The press release landed at 09:14 Pacific. The resignation letter had been circulating for six hours.
That timing is the story before the story. Anthropic published a set of economic scenarios for transformative AI โ three futures, labeled temperate, significant, and extreme โ and hours earlier, one of its own safety researchers, Jacob Coxon, walked out the door warning that the industry was racing toward self-improving superintelligence. One document describes what happens to the economy. The other describes why we might arrive there before anyone agrees on the rules.
I map the silence between the code and the chaos. In this release, the silence is loudest exactly where the numbers stop.
Context
Strip the branding and the artifact is not an AI model at all. It is a growth-accounting frame with AI capability treated as an exogenous input. The dial is not "how smart" but "how much of the work" โ three settings: AI the size of the internet, AI carrying half of knowledge work, and AI improving itself without human help.
Turn the dial and the macro outcomes move. GDP stretches from roughly $34 trillion toward $44.4 trillion. Labor's share of income slides from a mild dip, to 56.1 percent, to 45.2 percent. In the middle scenario, unemployment holds near 5 percent โ but knowledge-worker pay is flat. Growth doubles. Wages do not move.
The framing language matters as much as the arithmetic. Anthropic calls these outcomes "choices," not "predictions," and its interactive tool invites the public to submit their own forecasts. That is not a research paper. That is a database being seeded, ten thousand survey responses at a time, building a longitudinal baseline of AI anxiety that will be quoted for a decade. Goldman Sachs, working independently, traced the sharpest hiring contraction to junior technical roles โ the bottom rung of the knowledge ladder, removed first.
What the published material does not contain is just as telling. No model equations. No parameter calibration. No robustness checks. We can confirm that Anthropic released a scenario tool. We cannot confirm that its internals survive contact with an independent economist.
Core
Here is what I think the model actually says, and what almost nobody read correctly.
The variable that matters is not how fast AI grows. It is who holds the claim on that growth. A rising GDP with a falling labor share is not a growth story. It is a transfer. The economy gets bigger while the population that produced it loses its lien on the output. 56.1 percent, then 45.2 โ those are not decimals. They are the slow conveyance of the productive base from wages to capital.
In 2024 I sat with a compliance team at a mid-sized asset manager and translated cold storage architecture and hash-rate distribution into language they could underwrite. The exercise taught me one durable thing: institutions do not buy technology, they buy claim structures. So when I read a labor-share curve, I do not read an economic forecast. I read a collateral schedule.
Most consumer credit, most pension math, most household balance sheets assume labor income is stable and collateralizable. If AI carries half of knowledge work, the income stream behind a mortgage in a mid-tier city stops being a salary and becomes a residual claim on a model's output. Nobody has built the registry for that instrument. Nobody has even named it.
Crypto, awkwardly, is the closest thing we have. Not because tokens generate yield, but because a token is a machine-readable claim. I spent the past year inside what I call the Agency Economy โ analyzing roughly a hundred AI-driven protocols โ and the pattern is consistent: autonomous agents need three things, and only crypto supplies all three. Identity a counterparty can verify without a central issuer. Payment rails that settle between machines without a bank in the loop. Dispute resolution that does not require a jurisdiction.
Verifiable inference and decentralized identity are the only two primitives that let a machine's economic output be audited by anyone other than its owner. That sentence is the investable thesis of AI-crypto. It is also the answer to the question the Anthropic model declines to ask. If capital owns the intelligence, who audits the capital?
There is a technical catch nobody prices yet. If agents transact at machine speed, they inherit DeFi's oldest wound โ the oracle. Feed latency is where autonomous settlement breaks first, and a decentralized network of permissioned nodes does not fix a timestamp problem. My read is that blob markets saturate well before the decade turns, and agent logs will do to data availability what rollups did to block space. Auditability has a bandwidth cost, and the bill arrives late.
Notice, too, the shape of the middle scenario. Five percent unemployment with flat knowledge wages is a regime macroeconomics has no textbook for. The Phillips curve assumes a trade between inflation and employment. This assumes a trade between growth and participation. The models break before the society does.
Contrarian
The blind spot in this entire discussion is the direction of causality, and it is where crypto gets it backwards too.
Anthropic treats AI capability as the exogenous variable and the economy as the dependent one. The token market does the reverse. It prices the capability first, then back-fills the mechanism. I watched this happen in 2017, when I spent three months inside the Golem community observing technical skepticism curdle into ideological fervor over idle GPUs that mostly stayed idle. In the wild west, stories are the only compass โ but a compass pointing at a destination nobody has reached is not navigation. It is decoration.
In the last cycle, tokens sold the destination. This cycle, they have to sell the audit trail.
And there is a second blind spot, sharper. The extreme scenario is triggered not by a slope but by a threshold โ unaided self-improvement, a discrete event. That means every lab building toward general self-improvement is walking toward the same cliff edge on roughly the same schedule, coordinated or not. Anthropic publishing the worst case hours after a safety researcher resigns is not contradiction. It is documentation. A disclaimer filed in advance, by people who can read their own internal numbers.
This is where I have spent enough weeks sitting with the wreckage to be certain of one thing: truth hides in the bear market's quiet shadows, never in the liquidation candles. What survives the filter is never the pitch. It is the primitive.
Takeaway
The report ends by asking a better question than it answers: not what will happen, but who chooses.
In a market where labor is no longer the unit of account, that question has a technical answer, and it is being written in code right now. The next cycle will not be decided by which protocol claims to own the agent economy. It will be decided by which one can prove, to a stranger, that the agent did what it said it did.
Who signs the ledger when nobody is left holding a salary?