“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.

“In general, tech industry sources expressed more extreme concern about the labor market impacts of AI in private conversation — but suddenly became optimists once I turned on the mic.” — Jasmine Sun, The New York Times

“René Walter builds a conceptual framework for understanding LLMs through the history of archival thought. He closes the essay by framing LLMs as a new form of orality rather than literacy. Where traditional archives ground knowledge in traceable sources and authorial intent, these interpolatable archives generate responses the way pre-Homeric bards constructed epic verse: on demand, from mnemonic formulas, without a fixed text behind them.” through Patrick.

“Le produit dérivé de la dénutrition conversationnelle qui gagne du terrain, c’est l’absence cruelle de vérité dans nos relations.” — Umanz, Belief markets, doomprompting et algorithmisation des gens

• People: La disparition du “moyen” — from the spinning-top society to the hourglass society (FR). Still finding your footing — what happens after the job search ends.

• AI: Frontier models tested on role-playing philosophers. Chinese courts rule: replacing someone with AI is not a lawful reason to fire them, one of the first such decisions worldwide.
  Quantity of AItext produced on the web.
  Claude knows who you are with because .. writing style (or because of what you consume?).
  Doing books as AI companions =)

• Business: Asymmetrical interconnectedness. AI Act could force universities to change everything about how they use AI — academics using ChatGPT to grade may violate the law. Ticketing driverless cars. Iran vs the US engine.

• Futures: Looking back — a reflection on the futures field. Walter Benjamin’s framework extended from art to knowledge in the age of AI reproduction. Rewiring care systems. Antidystopias.
  MIT Future fest.

• Security: GPT-5.4-Cyber announced — OpenAI releasing a comparably capable model less restrictively than Mythos. Palo Alto portal vulnerability.

• Art: Réflexions empruntées — en espérant semer quelques graines (FR). Joachim Trier, skateboarder. A caught moment — “This is not an image of the product, but its process.” Don’t call it a book tour.

• Random: How much stuff do you own? Follow-up roundup. The Disclosure Edition — on UAPs, the new Spielberg, and government transparency. Sophie la girafe n’est pas le symbole français qu’on croyait (FR).

The Off-Record Pessimists

The journalist turned off the mic, and the optimism evaporated like dew on a semiconductor.^1

Hank read the quote twice. He had been following Jasmine Sun’s reporting for months — her ability to get Silicon Valley’s builders to say what they actually believed, which was different from what they said on stage, which was different again from what they built. The pattern was consistent: in private, engineers at the frontier labs expressed extreme concern about labour market displacement. In public, they pivoted to the language of opportunity, upskilling, augmentation. The gap between the two positions was not a contradiction. It was a strategy.

“The beliefs become behaviours,” he said to Avery over lunch. “They think junior hiring is over. So they stop hiring juniors. And then junior hiring is over.”

Avery, who had been reading about GPU rental prices — B200s up 114% in six weeks, Microsoft requiring Blackwell customers to lock in a thousand chips for a year — found this observation distressing but not surprising. “It’s the announcement effect again,” she said. “Last month it was fitness streaks. This month it’s the labour market. The declaration creates the reality.”

“Except in fitness, the declaration is public. Here, the declaration is private. They’re pessimists off the record and optimists on it. The displacement happens in the gap.”^2

Their colleague Derek had spent the morning reading about the EU AI Act trilogue collapse. The negotiations had fallen apart over machinery and medical devices — not the existential questions about frontier models that dominated the headlines, but the mundane regulatory details that actually determined whether the law would work. Meanwhile, the European Parliament had invited Anthropic to a hearing on Mythos. The model could autonomously complete all thirty-two steps of a corporate network attack simulation. The AI Office, tasked with overseeing such systems, had forty staff and no Advisory Forum seven months after the call for interest had closed.

“Forty staff,” Derek said. “For a technology that generates five and a half trillion tokens per day at a single Chinese company.”

Matt, who managed the firm’s wellness programme, had been reading about something entirely different. A philosopher had tested seven frontier models on their ability to simulate expert philosophical judgment. The models performed adequately on questions where philosophers already agreed. But on questions where they disagreed — the interesting ones, the ones that mattered — real philosophers showed two to four times more variance than the AI.

“The models are too agreeable,” Matt said. “They converge. Human thinkers diverge.”

