Musings, Meanderings, & Loops · Essay
October 22, 2026
I asked a strange question this week: is there an algebra to Bob Dylan? Not a metaphor — an actual formula, the way Claude Shannon found that the yes-and-no logic of a Victorian mathematician could describe the switches in a telephone exchange. Is there a notation that explains what those songs are doing?
The short answer is no, nobody has reduced Dylan to an equation. The longer answer turned out to be funnier, and it changed my mind about something I thought I already understood.
The closest anyone has come is a 2025 paper by Prashant Garg called "On Bob Dylan: A Computational Perspective." It builds on an essay by the legal scholar Cass Sunstein, who described Dylan's style as dishabituating — the way his lines take something you have heard a thousand times and set it next to something you have never heard at all, so that both suddenly look strange. Garg mapped fifty years of studio lyrics into a network of linked concepts and measured how much each song mixes the common with the rare. The finding is what any listener would predict: across the career, more metaphor, and more of that collision between the familiar and the peripheral.
Here is the part I love. To extract the concepts from the songs — to read them — the study used an AI language model.
So the best quantitative evidence that Dylan is unlike a machine was gathered by a machine. Keep that in your pocket; it comes back.
Shannon gives you the vocabulary for what the study found. He estimated that ordinary written English is roughly half redundant — that about half of what we write is predictable from what came before, which is why you can read a sentence with every third letter missing. Most songwriting lives comfortably inside that redundancy. Love rhymes with above, the night is long, the road leads home. The mid-1960s Dylan spends redundancy like a man who has found out it is money. The images in his longest songs from that period arrive from places the previous line gave you no way to expect — circus and scripture, Victorian ballrooms and the electric company, all in the same verse. Measured Shannon's way, he carries more surprise per line than almost anyone writing popular music.
The machine I asked about all this concluded, with great confidence, that Dylan is "the antithesis of AI." A language model predicts the most probable next word; Dylan, it said, systematically picked the least probable one.
It is a satisfying answer. It is also the easy one, and I think it is wrong.
Because before Dylan was surprising, he was saturated.
He arrived in New York at nineteen having absorbed Woody Guthrie almost to the point of impersonation, and he spent his first years in Greenwich Village soaking in the entire Anglo-American folk tradition — ballads, blues, spirituals, sea shanties, union songs, murder songs — thousands of them, learned by ear, sung in basement coffeehouses until they were part of his voice. When he accepted the Nobel Prize, the lecture he delivered was mostly about that: the folk songs and their whole vernacular living in him, and three books from school that never left — Moby-Dick, All Quiet on the Western Front, and the Odyssey.
And he did not just absorb the tradition. He reused it, constantly and openly. The melody of "Blowin' in the Wind" came from an old spiritual, "No More Auction Block" — Dylan said so himself. "Masters of War" rides the tune of an English folk song called "Nottamun Town." Decades later, readers found that lines on "Love and Theft" echoed a Japanese gangster's memoir, and that a later album borrowed phrasing from a nineteenth-century Southern poet. Dylan has always treated this as the way the folk tradition works: everything is common property, passed along and remade.
Look at that description again. An enormous corpus, absorbed whole. A memory so dense with prior text that it surfaces in fragments, recombined. Patterns learned so deeply they come out without being looked up.
That is not the antithesis of a language model. That is a description of one. Dylan may have been the most thoroughly trained one of his generation.
So if the corpus is not the difference, what is?
I came up with an equation for it in that conversation, half as a joke, and the more I sit with it the more I think it is right:
Bourbon, with a splash of Coke.
The bourbon is the corpus — the tradition, the structure, the rules of the form. It is the deep, reliable, amber part, the thing that makes the drink a drink. Both Dylan and the machine have plenty of bourbon. A large language model has more than any human could drink in a thousand lifetimes.
The Coke is the noise — the controlled disruption that cuts the predictability. Too little and you are drinking the same glass everyone else is drinking; this is the smooth, airless sameness of most machine writing. Too much and you have a sticky, formless mess, which is what you get when you turn up the randomness dial on an AI and watch it lose its mind.
I wrote in August that randomness alone is never the fix for dead prose — that human surprise is intentional, not statistical. Dylan is the test case that sharpens it. Even his wildest songs ride on old forms: steady verses, regular rhyme, melodies that often came straight from the tradition. The disorienting imagery lands because the structure underneath is so sturdy. It is mostly bourbon. The splash is small, and it is placed.
That is the whole difference, and it is not in the ingredients. The machine's randomness is poured blind: it loosens every word in the sentence by the same amount and hopes something interesting falls out. Dylan's surprise is poured by someone holding the glass — one image, in one spot, chosen because of what it does to everything around it, and tasted before the next line goes in.
The machine and the man share a bottle. Only one of them is tasting.
Here is where my builder's brain takes over, because I would genuinely like to build this.
The engineering exists. Researchers have ways to reward a model for reaching across distant regions of meaning instead of staying in the comfortable neighborhood of the last word, and newer approaches generate a whole passage at once rather than strictly left to right, which is closer to how a songwriter drafts — scatter the images, then shape them. You can build a system that measures semantic distance and pours controlled noise with real precision. You can make the splash exactly as big as you want.
What you cannot easily build is the tasting. A system can tell you that an image is rare. It cannot tell you that it is right — that this particular strange pairing opens the song instead of breaking it. I wrote two weeks ago about Shannon that the decision about what counts as signal is never mathematical; it is a judgment about what the thing is for. The same is true here. A surprise is only a splash of Coke if someone decides it improves the drink. Otherwise it is just spilled soda.
Which tells me what the tool worth building actually is. Not a machine that writes like Dylan — that machine would just be a faster way to make sticky glasses. The valuable tool is the bartender's assistant: something that hands a writer the strange options, the distant images, the collisions they would not have reached on their own, and then gets out of the way while a human tastes and keeps one. Friction on demand, with the pour left to the person.
Come back to the paper that started this, the one where an AI read Dylan to measure how un-machine-like he is.
I do not think that is a contradiction. I think it is the cleanest demonstration of the point. The machine was perfectly able to detect the surprises — to find, after the fact, which images were rare and which were common, to map the whole network and score the collisions. Detection is exactly what these systems are good at. What it could not have done is write them, because writing them meant choosing, and choosing meant knowing what each strange image would do to a song that did not exist yet.
Anyone can measure a splash after the drink is made. Somebody has to decide to pour it.