The Truman Show

THE NUMBER: 50 tons. That’s how many Japanese books, according to export records Nippon TV dug up, have shipped from Japan to the United States since used bookstores across the country started getting strange orders this summer. These weren’t bestsellers. They were philosophy, law, medicine, daily life in the Edo period and thirty-year-old political memoirs, ordered at machine-regular intervals from multiple accounts, and every box went to the same warehouse in Okayama. Nobody has claimed the buyer. The whole trade thinks it knows who it is anyway. Keep the number in your head. It’s the first price tag we’ve seen on a question that’s going to define the next decade: what does the machine want to read once it has read everything we typed?

A hundred books a day

Here’s the story as Nippon TV reported it on Tuesday. Starting around August, online used bookstores all over Japan (the kind that list on the big marketplaces, many of them descendants of the old shops in Tokyo’s Jimbocho book district) started getting bulk orders. One shop told the reporters it was selling a hundred books a day. Another said sales had hit five times normal on some days. A shop owner in the Kanto region who had sold more than a thousand books into the stream noticed the orders arrived at even intervals and guessed that someone had built a system to buy mechanically. He’d never sold that many books to one address in his life, and his friends in the trade were seeing the same thing.

What sold was the tell. A Tokyo seller called the mix “incoherent”: philosophy, history, medicine, law, books on Edo-era customs, books about politics thirty years ago. “Normally novels and comics sell,” he said. “Specialist books have never sold like this all at once.” Nobody collects that combination. What it adds up to is coverage, a buyer trying to own a little of every subject.

The boxes all went to one logistics center in Okayama Prefecture. The company that runs it declined to comment on individual transactions. When Nippon TV kept pulling the thread, it found export records for more than 50 tons of Japanese books headed to the US. A shop in western Japan got an email from an overseas company asking whether it could fill an order for tens of thousands of volumes. The Tokyo used booksellers’ co-op confirmed it was aware of the pattern and said the talk in the industry is that it’s for AI training.

One bookseller summed up the mood: glad about the sales, uneasy about what happens to the value of books. Me too.

The machine has read everything we typed

To see why someone would buy Japanese books by the ton, you have to see the wall the AI industry is running into.

Every large language model is built on text, and the cheap text is gone. The open web, Wikipedia, Reddit, digitized news, GitHub: all of it has been scraped, deduplicated, filtered and fed to every serious lab several times over. Epoch AI, the research group that tracks these things, estimated a couple of years ago that the stock of high-quality public human text would be fully used up somewhere between 2026 and 2032. We’re inside that window now.

Meanwhile, the price of processing text is falling off a cliff. Epoch’s newest numbers say the cost of a given level of AI performance has fallen about 47% a quarter for three years, roughly thirteen-fold a year. So you have a factory whose machinery gets thirteen times cheaper every year and whose raw material is running out. You don’t need an economics degree to guess which one gets more valuable.

The internet only holds the part of human knowledge someone bothered to type. Most of what humans ever figured out (our history, our arguments, our conclusions, the long record of people working things through) sits on paper, and most of that paper isn’t in English. Japan’s out-of-print specialist shelves are one of the richest untouched reservoirs on earth: a highly literate country, a century and a half of serious publishing, and a language barrier that kept most of it away from the English-speaking web.

And here’s the twist that makes the language barrier worth money instead of a problem. A model doesn’t care what language it reads in. Feed it a 1994 memoir of Japanese party politics and a 1960s textbook on Edo merchant households, and it gives them back in English, Spanish, Hindi or whatever you ask for. Every language in, every language out. Knowledge that was locked inside one language for a hundred years becomes available to everyone at once, and the shop that held it got the price of a used book.

Anthropic wrote the playbook

None of this is new in kind, and we know what the playbook looks like because it has been litigated.

In 2004, Google started scanning the collections of major libraries for Google Books. The authors sued, and after a decade of fighting the courts called the scanning fair use, because Google showed snippets and not whole books.

