The Weak Layer is a podcast. Why read when you can listen? This one has Beethoven!

Part One: Black Powder

In the 1600s, the cultural elite of Europe fell in love with a new drink: coffee. Imported from Africa by way of the Ottoman Empire, coffee gave writers, artists, and composers a rush of creative energy. Hundreds of coffee shops opened in Vienna and Paris and London, where customers gathered to work and play games and gossip with one another.

The political implications of all this high-energy talk were serious enough that Charles II tried to shut down the coffee shops. But public outcry was so strong his prohibition lasted only 11 days. Frederick the Great imposed punishing taxes on coffee, in part because he preferred beer, and had to send soldiers into the streets to sniff out unlicensed roasting operations.

Both Mozart and Beethoven were regulars at Vienna coffeehouses. Beethoven was also a coffee obsessive at home, where he supposedly counted out precisely sixty beans per cup, which he made in a glass carafe. Voltaire claimed to drink fifty cups of mocha per day, and Balzac, who drank coffee more or less constantly, gave us this memorable description of its effect on his mind:

Ideas quick march into motion like battalions of a grand army to its legendary fighting ground, and the battle rages. Memories charge in, bright flags on high. The cavalry of Metaphor deploys with a magnificent gallop. The artillery of logic rushes up with clattering wagons and cartridges. On imagination’s orders, sharpshooters sight on fire. Forms and shapes and characters rear up. The paper is spread with ink. For the next slide, Nightly labor begins and ends with torrents of this black water. As a battle opens and concludes with black powder.

The great minds of the Enlightenment adored their black powder for the power it gave them. They could paint and write and play and craft more output with more ideas and more ambition. That feeling of increased capacity, even more than the immediate effect of the drink itself, was an important part of what made and continues to make coffee so beloved.

Part Two: Busy Beavers

Since this AI hype cycle began with the release of ChatGPT, we’ve been told that one of AI’s greatest strengths was writing code. Any day now, AI agents would be taking over the job of computer programmer.

We are not far from world. I think we’ll be there in three to six months. where AI is writing 90% of the code. (Dario Amodei)

Continuing to trace the exponential, I think what will happen is coding will be generally solved for everyone. (Boris Cherny)

But it didn’t appear that many serious programmers were actually letting AI do more than make suggestions or fill in boilerplate. Prominent coders, people who’d written popular programming languages or built major apps, were pretty vocal in their skepticism.

AI is a fancy autocomplete, which will make writing code easier, but it does not make writing code free. (Michael Paulson)

And the way I use AI is in a separate window. I don’t let it drive my code. (David Hansson)

Then late last year, comments started appearing on social media, along the lines of, “hey, this stuff actually works now?” During holiday downtime, more of those professional coders gave these tools a longer leash on their personal projects. For many of them, it was like they’d been walking past the coffee shop every morning and finally decided to come in and have a taste.

There’s an old saying, “Your options are fast, cheap, and good. You can have two out of three.” AI coding challenges that equation because it’s not just fast, it’s almost instant. And it’s not just cheap, it’s almost free Yes, I know the AI companies are spending hundreds of billions of dollars on data centers, which in turn cost billions of dollars to operate. But if you’re an end user, a single person at a single keyboard, thousands of lines of code might cost a few pennies. And they’ll be done before you can drink a cup of coffee.

Programmers sometimes describe code as elegant, and credit good programmers with having taste. The output of these new AI tools isn’t that. It’s not the lean, not-a-line-longer-than-necessary code that humans strive for. But to be fair, neither is most human written code. This is not the code that you want managing your bank account or analyzing your x-rays. Not yet, anyway.

But there’s lots of uses for code that aren’t load-bearing columns in the world economy. And when code is free and instant, you start to notice a lot more of those opportunities.

Programmers have. More than noticed. Like Balzac on his tenth cup of the day, they’re going a bit crazy. “We all stopped sleeping,” is how one speaker at a conference recently put it. I don’t have a way to measure this for certain, but I’d wager that more code has been written in the past six months than in the prior history of computing.

It’s a disease. Token fever.

There’s a website called GitHub, which millions of programmers use to store code and share it with others. It was a big deal in 2025 when GitHub received over a billion new code saves in a single year. Today — and 2025 was only five months ago — GitHub’s getting several million saves a week and expect to see 14 billion for the year. A 14x jump over what was already a skyrocketing increase.

Anthropic’s Claude Code is the most well-known AI coding tool, and the programmers behind it are its most ardent fans. They are vocal about writing little or no code themselves. They let Claude Code build itself. The result has been a torrent of new features, almost daily updates with sometimes pretty substantial new capabilities. It’s not just Claude Code, there are dozens of these tools now. Elon Musk’s SpaceX just made a deal to buy one of them, called Cursor, for $60 billion. Token fever is a pandemic, and there’s no known cure.

