Hey guys, Non here, coffee in hand. Want to hear some interesting stories?
Somewhere tonight, a machine is cutting the spine off a book.
The pages go through a sheet feeder. They get scanned. Then the paper is thrown away. The book is gone, and a file exists instead.
This is happening at industrial scale, to train AI. And the argument about whether it matters pulled more than six hundred points and nearly four hundred comments out of Hacker News today.
I have a view on it. I'll give you the strongest version of the other side first, because that side is better than you think.
Before that: North Korea answered a summit invitation with a dozen ballistic missiles. After it: America crossed forty trillion dollars in debt, and the cheapest AI model on the market just learned to see. And at the end, an artist who spent ninety-nine years on the exact opposite of a sheet feeder.
Let's go.
North Korea first, because there is a summit on the table and a clock on it.
This week President Trump said he had asked Kim Jong Un for a meeting later this year, and had a positive response. He also cut short the annual military exercises between the United States and South Korea. They ended six days early on Friday. His reason, posted publicly, was that the drills are costly and send a hostile message.
That is a concession. A real one.
Here is how Pyongyang said thank you. On Thursday it fired about a dozen short-range ballistic missiles into the sea, launched from around the capital. That is South Korea's Joint Chiefs of Staff counting.
Then came the statement, from Kim Jong Un's sister. The shortened drills were, in her words, unworthy of comment. On the claim that Trump had been in touch with her brother, she said she was not aware of it at all.
And then, in the same statement, this. The relationship between the two leaders is still excellent.
Read those three moves together. Most coverage says North Korea slammed the door.
Wrong door. Wrong verb.
A political scientist in Seoul told NPR what is actually happening. She is raising the price of re-engagement, not closing the door on a summit.
Look at each piece as a price signal. The missiles say our leverage is live. Unworthy of comment says your opening offer was too small. And still excellent says the shop is open, come back with more money.
We even know the number on the tag. Since Trump returned to office, Pyongyang has repeated the same condition: recognition as a nuclear power. Washington keeps refusing, and the reason is simple. The moment you grant that, twenty years of policy becomes a sunk cost.
Meanwhile Seoul watches its own ally trim joint defence exercises as a bargaining chip in somebody else's negotiation. Experts there are warning that a transactional approach erodes the thing an alliance actually runs on, which is trust.
Read the incentives, and the missiles stop being madness. They start being a spreadsheet.
Now the main story, and it is the books.
A post is going round from Anna's Archive, and its title is the whole argument. AI companies destroy physical books. So let's scan the rare ones before it is too late.
Underneath it is one of the better fights of the year. And at the bottom of the fight is a fact so boring it decides everything.
Scanning a book without destroying it costs about ten times more than cutting the spine off and feeding the loose pages through. One commenter put it flatly. This is about cost. It is not about preservation.
Walk the incentive with me, because that is the whole story.
An AI company needs millions of books to train on. Used books are cheap to buy. Careful scanning is expensive to do. Destructive scanning is fast and cheap. So you buy by the pallet, you cut, you scan, you pulp.
Now compare that with Google's book-scanning project years ago. Google went the other way. It built elaborate machines specifically so it could borrow rare library books and give them back intact.
Why? Because preservation was in the deal. The books were borrowed, so they had to survive.
That is the difference the thread kept circling. When a book is borrowed, you must preserve it. When a book is bought, it is a commodity. And nobody checks a pallet for the one copy that happens to be rare.
That is the actual worry. Not that books die. That rarity becomes invisible at industrial speed. A pallet does not know it contains a last copy.
Now the other side, at full strength, because it was argued well.
Since the printing press, any important book exists in thousands of copies. Sometimes millions. Destroy one, and a digital copy survives that can be reproduced forever. In pure information terms, the trade is fine. Arguably the text is safer scanned than shelved.
If what you care about is the words, nothing was lost tonight. That is a serious argument and I am not going to pretend otherwise.
Here is where I land, and I land as a man who works on infrastructure.
The text is not the only thing a book carries. There is the provenance, and the marginalia, and the binding itself. There is the physical history of how knowledge actually travelled. None of that survives a sheet feeder.
