“Instructions for living a life: Pay attention. Be astonished. Tell about it.”
— Mary Oliver, “Sometimes,” via Rob Walker’s Art of Noticing (American Masters documentary, PBS)“Between the data and the reality falls the shadow. The shadow is the emotion, the stories, the messiness that humans have contained within us.”
— Rishad Tobaccowala, “Emotional Data,” via The Future Does Not Fit In The Containers Of The Past
People: 100 Years of Stafford Beer. Mary Oliver: Saved By the Beauty of the World. La France Solo (Futur(s), La Mutante): 38% of French households are composed of a single person (2022 data). Close Enough. Kids remember you worked late. The J.Crew Liquor Store. The Doxa Orange Edition.
AI strat: Emotional Data. LIWGY #784 — “The absurdity of the apocalypse”. The inflection point question. The absurdity problem .
Geopolitics: Le futur est imprévisible, mais pas illisible. Artificial Revolution: Power, Control and the Future of AI.
Foresight: Friday Interview with Charles Cumming.
Random: Do A Dull Thing with Style. Summer is cooking my heart. Nous avons tous tellement besoin d’autre chose (Umanz).
Jacques Cousteau became the US distributor in 1968. Clive Cussler put one on Dirk Pitt’s wrist and never took it off. The watch survived the quartz crisis, which took out most of its contemporaries. The orange dial was a functional decision that became an aesthetic one — legibility at depth, identity at the surface. WITI’s line, which earns its place: “Design gives you something to look at, stories give you somewhere to go.”
The 10% Problem
The product meeting was on a Tuesday. Beatrice had a spreadsheet. The question was whether to launch the notification center now or wait three more weeks for an additional data round.
“Three weeks,” Beatrice said. “We need the cohort to mature.”
Omar was looking at his phone under the table. He looked up. “Did you see the resignation? The Anthropic researcher. He says there’s a ten percent chance AI wipes out humanity by 2030.”
“That’s not relevant,” Beatrice said.
“It’s relevant to all decisions with timelines,” Omar said. “If there’s a ten percent chance—”
“More likely than the Bills winning the Super Bowl,” said Leila, who had been watching them both.
“That’s how another researcher put it, yes,” Omar said. “Which is a calibration, not a slogan. The Bills have lost four Super Bowls. The probability is empirically grounded.”
Beatrice returned to the spreadsheet. The cohort data would be more statistically significant in three weeks. That was simply true.
Leila had done twelve user interviews. In eight of them, the participant had tapped on the area where the notification center would appear. Not because they were told to. Because it seemed like the right place. She said this.
“That’s a sample,” Beatrice said. “Not a sample size.”
“Eight out of twelve people tapped on the same empty space without being prompted,” Leila said. “I think we can launch.”1
Beatrice looked at the numbers she had. They were not the numbers she wanted. The numbers she wanted would arrive in three weeks and would probably confirm what Leila was saying, at which point the notification center would still not have launched.
She thought about something she had read: that brands exist because humans are not computational, that people choose with their hearts and use numbers to justify what they just did, that the whole optimization stack is downstream of a desire that no model fully captures.2 She was not sure this was relevant either. She filed it in the same place as the Bills.
“We launch,” she said.
Omar looked up from his phone again, surprised.
Three weeks later the notification center was the most-engaged element in the product. Leila was correct. The cohort data, when it arrived, confirmed this.
Omar had a spreadsheet tab he had not mentioned to either of them. It was titled “10% Tracker.” A single cell, updated weekly, reflecting his best synthesis of AI safety researchers’ current consensus on human extinction probability by 2030. It had started at 10.2%. This week it read 10.4%.
Beatrice leaned over and looked at it on his screen.
“That’s going up,” she said.
“Yes,” Omar said.
“You should add a trend line.”
He added it. The trend line went up.3
1 Leila’s logic is not as thin as it sounds. Eight of twelve participants in a purposively sampled user study performing the same unprompted behavior is a meaningful signal. The case against waiting for larger samples is not that samples are wrong; it is that samples confirm what observation already knows, at a cost of time during which someone else may launch first.
2 Rishad Tobaccowala, “Emotional Data” (The Future Does Not Fit In The Containers Of The Past, September 2026): “Between the data and the reality falls the shadow. The shadow is the emotion, the stories, the messiness that humans have contained within us. Today in the idolatry age of data and AI, those who rely only on numbers and computation get it wrong and wrong and wrong.” Tobaccowala’s argument is not anti-data; it is that data is downstream of desire, and desire is not where most analytics teams are looking.
3 Jacob Coxon resigned from Anthropic the week of September 7, citing what he characterized as an unacceptable probability of AI-caused human extinction by 2030. Evan Hubinger, an Anthropic lead researcher, posted on X that he personally estimated the probability above 10% within the next decade. This is not a fringe position at the frontier of AI research. It is the stated belief of researchers who spend their working hours trying to make it less true. The trend line was a joke. The number it tracks is not. Separately: French commentator Laurent Alexandre posted that week that with GPT-6’s capabilities it was now “very difficult to find utility in employees with IQ below 130.” The researchers trying to prevent the 10% from becoming 100% did not, as far as anyone could tell, take a position on this claim. (Marie Dollé, “QI en crise?”, In Bed With Silicon, September 2026; LIWGY #784)