Crossgrain No.3 - Nothing in It Is Wrong
Is there something everyone in your industry repeats that you've never quite been able to put down? Have you ever caught it before someone pointed it out?
Usually we learn the answer first, and then feel we had known it all along. The problem is how natural that illusion feels.
Power and charisma don't arrive because you studied late in life. Intellect compounds. But if you mistake the direction of the compounding, decay compounds too. Lean on other people's interpretations long enough and your own circuitry goes dull.
Age into that without noticing, and you end up with more responsibility and someone else's head.
It was a dinner. The résumés around the table were impressive.
The topic was US Treasuries. They said China was dumping US Treasuries at scale. Ray Dalio's debt-cycle framework was cited. Taleb's tail risk made an appearance. A recent forty-page paper was produced as evidence. And for three hours, the same conclusion kept arriving: American debt is too large, China is exiting, the dollar's credibility is cracking.
I asked one question.
"That trillion dollars. Who bought it?"
I had inflated the number on purpose. Nobody corrected it.
If the house is on fire — if the dollar is collapsing — who is buying bonds denominated in a collapsing currency? If you believed America's future was dark, you would have no reason to buy. If someone is buying, the burning-house premise is wrong. Or the house is burning and there is a separate reason to buy at that price. Either way, "China sold, therefore America—" is not thinking. It's transcription.
For three hours, no one asked it. They analyzed the selling side and never once looked across the table. A trade is not just a sale. It means someone bought. Analyzing one side of a transaction isn't analysis. It's dictation.
It seems... Unless someone says look here, the field of vision doesn't open in that direction. After the fact, anyone can gather the data and explain. But at the time, the question didn't exist. Not avoided — the circuit that would have produced it was closed.
There was a scene that frightened me more.
I once asked a group of people I'd taught — graduates of top-tier universities, more than a decade into their careers — to retake the standardized exam that had admitted them. I covered the costs and the incentives. Twelve of thirteen scored below their own entrance results. All thirteen either skipped the quantitative section or got less than half of it right; the single exception merely matched his old verbal score.
A group already filtered by that exam, a dozen years into practice, now seeing less than the incoming students of their own cohort. You can explain it — motivation, format changes, age. The explanations don't blur what happened.
But turn it over: these are people with the courage to put their standing back in front of an exam. Rare people. This essay was sharpened by them.
That's exactly why it's frightening. Good people, smart people, conscientious people — lean on other people's interpretations long enough, and the circuitry dulls anyway.
People guard against wrong statements. That is precisely how they get taken. Correct statements draw less suspicion. They survive into meeting minutes, into reports, into recommendation letters and personal statements and strategy decks, repeated.
And at some point the writing is intact but the person has disappeared from it.
The student below started that way. His sentences weren't wrong. His industry read wasn't wrong. There was simply nothing in the text that anyone hadn't already approved.
Wrong sentences get caught by search. Missing questions don't appear in search.
That eye can be trained. What follows is what the training looks like — first on a student's draft, then on an official document, then on the one document you might not have read this way.
Part 1 — Coaching, before and after
One student's record, through session ten.
His first submission:
student's draft — session 1
With the rapid advance of AI, the HBM market is growing explosively, creating a critical opportunity for Korean semiconductor firms. Samsung and SK Hynix currently hold leading positions in the global HBM market, but maintaining competitiveness will require several strategic approaches.
First, R&D investment should be expanded to accelerate next-generation technologies such as HBM5 and HBM6. Second, manufacturing processes must be optimized to raise production yield, for which advanced equipment is essential. Third, partnerships with major AI companies should be strengthened to build a stable supply chain.
Through these strategic approaches, Korean firms can maintain leadership in the HBM market despite the intensifying US–China technology rivalry. Amid growing uncertainty in the global semiconductor industry, active government support policies will also play an important role.
My first feedback:
Every sentence in this is correct. R&D, yield, partnerships. And if I pasted this into a Micron investor deck, or into SK Hynix's business plan, no one could tell. Whose writing this is. Whose judgment. Where it was written.