And the digital copy inherits a single point of failure. It depends on a file format. It depends on a server. It depends on a company staying alive. Paper has quietly outlasted every storage medium invented to replace it, mostly by sitting still in the dark and asking nothing of anybody.
So the practical answer is the one in the blog post's own title, and it costs almost nothing. Check which ones are rare before the pallet reaches the blade. One commenter said exactly that — rarity is easy to identify.
The tragedy here would not be industrial scanning. Industrial scanning is fine, and mostly good.
The tragedy would be losing a last copy to save ninety cents. In an industry that sets fire to a billion dollars a week on compute. That is not economics. That is carelessness wearing economics as a costume.
Third story, short, with a very long number in it.
The United States federal debt passed forty trillion dollars this week. That is the Treasury Department's own daily ledger, and it is a record. Washington has added more than two trillion to the deficit already this year. Spending is expected to exceed tax revenue by more than two trillion again.
Now let me be honest about the limits of that. Debt on its own is not doom. It matters relative to growth, to interest rates, to who holds it. None of that detail is in tonight's reporting, so I am not going to invent a collapse for you.
But here is the part any project manager will recognise. A budget that overshoots by two trillion in a year is not a budget. It is a description written afterwards.
And the alarming line in that story is not the forty trillion. It is the phrase no signs of decreasing.
A system with no brake pedal does not need a cliff to be in trouble. It only needs time.
Fourth story. If you have ever sat in one of my workshops with no budget, you have heard me say the same thing. DeepSeek is the cheapest way through the door.
Well, DeepSeek has released an experimental model that can see images. Nearly three hundred and fifty points on Hacker News, and the thread is doing what that crowd does best. Refusing to take the press release's word for anything.
The billing works like this. Your image gets resized, converted into tokens by its dimensions, and you pay for those like text.
The early testing is a beautiful mixed bag. One person showed it a photo of a clock and asked the time. It answered five ten, confidently, and walked through its reasoning about the hour hand. It was wrong. A smaller rival model had nearly got it right.
Another user reported something better. The previous DeepSeek version used to assume it could see. And when it discovered it could not, it invented imaginary image-analysis tools and carried on anyway.
Which is the single most human thing I have ever heard software do.
So it cannot read a clock, and it used to hallucinate its own eyesight. Why am I still telling you to watch it?
Because of who is testing it. In that same thread there are working engineers saying precise image reading is the one thing they currently pay premium models for. If the cheap one gets close, the maths of an entire project changes.
Cheap and imperfect beats expensive and perfect for most real work. That has been true of every tool I have ever deployed for a government with no money.
Test it on your own clock before you trust it. That rule is free, and it applies to everything. Including me.
Before the last story: if you are getting something out of this, subscribe. The World in Twenty Minutes is on Spotify, on Apple Podcasts, on iHeartRadio, and anywhere else you already listen. It is free, it lands every morning, and subscribing is the whole reason it keeps finding people.
The last story is an obituary, and as always, I will tell you exactly how much I know. One line.
The Economist writes that Betye Saar, the assemblage artist who challenged black stereotypes, died on July 22nd, aged ninety-nine. The obituary itself is paywalled. So the name, the art form, the fight, and ninety-nine years is the whole of my material. I will not paint detail onto a real woman's life that I cannot see.
But look at what one line can hold.
Assemblage is the art of taking discarded objects, things thrown out as worthless, and arranging them until they say something the world cannot ignore. The summary adds that she believed thrown-out objects should have a second life.
Now put that beside where we started tonight.
An hour ago I told you about pallets of books, cut apart because careful handling costs ten times more. And here, at the end of the same episode, is a woman who spent ninety-nine years proving the opposite valuation. That the discarded thing, picked up by the right hands, becomes the most valuable object in the room.
She did it to challenge stereotypes. Taking the demeaning image, the thrown-away idea of a person, and rebuilding it into something with power.
Two philosophies of the scrap heap, in one evening. One cuts. One collects.
You already know which one gets remembered for ninety-nine years.
That's the show. Good night.