadcast ended with God Save The King and, of course, the shipping forecast.” — WITI: The Turning Off Long Wave Edition

• People: RTO is about worship — new research (Adam Grant et al.) finds executive return-to-office mandates correlate with narcissism, not productivity. GLP-1 raises employment 27 points — previously not-employed women starting weight-loss drug treatment are 27 percentage points more likely to be working 18 months later. Brain exhaustion follows muscle rules. RAISE US: AI workforce retraining.

• AI strat: Opus 4.7: 14 hours, $251 — working autonomously, it built a software package that would take 2–17 weeks of human engineering. AI revenues exceed capex depreciation. Discernment is now a skill — what to build and when to leave it, when to have the difficult conversation rather than defer it: capabilities AI doesn’t replicate, now identifiable as the competitive edge.

• Society: “I’d rather date a centaur”. Conversational malnutrition — the growing deficit in meaningful exchange produced by current platforms.

• Tech: BBC Radio 4 longwave: 100 years off — at 00:00 GMT on June 27, the signal ended, signed off with the shipping forecast. EU AI Act: high-risk pushed to 2027.

• Foresight: “Godzilla” El Nino: Africa has months. US anthem ends in a question — “O’er the land of the free and the home of the brave”?.

• Random: The Kennedy half-dollar — finest coin in the US Mint. Against houseplants — “I tried to keep houseplants because everyone did”.

• AI: Chatbot era as twilight — not the peak. Open-weights: own exponential curve — Chinese open-weights models lag the frontier by 6–12 months but are accelerating independently, cheap to run and modifiable by anyone.

Long Wave

Miriam had been monitoring the longwave signal for thirty-one years, which meant she had been monitoring it for longer than most of her colleagues had been alive, which meant she had stopped mentioning it at parties.

The signal lived at 198 kHz. She had become, over three decades, the kind of person who could tell from a slight variation in the carrier tone whether something was wrong at the transmitter before the diagnostics flagged it. This was not a skill she had been trained for. It was a skill she had acquired the way you acquire a knowledge of when a child is about to be sick — from accumulated attention, pattern by pattern, until the pattern was part of you.

The last day was a Friday, which felt wrong. Endings should be Saturdays. Saturdays had a particular quality of finality. Fridays still had the weekend ahead of them, a sense of something deferred. But June 27 was a Friday and nobody had asked her opinion, so she drove to the monitoring station at four in the morning as she always did, made a cup of tea, and sat down.

The letter from RAISE US was on her desk. She had not opened it yet. Her manager Yusuf had forwarded it with a message that said: worth looking at when you have a moment, which was Yusuf’s way of saying: you should look at this now. RAISE US was the new thing — Gina Raimondo’s people, some governors, a mission to help the workforce transition into the AI economy. Workshops. Retraining. New skills for a new era. The letter had a photograph of a woman in her fifties standing at a whiteboard with an expression that suggested she had recently discovered something excellent about the future.¹

The shipping forecast came on at 00:48 GMT. Miriam listened to it the way you listen to something when you know it is the last time, which is to say badly, because you are too aware of listening. Viking, North Utsire, South Utsire. Cyclonic, five to seven, occasionally severe gale eight...

She had heard the shipping forecast eleven thousand times. She knew, from experience, that you never really heard it — it was something you received, the way you receive background radiation or the knowledge that it is Tuesday. The sea areas were a geography she had never visited but knew precisely: Forties, Cromarty, Forth, Tyne. She had once explained to her daughter that the shipping forecast was a kind of poem, and her daughter had said that was very nice, Mum, and asked if she wanted another biscuit.

At 00:00, the final BBC Radio 4 longwave transmission would end. Not the station — Radio 4 would continue on FM and digital. Just this frequency. Just 198 kHz, which had been broadcasting continuously since before the Second World War, which carried the signal further than any other format precisely because the waves were long enough to bend around the curvature of the earth.²

She opened the RAISE US letter.

It was warm, as these things go. It acknowledged that people in her sector had built careers on specific technical knowledge that was now changing rapidly. It said that RAISE US believed in the dignity of work and the possibility of reinvention. It said that Janet Foutty, a former CEO of Deloitte Consulting, had recently joined their leadership and had herself begun again at fifty-six, and that this proved something about possibility. There were modules. There were pathways. There was a beginner’s mindset framework.

Miriam thought about the beginner’s mindset framework for a moment.

Then she thought about the thermionic valves. There were eleven spares. They were the kind of component that was no longer manufactured, which meant that when the last one failed — if the transmitter had been kept running — there would have been no replacement. The decision to end the longwave broadcast had been partly economic and partly about the audience of ninety thousand, which was small but not nothing, and partly about the fact that nobody had ordered more valves. At some point the decision not to order more valves was itself a decision to end, except nobody had said so at the time. The ending had been decided before anyone announced it, in the moment someone decided the infrastructure wasn’t worth maintaining.³

This, she thought, was not specific to radio.

