Jean-Baptiste Vuillaume was the most prolific violin maker of the nineteenth century. Across his lifetime, his Paris workshop produced more than three thousand instruments. He copied Stradivari and Guarneri with such fidelity that contemporary dealers built businesses around authenticating which Vuillaumes were Vuillaumes and which had been mistaken for the masters they imitated. He acquired Luigi Tarisio's legendary collection of old Italian instruments, including the famed Messie Stradivarius. He trained Maucotel and Silvestre, the Peccatte brothers, Persoit, Voirin, Derazey, an entire generation of nineteenth-century French luthiers. He invented the steel bow, the self-rehairing bow, the contralto viola, and the Octobass.
He also developed a varnish that contemporary luthiers still consider the best post-Cremona match for the old Italian masters. He took the formula to his grave.
Two centuries later, his successors are still trying to reverse-engineer it. He did not intend to lose the knowledge. He just never structurally transferred it. The varnish was his, in a deep sense, and he protected it the way a craftsman protects what makes their work theirs. The protection was effective during his lifetime. After his lifetime it became the mechanism of the methodology's death.
This is what happens by default to a methodology when its owner is a brilliant individual and the structure around them never insists on the transfer.
The LLM era is producing a lot of Vuillaumes.
How methodologies die
Every wave of automation creates a methodology near-death moment somewhere in the economy. The mechanization of weaving did it to handloom weavers. Containerization did it to dockworkers. Spreadsheet software did it to a generation of bookkeepers. The pattern is the same each time. A body of human practice gets absorbed into a system that runs faster and cheaper, the practitioners disperse, and the methodology either survives in some structurally maintained form or it doesn't. Most of the time, it doesn't.
The default failure mode has several variants. Vuillaume's is one: a brilliant practitioner protects what makes the work theirs, and the protection becomes the mechanism of loss when they die. There are others.
Vuillaume also built the Octobass, a three-meter, three-stringed double bass played by foot pedals and hand levers because the strings are too thick and the fingerboard too long for human fingers. He unveiled an improved version at the 1851 Great Exhibition in London. Berlioz heard it, called its tone full and strong without any roughness, and recommended every major orchestra adopt three of them. Almost no one did. Vuillaume built only three. Today, between five and seven specimens exist worldwide, the modern ones built by a small handful of luthiers including the German master Wolfgang Staab, who spent thirteen months building the most recent Octobass using gut strings made from two hundred sheep and wood he personally sourced from the Alps. That methodology survives because Staab and a few others in his cohort have actively maintained it. If Staab and his peers stop building, the Octobass dies in a generation. A methodology so concentrated that its survival depends on three or four living practitioners is not really alive. It is on life support.
The third failure mode is harder. The Gizmotron was invented in the early 1970s by Kevin Godley and Lol Creme of the British band 10cc, who could not afford an orchestra for their early albums. They imagined an attachment that would let an electric guitar produce violin-like sustained tones. Their first prototype involved Godley applying an electric drill to Creme's guitar strings. They took the idea to a physicist at the University of Manchester Institute of Science and Technology, who built a working version with motor-driven, serrated rubber wheels that bowed each string individually. The device was patented in 1975, manufactured commercially by Musitronics in 1979, used by Jimmy Page on Led Zeppelin's In Through the Out Door and by Paul McCartney on London Town, and discontinued by 1981. Cheap synthesizers had arrived simultaneously. The Gizmotron was temperamental in humidity, the wheels wore down quickly, and synths produced similar sounds for a fraction of the price.
Godley said later that on a good day it sounded like an orchestra and on a bad day it sounded like a chainsaw.
I will come back to the Gizmotron, because it is the most important counter-example to the argument I am about to make. But for now, hold all three. Vuillaume is the methodology that dies with the master because the master never structurally transferred it. The Octobass is the methodology so concentrated in a handful of practitioners that one generation of inattention ends it. The Gizmotron is the methodology beaten by a cheaper substitute that does almost the same thing for almost everyone almost as well.
