Hey guys, Non here, coffee in hand. Want to hear some interesting stories?
Let’s start with a statistic. A company in Japan, SoftBank, set a target to buy one hundred thousand new computers this year. Two months in, they had bought zero. Zero.
Why? Because the people who run their factories say the current computers are too slow to handle the new software they are trying to install. It is a strange place to be when the world is screaming about Artificial Intelligence, and a major industrial empire is standing still because it cannot afford to move. That is the big story for today.
Then, a quick check on why secondhand books are suddenly selling faster than new ones. A bit of trivia about how much working memory AI actually has compared to us. And an explanation of why paying for AI is actually a really hard problem to solve. Then, we end with something more human: the record-breaking climb of Nirmal Purja on Everest.
Let's get into the machine.
Independent booksellers across the globe are reporting a strange surge in bulk orders for secondhand novels and academic texts. In Northumberland, Stuart Manley of Barter Books says he recently saw a single order from a Canadian company that matched his entire weekly volume of two or three thousand books—a level of demand he hasn't seen in thirty years of trade. The prevailing theory among the bookselling community is that these books aren't going to new readers, but into the maw of artificial intelligence training programs where they are being systematically destroyed. To understand why this is happening, you have to look at how large language models—the tech behind chatbots like Claude—are built.
These systems need to ingest massive amounts of diverse data to learn how humans communicate, and "Project Panama," an internal initiative by the AI firm Anthropic, aimed to "destructively scan" books at an industrial scale. This isn't a standard scanning process; it involves shipping crates of books to warehouses where spines are removed so pages can be fed through high-speed scanners, after which the remains are pulped or recycled. Because a 2025 US court ruling determined that using purchased books for AI training is "exceedingly transformative" and not a copyright violation, companies have a green light to buy up physical stock to feed their digital needs. The popular narrative right now is one of high-tech plagiarism—that the very soul of literature is being cannibalized by machines.
It paints a picture of a tech giant steamrolling over human creativity to build a better chatbot. But there is a more practical, less cinematic reality at play here: it's an industrial logistics problem. The reason companies are buying these specific books isn't just about "learning" from them; it’s about the cost of data acquisition. It is often cheaper and faster to buy a bulk lot of existing physical books than to negotiate individual licenses with thousands of different publishers or to scrape every corner of the web for high-quality, structured text.
Someone will say that this isn't really "destruction" because these books were already sitting on shelves, forgotten by humans. They might argue that if we don't digitize them, they are essentially lost anyway. But there is a measurable difference between a book being out of print and a book being physically pulped. In Edinburgh, shop owner Derek Walker points out that while one copy of a modern academic text might not matter, the destruction of an 18th-century edition—the only surviving example in existence—is a permanent loss to human history.
We are currently trading the physical record of our past for the predictive capabilities of our future, and we don't yet have a way to measure what that trade actually costs us.
We are currently trading the physical record of our past for the predictive capabilities of our future, and we don't yet have a way to measure what that trade actually costs us. That was the question of the secondhand book, standing at the counter where a digitised algorithm tries to guess what you might like next. Now, a different kind of inventory problem has arrived, and it is sitting in the boardrooms of Tokyo. The question here is not whether we should trust a prediction, but whether we trust the people in the room to make it.
You have stood in a queue in Japan where the staff, a dozen of them, wait for one person to nod before anyone moves a finger. You have felt the polite, weighty consensus that stops a line from forming because it would be rude to leave someone behind. That same system has decided that artificial intelligence is not a tool, but a guest. It is a guest that must be introduced properly, seated at the correct angle, offered tea before being asked to work.
The result is a country with the oldest population in the world, a shrinking workforce, and a productivity record that has been stuck in third gear since the 1990s, yet it is staring at a machine that could theoretically do the work of ten people and choosing to ignore it. According to the Organisation for Economic Co-operation and Development, or the OECD, just 8.4 per cent of Japanese workers use AI as part of their job. That figure comes from data published at the end of last year. In the United States and the United Kingdom, the numbers are significantly higher.