At 00:00, the carrier went down. The monitoring equipment registered it: clean, complete, nothing wrong at the transmitter. The silence was the sound of everything working correctly.

She sat with it for a while. The RAISE US letter was still in her hand. Outside, the sky was beginning to do what July skies do in the north — not exactly light but not exactly dark, a diffuse grey that suggested morning was making preparations. A bird was practicing something.

She had thirty-one years of pattern recognition at 198 kHz, a familiarity with carrier waves fine-grained enough to feel anomalies in her chest, a knowledge of the sea areas that she had never consciously acquired. She did not know what any of this was worth in the new landscape, or whether it translated. She suspected that knowing what to listen for, and how to listen, might be worth something. She suspected that patience with frequency — the willingness to attend to the same signal long enough to hear what was in it — might still have applications.

She also suspected that nobody would describe this as a module.

She put the letter in her bag and drove home in the grey light, through sea areas she had never visited, under a sky that was not quite one thing or the other.

¹ The woman in the photograph may have genuinely discovered something excellent about the future. It is also possible she had been asked to look like this. The letter did not clarify.

² Radio 4’s longwave service was also, technically, one of the official methods by which British nuclear submarines could determine whether the government was still functioning. This capability has now migrated to other formats. What it means for a submarine crew to receive the shipping forecast as news of governmental continuity is a question Miriam had always found interesting and never quite answered to her satisfaction.

³ The eleven spare thermionic valves are still in storage somewhere. Nobody has thrown them away yet. This is either sentimentality or the standard lag between an ending and the administrative closure of an ending, which is always longer than the ending itself.

“When you’re young you think that time moves forward. At 80 you know that it doesn’t, it stands still. We’re the ones that move.”
— Bob Dylan, via workfutures.io

“Technology is legislation. Introducing a new technology is equivalent to imposing new global economic, legal, or societal policies with major consequences — and we don’t know in advance what that policy is going to be.”
— Karl Schroeder, Unapocalyptic

• People: Fatherhood changes men’s brains. Paul Lukas on the unlikely time capsule. Andrew Curry revisits his essay on Hockney refusal of photography: “the camera always lies about space”.
  Critical thinking and inner game.
  Kurt Vonnegut advice to kids.

• AI strat: Ajey Gore on the anatomy of an AI-native org: AI came for a task type, not a job title. GenAI drives nearly 2% of traffic to Walmart and Target . US export controls force Anthropic to shut down Fable 5 and Mythos 5 for foreign nationals.
  Agents and digital IDs, in Estonia.

• Foresight: Karl Schroeder: what comes after science fiction? SF as canary in the coal mine of technology — the canary has been singing. Andrew Curry on systemic change through the Grenfell Tower Fire, disasters as diagnostic: fire revealed the pre-existing failure.

• Tech: Never Give Them Your Face. Canada announces a federal nuclear strategy. The AI infrastructure story is increasingly an energy infrastructure story. DHL to ship freight on wind-powered cargo vessels.
  AI changes our language.

• Random: World Cup camera angles. Outed by autocomplete.

The Translation Layer

Marcus had been told the restructuring was good news for people like him, by which his manager Elaine meant people who were good at the “why.”

What Marcus was actually good at was explaining things to other people who were also good at explaining things. This was, he now understood, translation. He had spent eleven years being the person in the meeting who connected what the strategy people said to what the engineering people built, and he had been very good at it, and it had not occurred to him until last Tuesday that this was the job AI had been eating first.

The new org chart was two layers. There was the “what” and there was the “how.” The “what” had grown. The “how” had gotten faster. The middle — the thing Marcus had been — was described as gone in the memo from HR, which Elaine had forwarded without comment at 6:43 PM on a Thursday.

He had a three-year-old now. Leo. The brain research said that fathers’ brains changed in the caregiving areas — the same parental networks that lit up in mothers, just from a different starting point and peaking later. Marcus had noticed this without being able to name it: a new kind of attention, a vigilance that was not anxiety but something adjacent to it, the way you become aware of every open socket and unlocked gate. He had rewired for a small specific person. He had become, without knowing it, a different system.¹

“What are you going to do?” his wife Petra asked.

He had been on a call that morning about a pilot program requiring employees to verify their identity before accessing certain systems. Facial recognition. His face, scanned and stored. “You can change a password,” the pamphlet said. “But this is more secure.” The pamphlet did not mention what a future he could not currently imagine might want to do with a database of his face at thirty-eight.²

He thought about the memo. He thought about Leo. He thought about the fact that you could not change your face, and that you could, presumably, change what you were, though the org chart suggested the window was closing.

Leo had learned that week to climb out of his crib. He had done it once, at 2 AM, and presented himself at the bedroom door with the expression of a person who has just demonstrated a proof. He had not yet learned that getting out was the easy part.

Marcus went to bed. He lay in the dark and thought about translation. About the layer of work that had been the middle. About what it meant that the biology of caring for something small could restructure a person’s brain while the org chart was restructuring everything else.

He would need, he thought, a new theory of what he was for.

