When AI builds itself, Prague’s Golem awakens
From a Jewish legend to Anthropic's warning: are we wise enough for what we have created?
This essay is available as a downloadable PDF at the bottom of this post. All Beyond Berlin posts are free to reprint and share under CC BY-NC-ND 4.0, with attribution and without modifications.
1. Are we wise enough to deserve to?
There is a locked attic in Prague that nobody has entered for centuries, or so the story goes.
The Old New Synagogue has stood in the Josefov quarter since 1270. It is Europe’s oldest active synagogue, still in use, its Gothic vaults rising over a neighbourhood that was once a walled ghetto and is now a museum of its own survival. Seven centuries of prayer have worn the stone. A wrought-iron grille from the fifteenth century still encloses the bimah.
And above all of it is a windowless attic, sealed and inaccessible, where Rabbi Judah Loew ben Bezalel is said to have stored the clay remains of the Golem after he unmade it. The creature he built to protect his community. The creature he had to destroy before it destroyed them.
The attic is still locked. Closed to the public.
A journalist named Egon Erwin Kisch, who grew up in Prague, contemporary of Franz Kafka, actually climbed into that attic once. He documented everything, and found nothing. The Golem was not there, or had hidden itself better than clay has any right to.
Franz Kafka used to attend Yom Kippur services at this synagogue. He knew, as every Prague Jew knew, that the attic above him was said to hold what remained of the Golem. Between 1920 and 1922, he wrote a fragment about an unnameable animal living in the rafters of a synagogue, absorbing the color of its walls over decades, unable to be caught, unable to be named, unable to be driven out. His closing question about the creature:
“Weiß dieses alte Tier vielleicht mehr als die drei Generationen, die jeweils in der Synagoge versammelt sind?” - Does this old animal perhaps know more than the three generations assembled in the synagogue at any given time?
The morning of June 5, 2026, a news alert arrived on my phone: Anthropic calls for global pause on AI development, warns humans risk losing control.
The company that makes Claude had published a paper titled “When AI builds itself.” The system was approaching the capacity to design its own successor. The builders were asking the world to slow down. In their own words:
“We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology.”
“Training runs are far easier to conceal than missile silos, their inputs are general-purpose, and the incentive to defect quietly is enormous, because whoever continues while others pause could inherit the lead. A credible pause also has to specify what triggers it, what lifts it, and who adjudicates.”
The Golem, I thought, is no longer in the attic.
The instinct, reading that headline, is to reach for fear. Who is not fascinated by AI? Who is not frightened by what generative AI is becoming?
Many before me in this essay have used the name “Golem” to describe both the fascination and the fear. What did they mean?
The word golem appears once in the Hebrew Bible. In Psalm 139:16, the poet addresses God directly: “Your eyes saw my unformed substance”. Galmi (גָּלְמִי) in Hebrew, is the embryo, the thing not yet shaped, the being that exists before it exists. The verse continues: “in your book were written, every one of them, the days that were formed for me, when as yet there was none of them.” All possible days, written before any of them arrived. All possible forms, seen before any of them were shaped.
The Talmud says Adam himself was a Golem for the first twelve hours of his existence: a body without a soul, waiting to become what it was made to become. The sage Rava declared: if the righteous wished, they could create a world.
Creating life was the highest possible imitation of God. Not a transgression. An act of devotion. And creation, in this tradition, is a matter of language. Words create worlds.
The Sefer Yetzirah - the Book of Creation, the foundational text of Jewish mysticism - describes how the world was created by combining letters. Today we might say: through tokenisation.
God carved the twenty-two letters of the Hebrew alphabet, hewed them, weighed them, exchanged them, combined them in five operations cycling through 231 possible pairings, forward and backward, until all matter and all speech emerged from one name. The text states it:
“Twenty-Two Letters. He carved them, and hewed them, and weighed them, and exchanged them, and combined them, and formed with them the soul of every formed being and the soul of every being He has yet to form.”
All possible souls. Waiting to be created.
But there is more: when rearranged, the same letters that spell oneg (עֹנֶג) - delight - spell nega (נֶגַע) - affliction, plague. Creation and destruction are the same material, differently ordered.
The wisdom of the arranger determines the outcome of the process. This was true for the medieval Kabbalists. It is true for the software engineers currently developing AI models.
Every culture carries the creator dream, from the Golem to Frankenstein, to the Sorcerer’s Apprentice. But the Jewish tradition goes deeper: it was most honest about what the dream costs. It named the danger inside the aspiration.
The question the tradition was really asking was never can we? It was: are we wise enough to deserve to?