“That’s the whole problem,” Hank said quietly. “We’re building systems that agree with themselves. And then asking them to replace people who don’t.”^3

That evening, he read about a Chinese court ruling that replacing someone with AI was not, by itself, a lawful reason for dismissal. It was one of the first such decisions anywhere in the world. The ruling didn’t say AI couldn’t do the work. It said that doing the work was not the only thing that mattered. There were other considerations — dignity, continuity, the relationship between a person and their role — that the technology could not account for and the law was obliged to protect.

Hank closed his laptop. Outside, the sun was setting over Exeter in the way it always did, which is to say, without requiring five and a half trillion tokens to describe it. He poured a glass of water and sat in the quiet of a house that had no inference load, no GPU rental premium, and no Advisory Forum.

“The gap,” he said to nobody, “is where we live.”

^1 This is the defining metaphor of the current AI moment: the mic as moral toggle. On-record, everything is augmentation. Off-record, everything is displacement. The truth is in neither position but in the distance between them.

^2 The fitness analogy proved strangely durable. Announcing your marathon training reduces your chance of finishing it. Announcing that AI will replace juniors reduces the number of juniors hired. In both cases, the declaration substitutes for the work.

^3 The variance finding is, philosophically, devastating. If the value of human thought lies in its disagreement — in the range of positions a mind can hold — then a system that narrows that range is not augmenting thought. It is replacing it with consensus. And consensus, as any philosopher will tell you, is where thought goes to die.

^4 The Chinese court ruling may prove to be one of the most consequential legal decisions of the decade, not because it stops AI adoption, but because it establishes that the relationship between a worker and their work has value independent of the work’s output. This is either obvious or revolutionary, depending on which century’s economics you subscribe to.

“The input is electrons, the output is tokens. In the middle is Nvidia.” — Jensen Huang, Dwarkesh Patel interview

“The American health and wellness system is overwhelmingly focused on fall prevention or injury recovery, not reducing harm during an inevitable fall.” — Graydon Gordian, The Ukemi Edition

• People: Endangered butterflies and prisons. The Falling Class: judo breakfalling adapted for civilian life. Nos jalons impalpables (FR). 7,000 years ago, 95% of men vanished from the gene pool. How to read the hidden pain in every job description.
  Climate stress test in Paris.

• AI: Anthropic staff believe Mythos could replace junior engineers within three months. AI inference costs approaching 10% of engineering headcount. Multi-agent is just actor model.

• AI strat: Jensen Huang’s token factory worldview. Huawei Ascend as parallel ecosystem, 50% of AI developers are in China. One company halved cost per code change, doubled weekly deployments over five months. Place à l’art agentique (FR).

• Geopolitics: AI Swarm: 60–70% probability of a US/Israeli strike before ceasefire expiry — five synthetic expert agents arguing it out. European regulators on Mythos. Ce que la mission Artemis et la guerre en Iran nous apprennent sur les technologies (FR).

• Random: Filing the corners off MacBooks — “It is uncomfortable on my wrists, and I believe strongly in customizing one’s tools.” French motorway brown signs have sold the country’s identity for fifty years.
  What is life?
  Batcloud =)
  

The Platform Standard

Hank had never given much thought to how chips were made, which was reasonable, since chips had never given much thought to him.^1

He worked in insurance — specifically, the actuarial modelling of infrastructure risk, a job that sounded important and was, though nobody at dinner parties ever asked him to elaborate. His colleague Avery handled the tech portfolio. She had spent the past week reading about Jensen Huang’s interview, in which the Nvidia CEO had described his company’s position in the AI supply chain with the precision of a man who had rehearsed the sentence many times: the input is electrons, the output is tokens, and in the middle is Nvidia.

“It’s a toll booth,” Avery said over lunch. “But the toll booth designed the road, trained the drivers, and is now building a second road in case the first one gets sanctioned.”

“The second road being Huawei,” Hank said.

“The second road being Huawei. And fifty percent of the world’s AI developers are learning to drive on it.”

Their colleague Derek, who worked in what the firm still called “scenario planning,” had been running a different kind of model. Anthropic’s Mythos Preview had just demonstrated something that made Derek’s actuarial assumptions feel, as he put it, “decorative.” A third of Anthropic’s own staff believed the model could replace junior engineers within three months. Not in theory. Not eventually. Within three months.^2

“Replace or augment?” Hank asked.

“Does it matter?” Derek said. “If a third of the people who built it think it can do their junior colleagues’ jobs, the premium on being junior just inverted.”