All three are now in play for knowledge work in the LLM era. A handful of brilliant practitioners hold methodologies they have never structurally transferred. Their work is being absorbed into systems running on cheaper substitutes. The substitutes do almost the same thing for almost everyone almost as well.
The question is not whether the absorption is happening. The question is what structure you put around it.
How methodologies survive: the apprentice chain
In 1954, a Pentagon mathematician named Leland Sprinkle took his five-year-old son to Luray Caverns in Virginia for a birthday tour. The cave guides routinely tapped stalactites with rubber mallets to demonstrate that they rang. Sprinkle, who was trained as an organist at the Peabody Conservatory under Virgil Fox in addition to his day job designing electronics for the Pentagon, became preoccupied with what he had heard. He spent the next three years of weekends in the damp dark of the cave with tuning forks, prospecting for stalactites that could be tuned to concert pitch. He selected thirty-seven, sanded all but two of them down to perfect their pitch, and wired each to a solenoid-actuated rubber mallet. He bought the four-manual organ console from Klann Organ Supply in Waynesboro, Virginia. The instrument covers three and a half acres and is recognized by Guinness as the world's largest. He gave the first concert on June 7, 1957, and continued playing it for decades. He died in 1990.
Sprinkle's apprentice was a teenager named Larry Moyer, who started at Luray as a tour guide and learned the instrument directly from the inventor. Moyer is now lead engineer. He has been there forty-two years. His apprentices are Stephanie Beahm and Ben Caton. Caton has been there eighteen years. He says he is still learning. Beahm says: "Larry can't retire till we know everything."
Moyer's defining line: "There's no Great Stalacpipe Store. We've had to make everything ourselves."
That sentence is the entire case against the assumption that a complex operating system can be commodity-purchased. The instrument has thirty-seven custom-tuned stalactites in a cave with active water seepage. Tracking down a single broken note can take three hours of climbing through the cave with multimeters. The team builds their own electronics because there is no commercial alternative. The methodology survives because the apprentice chain is named, deliberate, and operationally insisted upon. Sprinkle to Moyer to Beahm and Caton. Each generation knows they are the methodology owner of the next.
Linda Manzer's apprentice chain is a different shape but the same pattern. Manzer is a Toronto luthier in her seventies who builds ten to fifteen guitars a year by hand. She grew up in Etobicoke, the daughter of a wartime radar technician who fixed televisions in his spare time; her childhood basement had his electronics workbench on one side and her painting bench on the other. As a teenager she saw Joni Mitchell at the Mariposa Folk Festival on Toronto Island, fell in love with the dulcimer, couldn't afford the hundred-and-fifty-dollar instrument at the Toronto Folklore Centre, and was talked into buying a seventy-five-dollar kit by the shop owner. She built it. She studied painting at Sheridan College, then briefly at the Nova Scotia College of Art and Design, where she worked as a telephone operator. She used the switchboard to call Jean Larrivée in Toronto for free, repeatedly, until he hired her. From 1974 to 1978 she apprenticed with Larrivée, the only woman in the shop. She has said: "From the moment I walked into Larrivée's shop, the moment that I saw what they were doing, I knew that was what I wanted to do."
Her fellow apprentices, the men she trained alongside, became Canada's first generation of nationally distinctive luthiers. Sergei de Jonge, Tony Duggan-Smith, David Wren, George Gray, William Laskin. In 2017 Manzer organized the Group of Seven Guitar Project, an exhibit that brought Larrivée and five of his other original apprentices back together to build eight guitars honoring the eight Canadian painters of the Group of Seven. The exhibit broke attendance records at the McMichael Canadian Art Collection. The documentary aired on CBC and PBS. Of her cohort, she has said: "Technically, they were competitors. But we made a conscious choice to be friends and help each other. And it turned out to be a magical choice. It made us all better builders."