In 2025 we got the AI version. In Bartz v. Anthropic, the court record showed that Anthropic had bought millions of print books, many of them used, cut off the bindings, scanned the pages and discarded the originals to build its training library. Judge William Alsup ruled that part fair use: you bought the book, and training on it was, in his view, transformative. Where Anthropic got into trouble was the other library, the pirated copies it had downloaded, and that part ended in a settlement of roughly $1.5 billion. The lesson the industry took away was simple. Buy the book and you can train on it. Steal the book and you’ll pay.

Nippon TV’s follow-up story is headlined around exactly that: paper books scanned for AI training and then thrown out. So picture the 50 tons. Someone bought them legally, at retail, one at a time from small shops, the way the Anthropic ruling says you should. The legal playbook and the Okayama warehouse are the same strategy.

There’s a darker version of the same hunger in this week’s news, too. Anthropic’s threat report, published Tuesday, detailed how Xiaomi pushed more than 400,000 requests through 1,500 accounts to distill Claude, and named seven Chinese labs doing the same thing at scale. Beijing responded by summoning those labs and asking why they’d routed their own users’ prompts and data through an American model. Distillation copies the answer key off someone else’s paper. The book buyers are going back to the original sources. As long as intelligence keeps getting cheaper, both trades pay. But only one of them is buying something scarce.

But books run out

Here’s where I part ways with the story as the Japanese press is telling it. The bookstores are a symptom. They’re what the end of a supply looks like: the last reserves of written human thought being bought up, one specialist shelf at a time. Once they’re scanned, they’re scanned. You only have to buy the Edo cookbook once.

Books capture what we concluded. They’re the edited, finished, footnoted version of human thought. What a model still doesn’t have is the messy part: how people actually behave when they’re arguing, deciding, hesitating, laughing, cooking dinner or losing their temper. That’s a completely different kind of data, and there’s a lot more of it.

So ask the obvious next question: where is the biggest pile of raw, unscripted human behavior sitting on a shelf?

The cutting room floor

Here’s the vault I’d be staring at: Warner Bros. Discovery.

Forget HBO and the movies for a second. Look at the cable side: Discovery, TLC, HGTV, Food Network, Investigation Discovery, Animal Planet, Magnolia. That’s the biggest reality-television operation in America. At its upfront in May, WBD said its cable networks would premiere more than 3,000 hours of programming in the 2026–27 season, most of it unscripted.

Now the part most people outside TV don’t know. Reality shows are made in the edit. A crew rolls cameras for days, and an editor cuts that mountain of tape down to forty-two minutes of story. The industry rule of thumb is about a hundred hours of footage for every hour that airs, and the Motion Picture Editors Guild’s own trade journal says a one-hour reality episode can mean 120 to 450 hours of dailies. Scripted dramas run a fraction of that.

Do the back-of-the-envelope math at the conservative end. Three thousand hours a year at 100:1 is 300,000 hours of raw tape a year. Run it back a decade and you’re into the millions of hours: people flipping houses, planning weddings, fighting over the check, crying in a kitchen, sitting in silence in a truck. The editor threw almost all of it away. The part that got thrown away (the unscripted in-between, the pauses, the ordinary behavior between the “moments”) is exactly what a model trying to understand humans wants.

We already know there’s a market for this. Early last year Bloomberg reported that OpenAI, Google and more than a dozen other companies were paying YouTube creators for their unused footage, usually $1 to $4 a minute, through brokers like Troveo and Calliope. Cutting-room-floor video went from worthless to an income stream overnight. That’s the price for one creator’s B-roll. Now think about the biggest reality-TV archive in the country.

And the vault is changing hands

The timing is what makes this more than a thought experiment.

On Monday, Paramount Skydance settled with twelve state attorneys general who’d sued to block its $111 billion all-cash takeover of all of WBD, Discovery networks included. The judge still wants a hearing on the settlement terms, but the deal could close within weeks. When it does, the combined company will carry roughly $80 billion of debt.

Nobody has announced an AI licensing deal for the footage. I want to be clear about that: this is an incentive, not a reported plan. But I’ve spent thirty years watching what happens after a leveraged deal closes. The new owner walks the building looking for assets that are on the books at zero and could be worth something. A vault of raw tape that has never generated a dollar and that AI labs would pay real money for is about as clean an example as you’ll find. Debt that size goes hunting for assets nobody has priced.