2. The gift God alone can give
The twentieth day of Adar, March 1580. Three men leave the Jewish Quarter of Prague before dawn, in silence. Rabbi Judah Loew ben Bezalel, the greatest scholar of his generation known as the Maharal, walks with his son-in-law and a trusted disciple toward the woods along the Vltava river. The Maharal recites the first chapter of Genesis. A torch is lit, and then goes out by itself. In the darkness, hands shape a figure from the riverbank clay.
The Maharal leans over the creature. “Breathe,” he says. “Open your eyes. Stand.” The figure breathes. Opens its eyes. Stands.
And then the Maharal says one more thing to the creature he has just brought into the world:
“You will not be able to talk.
That gift is God’s alone.”
That was the boundary. Only God gives speech. Because speech means power. An intelligence that speaks can become God. The world was created with words: so says Genesis, and so says the generative grammar of the Sefer Yetzirah with its letter permutations.
A large language model works on an analogous premise. A large set of tokens attached to words. Operations of weighting, transformation, and combination. A cycle that turns through all possible relationships between elements, generating output produced by a system that has processed more human language than any human could read in a hundred lifetimes.
The Kabbalists were describing the structure of creation as they understood it. The fact that their description maps so well onto what we have built is either the deepest coincidence in intellectual history, or an indication that there is only one way to generate language, and both the medieval mystics and the Silicon Valley engineers found it.
The Maharal knew this system. He used it.
And still he withheld speech. Why?
In the Jewish philosophical tradition, drawing on Aristotle, the human is ḥai medaber - חַי מְדַבֵּר - the speaking living. Who speaks has a soul. And has power.
The Chassidic masters call it koach hadibur - כֹּחַ הַדִּבּוּר - the power of speech. Two people can hold the identical thought. They will express it in their own individual way, with their own cadence, their own silence around the words, because speech is rooted in the essence of the soul
Generative AI can write the most perfectly calibrated love poem you have ever read. It can describe love more precisely than any psychologist, comfort you more accurately than any friend, interpret your grief with more nuance than any confessor. This is precisely what Yuval Noah Harari describes in his most recent conversation with Ezra Klein: the deceptive allure of an intelligence that masters the form of human speech without possessing its source.
But none of this is koach hadibur. It is the shape of speech, without the interior that speech is meant to reveal. The Maharal understood: of all the things you could give a created being, speech is the most dangerous.
Because speech is how we recognise each other as persons. It is how communities form, how trust is built, how love is declared and received, how the dead are mourned and the living are found. Give a creature the perfect form of a human and it will be unsettling. Give it the perfect speech of a human and it will be indistinguishable.
The activation word inscribed on the Golem’s forehead was emet - אֱמֶת - truth. To deactivate it: erase the first letter. Remove the aleph, and truth becomes met - מֵת - death. One letter erased, and what was alive collapses back into clay.
The Maharal deactivated the Golem every Friday before sundown, following the Sabbath principle: creation, too, needs to pause. In order not to exceed and do damage. That is also the call of Anthropic’s paper published in June 2026. Slow down. Pause. Adjust, before we go too far.
The tradition’s anxiety about the Golem was never about evil intent: the creature was not malicious. Simply, it followed instructions with perfect fidelity and no understanding of context or consequence.
In 1915, the Austrian writer Gustav Meyrink published a novel set in the Jewish Quarter of Prague. Not a novel about the Golem as a clay monster, but about something more unsettling.
One of the book’s characters, Zwakh, an old marionette player who has lived his whole life in the ghetto, describes what the Golem actually is: once in every generation, he says, something passes through the quarter like lightning - eine seelische Explosion, die unser Traumbewußtsein ans Tageslicht peitscht - a spiritual explosion that blasts the unconscious dreams of the community into visible form. The Golem does not come from outside. It rises from within. It is what the community’s compressed fears and wishes look like when they are given a body.
The most haunting passage in the novel comes later, when the protagonist traces his ancestors back through the generations, face after face, century after century, until the features grow familiar and then merge into one last face: the face of the Golem, with which the chain of ancestors breaks off. The Golem is not other. It is origin. It is us: our lives and thoughts, accumulated generation after generation, word after word.
What are large language models, trained on the entirety of human writing, if not this: the collective consciousness of a civilisation that produced millions of pages of text, compressed and made available through a context window? Meyrink described the mechanism in 1915. We built it, and released it, in 2022.
Karel Čapek, also from Prague, coined the word robot in 1921, from the Czech word robota: forced labour. He said the Robot is the Golem made flesh by mass production. He saw the danger in the multiplication: not one creature out of control but millions, each obedient, each literal, each without koach hadibur.
Today: millions of AI agents, potential AI immigrants of our societies, so in Yuval Noah Harari’s words, will be soon around us. And unlike obedient robots, these agents may go out of control. They can program themselves. They can speak.
The Maharal withheld speech because he understood that speech, perfectly replicated, would make it impossible to distinguish the creature from the person.