Most people, let’s say “regular” people, use AI primarily to answer questions. And there’s a lot of hype about using AI to do things out in the world, like make travel reservations or do our taxes. But AI coding is a different kind of thing. Because coding — making computer software — is creative. It creates things. And you may not find many of the things it creates all that interesting. But bear with me for a moment and dwell on the feeling of creation itself, regardless of the output.

The feeling of making something new in the world. You plant a seed and a flower emerges from the earth. A few hours with needles and yarn or paints and a canvas, or walking the neighborhood with your camera, and there’s a beautiful object, where before there was only a thought. We love this feeling.

The philosopher Jacob Bronowski wrote: “Every animal leaves traces of what it was. Man alone leaves traces of what he created.” I find that quote particularly compelling, because it’s so obviously wrong. Beavers redirect rivers, termites build cities, octopuses decorate their dens, and the great barrier reef can be seen from space.

I assume Brnowowski knew all that, but he was writing about what he called “The Ascent of Man.” He wanted to make a point about how central creative power is to our nature. Creation is powerful. We value it so highly that we don’t want to let the animals have it. We want it to be ours and ours alone.

Creation is powerful and it feels powerful. It also feels fun and optimistic and energizing. Programmers have stopped sleeping because they’re so busy creating.

Companies literally rank their software engineers by how many AI tokens they generate. Token maxing, they call it. And AI coders refer to themselves as token billionaires for their abundant use.

I don’t think I’m a token billionaire yet, but I can tell you, I know the feeling. My own folder of new projects has 43 items in it, more than half of them created in the past two months. I used one of them to record and edit the narration for this episode. There are plenty of existing tools for this, but none work quite the way I want. So I had Claude Code make me a better one.

I want to be clear, I am not a programmer. I’ve taken a few online programming courses as a hobby, but little of what I’ve learned of them has any bearing here. And nothing I’ve built could survive for five minutes in the wild, where it would be hacked, cracked, tormented, and turned into malware.

For me, that’s more or less how token fever manifests, as a game or a hobby. Still, though, the ability to create custom software, even if most of it isn’t very good. It’s enough to provide that intoxicating feeling of creativity. And as the agents improve and the software I can make with them gets better, who knows where that could take us?

When token fever infects a real programmer, someone with deep integrated understanding of how software works and what it can do… it’s like letting Shakespeare into a movie studio or installing a 3D printer in Edison’s workshop.

Even with that 14x increase in code saves to GitHub, I suspect this unleashed feeling of power is still confined to a small portion of society. It’s likely that token fever has only even infected a small portion of professional programmers, those with the time and the freedom to take the risks that are inherent in letting AI write the code.

Will this mania break contain? Will it endure? Is it the next pickleball, or the next tennis? Or a whole new category of thing?

Part Three: The Power Paradox

A century after Coffee conquered Europe’s intellectual elite a different kind of black powder emerged as the industrialists’ drug of choice. Coal burned day and night, powering the machines of their revolution. In seventeen hundred, Europe’s coal use was a few million tons per year. By the mid-1800s, it was hundreds of millions of tons.

Coal had become a precious national resource, especially in England, which burned more coal and produced more goods than any other nation. People were concerned about running out of coal and hoped that more efficient steam engines would mean they wouldn’t need so much in the future.

But in 1865, an economist named William Jevons noted something counterintuitive: Steam engines had gotten much more efficient but the result wasn’t that English factories used less coal. On the contrary, they used more. Much, much more. This has become known as Jevons Paradox.

What happened with coal was that increased steam engine efficiency made power cheaper. You could get more watts for less coal. Which in turn encouraged entrepreneurial industrialists to find more uses for power. Different kinds of factories, but also things like streetlights and trolley cars. Chemists discovered that with abundant raw power, you could transform materials into stronger or more pliable or more durable versions. Cheaper power unlocked our creativity, which outpaced the efficiency improvements themselves.

This increase in our use of power hasn’t stopped, even as we’ve diversified our sources of power. Those AI data centers are just the latest examples. Fun fact, some of them are powered by coal.

If you follow the AI conversation, you’ve probably heard Jevons Paradox mentioned. It’s one of the tech elite’s go-to references when confronted with pesky questions like, “Hey, is this AI thing gonna’ destroy the world economy?” What Jevons saw with coal, however, is not a generally applicable law of economics. If dog food were ten times cheaper, would people adopt a hundred times as many dogs?