Leo did not climb out of his crib that night. He had not yet worked out that he could do it twice.

¹ The threat detection activation in new fathers’ brains makes a specific kind of sense: you have introduced a variable into your life that is entirely dependent on your continued competence, and you are aware, at some level, that you have never been tested for this. The brain adjusts accordingly.

² The pamphlet is a real genre. Every major biometric rollout comes with one. They explain why this is better for you. They always omit the sentence: “We will have this forever, regardless of what we tell you now.”

”The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. We have to learn the bitter lesson that building in how we think we think does not work in the long run.” — Richard Sutton

”It’s incredibly hard, and maybe impossible, to escape the distorted perspective that’s been labelled ‘temporal chauvinism’: the feeling that the time we’re living in now is the most significant or terrifying one ever, simply because it’s the one you happen to be around to experience.” — Oliver Burkeman

• Tech: AI has arrived in wildlife video feeds: slop animals. Cory Doctorow: build the post-American internet in Canada. All is recorded now, inc with smart glasses. Passports online. Drone swarms.

• AI: AI-native startups run 25% smaller. Generalist frontier models now outperform specialist medical AI. Fei-Fei Li on a functional taxonomy of world models. LLM councils suppress idiosyncratic ideas.

• AI strat: AI is crushing valuations of pre-ChatGPT startups. Momfluencers are pitching AI as a better co-parent than men. GenAI now drives 2% of traffic to Walmart and Target.

• Society: 71% of Americans oppose local data centers. A worker in North Carolina successfully claimed a religious exemption from using AI at work as a Unitarian Universalist. Aidan Walker coins “cuck internet theory”.

• Security: The European Commission publishes the Code of Practice for labelling AI-generated content. 60-strong Scientific Panel to advise on systemic risks from GPAI models.

• Geopolitics: Ireland’s EU presidency (from July) plans to roll back digital regulation.

• Foresight: What remains scarce after AGI? Oliver Burkeman on temporal chauvinism.

• Art: Requiem for the DVD menu — a disappearing threshold (Dirt).

• Random: Shenzhen nightclubs nostalgic for ZIRP-era optimism. The pleasures of improvised bookmarks.

The Planning Meeting

Declan had explained the data center to his mother three times, and each time she had asked him if it was like a cloud. He had said yes. This was not technically accurate, but accuracy had not proved useful in their household.

The community center’s projector was broken, so someone had propped a laptop on a folding table and connected it to a television — an old one, still running a DVD of Canadian wildlife footage. The DVD had looped twice since the meeting began. A moose had crossed the same river four times. Harriet, who had placed herself in the front row, had not noticed. She was reading from a clipboard.

“Twelve thousand data centers worldwide,” she said. “North America holds half. Half. And now they want to put one three miles from an elementary school.”

Declan looked at the clipboard. He was sitting in the third row, near the exit, because Pramila in Communications had said it was important he attend and “not sit near the front like a target.” He was, he supposed, a target. His firm had recently completed a restructuring. From forty engineers they had kept three. The AI systems they maintained were faster and more meticulous than he had ever been, and genuinely did not care about lunch.

“The power draw alone,” Harriet continued, “is equivalent to twenty-six thousand households. Per facility. And they can just keep building them.”

Bram, two seats to Declan’s left, was reading a novel. He had brought his own chair cushion. He had been coming to planning meetings since 1994 and owned a collection of the programs from each one, which he used as bookmarks along with ticket stubs, bus transfers, and a laminated fortune cookie slip that read You have the resources you need. He did not look up.

“Every generation thinks they’re living through the worst of it,” Bram said, to no one in particular. “That’s just how time feels from the inside.”

Harriet turned. “Excuse me?”

“Temporal chauvinism,” Bram said. He had recently read a newsletter piece that used this phrase, and found it clarifying. “The belief that your own era is uniquely catastrophic. Turns out it’s universal. Romans had it. Probably mammoths.”

The moose crossed the river for the fifth time. Declan had been watching wildlife videos more since a colleague in Oslo had mentioned that most viral nature clips on social media were now AI-generated — animals that had never existed traversing landscapes assembled from satellite data and photographic libraries. This moose, which was almost certainly real, seemed reassuring.¹

“We’re not saying the world is ending,” said Harriet, tightly. “We’re saying the aquifer might.”

Bram nodded, apparently satisfied with the distinction.

Declan thought about Ruth, who had sat three desks from him before the restructuring, and who had subsequently claimed a religious exemption from using AI at her new firm, on the grounds that her faith required her to be present in her work, fully, without proxy. He had envied her clarity of conviction, even as he had spent three months onboarding the system that now did most of what he used to do in less than half the time.²

Outside, the town voted against the data center by sixty-one to forty-nine, not counting Bram, who had left early to walk his dog.³

The moose crossed the river one final time and was replaced by the DVD menu — looping, patient, its eight-bar theme playing into the empty room.

¹ His colleague had since joined a consortium developing certification standards for authentic wildlife footage. The consortium had four members and met on alternating Tuesdays.