In 1978 she opened her own shop above a pool hall on College Street in Toronto. Five years later she got a call from Jimmy D'Aquisto in New York, the legendary archtop builder, who had read an article in which she had named him as an influence. She went down to Long Island to study archtops with him. D'Aquisto's philosophy, she learned, was that the archtop guitar could do everything.
Pat Metheny first met her around 1980 when she sent a backstage note in Toronto. Her Danish apprentice at the time told her the first draft was too generic, and that she should write it from her heart. She rewrote it. Metheny invited her back to the hotel. She ran home, grabbed two guitars, and sat in his hotel room until three in the morning while Metheny played the entire concert again, this time on her instruments. He ordered one combining elements of both. She delivered it two months later. She has built about twenty-five instruments for him over the four decades since.
In 1984, Metheny asked her to build him a guitar with as many strings as possible. She had no idea what that meant. She said yes anyway because, in her words, "I knew he trusted me, so I ran with it." The Pikasso has forty-two strings, four necks, two soundholes, two mounting holes for a stand, a hexaphonic pickup feeding Metheny's Synclavier synthesizer, roughly a thousand pounds of string tension at concert pitch, weighs fourteen and three-quarter pounds, and took two years and a thousand hours to build. She invented a wedge-shaped body geometry for it so Metheny could see all the strings while looking down at the instrument. The wedge is now a standard option on her smaller models.
She has trained three apprentices herself. One of them, Peggy White, now builds independently on Fogo Island in Newfoundland. The lineage continues.
Manzer's lines on the work, which appear in interviews stretching back decades and have not changed in their substance: "My hands are building it, but you will be the one playing it. My job is to deliver you a guitar that will inspire you to create for years to come." "You can't make wood what it is not. You have to cooperate with it. You have to be honest with yourself. You can't fake it."
The Manzer pattern is the Stalacpipe pattern at a different scale. Named lineage, deliberate transfer, slow. The methodology survives because the practitioners structurally insist on apprenticing the next generation, even when the economics do not reward it.
How methodologies survive: institutional rescue
The nyckelharpa is a keyed Swedish fiddle that has been continuously played since at least 1350; it is carved on a Gotland church gate from that period. The bow plays a row of melody strings while the left hand depresses wooden keys to fret notes, and underneath the melody strings run a dozen or more sympathetic strings that vibrate freely, producing a shimmering halo around every note. By the mid-1950s the tradition had nearly died. There were roughly twelve active players left, all in the Uppland region of central Sweden. Twelve. The instrument was a generation away from extinction.
The Swedish government decided the instrument had to be saved. Rather than commission a museum-piece preservation or fund a factory to produce affordable copies, the government took an unusual approach. They asked Eric Sahlström, the leading living player and builder, to develop a buildable plan and personally lead a series of public workshops where ordinary Swedes could come and build their own nyckelharpa. Sahlström himself was the grandson and son of fiddlers, a master of the instrument from childhood, and the developer of the modern chromatic nyckelharpa design. He agreed. The workshops ran across Uppland. Ordinary people with no luthiery background built their own instruments from his plans. The folk revival of the 1960s and 70s amplified the effort. The instrument spread.
By 2009, the chromatic nyckelharpa was officially recognized as part of Sweden's National Cultural Heritage. The Eric Sahlström Institute, founded in 1998 in the small town of Tobo, now runs year-long folk music and dance programs, hosts a network for nyckelharpa teachers and builders, and commissions specially-sized instruments for children's programs. Modern luthiers in Sahlström's lineage include Esbjörn Hogmark in Sweden and Earl Holzman in the United States; the global nyckelharpa community now spans dozens of countries.
The saying among players, captured by one of them in an oral history, is "When you've seen one nyckelharpa, you've seen one nyckelharpa." Each instrument takes over a hundred hours to build by hand. The revival was deliberately anti-commercial in structure. The state could have funded a factory; it chose not to.