This is what the language of the order-taking side looks like. The side that gives orders manufactures questions. The side that takes them optimizes answers. The way Samsung tunes its yield to Nvidia's spec, the order-taker's thinking bends toward specs someone else wrote. So — who defines the spec here?
Building more HBM fabs with today's windfall profits extends the existing game. The moment you're earning the most is the moment of the most important choice.
Session five. He begins building his own hypothesis:
student's draft — session 5
Success with Nvidia hardens the market's belief that "next-generation memory means HBM" — and may, paradoxically, trap us inside our own technical excellence. Recall what Intel did to AMD in the 1990s: pin a competitor with better technology inside the market's frame. Rather than simply continuing the HBM race, we may need a strategic concession — deliberately throttling supply, the way oil producers do. Suppose we cut HBM output by some percentage while embedding AI-specific compute functions into the memory itself, offered exclusively to Nvidia. On the surface Nvidia gains; but if the function lifts training performance by fifteen percent or more, Nvidia becomes dependent on our memory.
Remember that in the 2000s, Intel ultimately had no choice but to rely on the DDR3 standard. If we obsess over making only better HBM, we remain Nvidia's captive supplier.
Throttling output and granting exclusive functions does risk short-term revenue and a major partner's trust. So we should also consider redesigning the GPU–memory interface protocol — a bus structure optimized for specific AI operations, licensed exclusively — to foreclose competitors' pursuit altogether.
My feedback:
Very good. Push it further.
Choices made hungry and choices made full have different structures. Hungry, survival is at stake, so you try new things. The problem comes after you've won. A full organization defends the present: expand capacity, hold share, hit guidance. That's instinct, not strategy.
Then who decides the company's fate? There's an owner, yes — but the executives who must make the long-term calls: are their terms long-term? Build an HBM fab now and the earnings land inside "my" tenure. Pour money elsewhere and the costs land in my tenure while the results go to my successor's successor. Which one gets my stamp?
There is a real dilemma. Move, and you get punished. Don't move, and competitors erode you. As long as you sit on the taking side, every choice happens inside someone else's permission.
So: start where Nvidia isn't looking, with an owner-backed unit on a long horizon. Find the territory Nvidia files under "we'll do that later ourselves." That unit must not build parts that attach to Nvidia's GPUs. Find what doesn't need a GPU — and doesn't eat the GPU market.
Session ten. He starts arguing back.
student's draft — session 10 (summarized, by request)
As AI inference grows, even fiercer competition is fine, because the market itself expands. Google builds its own chips; so does Amazon. The buying side stops being Nvidia alone. This is not the DRAM story again.
My session-ten feedback:
Does a larger buying side raise your bargaining power? Even with many buyers and only three sellers, if all three build to spec, that isn't leverage. Google doesn't call you and say "make us the thing only you can make." Google says "we have quotes from two others and Micron — match the price."
The market grows, buyers multiply, revenue rises. Isn't that the DRAM story exactly? The market grew for fifty years, and the memory makers spent fifty years crushed under the cycle. The size of the pie and the position of the knife are separate questions.
Then try this instead. Nvidia is a toll collector on computation — it earns wherever data must pass through its chips. But does all the data in the world need to pass through a chip like that? What kind of data was never that company's concern in the first place?
Try Qualcomm. Ask a different question first. What kind of data never touches a GPU at all?
This student changed his question in ten weeks. Most people don't change it in ten years.
Which is why, at the same moment, on the same material, everyone says the same thing. And careers stall.
Say only correct things and you'll never be wrong.
You'll also never be distinguishable.
Part 2 — Dissecting an official document
Vanguard. Ten trillion dollars under management. The inventor of the index fund. Your ETF and your retirement account sit on top of this system.
In August 2025, Vanguard published a research paper: "Setting the Record Straight: The Truths About Index Fund Investing."
Its conclusions: index funds do not inflate large-cap stocks. Index funds account for only about 1.2% of trading volume, so they do not amplify price distortion. Since 2000, index investors have saved roughly $503 billion in fees.
