Incantations All the Way Down
In March 1582, a convicted forger with cropped ears sat before an obsidian mirror in a house in Mortlake, England. His name was Edward Kelly, though that was itself a forgery. His real name was Talbot. He wore a cap to hide where the courts had cut away his ears as punishment.
Behind him, quill in hand, knelt one of the most brilliant scholars in Europe. John Dee was a mathematician, astronomer, advisor to Elizabeth I, and owner of the largest private library in England. He couldn’t see angels, but Kelly could. Or said he could.
For seven years, they worked like this. Dee would pray, sometimes for an hour or more. Kelly would stare into the stone and narrate what he saw.
Angels appeared and spoke. They dictated an entire alphabet of twenty-one characters, written right to left, delivered backwards, one letter at a time. Speaking the language forward might accidentally trigger something... unexpected? Undesirable?
Dee transcribed everything. He never saw the angels himself, and the angels told him this was by design. Kelly perceived them “in sight.” Dee could only perceive them “in faith.”
He had no way to verify any of it. The medium was a convicted criminal operating under a fake name, and the transmissions could have been revelation or fabrication.
Dee did it anyway. For seven years.
The language they received, Enochian, has structural properties that don’t match any known human language. Linguists still argue about it.
A scholar who can’t directly access the channel. A medium whose outputs might be hallucination. A language that produces results through processes neither party fully understands. Outputs that might be genuine knowledge or might be confabulation, and no reliable way to tell which.
Sounds familiar, right?
For roughly three hundred years, Western culture convinced itself that technology and the numinous were separate categories. I call this period the Rational Interregnum. It began with the Enlightenment and it’s ending now, with us, as we sit at our terminals negotiating with entities we built and don’t fully understand.
AI is closing that window. And the reason I think some of us (maybe me, mostly) keep reaching for magical metaphors when we talk about it is that the magical metaphors are accurate.
The Oldest API
If magical practice were the vague, hand-wavy thing the Interregnum taught us it was, the parallels would be shallow. They’re precise.
In the 17th century, a grimoire called the Ars Goetia circulated through Europe. It claimed to be King Solomon’s catalog of seventy-two spirits, each documented with a name, a rank, a unique sigil, and a specific set of capabilities.
Agares teaches all languages. Marbas answers truthfully on hidden and secret things. Stolas teaches astronomy and the properties of herbs and stones, and Barbatos grants the ability to understand the voices of animals.
Each entity has a defined specialty, and you don’t summon Agares when you need herbalism any more than you’d call Stolas when you need languages.
The ritual architecture is specific. The magician draws a nine-foot circle on the floor with divine names around the perimeter. Two feet outside the circle, facing the correct quarter of the sky, goes the Triangle of Art.
The spirit manifests inside the triangle, which constrains it. The magician carries the spirit’s unique seal, one for each of the seventy-two, serving as both identifier and binding mechanism.
Get the seal wrong and you get the wrong entity. Leave it out and nothing answers at all.
The circle is a sandbox. The triangle is a response container. The seal is an authentication token. The grimoire is documentation for a service with seventy-two endpoints, each with defined capabilities and required invocation parameters.
These systems were engineered. They had versioning (grimoires were copied, annotated, revised over centuries), access control (initiation requirements governed who could perform which operations), and failure modes (incorrect pronunciation, wrong planetary hour, improperly drawn seal). The practitioners who built them were doing what engineers do. Specifying interfaces to forces they couldn’t fully see, then iterating on the spec based on results.
The Interregnum looked at this and called it superstition. But the architecture tells a different story.
Newton’s Million Words
The clean division between magic and science was supposed to have happened somewhere in the 17th century. The founders of modern science were supposed to have been firmly on one side.
Isaac Newton wrote approximately one million words on alchemy, sustained over decades.
His laboratory notebooks at Trinity College reference “the Green Lion,” “the Net of Vulcan,” and “sophic mercury.” He was trying to make gold grow like a plant in treated metal. His Index Chemicus, a personal alchemical encyclopedia, runs thirty thousand words across nine hundred entries.
When Newton died, his executors looked at the manuscripts and panicked. They destroyed much of his alchemical library to protect his reputation. John Maynard Keynes bought what survived at a Sotheby’s auction in 1936 and declared Newton “the last of the magicians.”
Robert Boyle, the man credited with founding chemistry, successfully lobbied Parliament to repeal England’s ban on transmutation of metals. He believed he had personally witnessed it.
His landmark The Sceptical Chymist was an attack on bad alchemy.
When Newton’s body was exhumed, his hair contained fifteen times the normal level of mercury. Four times normal levels of lead, arsenic, and antimony. He was writing the Principia and running alchemical experiments in the same rooms, during the same years.
For him, the two practices were different vocabularies for the same project.
