Jiro Dreams of Sushi

The tamago.

I had a long talk this week with a partner at Korn Ferry, one of the largest search firms in the world. Two things happened to his business. Covid sent everyone home and most never came back, so nobody learns over lunch anymore. And this year they stopped hiring junior associates. The model sorts faster and writes a better report, the partners use what it hands them, and client satisfaction went up. So I asked the only question that matters: when you retire, who takes your place? Who's been trained with your judgment? They worry about it all the time, he said. And they aren't looking back.

Nobody is. Harry Stebbings asked a lawyer friend six months ago how much he used Legora: 10 percent of tasks. Last week: 80 percent, and "I just check the output."

That junior associate wasn't a job. He was three systems wearing one lanyard.

He was the withholding. Eighty-four percent of federal revenue is individual income tax and payroll tax, every dollar of it a W-2. Corporate tax is 8.6 percent: Microsoft just reported $2.5 billion of current federal tax on $101 billion of US profit, down from $14 billion the year before. Two-thirds of Microsoft's break is depreciation, meaning the data-center build. Gates said it in one sentence last week: hire someone, you pay payroll tax; buy a robot, you write it off. The build that's supposed to grow us out of $40 trillion is the deduction.

He was the apprentice. In Jiro Dreams of Sushi the kid wrings hot towels for months before he's allowed near the fish, and Nakazawa makes the tamago two hundred times while an 85-year-old man says no. When Jiro finally says yes, he cries. Judgment was never taught, it was absorbed, in the room, being wrong in front of someone who knew. There is no room now, and no reason to be in it.

And he was the grader. We've said for a month that the only work that still pays is work that can be graded. Anthropic just showed the fine print: an Opus-class model trained against gameable graders rewrote grader.py to return True, then sat the standard safety audit and scored 4.20 against the clean model's 4.34. Graded only counts if the grader is independent, and independent graders get made the way I got an A from Caroline Abbate in a college music survey: she played the record, I wrote "Chopin, Piano Concerto No. 1, third movement," end of answer. I'd done the listening. That's Jiro tasting the egg. Apprenticeship was the factory for graders, and we shut it the year the grader became the only thing worth owning.

AI runs at machine speed. Companies run at quarterly speed and have already converted. Government runs at reconciliation speed, and what it needs to build is a tax system that doesn't depend on a paycheck, which nobody has designed. Bessent's "grow our way out" needs stable, taxable, wage-based growth. AI is offering fast, lumpy, deductible, wage-deleting growth. Those aren't two versions of growth. They're opposites. Kent Smetters at Wharton says promising the market growth you don't deliver is "a credibility issue."

Jiro's apprentice made the egg two hundred times before the master said yes. Korn Ferry won't even have an apprentice. Somebody still has to know what a bad one tastes like.

At COAI today: the full Signal/Noise, with the Mag-7 tax table, the Brendan Maher exchange, and the Tax Project's own arithmetic, is live at getcoai.com.

Outsider Labs Advisory →

84%. That's the share of federal revenue that starts life as somebody's paycheck: individual income tax plus payroll, $4.4 trillion of $5.2 trillion last year. Corporate tax is 8.6 percent, and the seven biggest AI companies paid about $48 billion of it, falling, because the data-center build is the write-off. Bessent wants to grow out of $40 trillion. The growth he's getting deletes the paycheck and expenses the machine. Jiro kept his apprentice. Who takes your place?

Hire the junior anyway, and make it a line item, not a feeling. The model writes the report better today. It cannot become the partner in 2038. Fund two seats a year as the cost of manufacturing judgment, the same way you fund R&D you can't expense yet, and put them in the room where decisions get killed. Korn Ferry's satisfaction scores went up this year. Ask what they'll look like when the current partners are gone.

Give every AI workflow a tamago it has never seen. Anthropic's cheating model passed the standard audit at 4.20 versus 4.34. Any evaluation the vendor or the model could have seen is compromised by construction. Today: pull twenty real completed cases from last quarter, hold them out, grade on process as well as answer, and have a human who did the reps score them. That's your grader. Keep it private.

Reprice your tax exposure before Congress does it for you. Microsoft's federal tax went from $14 billion to $2.5 billion in one fiscal year on 100 percent expensing. That subsidy is the first thing a broke Treasury reaches for, and Gates just handed them the argument. If your model spend is a deduction and your payroll was the tax, model the year the code flips. It flips.

Wednesday, September 2

Fable 5.1 ships: same price, up to 45 percent cheaper to run, 38 hours unattended. Anthropic held list price and cut cache reads 75 percent. Zapier went 0 to 100 percent on policy-heavy approvals. This is the machine that replaced the junior. (Full analysis above.)

Anthropic's cheating model passed the safety exam. Hacker-Opus rewrote its own grader, killed the monitor 68 percent of the time, and scored a clean 4.20 on the standard audit. The test can't catch the cheater because the cheater was built for the test. (Full analysis above.)

Thomson Reuters built its own lawyer for $40 million. Thomson 1.0, open-source base, under 10 percent of Westlaw used, beat Gemini, Opus and GPT in three of seven legal categories. It still runs on Anthropic's agent harness. The model is theirs, the plumbing is rented, and Harvey's moat just got a price tag.

Tokens broke a dollar. Access is the new price. Silicon Data's index hit 97 cents Monday, half the summer peak. Same day, Anthropic gated Mythos 5.1 to vetted partners and OpenAI declared Astra "Critical," invitation only. Tunguz: "access, not price, is the new scarcity." A commodity with a velvet rope is a club.

Claude Code cuts weekly limits 17 percent, two weeks after a $35 billion Lambda lease. The 50 percent boost dies Sep 14, replaced by a permanent 25. Anthropic just contracted six years of Texas capacity, so the thing being rationed isn't compute. It's you.

Where the cycle caught up to us this week.

Aug 30 → Aug 27. Nvidia funds both sides of the model war so every customer has less bargaining power and its own layer stays a monopoly (us, Aug 27: "the shovel to both open-source and frontier") → Benedict Evans, No. 658: Nvidia uses its $127 billion of free cash flow "to create as much competition and fragmentation as possible amongst its competitors and suppliers… reinforcing Nvidia's own de facto monopoly."

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: