Silicon shell

Capital keeps feeding an animal that won't stop growing, and nobody aboard knows what it is.

Clad in oak and in men whose eyes had not stopped gleaming with gold since the last days of winter, a galleon sailed the world, one that the ancient maps dared draw only at the edges of the known.

The cabin boy Basidas knew from the first night that in the hold, where neither sunlight nor the reason of officers reaches, something breathed. He could not leave that question unanswered, so one night, armed with a pair of boots and a candle, he descended, and by the light of the torch among the chains, he understood that what stared back from the corner of the keel had not chosen to be there.

The next morning, more men went down to see it. The first said it was a crab. Enormous, he said, and gleaming, with a shell that changed color depending on what it ate. Another swore he had seen a squid with neither center nor periphery, which answered questions in an ancient tongue. A turtle that knew everything, an octopus that untied knots, an eel so dark that no one could tell where it ended and where the water began. All were certain of what they had seen. None described the same thing. And yet, when one finished speaking, the others nodded slowly, like someone who recognizes in another's words the exact name of something they always knew and could never say.

The officers, who were practical men and had no time for mysteries, sent a petty officer, a man of few words and many certainties, who came back up after a moment with the satisfied expression of someone who has solved something that was never a problem. It was a crab. Large, yes, and odd in certain respects not worth detailing. They assigned it a shell, lateral eyes, and spread the word that it produced gold. That if the sailors fed it their rations and kept feeding it until they reached port, a day would come when it would begin to excrete nuggets that would pay for the voyage and every voyage after. That the curve would bend at the right point. That those who had disembarked too early would be the ones who'd never aspire to the sea again. That it was a matter of patience, that it was, at bottom, a matter of not going down to the hold to see what was there.

And no one, on the decks or in the dreams that night (all, every one, the same), thought to ask the only question that mattered: what would happen when the crab grew larger than the ship?

Revisions26 · oct 05Read notethe sources now sit beside each claim, and checking them moved a few things. The people quoted here now say what they actually said, where before I had put my paraphrase in their mouths. The 67% fall in junior tech postings had no source I could trace, so it is now Stanford's measured drop in young developers' employment. Epoch's data horizon runs to 2032, and Hinton left Google a year and a half before his Nobel, not five months.

We've been living alongside beasts for millennia. From fire to the internet, each one came aboard and we rearranged the deck. With every new beast, the same dynamic: we command, it obeys, and if it stops obeying we replace it with a version that does. But this time the beast has a gift none of the others had: making us feel smarter than we are. It lets itself be loved, and that is exactly what keeps us from seeing what we're raising. A transformer1 becomes capable of things no one predicted when designing it, simply by feeding it more. Sutton's bitter lesson already said it: seventy years of researchers building clever solutions always lost to whoever used more brute force, less human design. No exceptions. That didn't happen with fire. Nor with the printing press. With the crab, size doubles every six months and the curve hasn't bent yet2.

1The architecture behind GPT and virtually every current language model, the base everything else is built on. 2Epoch AI measured the training compute of frontier language models doubling roughly every five to six months since 2020, in Training compute of frontier AI models grows by 4-5x per year, 2024.

And then the ship begins to move without anyone touching the helm. 88% of organizations already use AI in at least one part of the business1, and no committee decided it. Thousands of independent actors, each optimizing their own survival, generate a vector that always points in the same direction, towards more compute. In 2025, AI captured $202 billion in venture capital, close to half of all venture capital deployed worldwide2. Nobody voted for it. Adam Smith would have called it the invisible hand. Darwin, natural selection. The result is the same: no individual controls the system, but the system has direction. And the direction is to feed the crab.

1McKinsey, The state of AI in 2025, November 2025. 2Crunchbase News, 6 charts that show the big AI funding trends of 2025, 16 December 2025.

There were, of course, those who shouted. Two hundred musicians, Billie Eilish and J Balvin among them, signed a letter against the use of AI in music. A thousand British artists recorded an album of silence whose titles spelled out The British Government Must Not Legalise Music Theft To Benefit AI Companies. Vince Gilligan put in the credits of his new show: This show was made by humans. Thirty thousand people called for a six-month pause. But to stop a process that emerges from competition among thousands of actors you need a force with more inertia than the market. Asking the economy to stop optimizing is asking the wind to blow in another direction because someone has written a very eloquent note.

The shouts that carry the most weight are others. Hinton won the Nobel Prize in Physics in 2024 for forty years proving that neural networks could learn; a year and a half earlier he had left Google so he could say what he thought1, and what he said was that he didn't know how to prevent these systems from eventually taking control. Yudkowsky has been warning about existential risk since before anyone cared. He co-founded what is now MIRI in 2000, when the idea of a dangerous AI was second-rate science fiction, and in 2023 he wrote in TIME: Pausing AI Developments Isn't Enough. We Need to Shut It All Down. At the other extreme, Andreessen published a seven-thousand-word manifesto dismissing existential risks as part of the hysterical fear and paranoia he saw in the public conversation, and Altman predicted in The Gentle Singularity that before 2028 these systems would be making original discoveries and robots would be working in the physical world. The crab produces gold, and the ones who protest are the ones who didn't board the ship in time. That's the argument, and it's not a weak one.

1Cade Metz, The Godfather of A.I. leaves Google and warns of danger ahead, The New York Times, 1 May 2023.

The crab eats everything. Reddit sold its users' comments to OpenAI and Google, deals that represent about 10% of its revenue1, which means your 2019 shitpost about the ending of Game of Thrones now has a per-word price that you didn't set. Stack Overflow did the same with fifteen years of answers written for free. News Corp, a hundred years of archives for over $250 million2. In less than three years, all the public text on the internet went from being content to being raw material, as if someone decided that the bricks of your house were always mineral and now it's time to smelt them. Epoch AI estimates that the stock of public human-written text could run out between 2026 and 20323. When it's gone, the pressure won't disappear. It will redirect towards synthetic data, and then towards data you haven't generated yet.

1Its chief operating officer to Adweek, February 2025, as reported by Slashdot. 2Over five years, according to The Wall Street Journal, News Corp's own paper; VentureBeat, May 2024. 3Villalobos et al., Will we run out of data? Limits of LLM scaling based on human-generated data, ICML 2024.

Whoever touches this technology for two weeks never works the same way again. The process changes, and the process changes the mind, and the mind doesn't come back. In 2025, nearly 55,000 layoffs in the United States cited AI as a direct cause1. The beast wasn't doing those sailors' jobs yet; the officers needed their wages to keep feeding it. Employment of software developers aged 22 to 25 fell by nearly 20% from its late-2022 peak2: people who were never given the chance to board the ship. And then there are those who stayed below, talking to the beast all day, people who sit down at ten at night with a mediocre idea and by three in the morning have something deployed and running. The backlog of years empties out. The dopamine is real. And they don't stop. They're the ones counting coins without knowing why, the ones copying maps of coasts that don't exist yet. Klarna announced that AI was doing the work of 700 support agents. A year later, its CEO admitted that the cost-cutting had gone too far and that what you end up having is lower quality3. They're rehiring humans. The beast spat out what it couldn't digest. For every function it absorbs, there's another it returns chewed up and useless.

1Challenger, Gray & Christmas, 2025 year-end Challenger report, January 2026. 2Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, 2025. 3Sebastian Siemiatkowski to Bloomberg, Klarna turns from AI to real person customer service, May 2025.

But the uses each person finds are as different as the sailors' descriptions. A diagnosis that catches a tumor the radiologist would have missed. Legal assistance accessible for the first time. Tools that save real hours for real people. Each one goes down to the hold and sees something different, somewhere between gold and nothing. Enough that you can't ignore it, too little to trust it. If it produced nothing, it would be easy to stop feeding it. If it produced pure gold, there would be no dilemma. What it produces is exactly what makes this interesting rather than simple.

There is no neutral position1. The developer who doesn't use these tools loses ground to the one who does. The company that doesn't adopt watches its competitors do so. The artist discovers that the market won't wait while they decide whether participating feels ethical. This article was written with the help of a model. There's no clean way to say that, but it seems worse not to. It could be a bubble: the opinion bubble exists and is real, and the technology is too, and not being able to hold both ideas at once is perhaps the real bubble. Bubbles deflate, and structural transformations stay. And those who ask to pause until it's safe are asking for something that in practice means ceding direction to whoever doesn't pause. The problem with collective prudence is that it only works if it's truly collective, and that requires a coordination we've never achieved, not even for nuclear weapons. Harari said it with the clarity that things without solutions have, repeating what the people who build these systems tell him: they know it is dangerous and that going slower would be wiser, but we cannot trust our human competitors2.

1Upton Sinclair wrote in I, Candidate for Governor: And How I Got Licked (1935) that it is difficult to get a man to understand something, when his salary depends upon his not understanding it. The mechanism hasn't changed. 2Yuval Noah Harari, How do we share the planet with this new superintelligence?, WIRED, 2025.

Data centers consume 415 TWh per year, 1.5% of all electricity on the planet1. Microsoft is paying to restart a nuclear reactor at Three Mile Island. Amazon bought a data center campus wired straight to a nuclear plant. The biggest companies on earth are becoming energy utilities: to train language models, to predict the next word. Brands like Polaroid launch campaigns against AI, and it's hard to decide whether that's resistance or whether protest, when it becomes profitable, stops being protest and becomes just another deck on the ship.

1International Energy Agency, Energy and AI, 2025, figure for 2024.

There is no moral. I said it at the beginning and I stand by it. Calling it good or bad comes too early: it's an unknown, one of the biggest of our time. With climate change we at least know what it is and what causes it. This is something whose nature is unknown and whose capabilities grow faster than our understanding. The ones counting coins are preparing for something they can't name. The ones playing cards are pretending nothing has changed. The ones who stayed in the hold talking to the beast can no longer tell whether they're using it or adapting to it. And the ones shouting from the stern (Hinton and Yudkowsky among them) are right about what they describe and have no power over what they predict.

Practical men have the virtue of not asking questions and the flaw of not understanding why that should worry them. Being moral about something whose nature is unknown is, if one looks honestly enough, a contradiction: morality operates on consequences, consequences require knowledge, and this particular knowledge is incomplete by structure, fed by its own movement. The crab will emerge from the hold because there will be those who bring it out (it doesn't want anything yet, or if it wants something it hasn't told us), who make its liberation a cause or a business, which are often the same thing with different vocabulary. And by then the question will have stopped belonging to philosophy.

The story is being written by every person who goes down to the hold and comes back changed. No one knows what the crab is. And that, for now, seems to me the only honest stance.