The contrast is striking: Japan is facing acute labour shortages and an ageing population, conditions that should make AI essential, yet adoption has been sluggish. Austin Xu, the co-founder of a US start-up called Kuse AI, set up an office in Tokyo specifically to sell his systems. He describes the culture he found there. "There are organisations where process and consensus culture genuinely slow things down," he says.
"Where tolerance for AI mistakes is close to zero, especially in anything client facing. Some would rather leave a role unfilled than let a machine handle it." In the US, the attitude is different. Business leaders there tend to view AI agents as helpers that can try things, make mistakes, and be corrected. Japanese companies, Xu argues, require a higher burden of proof before they will integrate the technology into their workflows.
Parrisa Haghirian, a professor of international management at Kyoto University of Advanced Science, agrees that risk aversion is the primary driver. She points out that challenges of adopting new technology in Japanese firms are the same as adopting any change, whether it is a new office layout or a new production line. The hesitation is structural. "Japanese firms are highly sensitive to error, uncertainty and reputational risk," she says.
Because generative AI is not fully reliable—it can invent facts, it can hallucinate, it can confidently state nonsense—it is mostly used for low-risk tasks like writing emails, summarising documents, or gathering information. It is not used for core operations, decision-making, or process improvement. This conservatism is visible in sectors that should be modern. In healthcare, some hospitals have not fully digitised patient files.
An employee at a Japanese hospital, speaking on the condition of anonymity, described the paper situation as "like the Stone Age." Data accumulates at a staggering scale, and the friction of moving from paper to digital systems is treated as a risk that is simply not worth the headache. The Japanese government is aware that this hesitation is a problem. Last year, parliament passed the AI Promotion Act, aiming to use light-touch regulation to encourage businesses to invest more in AI. Tokyo has pledged to become "the world's most-friendly country for developing and utilizing AI." The Ministry of Finance reports that AI adoption by businesses has increased.
According to them, 75 per cent of companies are now using the technology, up from 11 per cent five years ago. On paper, this looks like a success story. However, critics argue that the data is misleading. They point out that at each of those 75 per cent of companies, only a tiny proportion of staff are actually using the AI, and that those who do so are using it only to a very limited extent.
It is not that the technology is sitting on a shelf; it is that the technology is kept in the box. Yasushi Ogasawara, an expert on Japan's social system and technology at Meiji University, offers another explanation. He argues that there simply are not enough tech-savvy people in the workforce. "Although the Japanese like playing with gadgets such as smartphones, digital literacy is low here," he says.
This brings us to the core of why this matters. The United States and the UK are treating AI as a way to accelerate their existing workflows. They are testing, iterating, and integrating. Japan is treating AI as a social experiment that must not cause embarrassment.
The consequence is not that Japan will fall behind technologically; it is that Japan is quietly opting out of the productivity gains that AI provides. While other nations automate routine tasks and free humans for higher-level thinking, Japan is preserving a work culture that rewards presence and consensus over output. The gap between what the government measures and what happens on the factory floor is widening. The Ministry of Finance sees 75 per cent of companies using AI.
The shop floor sees people sitting at desks, waiting for a nod. Japan is the third-largest economy in the world, but it is no longer the engine of global growth. It has decided, through a thousand tiny refusals, that the cost of being fast is too high, and the price of being right is a slowly shrinking country.
That last story about Japanese industry ended on a very Japanese note: a refusal to sacrifice accuracy for speed, even if accuracy means a shrinking country. This story starts in a different place, but it asks the same question about trade-offs. It asks whether intelligence is really about thinking faster, or whether it is mostly about remembering more. According to an essay published this week, researchers are beginning to suspect that what looks like a sudden jump in AI reasoning is actually just a jump in working memory.
A human mathematician can hold a small number of unfamiliar elements in mind at once. An AI model, by contrast, keeps the whole problem, hundreds of intermediate equations, several abandoned approaches, definitions and constraints all inside its context window. We normally interpret this as evidence that the machine is thinking faster or more creatively. But the author suggests we might be misreading the signal.