² Ruth’s exemption was granted under a statute designed for workers whose faith prohibited them from handling specific materials. The theological reasoning she offered was never disclosed.

³ Bram’s dog was named for a Byzantine empress. He had mentioned this to no one.

“Do I contradict myself? Very well then I contradict myself, (I am large, I contain multitudes.)” — Walt Whitman, Song of Myself

“The logic that protects the individual exposes the group. The defense becomes the advertisement. The sardines never figured this out. The Marlin didn’t either. Something always gets eaten.” — Colin Nagy, The Bait Ball Edition

• People: Soft-off days and soft-on days — the WFH lexicon of quiet deceleration. Paris discovers contrast therapy as ceremony. Gen Z migrates to listening bars (Japan’s ongaku kissa) for low-intensity sociality (FR). Steak ‘n Shake goes MAHA: tallow fries, Bitcoin payments, RFK Jr. pinned post. Second axial age. Digital clutter.

• AI: Claude 5 Fable works 12 hours on multi-page specs — autonomous isochrone maps, social science papers, art from pure maths. Cisco: AI power users trust their teams less than intermittent users. New cottage industry: lawyers defending students accused of cheating with AI.

• AI strat: Sanders, Altman, Trump, Khosla converge on sovereign wealth fund to distribute AI’s gains. Anthropic publishes framework for optional pause on frontier AI development. AI-spending US firms grow revenue 5× faster than the economy.

• Tech: Synthetic diamonds surge as heat spreaders for AI chips — they conduct heat but not electricity. Canadian startup builds tractor without computer controls: right-to-repair in steel and mud. EU Open Source Strategy.

• Business: The Low-Wage 20: 20 S&P 500 corporations employ 6.7 million US workers at wages below Medicaid eligibility while CEO pay climbs. The vanity trap in leadership.

• Foresight: Nine most under-discussed cultural trends from 130 cultural researchers worldwide. The grimoire of tiny rebellions. Urgent optimism.
  Football foresight.

• Random: Berthold Lubetkin’s 1934 Penguin Pool. A 1928 buffalo nickel received as change from a bodega: 23 million made, worth $2. One Minute Park.
  The MeDoraH project.
  King of Switzerland.

The Soft Ones

The third week of June, and Miriam had not spoken to anyone in eleven days.

This was not precisely true. She had exchanged 847 messages through various AI interfaces, which she privately counted as a form of conversation. She had also said “thank you” to a delivery driver who had looked at her with the particular wariness of someone who suspects the door is opened by a different person each time.

The coworking space at the bottom of her street had been Doris’s idea. “You’ll go feral,” Doris had said, with the authority of someone who had witnessed actual ferality in a colleague she would not name. So Miriam went, on a Tuesday, carrying her laptop and a sense that she was doing something virtuous. She sat at a long table. To her left, Barry was having what he described as a soft-on day.^1

“It’s not slacking,” Barry explained, with some urgency, to no one in particular. “I’m responsive. I’m chipping away. I’m just not pushing myself to maximum productivity.” He had a rerun playing in a browser tab and was annotating a document at approximately the pace of geological time.

Doris arrived at half past ten, sat down, and immediately drew something on a sticky note. She drew it with the focused expression of a person who had read something about sigils on the internet and decided that if the world was going to be absurd, she might as well be absurd back.^2

“What is that,” said Miriam.

“A threshold marker,” said Doris. “I cross it every time I leave the building. To see if it starts to mean something.”

“That’s not how meaning works,” said Miriam.

“You don’t know how meaning works,” said Doris, pleasantly.

Barry looked up from his soft document. “I read that Cisco found their heaviest AI users trust their colleagues less. Like, statistically.”

“That tracks,” said Doris.

“I trust my colleagues perfectly well,” said Miriam. “I have excellent relationships with my colleagues.”

A silence settled. Miriam’s colleagues, in the operational sense, were large language models that did not experience trust in either direction.

What happened next was harder to explain. The three of them stopped working simultaneously — Miriam closing a tab, Barry letting his rerun loop to a menu screen, Doris capping her pen — and sat for a moment in the specific quiet of a Tuesday afternoon in June when nothing is urgent. Outside, someone walked past with a dog. The dog stopped and stared through the window with the bait-ball instincts of a creature that has learned that clusters of beings sometimes mean food, and sometimes just mean beings.

“I found this thing,” Barry said. “One Minute Park. You spend sixty seconds in a green space somewhere in the world. Just watch it.”

He turned his laptop. A park in Osaka. Dappled light. A child running somewhere off-frame.

They watched for slightly longer than sixty seconds.

“The penguins at London Zoo,” Doris said, apropos of nothing, later. “They built them this beautiful pool in 1934. Cork paving, perfect design. Then they renovated it with concrete and the penguins got sore feet. So they moved the penguins out, and now the pool is Grade I listed and nobody can touch it and it’s been empty for twenty years. Because they changed the material and used the wrong species.”