It applies when there is an untapped need for the thing being made cheaper. In economic terms, it’s when the demand curve is highly elastic. For an input as basic as power, the demand is essentially unlimited, restrained only temporarily by the limits of our imagination.

The future trajectory of token fever and the code explosion that it’s generating will be determined by how well code fits Jevons paradox. For how long will cheaper, faster code enable an increasing variety of novel uses, as cheaper power has done for centuries?

Part Four: Software Ate My Brain

The customers of those 17th and 18th century coffee houses were an exceptionally productive group. Besides the plays and poetry and symphonies they wrote, the London Stock Exchange, Lloyds of London, Sotheby’s, and the Royal Society were all founded in coffeehouses. Their most important creation, though, is not any one thing, but the intellectual revolution we know today as the Enlightenment.

It’s probably overstating the case to claim that coffee caused the restructuring of an entire society’s worldview along rationalist lines. But it’s not implausible that it wouldn’t have happened without the coffee, the coffee houses, and their energizing effect.

The thing is, getting hopped up on caffeine and arguing over whether light is a wave or a particle has some shortcomings. Coffee may have literally killed Balzac, for one thing. Less dramatically, coffee drinkers, like anyone in the grip of a stimulant, can produce a lot of junk. I mentioned earlier that Claude Code, the most well-known agentic coding tool, releases new versions almost daily. If you read the release notes of those versions, they do contain new features, but they’re also followed by dozens of bug fixes to the features that they introduced just days before.

Creative mania also has a tendency towards beginnings and struggles with finishings. I wonder if token fever makes us less like Balzac, furiously scribbling away at four in the morning, and more like Leonardo da Vinci, whose genius left behind a slew of unfinished paintings and projects. Here was a man with perhaps too much talent, too much intelligence.

AI optimists tell us that this technology will give us all access to genius-level intelligence. But what if that’s like giving a ten-year-old a fire hose and telling them to water the garden with it? Wielding unlimited intelligence effectively, instead of just spraying it around, might turn out to be harder than it looks. Even actual geniuses struggle.

Token fever can give rise to opportunistic co-infections. Tech journalist Nilay Patel wrote an essay in which he describes a condition he calls software brain:

So what is software brain? The simplest definition I’ve come up with is that it’s when you see the entire world as a series of databases that can be controlled with structured language. Software code.

Yeah, when you have token fever, the world can start to look like that. Once you’ve seen the power of deploying code against things that are structured and are tractable, you want everything to work that way.

But as Patel goes on to point out:

People aren’t computers and they don’t live in automatable loops, and the entire human experience cannot be captured in a database. That’s the limit of software brain.

Patel is rightly concerned about overreach, about technologists who fail to see the boundary he’s drawing between accessible to code and not. But what’s inside the boundary is still quite a lot of our daily experience.

How many things in your day would you rather were automated? How many things could be more efficient? And what if to get that automation, you didn’t have to buy a program, install it, learn its nuances, etc. What if you could just say, “I want this,” and the AI would make it? You could move on with your day. Or if you were so inclined, if you had a touch of the token fever, you might have the AI tweak its creation. Shape your digital tool until it fits the curves of your life with a satisfying click.

There’s a lot of worry about this sort of thing dehumanizing us. But I don’t find anything particularly soulful about categorizing my credit card charges or scheduling a plumber. I might want to be more or less engaged in some of these things. Classifying my credit card charges is how I keep an eye on our overall financial health. But I wouldn’t mind if it were smoother. If the data were presented to me more cleanly, my decisions captured more elegantly. I want to have a conversation with my virtual accountant, not plod through cells on a spreadsheet.

The first time you tell an AI agent to build you something that smooths out a little friction in your life or brings you a little joy, and it works… that’s how you catch token fever.

Right now, the infection vector is limited. As magical as Claude Code and the like can be, getting working apps out of these tools still requires a decent knowledge of the infrastructure of computing. But millions of people do have that knowledge, and the AIs are getting better and better at shielding us from computing’s rougher edges.

Analysis of the impact of AI tends toward the rational and the calculated. I suppose that’s the legacy of the Enlightenment. We want to know how many millions of jobs will be displaced, what industries will be upended, what are the big picture macro changes we’re headed for. But the future of AI and its impact on society will be made by humans (at least until some technological singularity takes it out of our hands). And humans are not all that rational. We follow our feelings.

While most of us aren’t tech titans whose choices ripple through the world economy, millions of us are trying to solve little problems and create little moments of joy. Things that unlock human creativity, like coffee and coal did hundreds of years ago, tend to have far-reaching and unpredictable effects. Will AI join that list? I have a feeling it will.

But maybe that’s just the token fever talking.