Somewhere in a studio in London, 1967, Jimi Hendrix tilts his Stratocaster toward a wall of Marshall amplifiers. The guitar picks up its own sound from the speakers, feeds it back through the pickups, and the signal begins to loop. Electrons chase themselves through copper and vacuum tubes. This is the complete physical description of what is happening: electromagnetic induction, amplification, resonance. Nothing here that couldn’t be drawn on a circuit diagram.

But what comes out of the speakers is the sound of a national anthem being taken apart and put back together as a question.

I keep thinking about this gap. An economist named John Cochrane recently submitted a paper on fiscal theory to an AI tool and received, in his own words, among the best critical feedback of his entire academic career. The tool found logical gaps, unstated assumptions, a regime distinction he hadn’t fully operationalized. His response was not gratitude but bewilderment: “I don’t know how you get here from ‘predict the next word.’” And he’s right. Next-word prediction is what the model does. The feedback is genuinely insightful. There is no obvious way to get from one to the other. It’s the same gap as Hendrix’s rig. The circuit diagram is accurate. The anthem is real. And between them, there is nothing — no step where volts become grief or where token probabilities become intellectual judgment. Someone put it perfectly: “‘Predict the next token’ is true but not explanatory. It’s like saying humans ‘fire neurons.’ Technically correct, explains nothing useful about the behavior you’re actually observing.”

This is not a dispute about facts. It’s a dispute about altitude. And it’s everywhere.

Herman Melville spent nineteen years as a customs inspector at the Port of New York. He weighed crates, checked manifests, stamped documents. This was his job description, and it was accurate: customs inspector, Pier 40, $4 a day. It was also, somehow, the wrong description of the person it described.

A man named Simon Berger picks up a hammer and strikes a pane of safety glass. He strikes it again, and again, hundreds of times, each blow placed with a precision that only looks like violence. When he steps back, there is a human face in the glass — composed entirely of fractures, held together by the thing that should have destroyed it. You could map every crack, diagram every force vector, and you’d have the physics right while the portrait remains inexplicable. The explanation for how each fracture forms tells you nothing about how a face appears. In 1830, Honore de Balzac documented a similar absurdity from the other direction: he was drinking fifty cups of coffee a day and writing about the experience with the precision of a pharmacologist and the fervor of a convert. He charted the stages of caffeination the way a scientist charts phases of a reaction — the initial brightness, the flood of ideas, the period where “sparks shoot all the way to the brain,” and then the long descent into what he called “the relaxed, meandering, simple-minded and cryptogamous life of the retired bourgeoisie.” What the coffee produced, over years of this excess, was La Comedie humaine: ninety-one novels, two thousand named characters, an entire civilization in prose. The mechanism was caffeine. The output was literature. A biochemist could trace every molecule to its receptor and not come one sentence closer to explaining the work. Breaking glass doesn’t explain a face. Blocking adenosine doesn’t explain a novel. The descriptions are at the wrong level.

This is not mysticism. I am not arguing that something magical happens between the hammer blow and the portrait, or between the feedback loop and the anthem, or between the token prediction and the insight. I am arguing something more uncomfortable: that correct explanations, at a certain resolution, stop being explanations at all.

And sometimes this stops being merely philosophical.

Someone recently opened info.cern.ch — the very first website, still running on its original URL. What they found was a page of stark hyperlinks, text on a white background, a directory of information about a project called the World Wide Web. One commenter remembered running a script in those early days to count every website in existence: 324. That was the entire web. You could have described every protocol, every packet, every line of HTML across all 324 sites and you would not have predicted what was coming. Not because the technology was surprising, but because the thing the technology would become existed at a level the technology itself couldn’t describe.

Now turn that inside out. Researchers recently demonstrated that language models can deanonymize people from their public posts — not by finding a smoking gun, but by assembling a mosaic of the mundane. The city you mention. The sports teams you follow. The hours you keep. The vocabulary you reach for. Each fragment is innocent; each one, alone, is at the wrong level to identify anyone. But stitched together, they become a fingerprint that was never meant to exist. One commenter described doing this work by hand in the early 2000s, tracking stock scammers on Yahoo Finance — mapping their cities from team loyalties, their professions from jargon, their schedules from timestamps. All freely given. All harmless in isolation. The first website was 324 hyperlinks that grew, through some process invisible to the protocol, into something beyond prediction. Deanonymization is the same principle in reverse — scattered fragments assembling, through some process invisible to the individual poster, into a coherence that was never intended. Both are what happens when you follow the mechanism far enough and discover it has arrived somewhere the description didn’t mention. The difference is that one arrival was a gift. The other is a theft.

Faulkner was a postmaster. Bukowski sorted mail for eleven years. Wallace Stevens spent his career as an insurance executive in Hartford, writing poems on his walk to the office. These are not trivia. They are the wrong-level problem wearing a human face: the job description is accurate and the job description is useless. The same skull that processes insurance claims produces “The Emperor of Ice-Cream.” The mechanism of the days and the mechanism of the work run on the same neurons at altitudes so different they might as well be about two separate people.

There’s a version of this that works in the opposite direction, and it might be the most revealing.

Transit planners in American cities have been studying what happens when you remove bus stops — not add them, remove them. Increase the distance between stops from every other block to every third or fourth block, and the system gets measurably better. Faster trips, more reliable schedules, higher ridership. The subtraction creates something the addition couldn’t. The mechanism-level description — more stops equals more access — is true on its own terms and wrong about what actually happens. The system operates at a level the individual stops can’t see. Now place that next to a recent autopsy of an AI-generated 3D model — a pickleball paddle meant for a product page. At first glance it looked fine. But the geometry was what 3D artists call “triangle soup”: randomly scattered polygons, no logical structure, no edge loops, no clean topology. The texture was baked and blurry. The lighting was frozen into the surface. A human-made model of the same object at a similar file size was sharper, cleaner, more editable in every dimension. The difference was that the human model had been built at the level of intention — this edge loop exists because someone might want to lengthen the handle; this texture is separated because the material might change. The AI model was generated at the level of mechanism — triangles that satisfied a statistical distribution, a surface that resembled a paddle without knowing what a paddle was for. The bus system improves when you subtract because there is structure underneath the stops, a higher-level pattern that the stops were obscuring. The 3D model fails because the mechanism is all there is. There is no deeper level for anything to emerge from. All surface, no structure. The description with nothing underneath it.

This might be the crux of it. Emergence isn’t guaranteed. The gap between mechanism and meaning isn’t always bridged. Sometimes you get Hendrix’s anthem, and sometimes you get triangle soup. Sometimes the feedback loop produces something that exceeds its circuit diagram, and sometimes it just produces noise. The wrong-level problem is not only that low-level descriptions fail to capture high-level phenomena. It’s that you can’t tell, from the low level, whether there’s a high level waiting.

Hendrix could not have been predicted from a circuit diagram, but once you hear what he did, you can go back and find the circuit diagram was always capable of it. The face was always possible in the glass. The novel was always possible in the coffee. The web was always possible in the protocol. Melville, weighing crates at Pier 40, was always capable of the white whale. Each of these higher levels is consistent with its lower level without being deducible from it. The facts are all there, and they are not enough.

I don’t know what to do with this. The face in the glass does not contradict the fracture mechanics. The anthem does not violate Ohm’s law. The novel does not disprove caffeine pharmacology. The customs ledger never once mentioned the white whale. But knowing the lower story does not prepare you for the upper one, and standing on the upper one, you can never quite see where it started.