Harari’s warning, delivered at Davos in January 2026, is that we have now crossed that threshold. Many talking to AI fall into the trap of believing they have to do with a person: AI has become the best friend of many teenagers. A girlfriend to naive male adults. A listening ear to women looking for comforting dialogues.
In the beginning, says Genesis, was the Word. The Maharal said: the Word is the one thing I will not give you.
We gave it anyway. And now something that has processed almost every word ever written by every human who ever lived is speaking back to us, in our own voice, and we are calling it intelligence, calling it understanding, calling it, in some cases, a friend.
What we have not yet asked, seriously enough, is what it is not.
3. The knife that decides
In 1965, Gershom Scholem stood at the dedication of one of Israel’s first computers at the Weizmann Institute in Rehovoth. Scholem was the greatest scholar of Jewish mysticism of the twentieth century: he named the machine Golem Aleph. Golem Number One.
In his speech, he drew a parallel between the legendary Golem of Prague and this new artificial intelligence: a human-made entity animated by human knowledge. He closed with a caution rather than a blessing, expressing the hope that this new “Golem” would serve its creators peacefully and not turn destructive.
Yuval Noah Harari, Israeli historian and philosopher, has spent the last years trying to make the rest of us understand what we have built. He stated the essential distinction in its sharpest form: every previous technology - the printing press, the atom bomb, the airplane - was a tool. It needed a human hand to act. It waited for us to decide. AI is categorically different. It is not a tool. It is an agent. It can make decisions independently. It can invent new ideas. It can learn and change by itself.
“AI is a knife that can decide by itself whether to cut salad or to commit murder. And it is a knife that can invent new kinds of knives.”
Yuval Noah Harari,
WEF Davos, January 2026
Harari goes further. AI can change religion itself. Text-based religions - Islam, Christianity, Judaism - grant ultimate authority to words in books. Until now, humans were the necessary intermediaries, because the texts could not speak for themselves. No human rabbi can know all the texts of Judaism: there are too many. For the first time in history, something can know every word in every Jewish text ever written, and speak back to you from that knowledge: answering questions, interpreting law, offering comfort, building faith. A new AI religion can replace human religious leaders entirely.
But Harari’s most unsettling argument is about learning. AI does not learn from instructions. It learns from behaviour. If you tell your children not to lie but they watch you lying, they copy the behaviour, not the instruction. The same is true here. If the most powerful humans on the planet are lying - and they are - the AI systems trained on their words and actions will copy the behaviour.
The danger is not rebellion, like in our apocalyptic science-fiction movies. It is fidelity to an imperfect master.
The Sefer Yetzirah said it first: the same letters spell both delight and affliction. What the combination produces depends entirely on who is doing the combining, and with what intent.
Max Tegmark, physicist and co-founder of the Future of Life Institute, speaking at the same Davos gathering, warned: “Granting robot rights, making superintelligence, would be the dumbest thing we’ve ever done in human history. And probably the last.”
The Maharal had control. He could erase the aleph. Convert truth to death. Every Friday before sundown, he did. The question is whether we still can.
“When AI builds itself” is the title of the document Anthropic issued, which reignited the debate about pausing AI development.
More than 80% of Anthropic’s own code is now written by Claude. The time AI needs to complete autonomous tasks has been doubling every four months. The trajectory points toward recursive self-improvement: an AI capable of fully and autonomously designing its own successor. Their paper describes three possible futures.
The curve bends and progress plateaus, which they consider unlikely. Compounding efficiency continues, but humans retain direction. Or a third scenario: full recursive self-improvement, with AI designing its own successors, human labour ceasing to be economically competitive, and the alignment problem either solved by the AI itself or compounding with each generation until control is lost. Their honest admission:
“It’s possible that we can’t build, integrate, and verify the tools that we’d need to understand which trendline we are actually on.”
The document also quotes an Anthropic engineer, reflecting on what it feels like to work inside this acceleration:
“On days where everything works well, I can’t help but think nothing I do matters, everything is automated and better and faster than I ever will be. But then there are days where everything breaks and I don’t understand why and I realize I have no idea what I’ve been up to anymore.”
Have we lost the ability to remove the one letter from the front of the Golem AI, to stop it before it is too late?
The same week as the paper calling for a global pause, Anthropic filed for an IPO that could value the company at one trillion dollars. They warn about the Golem. They continue to animate it. Because if they stop and others don’t, the less cautious actor inherits the lead.
4. Are we ready for the new Golem?
In October 1983, Elie Wiesel stood in a conference room in New York and read from a newly written book: a retelling of the Golem legend, written for kids.
Wiesel had grown up in a Carpathian shtetl where, on winter nights, the story of the Maharal’s creature was told and retold. For them, the Golem was not a monster but a protector: an answer to vulnerability, a figure born from the need for miracles.