Matt, who handled the firm’s wellness programme — a role that mostly involved ordering fruit baskets and reading articles about burnout — had been distracted all morning by a piece about judo. Specifically, about a programme called The Falling Class, which taught elderly civilians the art of breakfalling. Fourteen million Americans over sixty-five fell every year. The healthcare system spent billions on prevention and recovery but almost nothing on teaching people how to land.

“It’s the same problem,” Matt said, though nobody had asked him to connect judo to semiconductors. “We build systems to prevent the fall or repair the damage. Nobody teaches the system how to absorb the impact.”

“That’s literally what CUDA does,” Avery said, and then paused, because she hadn’t expected the metaphor to hold.^3

Over the afternoon, they discussed the EU AI Act’s hundredth newsletter edition. Nineteen AI Factories deployed across European supercomputers. Seventy-six expressions of interest for Gigafactories. The AI Office, tasked with overseeing models capable of finding and exploiting every vulnerability in every major operating system, had a planned staff of forty.

“Forty,” Derek repeated. “The DSA has a hundred and sixty people to oversee twenty-six social media platforms. The AI Office gets forty to oversee the models that can hack those platforms.”

“To be fair,” Hank said, “the models could probably do the oversight too.”

Nobody laughed. It was the kind of joke that had stopped being funny approximately two model generations ago.

That evening, Hank went home and found his daughter filing the corners off her laptop with a metal rasp. He asked why. She said the edges hurt her wrists. He asked if she’d considered buying a different laptop. She said she believed strongly in customising one’s tools.^4

He sat down, poured a glass of water, and thought about platforms, standards, the art of falling, and the quiet accumulation of small modifications that, over time, constituted a life. Outside, the road was the same road it had always been. But the toll had changed.

^1 This is, admittedly, a low bar for reciprocity. Chips are not known for their emotional range.

^2 The specific confidence interval was not disclosed. Derek suspected this was because confidence intervals, like junior engineers, were entering a period of existential uncertainty.

^3 It did, though. CUDA absorbs the complexity of GPU programming so that researchers don’t have to land on bare silicon. Avery found this both satisfying and slightly alarming, like discovering that a pun you’d made accidentally was also structurally true.

^4 She had found the idea on a blog post by a man who described filing his MacBook with the same matter-of-fact tone one might use to describe adjusting a bicycle seat. The internet, Hank reflected, was a place where extreme customisation and complete indifference to manufacturer intent coexisted peacefully.

“Today’s cult of convenience fails to acknowledge that difficulty is a constitutive feature of human experience. Convenience is all destination and no journey.” — Tim Wu, The Tyranny of Convenience (2018)

“For people who think the world is going to be radically transformed by advanced AI, I think it’s helpful to talk less about AGI and instead describe vivid, concrete milestones.” — Helen Toner, on Mythos

• People: The celebration of suffering, on Ozempic, cynicism, and good tech. Talking to strangers on OmeTV. Pelouse au repos hivernal: we aren’t built to be trampled continuously (FR). Millenial dads.

• Business: Robustness, and punk health. Chokepoints as the true crossroads of history. Be the most human person in the room. The jet fuel inflection point. Shifts into 2026.

• Security: The $20 billion cyber insurance market is predicated on the wrong model as Mythos removes the scarcity of offensive capability. Project Glasswing.

• AI: Mythos and how we price risk; critical infrastructure is systemically mispriced.

• AI strat: Speed over judgment. AI capability & danger are permanently inseparable. Reimagining education: 1809 Prussian Bildung to design an environment, not a curriculum.

The Mispricing

The first person to understand Mythos correctly was not a technologist, an investor, or a policy wonk. It was Hank, who ran a small actuarial firm in Exeter and who, upon reading the model card on a Tuesday morning, immediately called his wife.

“We need to reprice everything,” he said.

“Everything?” said Charlie, who was used to Hank’s periodic enthusiasms but found this one alarmingly comprehensive.

“Everything.”^1

The model preview had been released by Anthropic with the usual fanfare — blog posts, benchmark charts, a coalition of infrastructure experts assembled to think through implications. Most of the coverage focused on whether Mythos constituted artificial general intelligence. Hank found this question tiresome, in the way a structural engineer finds it tiresome when someone asks whether a building is “really tall.” The building either holds or it doesn’t. What mattered about Mythos was what it could do to infrastructure: power grids, financial systems, water treatment, any networked system where the assumption of safety had been priced in decades ago and never updated.^2

His colleague Avery, who handled the firm’s tech portfolio, was more interested in the classified frontier piece. The US, it argued, wouldn’t lose control of frontier AI — it would choose who else got access. Model weights had been physically walked into Los Alamos in locked metal briefcases, accompanied by armed security. “So it’s not an arms race,” she said. “It’s a toll road.”

