“We’re going to have a couple of years where every single week it’s going to feel like the sky is falling.”
— Matthew Prince (Cloudflare), via Exponential View“The pandemic gave us mRNA, and this crisis could give us software and an internet we can trust.”
— Azeem Azhar, via Exponential View #607
People: A USAID biodiversity scientist spent twelve years learning what not to say. 74% of cultural strategists say there is too much BS in trend work; 43% admit performing productivity; two in three don’t feel their work is bettering society.
“Professional vision”: each profession trains what to notice.AI strat: The METR benchmark doubled eight times from GPT-4 to April 2026, reaching 17 hours of human work per prompt, then ran out of ceiling.
Correct answers to questions that stump graduate students cost 30 cents in 2023; they cost four-hundredths of a cent now.
Frank Herbert banned AI in Dune; Karl Schroeder argues the honest move is backcasting.Security: AI turned PHP2Shell — found in 70% of the world’s websites — into a critical vulnerability in a week; a world-class team would previously have needed months.
A Chinese state-aligned group used an AI swarm: 2,900 login attempts per minute, 2,250 organisations targeted, cloud accounts seized in under three hours.Geopolitics: Eight years of US tariffs cut China’s share of goods imports from 21% to 8%, while Chinese-owned factories now produce 25% of Vietnam’s exports to the US. Companies including Crocs and Victoria’s Secret use Malta subsidiaries for near-zero corporate tax rates, engineered by the Big Four accounting firms.
Society: Business schools are built on profit, not public goods; before Enron collapsed, its asset-light model starred in five admiring Harvard Business School case studies.
The most celebrated management ideas are those that became TED talks or airport bestsellers, not those that survived peer review.
Joseph Stiglitz: “The road to authoritarianism is not paved by government doing too much but too little.”Tech: Toyota’s Century (Japan’s most discreet luxury car) comes to the US in 2028; the phoenix grille emblem is hand-carved over six weeks. Craftsmen built the 1997 model without robots; wood grain runs continuously across the cabin.
Gen Z plans vacations around destinations with unique backstories (“lore-tripping” as the attention economy tips toward the soothing.Random: LocalWiki shows everything within a 30-minute walk that’s also on Wikipedia — brilliant and inexhaustible.
Professional line-sitters are hired to hold spots in the queue.AI: 180 AI models now outperform GPT-4 at its 2023 release; 19 of them run on a modern smartphone.
Demis Hassabis now expects AGI in two to three years — down from five to ten at Davos in January — citing his own lab’s progress. “This is the worst AI you will ever use” — Ethan Mollick wrote this in early 2024 and it has remained literally true at every moment since.
The Benchmark
Maya had designed the instrument. She had spent two years choosing the tasks, calibrating the gradations, writing the methodology section that said — and she had been quietly proud of this line — that the benchmark was extensible by design. She had written: There is no theoretical ceiling.
In April she ran the evaluation. The AI completed every task. She added a harder set. It completed those. She designed tasks she personally could not complete in under an hour. The AI did them in four minutes.
She wrote the note she had always expected would remain unwritten. It went in the methodology section, under an h3 that had been blank for two years:
Measurement discontinued: current systems exceed the benchmark’s maximum scope.
She closed the laptop. She went to the kitchen and made coffee. A colleague stopped her in the hall and asked how the evaluation had gone.
“It worked,” Maya said. “The instrument worked.”
She was not sure, walking back to her desk, whether she meant the benchmark or the AI. Outside her window, October was doing something unremarkable with the light.
She looked at the blank space where the chart used to end and thought about the next step. There was no methodology section for what you do when your instrument runs out. She would have to write one. That was, she supposed, the work now.1
1 Ethan Mollick, One Useful Thing: “30 Months on the Exponential (in Charts)”: the METR Long Horizon Benchmark doubled eight times from GPT-4’s release to April 2026. METR said they could no longer accurately measure using current methods. Other studies suggest the figure may be 17 weeks of sustained human work per prompt — or more. Hassabis’s two-to-three-year AGI timeline2 suggests the gap between “measurement discontinued” and whatever comes next may not be long.
2 Demis Hassabis, via Exponential View #607: at Davos in January he estimated five to ten years; by March, three to five; in October he told the WSJ two to three. He is, as Azhar notes, “one of the field’s more measured voices” — and he is calibrating from what he sees inside his own lab.