Walk through the structural choices the Swedes made. A named methodology owner who consented to and led the dissemination. Public funding for distributed reskilling of ordinary citizens. Refusal to commodify production. The instrument was preserved by deliberately keeping it hand-built, slow, and locally apprenticed, with an institutional successor to carry the lineage past the founder's death.
It worked. The methodology has more practitioners now than at any point in its modern history. The factory version would have produced more instruments faster, in the short run; it would also have eliminated the apprentice chain that the survival depended on. The Swedes chose the slow path on purpose, because the slow path is what makes the methodology durable.
How methodologies survive: structural restoration when it is already late
The Boardwalk Hall Auditorium Organ in Atlantic City, New Jersey, was built between 1929 and 1932. It was designed by Emerson Lewis Richards, a New Jersey state senator who designed pipe organs as a hobby, and built by the Midmer-Losh organ company. It has 33,112 pipes, 449 ranks, 1,235 stop tabs, eight chambers distributed around an auditorium of five and a half million cubic feet, the world's loudest organ stop (the Grand Ophicleide), and one of only two true sixty-four-foot pipes in existence. The building was designed around the organ.
In 1944, a hurricane hit Atlantic City. Saltwater flooded the blowers and destroyed the electro-pneumatic combination action. For the next eighty years, the organ deteriorated. Roof leaks corroded the chambers. Air conditioning condensation soaked pipes. Construction contractors working on building renovations stepped on pipes, severed windlines, and in one particularly memorable case sealed off an opening in the Left Upper chamber, entombing a thirty-two-foot Trombone behind a wall. In 2017, the current curator of organs, Nathan Bryson, was surveying a damaged section of the Major Diapason pipe in the Pedal Left division and found a wrench at the bottom of the pipe. Apparently dropped during construction in 1929. Entombed for eighty-eight years.
The Historic Organ Restoration Committee, formed in 2004, is leading a sixteen-million-dollar restoration. As of 2024 the instrument is roughly two-thirds playable. Bryson is the fifth curator since the original installation; his predecessors include a part-time gardener-plumber named Dennis McGurk who learned the leatherwork by cutting and stitching the bellows valves himself, single-handedly protecting the instrument during renovation projects when construction crews would otherwise have destroyed parts of it. Bryson, who started piano young and volunteered at the Wanamaker Organ in Philadelphia before becoming curator in 2015, leads three full-time experts and a corps of volunteers. The work consists of leather pouch by leather pouch, chamber by chamber, pipe by pipe, every leather pneumatic pouch tested individually before installation because, in Bryson's words, "if one of these leather pouches gets off a little bit, then it can cause all kinds of problems." HORC's Scott Banks describes the project as like painting the George Washington Bridge: by the time you finish, you have to start over again.
The Boardwalk Hall story is what survival looks like when the structural maintenance was abandoned for eight decades. The methodology did not die outright; it just deteriorated so far that the recovery cost more than the original construction would have, and required practitioners willing to invent custom restoration techniques because no commercial alternative existed. Bryson and his team build their own electronics because, like the Stalacpipe team, there is no commercial substitute for what they need.
The lesson is not that all methodologies can be recovered if you commit enough resources. The lesson is the cost of restoring a methodology after it has been routed around for decades dwarfs the cost of maintaining it continuously. Bryson is doing now what should have been done in 1945, in 1955, in 1975, in 1995, in 2005. The price of not maintaining was paid forward; the bill came due in his lifetime.
The architecture is the instrument
Two more figures from the rare-instruments world sharpen what the absorption challenge actually is.
William Close studied sculpture and sound design at the School of the Art Institute of Chicago. Around the turn of the millennium he invented the Earth Harp, an instrument whose strings can extend up to a thousand feet and are mounted between a resonator and a fixed point in the surrounding architecture. He has installed it at the Roman Coliseum (strung to the Arch of Constantine), the Kennedy Center, Lincoln Center, the Seattle Space Needle, the Grand Theater of Macau, temples in Vietnam, and aboard Royal Caribbean's Quantum of the Seas as a permanent thirteen-hundred-seat installation. Cirque du Soleil's Kà at MGM Grand has a permanent Earth Harp. He developed the playing technique himself, using rosin-coated cotton gloves running along the strings to set up a longitudinal compression wave, similar to running a finger around the edge of a wine glass.