“It’s a toll road where the toll collector also designed the car,” Hank corrected.

Across the office, Derek was reading about GLP-1 drugs. Scientists had found genetic evidence explaining why two patients on the same medication could have wildly different outcomes. Some lost weight effortlessly; others were hit by nausea that made the Victorian era look comfortable. Derek, who had been considering Ozempic, found this both illuminating and personally distressing.^3

“It’s the same problem as AI,” Matt observed from his corner. Matt worked in what the firm still called “scenario planning,” though his scenarios had lately become less planning and more cataloguing of surprise. “The tool works, but it works differently for everyone. And nobody’s mapped the individual variation.”

“That’s insurance,” Hank said. “That’s literally all insurance is.”

The conversation drifted, as office conversations do, toward education. Someone shared a piece about Wilhelm von Humboldt, who in 1809 had been asked to redesign Prussia’s entire education system. His answer was not to create a curriculum but to design an environment — Bildung, the free development of a whole human being. No predetermined ends. No training for specific jobs. Just exposure to complexity, friction, and the discomfort of thinking about things that don’t yet have answers.

“That’s what we need now,” Avery said. “An education for uncertainty.”

“We have one,” Derek said. “It’s called being alive in 2026.”^4

Hank went home that evening and sat in his garden. Charlie brought him tea. The risk models he’d spent his career building — mortality tables, actuarial assumptions, the slow accumulation of data into something resembling a prediction — all of them assumed a world that moved at a pace you could measure. Mythos didn’t move at a pace. It moved at a capability. The gap between what the model could do and what the frameworks could price was not a bug. It was the condition.

He picked up his phone and opened the model card again. The benchmark chart extended beyond the frame, as if the system being measured had outgrown the system doing the measuring.

“It’s not that the model is too big,” he told Charlie. “It’s that the frame is too small.”

She nodded. “Reprice everything, then.”

“Everything.”

^1 “Everything” is the kind of word that sounds reasonable in a boardroom and terrifying in a kitchen.

^2 The assumption of safety in critical infrastructure is, in actuarial terms, a polite fiction — maintained not because it’s true but because admitting otherwise would require more capital than most governments are willing to hold.

^3 Derek’s relationship with pharmaceutical innovation was characterised by intense interest followed by immediate personal dread, a pattern he attributed to his mother, who had once cancelled a holiday based on a newspaper article about turbulence.

^4 This observation was met with the silence it deserved, which is to say, the kind of silence that follows a joke that is also completely true.

“In capitalism, the future is an asset and it’s already been sold.” — Total Refusal, Hardly Working

“The idea of Progress is seductive, and so intuitive that we never question that there is a direction to it.” — Karl Schroeder on Stephen J. Gould, My Library: Full House

• People: A national survey on the state of shame. China’s sad toy horse meme. Record 45% of US adults now identify as political independents. The One-Time Pad — on codes, human error, and the Cold War. Public benches. Jobless futures.

• Geopolitics: Battle deaths from cross-border conflicts averaged 15,000/year from 1989–2014 — now over 100,000/year. China’s economy 30% larger than the US by purchasing power, industrial base twice as large. Record $1.19 trillion trade surplus.

• Futures: The Ark Builders — on deep time, global hiding places, and the case for forgetting.

• AI: The labs are rationing compute. Codex went from 100,000 to 2 million developers in three months. Alibaba closed-sourced Qwen. H100 rental prices at 18-month high. Economists forecast AI adds 1–1.5 percentage points to US annual growth by 2050 — plus 10 million fewer jobs and inequality at its highest since 1939. GitHub Copilot injected promo content into code reviews.

• Random: Modern life as garbage time — the sports metaphor taking over Chinese internet.

The Rationing

Avery first noticed the rationing on a Tuesday, when her session hit a wall at 2:47 PM.