16 pages · eleven charts · twelve cited papers
Most readers ask: what's the evidence?
= Trading volume. Index funds are only 1.2% of it, so they don't distort the market.
Don't stop there. We can go further.
Wait — trading volume? What is that?
= The act of buying and selling.
↓
Is that all? Is buying and selling the only thing in the structure?
= …
(No. Something follows the trade: the state of owning — holding the company's shares. Vanguard measured only the act. 1.2%. Negligible, it says. But the ownership?)
↓
So what happens next? Do I become an owner of Nvidia?
= …
(No. When you buy $1,000 of a Vanguard S&P 500 ETF, Vanguard buys Apple, Microsoft, and Nvidia with it, at index weights. Those shares are held in Vanguard's name, in trust. Trillions flow in worldwide. BlackRock runs the same structure. So does State Street. With the world's pensions and retirement savings pouring into index funds, the median stake these three hold in an S&P 500 company is around 22%. With your money. Without your name. Even while trading only 1.2%.)
↓
So — is it good or bad that they hold my shares for me?
= …
(The votes change hands. CEO appointments, executive pay, merger approvals — the right to vote on all of it. The voting power attached to your money is exercised by Vanguard, by BlackRock, by State Street. They protest that they don't intervene in management. Hold that much of a company, and you have de facto say regardless.)
Knowing that "every document has bias" is common sense. Naming specifically what is missing, and why it is missing — that is a skill.
This document isn't sloppy because it's corporate. IMF reports are no different. Neither is Harvard Business Review, nor the analysis on a Bloomberg terminal. Much of what circulates under the name "insight" is narrative. This particular research paper announces it will set the record straight about index funds — and pushes the biggest change index funds ever created outside its own frame: the concentration and delegation of voting power. Everything the paper says is true. But the declared subject and the analyzed scope don't match. A ten-trillion-dollar manager does not misunderstand its own product.
Private firm, public institution, academic journal — none of them will help the reader see the whole structure. The only difference: the higher the authority, the harder the missing question is to notice.
Part 3 — Crossgrain
You just watched me do it to a ten-trillion-dollar manager. Allow me to do more.
The BIS and the IMF report that major financial institutions' exposure to circular financing is modest.
— But the institutions that aren't major?
The World Gold Council reports that central-bank gold buying has doubled.
— So when I want to sell, is a central bank on the other side of my trade?
The story goes around that Petrobras has the lowest oil production cost in the world.
— Is production the only cost a barrel carries?
Notice: these are all the same move. The dinner question again — who is on the other side, what sits outside the frame, which number was chosen for you. None of the counter-questions requires data. Each requires only refusing to let the document choose your field of vision.
Now point the same reflex at a career.
The memory industry is loud right now. Micron, Samsung, and Hynix can fire, pivot, and wait out a cycle — that is what companies are built to do. But the engineer in Texas or San Jose who jumped in because the industry was loud: what does he do when it goes quiet? We have run this experiment before. The coding boom ran it. The bio boom ran it. It ends the same way every time — a market overflowing with talent, price freshly cut, holding a spec that someone else no longer needs.
An industry's noise is a narrative like any other. It, too, has a side that manufactures the question and a side that optimizes the answer.
The people I coach have decided not to run that experiment with their own careers, because they'd rather hold the center of their career than rent it from whatever is loud this cycle.
The diagnostic — $300 USD
One session, 60 minutes, on your current draft. Google Meet. Submit the draft before the session. Bring one document you wrote — a résumé, a self-review, a strategy deck, an MBA essay, an interview memo.
What the diagnostic covers:
- Which question your draft currently answers — theirs, or the easier one (Y)
- Whether a committee could tell your essay from the other applicants' (Y)
- Whether what's missing is fixable in four sessions (Y)
- How to ask better questions (Y)
- What's missing, and how to fill it (N)
The program — four sessions, $3,000 USD
One session per week, 60 minutes each, Google Meet. Total with the diagnostic: five sessions, $3,300.
Nothing in it is wrong.
Something is missing.
— Crossgrain, by Jung.