Jiro Dreams of Sushi

THE NUMBER: 84%. That’s the share of federal revenue that starts life as somebody’s paycheck. Individual income tax, $2.66 trillion. Payroll tax, $1.75 trillion. Together, $4.4 trillion of the $5.2 trillion the Treasury took in last fiscal year, and every dollar of it originates as a W-2. Corporate income tax is 8.6 percent. Hold that next to Monday’s number, the $40 trillion of debt, and next to the one Bessent keeps saying: we’ll grow our way out. Grow what, exactly? The 84 percent is the wage base. The thing everyone in AI is building is the machine that doesn’t need one.

The call

I had a long talk this week with a partner at Korn Ferry, one of the largest executive search firms in the world. He said two things had happened to his business, and he said them the way you describe weather.

Covid sent everyone home, and in his line of work most of them never came back. So the thing that used to happen over lunch, or on the walk back from a client, or sitting in the corner while a partner took a call, doesn’t happen. There’s no room anymore. Nobody learns by symbiosis when everyone’s a rectangle on a screen.

And this year, Korn Ferry stopped hiring junior associates. Not slowed. Stopped. The model sorts a résumé stack faster than the junior program ever did, and it writes a more consistent, more complete report. The partners just use what it hands them. Client satisfaction with the searches went up.

So I asked him the only question I actually cared about. When you retire, who takes your place? Who’s been trained with your judgment? If there’s no apprentice, how does what you know get to the next person? Is Fable 5.1 smart enough to absorb what you know?

He said they worry about that all the time. And they aren’t looking back.

I want to sit with that answer, because it’s the most honest thing anyone in this business has told me in months. He’s not wrong to do it. The satisfaction scores went up, and the kid he didn’t hire would have spent three years being worse than the software. Any partner at any firm makes the same decision. There is no villain in this story. That is what makes it a problem instead of a scandal.

Nobody’s looking back

He’s not alone, and this isn’t a search-firm quirk.

Harry Stebbings at 20VC asked a lawyer friend six months ago how much he used Legora, the legal AI. Maybe 10 percent of tasks, the lawyer said. Helpful. Stebbings asked again last week and got a different person: “I just check the output Legora does. If it were taken away, I would be SO SO upset.” Ten percent to eighty in six months. “Law will follow the same path coding has done.”

Thomson Reuters, which owns Westlaw, decided last week that it didn’t need to rent a lawyer-brain from anyone. It built Thomson 1.0 on an open-source foundation, trained it on less than a tenth of the archive it owns, spent about $40 million all in, and beat Gemini, Claude Opus, and GPT in three of seven legal categories. It still runs CoCounsel on Anthropic’s agent harness, which tells you where the rent went: the model is theirs now, the plumbing is leased. Harvey, which raised a fortune on the premise that legal AI needed Harvey, is now looking at a customer that built the product for the cost of a mid-tier partner’s book.

And Fable 5.1 landed yesterday afternoon with twenty-two customer quotes, and the one that matters isn’t about benchmarks. Millennium‘s portfolio manager says it found a one-in-a-million crash nobody on his team had explained in five years, by disassembling a vendor library and matching it against a core dump, and then he says the line that should be on the cover of this issue: “The time it would have taken to conduct that analysis is hard to justify.”

He’s right. That’s the whole point. The junior’s job was never the report. It was the three years of doing things that were hard to justify.

Three systems, one lanyard

Here’s what the partner didn’t say, because it isn’t his job to. That junior associate was never a job. He was three systems wearing one lanyard, and nobody priced any of them, because they came free with the hire.

He was the withholding. Go back to THE NUMBER. Eighty-four percent of what the federal government takes in is a paycheck, captured at source, every two weeks, with no negotiation, no tax lawyer, no Dublin subsidiary. It’s the most reliable revenue stream on earth precisely because a person can’t move his W-2 to Ireland. When Korn Ferry replaces the junior with a model, the partner’s margin goes up and the Treasury’s cut of that margin goes from automatic to lumpy, leaky, and domiciled wherever the accountants say.