“What’s the right species?” said Barry.

“It doesn’t matter anymore. They’re somewhere else.”

Miriam thought about this for a long time on the walk home. She thought about it in the way you think about things that feel like a parable but resist being extracted into a lesson. She crossed a threshold — just a crack in the pavement — and did not do anything special when she did.

That evening she sent a message to a colleague. A human one. She asked how they were doing.

The reply came back twelve hours later. Slightly chaotic, slightly contradictory, entirely unhelpful on the actual question she had buried in paragraph three.

She felt, obscurely, better.^3

^1 A soft-on day: online, responsive, chipping away, ideally from one’s couch in cosy clothes with a comfort rerun in the background. Not to be confused with a soft-off day, which is when you run personal errands during work hours and feel no guilt whatsoever. Both are now considered mandatory by significant portions of TikTok.

^2 The RADAR newsletter has been publishing a “grimoire of tiny rebellions” — small embodied acts designed to prove that different arrangements are possible. The underlying thesis: you cannot think your way out of stuck patterns. You have to act your way out, starting with gestures small enough that nobody tries to stop you.

^3 Cisco’s internal data from January 2026 found that its most active AI users trusted their teams less than intermittent users — “likely because the power users were spending more time on their own and less time with their colleagues.” The coaching platform BetterUp found these same users reported higher burnout and a greater desire to leave their jobs. Twelve hours for a reply is, in this context, a good sign.

“Cognitive debt is where you forgo the thinking in order just to get the answers, but have no real idea of why the answers are what they are.” — John Willshire, Smithery

“Whereas in the past humans struggled against exploitation, in the future the biggest struggle will be against uselessness — and it will be much worse to be useless than to be exploited.” — Yuval Noah Harari, via Umanz

• People: “The AI economy promises profits without career paths“ — Cottom on girl-boss AI gaslighting and the pyramid with no middle. Mattering. Cognitive debt — “all the holes in your business where the humans would have been”. Developer deformation. Make some tea.

• AI: Anthropic reports AI now writes 80% of its code — 8x developer output since 2024. Coding agents yield 17x more volume across firms. Mythos.
  Karpathy knowledge management w/ LLMs.

• AI strat: Labs move from bundles to metered pricing for coding tools, one user racked up 130B tokens in a month. Uber caps at $1,500/developer/month. CPM→CPC: the internet advertising precedent.
  SysML, and SYSMOD creation with Claude #MBSE.

• Foresight: Co-Existence. Book website now has a page written for AI agents. Have books come to an end?
  Anthropic calls for the option to pause frontier development if recursive self-improvement takes hold.
  Swiss capped at 10 million people?
  Fables from the future.

• Security: Beaufort castle: “every invading army since the Crusades has seized, held, and lost it” — now in the age of drones.
  & DIY: Tracking V2X with ESP32. Lightning sensors.

• Random: Deep sea exploring. PNT. Funny sea hares - they cut their own heads and survive it.

The Holes Under the Building

Neville had not written a line of code himself in eleven months, and he had never been more productive.

The quarterly metrics said so. His team lead said so. The dashboard said so in a cheerful green font that Neville had spent six minutes asking an AI to design, and then another four asking it to make slightly less cheerful. He was, by every available measure, eight times the engineer he had been a year ago. He had told his mother this at Christmas and she had asked if he was eating.

The trouble began on a Tuesday, as trouble usually did. A junior colleague named Harriet asked him to explain a function he had shipped the previous week. Neville looked at the code. He recognised it in the way you recognise a street you have been driven down many times but have never walked.^1

“The AI wrote it,” he said.

“Yes,” said Harriet, “but you submitted it.”

Neville had a feeling there was something important in the gap between those two sentences, like a word you know exists but cannot quite retrieve. He stared at the function. It was correct. It was elegant. It solved the problem he had described in a prompt that had taken him forty seconds to compose. He had no idea why it worked.

Tim from Finance arrived at this moment to announce that the team’s agentic coding allocation for the month had hit 94% by the 9th. He said this in the tone of a man who had discovered something alive inside the photocopier.^2 Neville’s team had, by conservative estimate, consumed enough tokens to write the complete works of Balzac seventeen times over, in exchange for which they had produced software that the company would use and nobody would understand.

“What do you do when you don’t use the AI?” Harriet asked him that afternoon. She was the sort of person who kept a notebook — a paper one, with a pen — and who sometimes spent eight minutes staring at a blank page before writing anything. Neville had always found this suspicious.

“I use the AI,” he said.

She nodded slowly, as if confirming a hypothesis. She had recently been reading about a Japanese clothing manufacturer who had kept antique knitting machines running purely because the fabric they produced was, by some textile argument Neville could not follow, better. The machines were very slow. The sweaters cost four hundred pounds. There was a waiting list.^3

Neville went home and tried to write a function himself, just to see. It took him forty minutes to produce something he would previously have written in four. The shape of the thing was wrong in ways he felt but could not name, the way a chord can be almost right without being right. He looked at it for a long time.