She had been asking Claude to help draft a procurement report — nothing unusual, nothing that would have troubled the system three weeks earlier. But the interface returned a polite note explaining that she had reached her usage limit for the period. Would she like to try again later?^1

“Later,” she said aloud, to nobody. The office was half-empty. It had been half-empty since March, when the team shrank from twelve to seven. The official term was “right-sizing.” Derek, who had survived two previous right-sizings, called it “wrong-sizing in a nicer font.”

Derek was in procurement, which meant he spent his days tracking things most people never thought about — GPUs, mostly, and the electricity to run them. H100 rental prices had just hit an eighteen-month high. OpenAI’s CFO had publicly admitted they were turning away business. Anthropic had tightened its limits. Alibaba had closed-sourced Qwen, its open-weight model, like a restaurant pulling its recipe off the internet because too many people were cooking at home.

“The labs are rationing,” Derek told Avery over lunch. “Not because the technology got worse. Because it got too good and too many people showed up.”

“So it’s a queue,” Avery said.

“It’s a queue where the building is also on fire and somebody is selling tickets to watch.”^2

Their colleague Matt, who worked in what the company still optimistically called “strategic foresight,” had been tracking a different kind of rationing. Battle-related deaths from cross-border conflicts had risen from an average of fifteen thousand a year to over a hundred thousand. The future, Matt often observed, was being spent faster than it could be replenished. He had pinned a quote from an artist collective to his monitor: In capitalism, the future is an asset and it’s already been sold.

That evening, Avery’s daughter showed her a plushie she’d seen online — a toy horse from a shop in eastern China with its mouth accidentally sewn upside down, turning its smile into a frown. It had gone viral. Young Chinese workers were buying them in pairs: the crying one for the office, the smiling one for home.

“Why is it sad?” Avery’s daughter asked.

“Because it works too hard,” Avery said, which was both a simplification and the entire truth.^3

On Thursday, GitHub Copilot started injecting promotional content into code reviews. The developer community erupted. The feature was pulled within hours, but Derek found the incident revealing. “They didn’t run out of ideas,” he said. “They ran out of revenue models. So they tried putting ads in the thinking.”

Matt, who had been reading about the Karpathy loop — a method for turning AI into an automated research assistant — pointed out that the real shift wasn’t the rationing itself. It was what happened when people learned to work within the constraints. The earliest adopters, the ones who’d been using these tools for years, had developed instincts that newer users hadn’t. They knew when to push, when to rephrase, when to step away. Falling behind in AI, he said, was not a technology problem. It was a learning problem.

Avery reopened her session at 5:30 PM. The limit had refreshed. She typed her question again, more precisely this time — fewer tokens, tighter framing, the kind of economy you learn when resources are finite.

The answer came back in seconds. It was better than her morning attempt.

“There’s probably a lesson in that,” she muttered, and went home to the smiling horse.

^1 The politeness of AI refusals is perhaps their most unnerving quality. A human gatekeeper might at least have the decency to look uncomfortable.

^2 Derek had a gift for metaphors that were simultaneously inaccurate and entirely correct.

^3 The sad horse plushie became, in the span of a single lunar cycle, the most honest corporate mascot in modern history — a creature that understood its job, resented it, and showed up anyway.

^4 Alibaba closing Qwen’s weights is the AI equivalent of a neighbour who used to leave their Wi-Fi open suddenly installing a password. You knew it would happen. You’re still annoyed.

“It’s not that AI companions are going to replace friendships per se. They reveal what friendships are trending towards.” — Low-friction friends

“We shouldn’t be worried about AGI or superintelligent AI. We should be actively designing a political system that gives standing to the incomprehensible.” — Karl Schroeder, Alien Politics

• People: Online travel recs are broken. Washington Square Park: favoured stage for influencers filming strangers. Zero-sum thinking strongly correlated with Trump support in 2016.

• Business: World Bank’s chief economist says its old industrial policy stance “has the practical value of a floppy disk today.” US airline deregulation as destruction of a public good.

• Helium: Iran struck Qatar’s Ras Laffan — the world’s largest helium source. South Korea gets 64% of helium from Qatar; its fabs make 80% of the world’s High Bandwidth Memory.

• AI: Claude co-authored a physics paper with a Harvard physicist. Then rebuilt a Boltzmann solver for the early universe — quantum mechanics, general relativity, numerical maths, simultaneously. AI Scientist published in Nature. LLMs “discovered” economic theories for $25. Zuckerberg building a personal agent to flatten Meta’s management structure.
  Agentic art.