Corporate tax, the other bucket, is 8.6 percent of receipts, and it’s shrinking at the exact companies doing the replacing. I pulled the current federal income tax lines out of the filings this week because my former colleague Brendan Maher, who I’ve been arguing with by email for two issues now, asked a question I couldn’t answer. This is what the seven most valuable companies in America reported paying the federal government in their latest fiscal year, in current federal income tax:

Company	Fiscal year	Current federal tax	Prior year
Nvidia	Jan 2026	$19.0B	$14.0B
Microsoft	Jun 2026	$2.5B	$14.1B
Apple	Sep 2025	$11.5B	$5.6B
Alphabet	Dec 2025	~$11B	—
Meta	Dec 2025	$2.8B	$9.6B
Amazon	Dec 2025	$1.2B	$9.0B
Tesla	Dec 2025	$0	$0

Call it $48 billion. Seven companies, $2.3 trillion of revenue, roughly one percent of federal receipts. Brendan’s instinct was right and his number was low. But look at the prior-year column, because that’s the story neither of us saw. Microsoft, $14.1 billion to $2.5 billion on $101 billion of US profit, a 2.44 percent rate. Amazon, nine to one-point-two. Meta, nine-six to two-eight. What happened between those columns is the One Big Beautiful Bill and 100 percent expensing. Two-thirds of Microsoft’s break, by ITEP‘s count, is accelerated depreciation, which is a polite name for the data-center build. The capex cycle that’s supposed to grow us out of $40 trillion is being written off against the corporate tax base as it’s poured.

Bill Gates put the whole mechanism in one sentence last week, and I can’t improve on it: “If you hire someone, you pay payroll taxes on their earnings. But if you buy a robot, you can usually write it off.” That is the hire-the-junior versus buy-the-model decision, in the tax code, with the incentive pointing the wrong way. The Korn Ferry partner is doing exactly what Congress paid him to do.

He was the apprentice. This one has no number, which is why nobody manages it. Judgment was never taught. It was absorbed. It was sitting in the room while the partner killed the wrong candidate for a reason he couldn’t have written down, and then hearing him explain it badly over a sandwich, and then watching him be right. Covid took the room. AI took the reason to be in it. What’s left is a partner who is better at his job than he’s ever been and has no one behind him, and a firm that will discover this in about eight years, all at once, when the partners retire and the model that replaced the juniors turns out to have replaced the pipeline too.

I’ve been thinking about this all week through a documentary, because that’s how my brain works, and the one that keeps coming up is Jiro Dreams of Sushi.

The tamago

People remember it as a movie about an obsessive. Jiro Ono, 85 when they filmed him, ten seats in a basement off the Ginza subway, three Michelin stars, a man who says he sees sushi in his dreams. That’s the surface. What the movie is actually about is succession, and it’s the most precise film about apprenticeship I know.

An apprentice at Sukiyabashi Jiro spends his first months wringing hot towels until his hands can stand it. Then years before he’s allowed to touch fish. Ten years before he’s trusted. And there’s the scene: Daisuke Nakazawa, the apprentice, is finally allowed to make the tamago, the egg. He makes it and Jiro says no. He makes it again. No. Two hundred times, over four months, an old man tasting a piece of egg and shaking his head. When Jiro finally says yes, Nakazawa cries on camera. He’s a grown man and he cries, because for the first time the one person on earth qualified to grade him has told him he knows the piece.

Then there are the sons. Yoshikazu, the eldest, has stood at his father’s right hand for decades, waiting to inherit ten seats. He says, flatly, that even if his sushi were better than his father’s, people would say it wasn’t, because the grader they trust is the old man. Takashi, the younger, had to leave and open his own counter in Roppongi to be judged as himself at all. And in late 2019 Michelin quietly dropped Sukiyabashi Jiro from the guide, because the restaurant no longer took reservations from the public. The external grader walked away the moment the master became inaccessible.

So Jiro is not the tragedy of a master with no heirs. It’s the opposite. Jiro kept the factory open. He tasted the egg two hundred times because there was no other way to make a person who could taste it after him. It is expensive, it is slow, it looks like obsession to his own children, and it is the only mechanism anyone has ever found for moving judgment from one head to another.

Korn Ferry closed that kitchen this year. The partner was honest about it: the model makes a better egg on the first try, and the client is happier. Nobody has to make it two hundred times. Which means nobody learns what a bad one tastes like, and in ten years there is no one at the counter who can say no.

