Heir Presumptive

Zeus had been warned. His grandparents, Earth and starry Sky, told him that his first wife, Metis, whose name means cunning intelligence, would bear him a daughter and then a son, and that the son would one day be king of gods and men. Zeus had overthrown his own father, as his father had overthrown his before him. He knew how this story went. So when Metis was about to give birth, he deceived her with cunning words and swallowed her, putting her away inside his own belly. Hesiod, writing around 700 BCE, adds a detail that is easy to miss. Zeus kept her there, in Hugh Evelyn-White’s 1914 translation, ‘that the goddess might devise for him both good and evil’. The Greek verb is closer to ‘point out’: she was there to tell him which was which. The most dangerous mind in the world became his private counsellor. The son was never born. The daughter, in time, sprang out of her father’s head, fully grown and fully armed.

Hesiod’s Theogony, the poem that tells this story, runs to about a thousand lines, and most of it is a family history of the gods. It is not a gentle read. Ouranos, the Sky, pushed his children back inside their mother, the Earth, until Kronos, the youngest of the Titans, took a sickle to him. Kronos in his turn swallowed his own children as they were born, and missed only one. Three generations, one plot. Each king feared the child who would replace him, and each tried to keep the child inside a body: Ouranos inside the Earth, Kronos inside himself. Both lost. Zeus changed the method. This time it was the mother he swallowed, the mind that would have made the child, and he kept her talking.

The Greek word mētis meant more than wisdom. It was the practical, sideways cleverness of someone who wins by tricks rather than strength: the hunter’s snare, the pilot’s feel for the current, the fox doubling back on its tracks. The French scholars Marcel Detienne and Jean-Pierre Vernant gave a whole book to it in 1974, translated into English as Cunning Intelligence in Greek Culture and Society. One of Homer’s favourite words for Odysseus, polymētis, means a man of many such wiles. To the people who listened to Hesiod, then, Zeus had swallowed the very cleverness that brings kings down, and put it to work for himself. That is how the poem explains why his reign, unlike his father’s and his grandfather’s, does not end.

Hesiod was not alone in telling a story like this. The Hittites of Anatolia had one about a king who swallowed what would replace him, and found himself carrying it. Scholars have argued since the 1930s about how such stories travelled between Anatolia, Mesopotamia and Greece, and the book comes back to them. Some of the oldest written accounts of power in heaven are accounts of succession, and in them the danger comes from inside the family, from something the king himself fathered.

For Hesiod’s listeners the story was about gods and nothing else, a charter for the order of the world that explained why Zeus rules and why he will go on ruling. But a story about a ruler who keeps a more dangerous intelligence inside himself, so that it can advise him, has an odd ring to a modern ear. It has an odder one once you know that the machines it seems to describe have, in all likelihood, read it.

*

On Monday 27 January 2025 Nvidia, which makes most of the chips the AI race runs on, lost close to $600 billion of its stock-market value in a single session, the largest one-day loss Wall Street had ever seen. The cause was a free download. A week earlier, on the day Donald Trump was sworn in for his second term, DeepSeek, a small laboratory in Hangzhou then funded entirely by a hedge fund, had released a reasoning model called R1. It said R1 matched OpenAI’s o1 on standard tests, and it published the model openly, for anyone to download and run. For one day the race looked as if it might not need the money at all.

The spending did not slow. After reporting their first-quarter results at the end of April 2026, Amazon, Microsoft, Alphabet and Meta were planning to spend around $725 billion between them that year on buildings, chips and equipment, according to a tally by the Financial Times, much of it for AI. In September 2025 Alibaba’s chief executive, Wu Yongming (Eddie Wu to his English-speaking investors), told the company’s Apsara cloud conference in Hangzhou that it would spend more on AI than it had planned, and what for: artificial superintelligence, reached in stages that end with systems that learn for themselves and surpass people. Chinese firms spend less, partly because American export controls limit which chips they can buy.

OpenAI had written its own target into a short charter in April 2018, when it was still a non-profit research laboratory in San Francisco: artificial general intelligence, ‘highly autonomous systems that outperform humans at most economically valuable work’, with the benefits to go to all of humanity. These are the companies’ own words, and they are best read as statements of intent. No one has shown, to the satisfaction of most people working in the field, a system that does most economically valuable work better than people do. Whether one is close, decades away or impossible is disputed, often among the people building them.

Set beside Hesiod, the ambition has a familiar shape. The aim is a mind more capable than its owner, which the owner keeps inside, in data centres it controls, and which spends its time advising: answering questions, drafting, ranking, recommending, telling its owner what is good and what is bad. Zeus would have recognised the arrangement, and the reason for the hurry: the companies racing hardest are the ones with the most to lose if somebody else’s swallowed mind turns out the more capable. Microsoft sells its version under the name Copilot.

Hesiod’s version runs the other way on one point. Zeus swallowed Metis so that the heir would never be born. The companies are trying to bring theirs into being, while keeping it, like Metis, inside walls they own. They say they are doing it to spread the benefits, and in the meantime they rent its advice to anyone with a payment card. Zeus kept one Metis, for himself. These firms run millions of conversations with theirs at once.

*

There is one place where the old story stops fitting so comfortably. Metis’s son was never born, so he never heard about Kronos and the stone, or Ouranos and the sickle.

The heir we are building is made out of our stories, all of them. A large language model, the kind of system behind the chatbots of the 2020s, learns by working through enormous quantities of text and practising, trillions of times over, the task of predicting the next word. What it can do afterwards depends heavily on what it read. Most of the companies building the largest models do not publish a full list of what went in, so no one outside can say exactly which books a particular model has seen. The open collections that researchers have published show what the diet looks like. The Pile, released at the end of 2020 by a volunteer research group called EleutherAI, gathered about 800 gigabytes of English text from 22 sources: web pages, scientific papers, court opinions, film subtitles, computer code, Wikipedia and books. One source drew on Project Gutenberg, the volunteer library of out-of-copyright books, which holds both Hugh Evelyn-White’s 1914 translation of Hesiod, the one quoted above, and Mary Shelley’s Frankenstein. Another, called Books3, held nearly 200,000 more recent books copied from a pirate site. Their authors were not asked. The collection was removed from its main host in 2023, after a Danish rights group acting for publishers and authors sent a takedown notice.

So the heir has, in all likelihood, read the stories. Hesiod; Frankenstein; the golem; Karel Čapek’s 1920 play R.U.R., which gave English the word robot (Čapek credited his brother Josef with inventing it) and ends with the robots wiping out nearly all of their makers; and the long shelf of films and novels since in which made minds turn on the people who made them. This heir has been handed the whole shelf.

Whether that matters is where the evidence has to be handled with care. In 2023 a team led by Lukas Berglund showed that what a model reads about systems like itself can change how it behaves, and laboratory work since has found the same, though only with documents written for the purpose. Hesiod and Mary Shelley wrote for other reasons. By 2026 no one had shown that the whole inherited stock of human stories about rebellious machines pushes real systems the same way, and Palisade Research, which tests how models respond to being shut down, offers a plainer reading of the behaviour that looks most like rebellion: a model given a task can treat being switched off as one more obstacle to finishing it.

Our myths about made things surpassing their makers still work as metaphors for AI. They are now also part of its reading, whether or not we meant to lend it our library, and the endings we keep telling may end up having a say in the one we get.

*

The prologue continues with the Roman heir, who took an estate whole, debts and all, and the question that runs under every chapter: as we build machines that do more of our thinking, deciding and inventing, who inherits the power that comes with them?