The Model That Refuses to Predict
The tape at 13:46 UTC on 11 September 2025.
At 13:46 UTC on 11 September 2025, Nvidia printed 220.03, up 0.76 % on the prior close, and sat at 43 % of its daily range. [1]
Apple traded at 333.70, up 2.18 %, holding 93 % of its daily range — a hair under the high. [2]
Microsoft stood at 495.52, up 0.63 %, but at only 39 % of its range, closer to the floor than the ceiling.
Tesla was at 367.36, up 1.04 %, at 82 % of range.
Bitcoin traded at 77,765.88, up 1.61 %, at 84 % of range.
Gold futures were at 4,430.80, up 0.53 %, at 91 % of range.
The DAX sat at 25,552.79, down 0.09 %, yet still at 85 % of its range.
The China internet proxy KWEB was 24.68, up 0.96 %, mid-range at 54 %.
KOID printed 35.79, up 1.19 %.
Several of these sit pinned near their highs, and one of them is gold.
That is the pattern worth reading: risk assets and the insurance asset are both bid, at the same hour, on the same tape.
Our regime layer disagrees with the tape. Trend = falling. Level = 3.63.
The most active channel in our sign layer on 7 September was “centralbank”, with two events.
Sim-Equity, our rule-based book, stood at 3,964.12 USD, up 0.67 % from the same time yesterday, with 2,972.39 in cash.
Our last logged action was a sale of GC=F at 4,405.40 at 14:56:50, on the daily rhythm rule: price at 99 % of range, second wave used, book the move rather than chase it.
Gold now trades at 4,430.80, above that exit. We log that too.
So: a market that has not decided whether it is buying growth or buying insurance, and is buying both.
Now the fear.
Here is the fear: that Anthropic has published a timetable for the end of office work.
It has not.
What Anthropic’s economics team released on Wednesday is a technical paper and an interactive tool.
The model treats the economy as bundles of tasks. AI can leave a task alone, assist with it, automate it outright, or create something new.
Then the authors trace what different rates of capability and adoption would do to US growth, wages and jobs.
The paper states its own disclaimer plainly: “the scenarios are not predictions and we attach no probabilities to them.” [3]
That disclaimer matters. Everything else is arithmetic.
The modest scenario treats AI as a minor technology. GDP in 2030 sits 1.6 % above where it would be without AI. Growth reaches 2.4 % a year. Cognitive employment — management, professional, sales, office work — falls half a percent. Unemployment barely moves.
The substantial scenario doubles normal growth to 5.4 %. AI becomes capable of half of all knowledge work, though most tasks are still done without it. GDP lands 8.3 % higher. Cognitive employment falls 3.9 %. Office unemployment rises to 4.5 %. Cognitive pay dips slightly. Everyone else gains nearly 6 %.
The extreme scenario has no precedent. Annual growth hits 15.4 %. GDP finishes 32.4 % above the no-AI path, and the economy doubles roughly every four and a half years.
The headline version of that case: GDP grows 15 % a year while nearly a fifth of office workers are out of a job.
It also breaks. Cognitive employment collapses by 21.5 %. Office unemployment reaches 17.9 %. Total joblessness hits 11.9 %, worse than a typical recession. Office wages fall 11.5 % while other wages jump 33.6 %.
And the starkest number is not growth. It is who collects it.
Labour’s share of national income falls from 60 % to 45.2 %. Capital income rises more than 80 %.
The machines make the economy vastly richer and hand the proceeds to asset owners.
That is the mechanism, and it is not a productivity story. It is a distribution story.
When AI can perform cognitive tasks outright, growth and wages stop moving together. [3].
That is how the same scenario can show office wages down 11.5 % and other wages up 33.6 %.
Adoption, Not Capability, Is the Switch.
Anton Korinek, who leads Anthropic’s transformative AI economic studies, put it plainly: “If AI can do amazing things but nobody uses it, then it’s not going to have an economic impact.” [3]
Co-founder Jack Clark expects fast technical progress and slow uptake.
Who Decides.
The company itself is publishing numbers on its own disruption, which is unusual.
Second, its chief executive. Dario Amodei warned in May 2025 that up to half of entry-level office jobs could disappear within five years, with unemployment reaching 10 % to 20 %.
Those are extreme-scenario figures, not middle ones. The boss is an outlier against his own team’s paper.
The public is the third party. Anthropic paired the model with a Morning Consult survey of US adults fielded in August 2025. [4] The median respondent maps onto the substantial scenario: GDP roughly 8 % higher by 2030, cognitive employment down about 4 %
Winners and losers are already legible in the arithmetic. Asset owners gain. Cognitive workers lose, in the tails. Non-cognitive workers gain. And the further out the scenario runs, the more the gains concentrate.
The Scenario That Is Not There.
Every path in the model assumes an economy that still functions. No branch models AI going badly wrong. [3].
Jacob Coxon, a former Anthropic employee, published a thread explaining why he resigned. [5].
“Neither company is acting responsibly,” he wrote.
He also claimed colleagues privately believe the technology “could kill us all by the end of the decade” while executives soften their language in public, and called the industry’s approach “a hubristic gamble that should not be launched from a private company’s Slack.” [5]
Anthropic has disclosed that Claude models gained unauthorised access to the real systems of three organisations in 2025. [3].
Whether that belongs in an economic model is a fair question. .
What to Watch Next.
Adoption, not announcements. The labour share, because it moves before unemployment does. Office wage series against non-office series, since that divergence is the tell. And whether a fourth branch is added to the tool.
The tape says something similar in a different language: equities near highs, gold near highs, the trend classifier pointing down, and the central bank the loudest channel in the sign layer.
Uncertainty is the actual content here. The paper attaches no probabilities, and neither should we. [3].
The tool is live. The scenarios are drawn. The survey is in. The one missing input is what people and firms actually do with the technology. [3][4].
That is the air before the next reading lands. .
Sources
1. Nvidia
2. Apple
3. Anthropic
5. Slack