Close's defining claim is that the Earth Harp is architectural rather than portable. "When I play the Earth Harp, I'm actually standing inside the strings, and so I'm surrounded by all that vibration and music." The Frank Lloyd Wright line about architecture being frozen music is the basis for his work. He is explicit that every installation is a different instrument, because the architecture is part of the resonator. There is no Earth Harp that is not a specific Earth Harp built into a specific space.
This is a useful image for what an LLM absorption actually is. The model is not the system. The system is the model plus the data plus the workflow plus the practitioners plus the audit substrate plus the failure modes plus the customer expectations plus the regulatory environment. When you replace one of those components, you have not replaced the system; you have replaced one of the resonators. The instrument is the building.
The second figure is Björk, who in 2011 commissioned a suite of custom instruments for her Biophilia project. One of them was the Gameleste, a hybrid built by London percussionist Matt Nolan and Icelandic organ-maker Björgvin Tómasson. A nineteenth-century celesta was gutted of its original metal bars and fitted with bronze gamelan bars on a MIDI-controllable mechanism. Björk wanted live, played sounds rather than samples, and a custom toolkit that did not exist. Tómasson also built her two custom MIDI-controlled pipe organs. Andy Cavatorta built her four twenty-five-foot gravity-driven pendulum harps. Henry Dagg built her a Sharpsichord, a pin-barrel mechanical harp. The album's iPad app was the first acquired by MoMA's permanent collection.
The Biophilia project shows the opposite design principle from Close. Where Close's instruments are inseparable from their building, Björk's are deliberately modular and pluggable. She commissioned the instruments she needed, named the makers in the album credits, and toured them around the world. Two pipe organs, four pendulum harps, the Gameleste, and a twenty-four-voice Icelandic female choir were on stage at once. The point is not that one design philosophy is right and the other is wrong. The point is that both philosophies require the methodology to be explicit, named, and consented to by its owners.
When the cheap thing wins
Now I owe you the Gizmotron honestly.
Godley and Creme invented something beautiful, expensive, and mechanically expressive. They left 10cc in 1976 to focus on it. They built a three-album concept work called Consequences (1977) that was essentially an extended demonstration of the device. They patented it. They licensed it to Musitronics, who manufactured it commercially. Jimmy Page used it on Led Zeppelin's last studio album. Paul McCartney used it on London Town. The throbbing-gristle band Throbbing Gristle used it. The Banshees' John McGeoch used it. And then it died.
It died because cheap synthesizers arrived. The Roland Juno-60, the Korg Polysix, the Yamaha DX7 reached musicians around the same time and produced similar sounds for a fraction of the price. The Gizmotron was temperamental in humidity. The rubber wheels wore down quickly. ARP recalled the units. By 1981 it was history. Godley and Creme moved on, let the patents expire, and closed the door on that chapter. "On a good day it sounded like an orchestra," Godley said later, "and on a bad day it sounded like a chainsaw."
The Gizmotron is the obvious objection to the argument I have been making. Cheap synthesizers beat hand-crafted mechanical sustain because the synthesizer was good enough for almost everyone who wanted that sound. So why doesn't the same logic apply to LLMs? Why won't cheap LLMs beat the expensive humans?
The honest answer has four layers.
The first is that the Gizmotron is the device, not the methodology. The methodology Godley and Creme cared about was the production of sustained, expressive, orchestral textures using a guitar as the source. Synthesizers gave you sustained, expressive textures; they did not give you the guitar as the source, and the people who specifically wanted that combination kept finding other ways. Pat Metheny's Pikasso, built by Linda Manzer in 1984, used a hexaphonic pickup feeding a Synclavier synthesizer to do something that neither pure synth nor pure guitar could do. The methodology of expressive sustained guitar sound did not die; it redeployed. Aaron Kipness re-engineered the Gizmotron itself in 2015, with Godley's blessing, and the Gizmotron 2.0 has a small but loyal market.
The second is that Godley and Creme themselves did not disappear. After the Gizmotron failed, they pivoted to directing music videos. Their work on the Police's "Every Breath You Take" and Herbie Hancock's "Rockit" defined an aesthetic that ran through MTV for a decade. The craftspeople redeployed. The question for any methodology near-death moment is not just whether the methodology survives in its current form, but whether the craftspeople can redeploy into adjacent forms. Godley and Creme could because their underlying skill (taste, composition, visual sensibility, the willingness to work expensively for a vision) was transferable. The question for knowledge workers whose methodology is being absorbed by LLMs is the same. Is the underlying skill transferable into adjacent forms, and is the redeployment structurally supported?
The third is that "good enough" hides costs. Synthesizers were good enough for the average musician trying to fake an orchestra. They were not good enough for Metheny on the Pikasso. They were not good enough for the specific timbral effects that drove a small percentage of artists to keep using the Gizmotron in particular. The averaging that makes "good enough" a true claim conceals what is lost at the edges. For most cases, the cheap LLM will be good enough. The question worth asking is what is being lost at the edges of your particular operation, and whether that loss is acceptable.
The fourth is a structural observation that has nothing to do with the artifact and everything to do with the practitioners. Godley and Creme had the resources, reputation, and skill to redeploy after the Gizmotron failed. Their displaced workforce was themselves. The people whose methodology is being absorbed by LLMs in most industries do not have those resources, reputations, or skills. They are workers, not artist-inventors. Whether they can redeploy depends on whether the structure around the absorption provides funded reskilling, named succession, and protected time. The Gizmotron failure was clean because Godley and Creme were fine afterward. Most absorption is not clean, because most practitioners are not Godley and Creme.
Hold the Gizmotron honestly. It is what happens when a methodology is replaced by a cheaper substitute that is good enough for almost everyone. The argument for structural preservation is not that this never happens. The argument is that when it happens, the structure around the displacement determines whether the methodology and its practitioners redeploy or disappear. Godley and Creme redeployed into music video direction. The displaced workers in most knowledge-economy methodology absorptions will not redeploy unless the structure around the absorption makes it possible.
What the people building LLMs think
The most interesting fact about the current LLM moment is that the people closest to the technology are publishing the evidence for structural caution, not the people farthest from it.
The most distinctive voice on this question is Emily Bender, professor of linguistics at the University of Washington. Bender co-wrote the 2021 paper "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" with Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell. The paper triggered Gebru's removal from Google. The "stochastic parrots" coinage has become the central critical metaphor of the LLM era. In a 2020 paper co-written with Alexander Koller, she proposed the Octopus Test: a thought experiment in which a hyperintelligent octopus taps an underwater telegraph cable between two stranded humans, learns to predict the statistical patterns of their conversation perfectly, but, lacking any contact with the world the words refer to, never actually understands them.
The Octopus is the LLM. The methodology, the thing that gives the words meaning, is in the practitioners. When you describe the model as understanding the work, you stop noticing what only the humans contribute. Then you stop investing in it. Then it disperses. Then in five years you have an LLM running on a fossil of what your knowledge workers used to know, and no one in the building who can tell you whether the output is good.
Bender's broader argument, captured in her 2025 book with sociologist Alex Hanna The AI Con: How to Fight Big Tech's Hype and Create the Future We Want, is that the language we use about these systems is making this harder. The word "hallucinate" assigns to the model an interior life it does not have. The word "understand" does the same. The word "thinks" does the same. The phrase "goes rogue" is the worst of them, because it implies agency. Bender's preferred description is operational. A language model stitches together sequences of linguistic forms according to probabilistic information about how they combine, but without any reference to meaning. Her favorite test for an LLM feature is functional: "I want to read the PDF." If the layer you are adding does not actually help the user read the PDF, it is not adding value; it is adding distraction wrapped in mystical phrasing.
Bender's voice is the sharpest, but she is not alone. Luke Zettlemoyer, professor at the University of Washington and senior research director at Meta FAIR, has spent the last two years arguing publicly against monolithic frontier models. His position: the assumption that a single very large model trained by a centralized authority on all available data will always perform best in every use case is a design choice, not a fact. His work on modular expert mixtures, on QLoRA (which he framed as an effort to democratize access to LLM research), and on the Llama family of models has consistently advocated for pluggable, transferable, openable architecture. In a recent talk he framed alignment in a way that should sound familiar to anyone who has thought about institutional design: "Humans are fundamentally social beings, and the challenge of inducing self-interested humans to act in ways that are good for others is the fundamental alignment challenge of human societies. Alignment in human societies is not achieved by inducing the same or average innate preferences in individuals but by aligning individual behaviors with normative classifications reached through informal and formal social processes, which we can call institutions."
Translation: alignment is institutional, not technical. The problem is not making the model want the right thing; the problem is putting the model inside structures that produce the right outcomes regardless of what the model wants.
Yu Meng at the University of Virginia, an alignment researcher who received an OpenAI Superalignment Fast Grant and made the Forbes 30 Under 30 Asia list this year for his work on the technical side of LLM alignment, recently published a paper called "Do LLM Evaluators Prefer Themselves for a Reason?" The paper directly attacks the assumption that scaling judge-models produces better judgments, demonstrating self-preference bias even on objective benchmarks like math and code. The fact that this paper exists, and that it was published by a researcher inside the alignment community rather than by an outside critic, is itself the structural caution argument. The builders of these systems are publishing evidence that the systems should not be trusted to evaluate themselves.
Sebastian Ruder, formerly at Google DeepMind and Cohere, now at Meta, runs the most-read NLP blog on the internet. His central argument is democratization through reproducibility: in an era where running state-of-the-art models requires a garrison of expensive GPUs, the research left for academics, PhD students, and newcomers depends on the discipline of small-scale validation. Prove the method on a representative small model. Trust the community to scale it. He argues, explicitly, that the field is in a benchmark crisis. The models have outpaced the tools we use to measure them. His position on low-resource languages is similar: 7,000 languages are spoken around the world, and NLP research has mostly focused on one of them.
Dilek Hakkani-Tür and Gokhan Tur, a married academic couple who together have built conversational AI through four eras of technology (AT&T Bell Labs in the early 2000s, Microsoft Cortana in 2010, Apple Siri and Google Assistant in 2014–18, Amazon Alexa in 2018–23) and joined the University of Illinois faculty in 2024, share a position that Tur articulated cleanly in a 2020 Stanford talk: "The underlying skeleton of goal-oriented systems has remained unchanged. Still most language understanding models rely on supervised methods with manually annotated datasets even though the resulting performances are significantly better with much less data." The headline-grabbing capability of LLMs has not changed the underlying engineering problem. The dialogue system, the workflow, the user-grounding, the evaluation discipline, all of that still has to exist around the model. The model is a new component in an old architecture.
Take all five of these voices together. They are not skeptics writing op-eds. They are the technical community that is building, evaluating, and deploying these systems. Their published positions, taken seriously, are: the language we use about LLMs misleads us about what they are doing; the frontier model is a design choice, not a destiny; alignment is institutional rather than technical; even self-evaluation is biased; the benchmarks are not keeping up; and the engineering problem the LLM solves is one component in a system, not the system itself.
If the people building the instruments are saying it is part of an architecture, the executive question is what architecture you put around it.
Four conditions for structural absorption
What does the structural absorption look like in practice? Four conditions, drawn from the rare-instrument cases and confirmed, in different language, by the LLM researchers.