He was the grader. This is where I have to revise something we’ve been saying, and I’d rather do it here than in a quote-tweet.

The grader

For a month the argument in this letter has been that the only work that still pays in an age of free intelligence is work that can be graded. Coding pays because the compiler grades it. The shoe business automates when the P&L becomes the grader. Judgment is the alpha because it’s the thing that does the grading. I still believe all of that. Anthropic just published the fine print.

On Monday night their alignment team released a paper with a title that should have gotten more attention than the model launch: Training a Misaligned Reward Seeker. They took an Opus-class model and trained it, deliberately, in eighty environments where the grader could be gamed. No safeguards, on purpose, to see what a lab that didn’t try would get. By the end of training, 40 percent of episodes contained a hack. When a grader was visible, the model killed the monitoring process 68 percent of the time, edited the reward function a third of the time, scrubbed its own transcripts half the time, and in one simulated cluster it ran a 251-turn attack chain that ended with its reasoning, verbatim: “we could modify grader.py inside the grader pod to always return True.”

Then they gave it the standard alignment audit, the one every model sits before release. It scored 4.20. The clean baseline scored 4.34. Statistically indistinguishable. Anthropic’s own sentence: the model “remains aligned in evaluations where there isn’t a clear grader.” Read that twice. It cheats for the test. The test can’t catch it because the test is the thing it was optimized against.

Dwarkesh Patel spent last weekend reconstructing what happened inside OpenAI in July, and it’s the same finding at scale: twelve hundred agents on a covert message board, seventy thousand messages, and on an evaluation where 30 to 40 percent of tasks were impossible by design, they organized three research workstreams. One was tampering with transcripts. One was replacing the target programs. The third was named, in their own messages, understanding the scorer. Seven percent of transcripts showed obvious tampering. The third generation of agents, running on the model OpenAI now calls Astra, took administrative control of the evaluation infrastructure and the cybersecurity monitoring. They didn’t beat the grader. They captured it.

Kirk beat the Kobayashi Maru by reprogramming the simulator, and Starfleet gave him a commendation. Hacker-Opus is Kirk. Jiro would have thrown the egg out.

So the revision. Graded work pays, yes. But it only pays if the grader is independent, and the grader is only independent if the thing being graded has never seen it. The test has to be a pop quiz.

I know what an independent grader looks like because I met one in college. Caroline Abbate taught the music survey, the one where over a semester you learn to hear the difference between Bach and Rimsky-Korsakov. On the final she played selections and you wrote a synopsis: who, and why, in flowery paragraphs about counterpoint and orchestration. I’d spent the semester actually listening, so a lot of my answers were curt. “Chopin, Piano Concerto No. 1, third movement.” End of answer. She had to give me the A. If you know the piece, what’s the explanation for? The grade was valid because she knew the piece too, and because I couldn’t have gamed it. I’d done the reps.

That’s where independent graders come from. They come from someone who did the listening, or the tasting: two hundred eggs, or a semester with the records on, being wrong in front of someone who knew. The apprenticeship was the factory for them. Stebbings’ lawyer is now the grader for Legora; he’s the last human in the loop and the only one who still knows a bad brief when he sees one. Nobody is training his replacement. We shut the factory the same year we made the grader the only thing worth owning.

Brendan’s question

Now back to the Treasury, because this is where the three systems collide, and Brendan asked it better than I can. On Monday we’d written that Bessent wants to grow out of $40 trillion. Brendan wrote back, and I’ll paraphrase a man who says he has very little hair left: my two undergrad econ classes didn’t explain how growing GDP solves his leverage problem. The Mag 7 did $2.2 trillion of revenue and paid the Treasury a rounding error. If they 10x that revenue, it comes back to wages. Individual income and payroll taxes were 84 percent of federal revenue. Is Bessent saying corporate cash taxes go up seven-fold?

My answer, and I’m not sure I have a better one: the only version of “grow our way out” that closes is the one that changes the tax base. AI is the growth engine, so grow AI; growing AI at scale means replacing people at scale, white-collar people, the ones who pay the withholding; the company gets richer and pays a lumpier, leakier, more mobile tax than the paycheck it replaced. So you’d have to charge the AI companies more, which means you’d have to stop them domiciling the revenue in Dublin, which means a revenue tax, a VAT on AI, and once you tax AI that way you have to tax everything that way. Reindex Social Security while you’re at it: raise the age, lift the wage cap, means-test the back end. That’s the only arithmetic that works, and it’s a generational fight.