Two honest caveats. First, not every hour of raw tape belongs to WBD. Plenty of reality shows are made by outside production companies, and who owns the dailies varies contract by contract. Second, the people on camera signed releases years ago that never imagined their argument over a kitchen backsplash would train a model. Expect that fight. But expect it after someone prices the vault, not before.

The purest trove is the one you’re wearing

There’s one more problem with reality TV as training data, and it’s the reason the story doesn’t end at Paramount. Reality TV is still people performing. They know the camera’s there. They’re cast because they’re loud, and they’re edited to be louder. It’s behavior, but it’s behavior for an audience.

The purest record of human behavior is people who have forgotten the camera is there. And on Wednesday Mark Zuckerberg stood on stage at Connect and moved that camera onto your face.

Meta’s Muse agent is coming to its glasses in the coming months, with live video and real-time conversation. Muse will be free “for a huge number of tokens,” Zuckerberg said, with Meta planning to profit over time from small fees on transactions. Meanwhile, the fine print is already written. Since last spring, the AI on Ray-Ban Meta glasses has been on by default, and Meta keeps voice recordings for up to a year to help train its AI. You can delete recordings one by one. You can’t opt out of the training. Meta says a “Private Processing” mode is coming by year-end, and I’ll believe it when I read the terms.

Think about what that stream is. It’s first person. It’s continuous. It’s every room the wearer walks into, and every person in those rooms, none of whom signed a release. It’s in every language, because people wear glasses everywhere. Books gave the machine our conclusions. Reality TV gives it our behavior when we know we’re being watched. The glasses give it our behavior when we don’t.

That’s the ladder, and every rung is worth more and consented to less than the one below it.

We accept the reality we’re presented with

In The Truman Show (1998), Jim Carrey plays Truman Burbank, a man whose entire life, from birth, has been a 24-hour television show. Five thousand hidden cameras, a whole town full of actors, and a producer named Christof who explains it with one line: “We accept the reality of the world with which we’re presented.” Truman is the last one to find out. Everybody else is watching.

The movie was a satire of reality TV. It turns out it was a forecast. The show isn’t going to be one man’s life. It’s going to be everyone’s life, a little at a time: our books first, then our tape, then whatever our glasses see. And the audience isn’t a bar full of fans in Seahaven. It’s a model.

I’m not a Luddite about this. The Japanese books will make models better at Japanese history, Japanese law and Japanese medicine, and that’s good for everyone who asks. The raw tape will help machines understand people, which is what we keep telling them to do. But every rung of the ladder has the same pattern: the material is priced by the buyer, the seller finds out afterward, and the people in the pictures find out last. The bookseller got retail. The YouTuber got $2 a minute. Truman got nothing.

What to do about it

Three moves, today:

Inventory your unread books. Every business has a pile no model has seen: paper files, recorded sales calls, training videos, field notes, raw footage from the marketing shoot. Somebody is paying by the ton for Japan’s version. Make the list, then go read the “we may use your content to improve our products” clause in every tool your company uses, because some of them are already taking your pile for free.

Set a glasses rule before your first client walks in wearing a pair. On-by-default AI, a year of retained voice, no opt-out from training, and live video coming. Write one line for your team today: are recording wearables allowed in client meetings, and who has to be told? It’s a two-minute decision now and an awkward one later.

Put an AI clause in every release you sign from here on. Talent releases, vendor contracts, event recordings, customer calls. One sentence: can this footage or transcript be used to train AI, and at what price? The people on a 2014 reality show didn’t have that sentence. Their tape is about to change hands anyway.

And one for the allocators, with the usual caveat that this is analysis, not advice: if intelligence gets thirteen times cheaper every year, the thing to own isn’t the intelligence. It’s the stuff it hasn’t read yet. Archives, libraries, media vaults, newspaper morgues, proprietary data. Anything original, human and not yet scanned is being repriced upward while the model is being repriced down.

Good morning, and in case I don’t see ya

Truman’s catchphrase was a greeting to his neighbors: “Good morning, and in case I don’t see ya, good afternoon, good evening, and good night.” He said it every morning to people who were paid to be there, in front of cameras he didn’t know about.

They bought our history by the ton this summer. The tape is next. Then the day you’re living.

Own your tape. In case I don’t see ya, the glasses will.