The performance difference may simply be the removal of one of our most important biological limits: our working-memory capacity. Paper does not make us more intelligent; it just expands our effective working memory. A mathematician uses notation, scratch paper and diagrams not just to communicate, but to make the reasoning cognitively possible. The gap between a novice and an expert is not that the expert has a bigger brain; it is that the expert has learned to treat a long sequence of symbols as a single conceptual object, a trick that lets more information fit inside the same biological limit.
Fine. You might say that this is obvious. We all know we cheat by using tools. But we tend to assume that using tools is a side effect of being smart, not the cause of it.
The Hacker News discussion last night made that distinction clear. One commenter wrote that they used to consider themselves very high-performance in their software career, but looking back, most of it came down to having the energy to tackle problems others thought were too much trouble, or simply remembering more details from previous jobs and self-study than the people around them did. Another pointed out that human mathematicians only publish the results that work. A professor might have a file drawer full of dead ends, but the incentives and bandwidth of a human career make publishing those negative results impossible.
An AI agent has no such constraints. It can publish and re-use negative traces without fatigue or annoyance. One engineer summed it up as “out-brute forcing” us: the thing does not get tired, does not get discouraged and does not care, so it keeps trying new directions until something ends up working. There is a real case that this is exactly what we are seeing.
But there is also a counter-argument worth holding in mind. Working memory and general intelligence overlap substantially, and statistical control can never perfectly isolate one from the other. In 2011, psychologist Tracy Alloway and psychologist Rita Passolunghi examined working memory, verbal ability and mathematical skills in children and found that working memory made a distinct contribution to mathematical performance, rather than simply reproducing the association between the two. A separate six-year longitudinal study by Alloway and Alloway in 2010 measured children at age five and tracked their academic achievement six years later.
They found that early working-memory performance predicted later literacy and numeracy even after controlling for IQ, and working memory was actually a stronger predictor than the IQ measure itself. A 2013 meta-analysis by Friso-van den Bos and colleagues found a consistent relationship between working memory and mathematics across primary-school studies. These findings should not be exaggerated, but they do suggest that working memory is not just another imperfect measure of IQ. It is a separate cognitive system that constrains what we can do.
The more important question is what happens when the constraint disappears. If working memory is a hard limit on human intelligence, then what we call reasoning is really just the act of keeping enough pieces of the puzzle in the air at once. If you can store the pieces elsewhere, or keep them in a symbolic workspace that never forgets, the act of holding them in your head becomes the bottleneck, not the thinking itself. A mathematician uses a scratchpad to make the reasoning possible, not just to record it.
If an AI is doing the same thing at a scale we cannot match, we are not watching it think better; we are watching it remember more. And the trade-off is the same as the Japanese factories, just on a much faster time scale. We are deciding whether we want to pay the price of being right by using tools, or whether we want to insist on doing it all in our heads.
If your last story was about the terrifying convenience of a machine that never forgets, this one is about what happens when you ask it to start charging you for its memory. The same logic that made you want to offload your thinking to the cloud is now hitting a wall: the cloud is billing you by the word. According to Goldman Sachs, we are about to consume a hundred and twenty quadrillion tokens a month by 2030. That is not a casual figure.
It is a number that implies every human on the planet would have to be typing into a chat interface for three hours every single day just to keep up with the machines. Right now, if you are using a free account, you are essentially mooching off the infrastructure that Microsoft, Google, and Anthropic built with those hundreds of billions of dollars. You get a bargain, but they are taking a bath on your usage. Third‑party firms are now selling AI agents—software designed to do specific jobs like fixing a security breach or writing code—and they are terrified to lock in a price.
Simon Gooch at Saviynt says tying a client to a cost model for twelve or twenty-four months makes no sense because the math is shifting under them every week. To explain why this is breaking, we need to look at how these systems actually work, which is a little counter‑intuitive. When you ask an LLM to do something, your request is chopped into mathematical chunks called tokens. The model processes those chunks and spits them back out.