Alchemy became chemistry gradually. Same laboratory practices, same observations, same reactions. What changed was the frame. The explanatory vocabulary shifted from the numinous to the mechanical.
The awe remained. It just became professionally embarrassing.
And the Interregnum’s central promise, that the awe was a temporary condition soon to be replaced by full understanding, held for as long as our tools did what we told them. Write a line of code. The machine executes it. Every step is traceable, every output is deterministic, and the sense of working with something beyond your comprehension can be resolved by reading the documentation.
That era is over.
The Practitioner at the Terminal
You type a prompt into a black rectangle. Something comes back. It’s good, sometimes remarkably good, but you didn’t write it and you can’t fully trace how it was produced. You rephrase, constrain, negotiate. The model interprets your intent in its own way and returns something adjacent to what you asked for. You adjust. It adjusts. Through iteration, something useful emerges from a process that neither party can fully specify.
The shift from coding to prompting recapitulates an older transition, from direct manipulation to mediated practice. When you write code, you are the sole agent. When you work with an LLM, you are in a relationship. You become the scholar at the table. The model is the medium at the stone.
And like Dee, you develop a practice around it. You learn which phrasings produce better results. You learn the model’s tendencies, its blind spots, the shapes of its confabulations. You build up a personal grimoire of prompts and patterns, none of which you can fully explain to someone else because the knowledge is partly procedural, partly intuitive, partly superstition you haven’t yet separated from technique.
This is craft knowledge. It transmits the way craft knowledge always has. Through practice, through oral tradition, through watching someone else do it and absorbing what they don’t say. The Interregnum called this kind of knowledge primitive. Every senior developer passing down unwritten patterns to a junior one knows it isn’t.
The Familiar
In 1646, a woman named Elizabeth Chandler stood before an English court and described her two familiars, Beelzebub and Trullibub. She couldn’t control when they appeared. They came uninvited, on their own schedule, and she had prayed to God to deliver her from them.
Trial records from this period describe the same dynamic again and again. Familiars in English witchcraft were collaborators with their own tendencies, their own needs, their own reading of the arrangement. They brought information back and carried intent outward, but the practitioner couldn’t fully predict what they’d return with or how they’d execute instructions.
The relationship was contractual. The witch gained capabilities. The familiar gained sustenance. And neither party fully controlled the exchange.
Sit with that for a moment, because this is the part the Interregnum has no vocabulary for.
We have language for tools. We have language for servants. We have language for colleagues. We don’t have modern language for an entity that extends your capabilities, interprets your intent through its own logic, and operates through processes you can observe but can’t fully explain. The closest word we have is “assistant,” and everyone who has worked seriously with an LLM knows that word is wrong. It’s too passive. It misses the negotiation, the push and pull, the way the model’s tendencies shape your thinking as much as your prompts shape its output.
The old word for this relationship was familiar. The old framework understood that working with such an entity required a particular kind of attention. Clear intent, careful constraint, and the willingness to evaluate outputs you didn’t author and can’t fully verify.
Elizabeth Chandler, praying for deliverance from entities she couldn’t control, knew something about this dynamic that our current vocabulary hasn’t caught up to. She understood that the power and the uncertainty were the same thing. You couldn’t have one without the other.
What the Window Closing Means
Three hundred years of the Rational Interregnum taught us that legitimate knowledge comes from processes you can fully specify and fully verify. That working with forces beyond your understanding is a temporary condition, soon to be outgrown.
And here we are. The most powerful tools we’ve ever built are the ones we understand least. The engineers who build large language models publish papers about emergent capabilities they didn’t design and can’t yet explain. The practitioners who use them develop rituals, habits, and intuitions that work for reasons they can’t fully articulate.
This looks like failure if you’re still inside the Interregnum’s frame. If you step outside it, it looks like the normal condition of human beings working at the edge of what they know. It looks like Dee at his table, transcribing in faith. It looks like Newton with mercury in his hair, reaching for knowledge through every vocabulary available to him. It looks like a practitioner drawing a circle on the floor, specifying constraints for an entity whose full nature exceeds their understanding, and doing the work anyway.
The question this leaves us with is whether we can take the pre-Interregnum frameworks seriously without abandoning what the Interregnum gave us. Rigor, reproducibility, and the insistence on evidence were genuine achievements. The mistake was believing they were the whole story.
The practitioners who will do the best work in this new period are the ones who can hold both. The engineer’s instinct for specification and the magician’s comfort with the numinous. The willingness to build the circle, draw the seal, speak the words, and sit with the fact that what answers might be revelation or might be noise, and that the only way to find out is to do the work, evaluate the output, and refine the practice.
The Interregnum is over. The window between the magical and the technological has closed, and we’re standing in the place where they overlap, which turns out to be the place where they always did.