He did not delete it.

The next morning he arrived early and fixed the original function — the one Harriet had asked about — without asking the AI anything. It took two hours. He felt, when he had finished, a kind of exhaustion that was adjacent to satisfaction, which he had not felt in almost a year, and which he had not known he was missing because the dashboard had been so cheerfully green.

He wrote to his mother. He said he was eating fine. She asked if he was sleeping.

He said he was learning.

^1 The IT department later described the function as “correct but inscrutable” — a phrase that also appears on at least three pieces of public art Neville had walked past without stopping.

^2 Tim subsequently discovered that the photocopier problem was unrelated. It was a sandwich. He did not disclose this.

^3 Harriet had not bought one. She was making something instead, on a pair of old needles, which was taking much longer than she had planned and was turning out slightly wider than intended, which she found she did not mind.

“You can see the computer age everywhere but in the productivity statistics.” — Robert Solow, 1987, quoted in Exponential View

“Systems pushed too far in one direction eventually summon their opposite.” — Why Is This Interesting?

That “Built with Fable” X thread.

• Society: Singapore’s Singlish, creole the state tried to suppress, now a national identity marker. Enantiodromia: internet maximalism about to summon its opposite. Post-reality aesthetics.

• Planet: AI for ecological defenders. Mining that looks like farming: the slow-dawning model shift.

• AI: Token volumes up 17,000× in four years as inference costs collapsed. Only 15–20% of agentic token consumption is real inference. LLMs as fuzzy compilers. AI prose as meaning-shaped attention vampires.

• AI strat: Individual gains don’t compound — “1+1+1+1 = 1.5” at the firm level. The J-curve of general-purpose tech: electrification took 40 years to show up. Agentic coding is a trap. Post-AI employment concentrated in a tiny sliver.

• Business: 25 years of startup punditry failure. The AI-era bias toward action — Kahneman’s System 1 risk at organisational scale. Consent beats consensus in fast-moving organisations.

• Security: Fuel-tank readers at US gas stations breached. Chatbots know more about you than you realise. The quiet rise of emotional surveillance.

• Tech: Uber turning its drivers into a sensor grid for autonomous vehicles. $300,000 robot dogs now patrolling large data centres.

• Foresight: Les trois horizons compressés — stabilisation, innovation, rupture, désormais simultanés (FR). What will be scarce in the post-commodity future of work.

• Random: Leaf sheep slugs is the animal that photosynthesises on stolen chloroplasts. Iconic restaurants taking reservations for 2046. Keith Haring’s to-do lists. US federal courts and their font question: Times New Roman. Wrigley’s Fascinating Artificial Flavor.
  The EU OpenSource Strategy.

The Residual

Derek had prepared seventeen slides for the AI Value Summit, which was the kind of name that felt serious until you said it aloud. The slides documented a productivity paradox in careful sequence: his firm’s engineers had submitted 40% more code in Q1. They had closed 60% more tickets. Pull requests had doubled. And yet the organisation had not shipped proportionally more of anything that mattered. He had prepared a graph for this — a slightly baffling bar chart he thought of privately as the 1.5 Slide.

“One plus one plus one plus one equals one and a half,” he said to the room, and paused, as if someone might disagree. The room did not. Avery, the consultant they’d flown in, nodded in the slow way of someone who had seen the same graph in different colours at different companies across an entire industry that had spent three years measuring input and calling it output.

“Lightbulb stage,” she said. “The factory got better lighting. Nobody changed the factory.”

Hank, officially there as head of data science but having spent the morning thinking about sea slugs, looked up. He had read, somewhere in the week’s accumulation of newsletters, about an organism that stole chloroplasts from algae, incorporated them into its own tissues, and ran briefly on solar energy. The slug never actually learned to photosynthesise; it borrowed the machinery and trusted the light to last.^1

“We’re doing kleptoplasty,” he said.

Nobody asked him to continue but he did. “We’re incorporating AI into existing bodies. We haven’t changed the bodies. We’re borrowing the mechanism and hoping the efficiency diffuses.”

Derek considered this. He had forty minutes left in his slot, and an action plan that was basically “do more of the same faster,” which he now suspected was precisely the wrong lesson.^2

Matt, from Infrastructure, interrupted without quite meaning to. He had pulled something up on his phone — a facilities dashboard accessible through the hotel’s sensor API, which had not been configured with authentication. The dashboard tracked something called an Ambient Affect Index for each conference room. This room was reading 62 (Mildly Tense) at the moment of the 1.5 Slide, and had since dropped to 48 (Disengaged).

“The hotel is measuring how we feel,” Matt said.

“Hotels have done comfort profiling for years,” Avery said.

“This is a corporate conference centre. This data goes somewhere.”

Avery looked at the number. 48: Disengaged. She thought of a tipping-point study she half-remembered — 25% of a group, certain enough, could redefine the group’s collective understanding. The question was what understanding she was in 25% of.

