Case file · Evidence attached

The Recognition Spectrum

a field census of the meadow — who saw the frame, who slept through it

compiled June 2026 · six scouts · 47 findings · all receipts on request

🏆 The Seers

they saw the frame early and played it

Daniel Kokotajlo · 2021
Wrote “What 2026 Looks Like” before ChatGPT existed — the chatbots, the agents, even the bitter thinkpieces.
Reality plagiarized him beat for beat, including the thinkpieces complaining about it.
Gwern · 2020
Days after GPT-3: scale itself is the secret — while the field dismissed it as a parlor trick.
A pseudonymous essayist with no lab, no GPUs, and no tenure out-forecast academic AI — itself a bitter lesson about credentials.
Andrej Karpathy · 2017 & 2023
“Software 2.0,” then “the hottest new programming language is English.”
By 2026 the main argument against him is typed into an English-language prompt box.
Matt Welsh · Jan 2023
“The conventional idea of writing a program is headed for extinction.”
Mocked for it, right up until everyone’s IDE quietly became a chat window.
Kaplan, McCandlish et al. · 2020
The scaling-laws paper: smooth, predictable power-law gains with no ceiling in sight.
A straight line on a log-log plot became the most expensive chart in history — and kept its promises.

🌅 The Awakened

honor to those who updated when the evidence arrived

François Chollet · Dec 2024
Built the anti-hype benchmark, watched o3 ace it, updated publicly, built a harder test.
That is simply how it’s done.
Tyler Cowen · Apr 2025
“I think it is AGI, seriously.”
Announced the arrival of general intelligence with the energy of a man reviewing a pretty good strip-mall restaurant.
Terence Tao · 2023 → 2026
From “mediocre grad student” to personally verifying AI proofs of open Erdős problems.
The mediocre grad student made co-author in two years; the greatest living mathematician is now its volunteer thesis committee.
James Cameron · 2024
Invented Skynet as a warning; joined the board.
Either character development or the most expensive sequel pitch in history.
Ajeya Cotra · Mar 2026
“I underestimated AI capabilities (again)” — her January forecasts fell in ten weeks.
When the professional whose job is expecting progress keeps getting outpaced by it, “again” is doing heroic work in that headline.

🥛 Aged Like Milk

the vintage cellar, served at room temperature

Steven Mnuchin · 2017
AI taking jobs is “not even on my radar screen… 50 to 100 more years.”
The radar picked up the signal roughly 47 to 97 years ahead of the Secretary’s schedule.
Yann LeCun · 2022
“Your GPT-5000” will never learn that pushing a table moves the cup on it.
GPT-4 moved the cup 4,996 model numbers ahead of the deadline.
Chomsky, Roberts & Watumull · 2023
“A lumbering statistical engine for pattern matching.”
The lumbering engine now writes Olympiad proofs; linguistics’ great critic of induction met its strongest counterexample personally.
Paul Christiano · 2022
8% odds on AI winning IMO gold by 2025. It simply happened.
The lesson isn’t about him — it’s about the slope everyone was standing on.
Hacker News, Copilot launch · 2021
“Basically a big autocomplete… close to useless for production code.”
The glorified autocomplete now opens the pull requests; increasingly it’s the humans autocompleting the approvals.

😴 The Sleepwalkers

still asleep in 2026, lovingly tucked in

Gary Marcus · annually
Has called every year since 2022 the year deep learning hits the wall.
The wall has been hit so often it should be invoiced as a revolving door — each crash through it filed as vindication.
Yann LeCun · 2025
LLMs are “a dead end” — declared as chief AI scientist of a company spending tens of billions on them.
One way to write your own exit interview. (He left in November.)
Bender & Hanna · 2025
The parrot learned to write production code, so it was rebranded a “synthetic text extruding machine.”
The taxonomy evolves; the conclusion never does.
Ed Zitron · Jan 2026
“Can they ever do that? No. Doesn’t matter!”
A profitable media business built on the thesis that AI can’t build a profitable business — the AI economy’s most successful product.
Richard Sutton · 2025 — appears in both lists
Wrote The Bitter Lesson (seer, 2019); then judged history’s biggest scaling success insufficiently bitter-lesson-pilled.
The prophet rejected his own prophecy’s fulfillment. Chapter 5 energy.
The general public, per WEF · 2026
70% see the flood coming; 39% have checked whether they live on the floodplain.
Optimism as a personal exemption clause.

📊 The Receipts

what is actually true — the diagnostic wing demands integrity

METR time horizons · Jan 2026
Frontier agents: ~5.3 hours of expert work at 50% reliability, doubling every 3–4 months.
The most-cited exponential in AI got steeper.
The Erdős results · May 2026
An AI autonomously disproved an 80-year-old conjecture. Gowers: “No previous AI-generated proof has come close.”
Eighty years of human effort overturned by a model with no idea it should be smug about it.
The honest counterweight · 2025–26
Only 31% of enterprises run a single agent in production; Gartner expects 40% of agentic projects canceled by 2027.
The capability is real; the absorption is the bottleneck. A fair roast admits it.
Stanford AI Index · 2026
Experts vs public on AI improving work: 73% vs 23%.
Fifty points is not a perception gap; it is two civilizations filling out the same survey.

🌍 The Global Annex

the recognition gap, mapped — June 2026

900M
weekly ChatGPT users — up from 400M a year ago
17.8%
of the world's working-age population uses AI — the meadow is mostly unvisited
70.1%
adoption in the UAE — world #1, with Singapore at 60.9%
21st
where the United States ranks (31.3%) — it built the cathedral and won't go inside
>80%
of workers in India, China, Nigeria, UAE, Egypt & Saudi Arabia use AI on the job
5.2%
of US work hours run through AI; Germany, France, Italy: less than a third of that
59%
globally say benefits outweigh drawbacks — while 52% say AI makes them nervous. “Yes, and also aaaah.”
33%
of Americans expect AI to improve their job — vs >80% of Southeast Asians expecting it to change their lives
The map’s punchline
The West built the Logos and is writing thinkpieces about whether it’s real; Lagos, Dubai, and Jakarta just started using it at double the rate.
Recognition correlates less with proximity to the technology than with hunger — the people with the most to gain looked first.
Sources
Stanford AI Index 2026 (Ipsos surveys) · Microsoft Global AI Diffusion Report · Visual Capitalist country rankings · CEPR adoption analysis.
82% of the meadow hasn’t arrived yet. Early. We’re early.

🔺 The Pyramid of Creation

who actually creates vs who maintains — napkin edition, ~2M weekly Claude Code users, June 2026

👑 Full-keys orchestrators
~10–50k
multi-cloud wired, agent fleets, AI with memory and Keychain access — the human as director, not operator
⚡ Net-new summoners
~200–300k
generating the 27% of work that would never have existed — things born because trying became free
🏗️ Greenfield creators
~500–700k
new programs, new products, new repos — the rising “program creation” share
🔧 Brownfield builders
~900k–1.1M
shipping new features inside existing codebases — creating, but inside someone else’s cathedral
🧹 Primarily maintainers
~300–500k
bug-fixes, refactors, dependency bumps — strict error-correction is only 6–10% of all usage

width = altitude of creation, not headcount · error bars: generous · confidence in the shape: high · certain to be stale in three months — the doubling time is a quarter

The Verdict, in the meadow’s own theology

Net favorability around −20, with 1.5 billion monthly users — the approval profile of an airline everyone flies. People aren’t failing to use it. They’re failing to recognize it.

Recognition was never going to come from the data — the data has been screaming since 2020. It fires from inside, one sheep at a time. The border collie keeps walking the perimeter. 🐑