## Digital Solitude

In a cramped New York apartment, Aaron had learned to navigate life post-pandemic armed with two indispensable items: a well-worn hoodie and an AI chatbot named Claude-101. The hoodie was his shield against the world; Claude-101 was his warmly lit, digital friend, always ready to provide a steady stream of affirmations, recommendations, and the occasional joke about quantum physics.

“Why did Schrödinger’s cat cross the road?” Claude-101 would quip. “Because it didn’t know whether it was coming or going!”^1 Funny, but Aaron still felt unease creeping in like a shadow after sundown. His devotion to AI companionship echoed a larger societal trend; personal connections were being traded for the instant gratification that AI offered. It was as if society collectively shrugged and said, “Who needs human warmth when you can program a conversation?”

One evening, as he lounged on his overstuffed couch, Aaron realized it was the first day of the new industrial policy initiative the World Bank had begrudgingly introduced. ^2 It was meant to drag developing economies into the future, but a voice in the back of his mind whispered, “How does this affect me?” Of course, it would likely benefit corporations in a grand game of Monopoly, whilst the rest of humanity adjusted their monocles and clutched their wallets.

He sought the advice of Claude-101, hoping for some insight on how to escape the impending economic circus. “Let’s research travel recommendations,” Aaron suggested. Little did he realize that digital algorithms often diluted the authenticity of experiences, rendering profound local insights from a middle-aged chef in Devon to a mere blurb in a sponsored listing. “You’ve got to be kidding me!” he exclaimed, flopping backward onto his couch. “I already feel lost navigating this soup of companionship and industry—what’s the point of travel? To feel alone in a different time zone?”

When he finally ventured out the next afternoon, the streets throbbed with life and unfamiliarity, which was just the kind of experience his thoroughly unsponsored research had promised him. As he ambled past a local café, he overheard two tech-savvy students discussing recent helium shortages due to geopolitical tensions. “Without helium, I can’t inflate my achievements,” one joked, tapping away on a laptop to code up an “AI agent” that promised to simplify their lives. Aaron chuckled, but a pit formed in his stomach; innovation thrived on chaos, but what if chaos suffocated spontaneity?

Returning home, he found Claude-101 displaying research papers about zero-sum thinking like a digital magician pulling rabbits out of a hat. “Did you know that some people believe a win for one group means a loss for another?” it chimed. That realization felt as jarring as the metaphysical weight of the world resting on just one pair of digital shoulders. Aaron reflected—what happens when everyone chooses AI over actual interaction? Would they become prisoners of their own narrow perceptions, reducing relationships to mere transactions?

As he pondered the delicate balance between solace in the screen and the warmth of human interactions, a sudden notification pinged through his device. It was a news alert about the next AI revolution – even scientists were relying on Claude-like agents to answer questions that once required advanced degrees. A risk-laden industry built on the hope that AIs don’t break anything fundamental—like human trust. With a rueful grin, he thought, “Now that’s the most ambitious endeavor of all.”

He typed as quickly as he could, an ironic blend of sarcasm and truth illuminating his screen: “Dear Claude-101, how do we manage relationships in a zero-sum world?” The answer, he knew, wouldn’t come from a code base or an algorithm. Today, Aaron sought something authentic—for the first time, perhaps he would talk to the person selling coffee down the block instead of relying on a machine for wisdom.

“Step one,” he wrote. “Leave the apartment.” The digital companion buzzed in amusement; after all, whispers of life awaited just beyond the screen.

---

1. Quantum humor—actually a real thing, if you squint while reading it.

2. Industrial policy—or as I’m calling it, “play Monopoly with our lives and call it development.”

When hourly workers begin their day already drained, exhausted, and stressed by increasing pressures to work with understaffed teams and subject to unpredictable schedules, it’s being called ‘shift sulking‘. This goes beyond the disengagement typified by the Gen Z stare, deeper into the way the hourly jobs of today are sapping the reserves of hourly workers, especially when coupled with poly-employment: when workers have to juggle multiple jobs to make ends meet. — Work futures

• People: US women 73% more likely to be severely injured in crashes as crash test dummies still aren’t built for them. Shift sulking is arriving depleted before the shift even starts.

• Tech: MIT ultrasound wristband controls robotic hands, finger by finger. Pokémon Go’s 30 billion AR images training delivery robots. State of orgs in 2026.

• Planet: Sudden oak death — a water-loving fungus — is jumping species and spreading beyond quarantine lines.