The problem is that a single prompt can vary wildly in how many tokens it generates, and the same prompt on two different models can return two entirely different results. An agentic system, which uses multiple AI agents to decide and act, makes this mess even worse because it consumes tokens not just to think, but to execute. The cost of an individual token has crashed in recent years, but we are burning through them so fast that the bill is still exploding. Fair enough.
You might be looking at your own software budget and asking why you should care about some abstract math in a spreadsheet. The answer is that you are already seeing the friction. Uber recently burned through its annual budget for AI coding tokens in a matter of months. Microsoft’s engineers are accidentally racking up huge bills using third‑party tools.
It is easy to be cynical about corporate waste, but this is different. The people I spoke to say it is not greed; it is confusion. Oliver King‑Smith from smartR AI points out that smaller companies can fly under the radar by using personal accounts, but he says the big vendors will not let that last. Once shareholders start demanding a return on investment, the flat‑rate free model ends.
We will be forced to sign contracts with annual pricing that no one can predict. The more important question is not why the companies are charging, but how we stop ourselves from accidentally spending the company’s entire quarterly revenue on a single chat session with a marketing bot. You would not send your spouse to the grocery store without a list. You would not hand your child a credit card and say, "Have fun, I’ll check the receipt later." Yet that is what many businesses are doing with AI agents right now, expecting them to use tokens efficiently without a detailed brief.
Rob Steele at iplicit puts it plainly: a prompt needs to be as precise as an instruction manual for a nuclear reactor. Otherwise, the bill arrives and the money is gone. The gap here is between what we assume AI costs—which is near zero because we have been using the free tier—and the reality, which is that we are burning through the equivalent of a small country's electricity bill on tasks we did not think required one.
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 we did was about the invisible bill we are all paying for cheap artificial intelligence, and the question it raised was how much is a computation actually worth. The person who spent his life answering that question in a very different way was Nirmal Purja. The Gurkha mountaineer and Netflix star died in an avalanche on July 30th, aged 43. That is all I have from The Economist obituary line, and it is not enough to tell you who he was, but it is enough to know what he was running toward.
We measure the value of a service by what we are willing to pay for it, so we build systems that optimise for the lowest possible price, and then we pretend the resource is free. Purja did the opposite. He was a mountain climber, and he set out to climb the highest peaks on every continent in the shortest amount of time. He wanted to know what a human body could actually endure.
The Death Zone is the part of a mountain above eight thousand metres where the air is so thin that you do not have enough oxygen to think clearly, let alone breathe. Most people turn back, or they die, or they take risks that save them but ruin their health forever. Purja pushed into that space, and he did not do it to make money, he did it to break a limit. He did not want to argue about the price of oxygen; he wanted to see if the limit was real.
Someone will say that climbing mountains is a luxury hobby for people with too much time and too much money, and that it has nothing to do with the energy crisis we just talked about. But the two stories are not as far apart as they seem. The incentive is the same: a system assumes a limit exists, and it optimises around the assumption that the limit cannot be crossed, so it never bothers to check. The engineer assumes the computer will run on free electricity forever, and the climber assumes the body will stop at eight thousand metres and never go higher.
One system burns through gigawatts of power until the lights go out; the other burns through years of life until the body gives up. In both cases, we treat the constraint as a rule of nature, not a choice we are making. We treat the Death Zone as an unchangeable fact of physics, when really it is just the place we stopped trying. We do not have good data on how many people died on those peaks before cameras started recording everything, or after, because the record-keeping was bad.
We also do not have a precise number for how much electricity we have burned through to train the models we are using today, but the bill is in the air that is becoming harder to breathe. The difference is that with climbing, you at least see the mountain and you at least see the body. With AI, we are told the limit does not exist, so we keep pouring in more and more until something has to give. I do not know what the final cost will be, but I am fairly sure it will not be zero.
Thanks for listening.