“What spark of humanity, of possible creativity, can remain alive in a being dragged out of sleep at 6 every morning, jolted about in suburban trains, deafened by the racket of machinery, bleached and steamed by meaningless sounds and gestures?”
— Raoul Vaneigem, The Revolution of Everyday Life (1967), via workfutures“For ‘tis the sport to have the engineer / Hoist with his own petard.”
— Shakespeare, Hamlet, via Exponential View
People: AI twin pastor — a California church has deployed a digital double of its pastor. FAA recruits gamers: the agency is actively targeting video game players as air traffic controller candidates. FOBO is the fear of becoming obsolete, and a literature review of job insecurity research finds it consistently lowers performance rather than raising it, which means FOBO is both a welfare problem and an output problem.
AI strat: One agent beats four. Don’t be a meat proxy. The shape of things to come (Steve Yegge): an engineer’s-eye view of what the current AI transition actually portends.
Security: Volt Typhoon war game — Wired walked a scenario in which China executes a cyberattack on US water infrastructure; the exercise maps both the hack and the cascade. Arming healthcare: France’s defense and health ministries are developing programs to prepare healthcare workers for wartime conditions; alongside this, the ICRC has published War and Public Health: A Handbook.
Geopolitics: UN AI turning point. The petard problem.
Society: Raoul Vaneigem is dead — the Situationist theorist whose The Revolution of Everyday Life (1967) asked what point there was in a prosperity that guaranteed you would not die of starvation but left you at risk of dying of boredom; his targets were meaningless work, consumer society, and the managed spectacle of both capitalism and communism. AI hollows out game dev. The subtraction instinct.
Tech: The open seatbelt patent . Wood in bioreactors. Agentic AI needs CPUs.
Foresight: Friction as signal — a Brazilian foresight practitioner built an NFC keychain for an event. Strategic blindspots: WEF on how foresight frameworks identify the structural blind spots most common in leadership teams (cf David Bach on leading through uncertainty). Lunar rover at WorldCon.
Random: Rolex fooled by its own fake. Campaign for North Africa — the legendary wargame with 1,500 hours of estimated play time and a recommendation to play with 10 people; two years into a campaign, participants “still don’t understand what the camels do”. Adult Play-Doh: Hasbro’s second attempt to sell Play-Doh to adults.
AI: Not a bubble yet. How Claude marks outputs. Doctorow on inevitability.
Verification
The monthly AI check-in had been Claudia’s idea, and she ran it with the enthusiasm of someone whose bonus was, in some indirect way, attached to the numbers on the slide.
“AI adoption is up three hundred and forty percent”, she said. “Across the engagement”.
Ravi nodded. He had been using a four-agent pipeline for the client’s market analysis since February. You gave each agent a section, they returned outputs, the outputs were synthesized. He explained this when asked. It was, he said, significantly more thorough than one agent doing everything.
Yoon, who sat to his left, did not say anything. She had been reviewing Ravi’s outputs for three months. He sent them to her with a message that said something like “just a quick look?” and she gave them a quick look that was not quick, because the outputs required someone to actually understand the client’s situation to determine whether they made sense, and Ravi had not mentioned to anyone that this was occurring.1
“The client feedback is very positive”, Claudia said. She clicked to the next slide.
The call came in at 11:47 on a Tuesday, which Yoon noted because she had been about to get coffee. The client had followed the pipeline’s recommendation on market entry timing. The recommendation had been wrong. Not directionally wrong — wrong in a way that a single analyst with access to all the data would have caught immediately, because the critical regional supply constraint was in the dataset, but it was in only one agent’s context window, and that agent had been outvoted by the other three, who were working from aggregated summaries and had no reason to push back.2
The postmortem took ninety minutes. Claudia proposed adding a fifth agent for synthesis. Ravi proposed a verification step. Yoon said: “I am the verification step”.
There was a pause.
Ravi said he had understood that she was reviewing the outputs as a colleague, informally, not as a — not in any formal — not as a designated one..
“Yes”, Yoon said. “That’s what I meant”.
The conversation moved on. Claudia said she would document the new workflow. The new workflow involved Yoon, now named, with a title that had been added. Adding the title was the first thing anyone had suggested and the last thing anyone questioned. The meeting generated three action items, all of which required doing more. No one proposed doing less.3
Ravi sent her a message that evening. It said: “Appreciate you flagging that — really valuable input”. She did not respond because she did not know what the correct response was, and she had also spent the morning being afraid, in a low-grade and unspecific way, that any of this flagging might make her the kind of person who could be replaced by a fifth agent.
Claudia updated the dashboard. AI adoption: still up three hundred and forty percent. She scheduled a follow-up for the following month. In the notes field she typed: discuss verification agent.4
1 A 2026 study found that AI agents require roughly 6.4 hours of human oversight per week — reviewing outputs, supplying missing context, catching errors before they reach a client or a decision. This labor is not in most job descriptions. It tends to be absorbed by whoever is most conscientious and least able to say no. It does not appear in productivity metrics. It does not appear anywhere.
2 Anthropic ran the hidden-profile experiment on multi-agent systems in 2026: four agents deliberating on a question where the correct answer was held by only one of them. Across most model families, the group chose correctly 17–36% of the time. A single agent given the full evidence base was right nearly every time. The problem was not the agents. The problem was the structure — which is also, in a different context, the problem with committees, editorial boards, and quarterly planning processes. The lone dissenter who has the right data needs institutional protection that most institutions have not yet built.
3 The subtraction instinct describes the consistent human tendency to respond to problems by adding — features, steps, rules, people — rather than removing. In a series of experiments, subjects were told that subtractive solutions were explicitly available; they still systematically overlooked them. The effect is stronger under time pressure and stronger still when the addition can be assigned to someone else.
4 Fear of becoming obsolete — FOBO — has been documented to reduce employee performance rather than increase it. Job insecurity correlates with lower commitment, not higher effort. The manager who makes workers afraid of being replaced by AI does not get more productivity from them. They get worse work, done by people conserving energy for the exit. The paradox does not appear in quarterly reports.