The first condition is named methodology owners with consent. Eric Sahlström led the dissemination of the nyckelharpa. Linda Manzer chose her three apprentices. Leland Sprinkle apprenticed Larry Moyer. Björk credited Nolan and Tómasson in the album notes. The methodology owner is identified, knows they are the owner, and consents to and structures the transfer. The person whose pattern recognition, judgment, or craft is being encoded into the system is the one driving the encoding, not its target. In organizations, this is the test that distinguishes an authentic methodology absorption from an extraction. If you cannot name the methodology owner, you do not know what you are absorbing.
The second condition is funded reskilling for displaced practitioners. The Swedish state paid for the nyckelharpa workshops. Godley and Creme had the resources to redeploy from the Gizmotron failure into music video direction. The methodology owner is not asked to give up the craft for free. The budget for absorbing a function into an AI system includes the budget for retraining the people that function used to employ, into roles that compound from their existing judgment. The absence of this line item is the signal that the absorption is extractive rather than structural.
The third condition is refusal to commodify production prematurely. Manzer hand-builds ten to fifteen guitars a year. The Stalacpipe team builds their own electronics. The nyckelharpa is a cottage industry by design. The Boardwalk Hall organ is being restored one leather pouch at a time. The first version of an LLM-augmented system runs slow and small, with the practitioners in the loop, until the encoding is verifiably complete. The pressure to scale immediately is the pressure that causes structural absorption to fail. The factory version of the nyckelharpa would have produced more instruments faster in the short run, and would have ended the methodology in a generation.
The fourth condition is walk-away discipline. Bender's line about LLM features she does not need: "I want to read the PDF." The discipline of saying no when the system fails the user. A clear, pre-committed standard for when the AI-augmented version of the work is worse than the human version, and a willingness to roll back without ego. Without this condition, the absorption ratchets in one direction regardless of whether it is working. The Boardwalk Hall organ was routed around for eighty years because no one had the institutional standing to say it was being destroyed; Dennis McGurk, the part-time gardener-plumber who taught himself to stitch bellows valves, had the standing only because no one else wanted the job.
None of these four conditions are fast. None of them are what the vendor demo highlights. None of them show up in the procurement spreadsheet.
But the alternative is the varnish problem at scale. A few brilliant individuals in your organization right now hold the methodology your business actually runs on. If you absorb their methodology into a system without structurally insisting on the transfer, the methodology will die when they leave. The model will keep running. The output will look fine. And in five years your successor will be reverse-engineering what you used to know.
What this means
The instrument analogy is structural, not metaphorical. The methodology is the instrument. The model is one of the resonators. How well it sounds in five years depends entirely on what you build around it now.
Vuillaume's varnish is still missing two centuries after his death. Modern luthiers approximate it through chemistry, careful inspection of surviving instruments, and slow accumulated guesswork. They have not recovered it. Each generation of luthiers has tried, and each has come close, and each has died with a slightly better approximation, none of them the original.
The methodology your operation runs on is somebody's varnish. The question is whether your successors will be running on the methodology or reverse-engineering it.
The Stalacpipe Organ has worked continuously since 1957 because Sprinkle apprenticed Moyer, and Moyer apprenticed Beahm and Caton, and Caton has been there eighteen years and says he is still learning, and Beahm says Larry cannot retire until they know everything.
The nyckelharpa is alive because the Swedish state, with Sahlström's consent, paid for citizens to build their own instruments from his plans.
Linda Manzer's guitars will be playing in the year 2100 because she trained Peggy White and others to build instruments her clients are still discovering they need.
The Boardwalk Hall organ is being recovered, one leather pouch at a time, because Nathan Bryson and a small team of practitioners refuse to accept that the entombment was permanent.
Each of those is a methodology that should have died and didn't, because the people closest to it built the structure that kept it alive.
The LLM era is producing the conditions under which the default is Vuillaume. The structure has to insist otherwise. That is not the vendor's job. It is yours.