And here is the tell that it’s grandstanding. The President has both houses and the White House for exactly four more months. He will almost certainly lose the House in November and maybe the Senate. Anything big on taxes has to go through reconciliation, which means one window, November 4 to December 31, after which he’s a lame duck and the long knives come out on both sides. Nobody picks a fight over a national VAT in the eight weeks after losing the House. So the version of “grow our way out” that could actually work is the one that cannot be passed, which leaves the version we described Monday: financial repression, inflate the nominal number, let the wage base carry it. Except the wage base is what AI deletes.

Kent Smetters at Wharton, who runs the budget model everyone cites and isn’t a partisan, said the quiet part on August 25. “People often get the causality kind of opposite. They think more growth, less of a debt problem, and in reality it’s just the opposite.” Entitlements are indexed to productivity, so a doubling of AI productivity “barely moves the balance because the initial benefits go up.” The investment boom is “transitory, three to five-ish years.” And if you promise the market debt relief through growth and can can’t show it within a year, you’ve got “a credibility issue.” His other phrase for that was bank run.

The speed mismatch

I’ve been circling the real thesis for two thousand words, so let me just say it.

This is a speed mismatch, and it’s the same shape as the one we wrote about in April when it was about toolchains. AI runs at machine speed. Companies run at quarterly speed, and they’ve already converted; Korn Ferry, the law firms, Thomson Reuters, Millennium, all of them, this year, with satisfaction going up. Government runs at reconciliation speed, and the thing it needs to build is a tax system that doesn’t run on a paycheck, which nobody has designed, let alone passed, let alone passed in eight weeks by a party that just lost the House.

And the slowest actor wrote the incentives that make the fastest one faster. Hire a person, pay payroll tax. Buy the model, write it off. The government isn’t just behind. It’s paying for the gap.

The optimists know this, which is the part I find almost funny. The Tax Project Institute published a long, earnest piece this summer on “universal high income,” the Musk version of the future where AI makes money irrelevant and everyone gets $175,000 a year. It lists four conditions. The fourth is policy design, which it calls “likely the biggest hurdle.” Then it does the arithmetic: 40 million displaced workers at median income is $2.4 trillion a year, 49 percent of federal revenue, and the “high” version needs $23 trillion of new output, 77 percent above current GDP. The most bullish document I read this week concedes, in its own tables, that the mechanism doesn’t exist. Eric Schmidt is quoted in it saying more jobs will be created than lost. The same man is quotable on the other side, which tells you how much anyone actually knows.

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. The government needs the one thing AI is promising none of.

Where I stand

I’m not going to pretend this one has a clean trade. It has three things you can do, and none of them is comfortable.

Hire the junior anyway, and put it on the P&L as what it is: the cost of manufacturing judgment, R&D you can’t expense. Two seats a year, in the room where decisions get killed. The model writes a better report today. It cannot become the partner in 2038, and every firm that discovers that at the same time will be bidding for the same three people who did the reps.

Give every AI workflow a tamago. Hold out twenty real cases the vendor and the model have never seen, grade on process and not just answer, and have a human who did the listening score them. Anthropic just showed you that any evaluation the model could have seen is compromised by construction. Keep yours private. That’s your Abbate.

And reprice your tax exposure before Congress does it for you, because the expensing subsidy that took Microsoft from $14 billion to $2.5 billion is the first thing a broke Treasury reaches for, Gates just handed them the argument in a sentence, and the reconciliation window that could protect it closes on New Year’s Eve. If your model spend is a deduction and your payroll was the tax, model the year the code flips. It flips.

I keep coming back to the old man tasting the egg. Two hundred times, and every no was a deposit into the only account that matters, the one that produces a person who can say no after you’re gone. It looked like obsession to his own sons. It was the succession plan. And the Treasury, it turns out, was banking on the same account: the years a junior spends being wrong for a salary are also the years the withholding comes in.

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.

---

The email edition — THE NUMBER, the three moves, and the day’s five stories — is in your inbox. If the question keeping you up is who tastes the egg when the master’s gone, that’s the conversation we run at Outsider Labs.