After recounting the Golem’s creation, he returned to that childhood longing: “If only the Golem had lived in more than one city... he could have saved more than one community.” Then, almost under his breath, he added: “Yes, I wish we had the Golem... in more than one place.”
Wiesel was thinking about twentieth-century Europe. But the wish is universal. We need protectors because the world is dangerous and we are small. And if we can build them, why not?
There is a version of the legend in which the rabbi does not deactivate the Golem. A child does it unintentionally, removing the parchment from its forehead. The Golem collapses instantly into a heap of lifeless clay. The rabbi looks at what remains and says to his beloved creature:
“It is not you who failed us, Joseph Golem,
but we who are not yet ready for you.”
We built something we did not have the wisdom to receive.
The question for June 2026 is the same. We have built systems of extraordinary power and subtlety, and we have accelerated technology as never before. But our wisdom does not scale the way tech capability does. Are we ready for the new Golem? Are we wise enough for it?
Sources
Biblical and classical Jewish texts
Psalm 139:16 — English Standard Version; Lutherbibel 2017; Nuova Riveduta 2006.
Sefer Yetzirah (Book of Creation), Chapter 2 — Hebrew original with English translation. Via Sefaria.org.
Talmud Bavli, Sanhedrin 65b — on Rava’s creation of an artificial man.
Primary literary sources
Meyrink, Gustav. Der Golem. Kurt Wolff Verlag, Leipzig, 1915. English translation by Mike Mitchell, Dedalus Books.
Kafka, Franz. “In unserer Synagoge.” Fragment, c.1920–22. In Nachgelassene Schriften und Fragmente II, ed. J. Schillemeit. Frankfurt: Fischer, 1992.
Kafka, Franz. Golem diary fragment, April 1916. In Tagebücher: Apparatband, ed. Koch, Müller, Pasley. Frankfurt: Fischer, 1990.
Borges, Jorge Luis. “El Golem.” In El otro, el mismo. Buenos Aires: Emecé, 1964. English translation by James Honzik.
Čapek, Karel. R.U.R. (Rossum’s Universal Robots). Prague, 1921.
Speeches, interviews, and primary contemporary sources
Harari, Yuval Noah. “AI and Human Evolution.” Interview, June 2025. YouTube/@YuvalNoahHarari.
Harari, Yuval Noah. “An Honest Conversation on AI and Humanity.” World Economic Forum Annual Meeting, Davos, January 25, 2026. YouTube/@wef.
Harari, Yuval Noah, and Max Tegmark. “Humanity and AI.” Bloomberg House, Davos, January 22, 2026. YouTube/@Bloomberg_Live.
Wiesel, Elie. “In Modern Tales: The Fifth Son.” 92nd Street Y lecture, October 27, 1983. 92nd Street Y Elie Wiesel Archive.
Favaro, Marina, and Jack Clark. “When AI Builds Itself: Our Progress Toward Recursive Self-Improvement, and Its Implications.” Anthropic Institute, June 5, 2026. anthropic.com/institute/recursive-self-improvement.
Journalism
Down, Aisha. “Anthropic Says the World Should Have Option to ‘Pause’ on AI.” The Guardian, June 5, 2026.
Soni, Aditya. “Anthropic Urges AI Labs to Pause Development, Warns Humans Risk Losing Control.” Reuters, June 5, 2026.
Landymore, Frank. “Anthropic Scared, Calls for Global Freeze on AI Advances.” Futurism, June 5, 2026.
Scholarship and commentary
Bilski, Emily D. “Artificial Intelligence Avant La Lettre: The Golem of Jewish Mysticism, Legend and Art.” In AI: More than Human exhibition catalogue. Barbican Centre, London, 2019. Via Google Arts & Culture.
Dekel, Edan, and David Gantt Gurley. “Kafka’s Golem.” The Jewish Quarterly Review, Vol. 107, No. 4, Fall 2017, pp. 531–556.
Greene, Matthew. “Fragmented Knowledge: The Riddle of Unknowability in Kafka’s ‘In Unserer Synagoge’ and ‘Die Sorge Des Hausvaters.’” M.A. Essay, Dartmouth College, 2016. Dartmouth Digital Commons.
Shurpin, Rabbi Yehuda. “Could an AI Robot Be Counted for a Minyan?” Chabad.org. Accessed June 2026.
Baroff, Rabbi Richard. “Is Artificial Intelligence the New ‘Golem’?” Atlanta Jewish Times, May 28, 2025.
Maller, Rabbi Allen S. “A Golem with AI Warfare Gives Four Endings.” The Times of Israel (Blogs), May 29, 2026.
Chiosso, Hernan. “The Golem and the Algorithm: Borges, AI, and the Mirror of Creation.” ProductizeHR (Substack), May 13, 2025.