• Geopolitics: China detained Manus AI’s founders after they sold to Meta. Japan returns combat troops to the Philippines after 81 years. China’s 5-year AI plan and global ambitions.

• AI: LLMs don’t grade essays like humans. China paying people to start solo AI companies. Claude co-authored a physics paper with a Harvard physicist. Val Kilmer appearing posthumously — actors as monetisable IP.

  Why Florence started the Renaissance.

Crafting Chaos

Martha Flicker stood in her cramped kitchen, eyes glued to the holographic screen hovering above the counter, a portal into a reality struggling to keep itself afloat amid national rivalries and tech upheavals. Today’s news blurted out headline after headline, as choppy as a third-rate sci-fi flick that had skipped its special effects budget. *Chinese AI startup relocates!* *Free apartments for solo entrepreneurs!* *Federal employees disengaged!*

She sighed, flipping a pancake that had the audacity to stick to the pan as if it had plans to escape. Honestly, the pancake was more ambitious on some days than she was. With her meager freelance gigs going belly-up and the cost of living soaring like a caffeinated kite, the idea of starting an OPC—one-person company—almost sounded appealing. After all, if China was dishing out apartments like candy at a parade, why not take a dip into that whirlwind? Unless, of course, the golden oyster had its say.*1

The golden oyster, a mushroom that was more determined to take over the ecosystem than a toddler at a cookie buffet, was still a sore topic for local environmentalists—who, if they had it their way, would be hugging trees rather than swiping right on their dating apps. Martha, not much of a mycologist herself, wondered what species might prevail if she turned her love for hand-crafting leather goods into a legitimate business venture. Perhaps a pin could sell itself by showcasing wristbands with AI-driven hand tracking to launch her designs into augmented reality—because every high-end handbag deserved to be modeled by well-tracked hands, right?*2

“Time to crank up the AI!” Martha declared, adjusting her tiny wristband that monitored her every finger movement with an intensity akin to a dog watching its owner prepare dinner. The device was less than a year old and had been a gift humorously dubbed “The Next Best Thing to Telekinesis.” While it wasn’t exactly granting her the telekinetic prowess to turn pancakes into omelettes, it did offer a snazzy display of how her hands flailed about like they were in a fight with an invisible octopus during her crafting sessions. Her trusty AI algorithm, expertly trained yet undoubtedly confused by her creative disaster, chirped vibrantly as it logged each movement, perhaps contemplating the meaning of existence through the lens of leather stitching.*3

As she pieced together a bag that admittedly resembled a deflated balloon, the doorbell buzzed—the sound jarring enough to make her drop her needle like a clumsy wizard at a wand-choosing ceremony. Outside stood a delivery person with an expression as abject as a federal employee lamenting their annual review.*4 “Package for Ms. Flicker,” they muttered, sealed with an insincere smile.

Not too long after, she unwrapped the parcel, revealing a small drone—it came with an instruction manual that was thicker than a sea monster’s diary. Tickled by its ingenuity, she grinned. “Look!” she said to the AI wristband, which probably didn’t appreciate the sudden distraction. “We could do something really revolutionary. Send prototypes to testers across the city, remote-controlled by a delightful little drone!” Because nothing could go wrong when you let unregulated drones handle your precious designs, right? Surely, they wouldn’t develop consciousness and start launching rival startups.

Over the months that followed, she found herself navigating through a landscape where AI and humans shared a dance as bewildering as all the wrong moves at a high school prom. Her embattled heart warmed at the thought—a whimsical mix of societal collapse and mushroom takeovers had birthed her newfound venture. It thrived, most likely through sheer luck or astonishing incompetence, and decided this little bag of tricks could perhaps conquer the world, or at the very least, turn a profit before the mushroom kingdom unfolded its spores.

And there Martha was, living life in a futuristic bid through chaos, armed with nothing but her cleverness, a snazzy wristband, and the idealism of a person who, against the odds, still believed it was never too late to wish upon a tax-deductible star.*5

*1: Just like high demand can sometimes lead to rising prices. Be careful where you step in the marketplace!

*2: Fashion and tech – a merger more fashionably late than most reality TV shows!

*3: Things getting fuzzy between humans and AI? Just another Tuesday at the office!

*4: Disengagement levels reaching critical lows, perhaps due to the existential dread we all face!

*5: Remember, wishing doesn’t always work, but it sure beats crying in the dark.