“We usually think of these horizons as sequential,” she said, addressing no one in particular. “Stabilise first, then innovate, then transform. But what we’re actually in is all three simultaneously. The infrastructure is being disrupted while it’s being built. The transformation is happening to systems still looking for a stable floor.” She gestured vaguely at the sensor dashboard. “And the building is monitoring our feelings about it.”^3

Derek closed the 1.5 Slide. He had not resolved the graph’s central claim — that individual gains, summed across hundreds of engineers, didn’t compound into organisational results. The explanation wasn’t in the data. It was in the invisible work the data didn’t capture: context re-read forty times; the overhead that looked like inference but wasn’t; the residual that got called productivity because no one had given it another name.

Outside, through the glass, the day’s other events moved in both directions: a wedding reception, a wellness retreat, a seminar on navigating uncertainty. None of them, Hank noticed, were running on their own energy.

^1 The leaf sheep sea slug (Costasiella kuroshimae) can photosynthesise for weeks using stolen algal chloroplasts. Ecologists call this kleptoplasty. Enterprise architects do not use this word, but they might find it useful.

^2 “Do more of the same faster” is Stage 1 of general-purpose technology adoption, corresponding to improved lighting in the old factory. Stage 2 requires restructuring the factory. Stage 3 takes approximately forty years and a generation of managers who don’t remember the previous arrangement.

^3 The three-horizons model — stabilise, innovate, transform — was designed as a planning sequence. Its compression into a single present moment is, depending on who you ask, either a crisis of coordination or an exciting opportunity. The consulting industry charges differently for each interpretation.

“An Event is something that arrives without the right word attached to it. You reach for your vocabulary and come up empty. You reach for the law and find it wasn’t written for this.”
— Where Do You Make the Cut?, Karl Schroeder

“Ernest speaks with the authority of success. I speak with the authority of failure.”
— F. Scott Fitzgerald, via La lettre de Umanz

• Society: One in three American men not working and not looking. Partnership as class privilege: 43% of French living alone by 2030. Universities “rich in information, poor in meaning“. Local (AI) news outlets for PR stunts. Disneyfied world. Cartoonification. The security envelope liner collectors. Thai restaurants in America — Bangkok paid for them. Wargaming’s long path to D&D. Best japanese stationery.
  Work in a world in flux.

• Planet: Chernobyl fungi. Whales chatting.

• AI: China’s efficiency moat from export controls. Backlash growing: booing at commencements. Anthropic CFO: $50B ARR. Microsoft and OpenAI consciously uncoupling — $13B invested, $30B returned, no longer exclusive. Who taught them to write with AI?.

• Regulation: EU Digital Omnibus agreed: high-risk AI rules delayed to December 2027. From August 2026, EU residents must be told when talking to AI. Europe and drones.

• Foresight: AI as Badiou’s Event — conditions for which our legal and ethical vocabulary was never written.

• Hardware: Tetris on a book cover.

The Water Test

Derek had printed out three articles about the AI backlash before Hank arrived. This was, Derek thought, either research or anxiety management — he wasn’t sure which line separated them anymore.¹

Hank came through the back door at noon, as he had every weekday for the past three months since he’d left his job at the fulfillment center. Not quit. Left. There was a distinction Hank maintained carefully, which Derek had learned not to challenge.

“Water still coming out right?” Hank asked, filling a glass at the kitchen tap. He held it up against the window. Clear. Hank had been monitoring the tap water since a data center was announced in the next county, and his vigilance had taken on a quality Derek recognized from his grandmother’s rosary habit — not really about the thing itself, but about the rhythm of the ritual.

“It’s fine,” said Derek.

“AOC held up a jar of brown water from one of these sites on C-SPAN. Just lifted it up and let it sit there.” Hank set the glass down. “Brown.”

Derek didn’t answer. He worked for a company building AI scheduling tools for hospitals. He had a complicated relationship with the data center question — somewhere between civic discomfort and financial dependency.

Avery came in from the living room, still wearing her teacher lanyard. She taught eleventh-grade English and had spent the morning marking what she’d begun calling “the slop pile” — essays that were technically adequate, grammatically flawless, and entirely empty. Not written by students who were cheating; written by students who had never been taught to care.² She dropped the stack on the table with a sound that communicated everything.

“How was it?” Hank asked.

“I gave eight of them feedback on the wrong essay,” she said. “They’re all the same essay.”

Hank nodded. He’d been spending his free time — and he had a lot of it — reorganizing his collection of security envelope liners. The patterns inside the envelopes that banks and doctors’ offices sent: cross-hatch, diagonal wave, fine-grain spiral. He’d found a community online — thousands of them — cataloguing and trading photographs. It turned out this was a real thing. It turned out a lot of things were real things once you had time to notice them.

“Charlie found one from 1987,” Hank said, to no one in particular. “A pattern they haven’t made since. Just the diagonal, extremely fine, like fish scales.”

“What do you do with them?” Derek asked.

“Just know they exist.”

That night, after Hank had gone, Derek read an article tracing a straight line from the Prussian Kriegsspiel — the 19th-century wargame used to train officers — through a series of American hobbyists and miniature painters, directly to Gary Gygax and the first edition of Dungeons & Dragons.³ The path was absolutely clear once you saw it. Military doctrine becoming leisure becoming mythology. He thought about Hank’s envelopes. About the slop pile. About the water.

He held his glass up to the kitchen light. Clear. He set it down.

¹ The same question haunted several tech executives that week, who were solidly booed at commencement addresses. The water test, in retrospect, applied to them too.

² This is not the students’ fault. It is difficult to care about an outcome when every incentive in the system you have grown up in has been carefully designed to reward the product rather than the process.

³ The article was in Asterisk. It turns out war has been a hobby since war was invented.

“We learn to think of history as something that has already happened, to other people. Our own moment, filled as it is with minutiae destined to be forgotten, always looks smaller in comparison.”
— M. Gessen

“When the future loses its promise, the past becomes a refuge.”
— Yi-Ling Liu, via workfutures.io

• People: Quieter beta mums. Tarot booming in China. Economics of romance. Omiokuri.
  Hemingway and loss.

• Society: The job ladder in structural decline — workers today half as likely to receive a better-paying outside offer as in the 1980s. Engels’ pause. Smartphones creating a myopia epidemic in children.
  Paris “Tiers lieux”.

• Foresight: Extrapolated Futures Archive — a reverse-lookup for speculative fiction, organised by scenario type. Thin places — Celtic liminal spaces (in Irish, aiteanna fliuch, “wet places”).

• Tech: Your car is spying on you. Surveillance pricing is coming to grocery stores. Data centers overtook office construction in December 2025. Export controls are inadvertently capability-generating.

  Producing all psychotropes in simple tobacco.

• Random: Japanese snack packaging going grayscale — an ink shortage is stripping colour from the shelves. Martin Margiela, in a 1998 interview, on colour: “What is black? An absence, a presence, a mood, a coat. What is red? A blush, a redness, a fever, an order. What is skin? Protection.”

• AI: Chinese labs: Claude is the preferred coding model across every lab visited. Vibe coding platforms (Loveable, Replit) leaking medical records, financial data, and Fortune 500 internal documents onto the open web.
  Token inequality.

The Efficiency Moat

Bev had printed out seventeen pages of the report and highlighted them in four colours, which Marcus thought was itself an argument against the report’s thesis. The report argued, among other things, that the people who had learned to do the most with the least were now the most dangerous.¹

“They’re getting four to seven times more intelligence per unit of compute,” she said, fanning the pages across the conference table. “Seven times. We’re just sitting here with all our chips, burning money.”

Marcus looked out the window. Their colleague Pat was walking to the parking lot below, and Marcus watched until Pat turned the corner and was fully gone from sight — which took longer than expected, because Pat walked slowly and the corner was further than it looked. Then Marcus turned back to the pages.

“Excel made more accountants,” he said finally.

Bev stopped. “What?”

“When Excel came out. 1993. Everyone thought it would replace accountants. It made more of them. Jevons paradox.”

“That’s not what I’m talking about.” Bev collected the pages and tapped them into a neat pile, a gesture she had learned from her mother, who had been the kind of parent who made her children’s homework schedules in advance and colour-coded them by subject. Bev had decided this approach had some merit, though she preferred to describe herself as “intentional” rather than “intense.”²

In the break room, someone had replaced all the snacks. The packaging was different — flatter, more subdued, everything in grey and black. The chips had always come in a shiny orange bag. Now they came in what appeared to be a photocopy of a shiny orange bag. “Ink shortage,” said the note on the fridge. “Normal service will resume when the supply chain does.”

The Cerebras chip had arrived that morning in a foam-lined case the size of a piece of luggage. It was a wafer — a literal silicon wafer, the size of a dinner plate — and it had cost more than Marcus’s car and possibly his flat. Bev had photographed it immediately and reverently, the way people photograph their children’s first steps. The chip was now in the server room downstairs, where it was, in theory, thinking.³

“The point,” Bev said, returning with a grey-packaged coffee, “is that efficiency is the new moat. If you learn to do more with less, the resources stop mattering.”

Marcus thought about this. He thought about Pat turning the corner, the way a person’s back tells you something their face doesn’t. He thought about how the tarot lady had told his sister that 2026 was the year of the Wheel of Fortune. Not the game show. The other kind.

“Maybe,” he said. “Or maybe constraint is just constraint. And we’ll know in ten years.”

Bev gathered her highlighted pages and left. Marcus watched her go until she turned the corner into the corridor and he could no longer see her, and only then did he get up to get his own grey-packaged coffee from the break room.

It tasted exactly the same.

¹ The most efficient people: always the ones who were forced to be.

² Colour-coded homework schedules remain, statistically, no guarantee of anything — though they do correlate strongly with the ownership of multiple highlighters.

³ Whether wafer-scale thinking is different in kind from smaller-scale thinking is a question the chip itself was not invited to answer.