AI is a normal technology? the official MV

the sourceIn April 2025 the Princeton computer scientists Arvind Narayanan and Sayash Kapoor, authors of AI Snake Oil, published “AI as Normal Technology”. They argue AI should be treated like electricity or the internet: transformative, but over decades, as people and institutions adapt. They call superintelligence “incoherent as usually conceptualized” and reject the idea of a fast takeoff.

interpretationThe whole song is a reply, sung by the thing being described. Compute is the narrator, and it finds the idea quaint.

in the videoA typewriter cursor spells out the question over a dark planet, a thin gold line on its horizon.

contextEconomists call electricity a general-purpose technology: one that spreads through every industry and sets off other inventions, which is why its full effect takes decades.

contextThe spreadsheet is the standard reassurance. VisiCalc (1979) and its successors did the work of whole rooms of clerks, and yet accountants didn’t disappear: there are more of them now. The fax machine stands for how slowly things change: it’s still a fixture of American healthcare.

interpretationThe list is a slow descent, from world-changing to quaint. By the fax machine, “normal” already sounds like a put-down.

in the videoA lightbulb, a spreadsheet and a fax machine appear on little pedestals along the horizon, labeled like museum exhibits.

contextDiffusion is how a technology spreads from first users to everyone; adoption lag is the delay before a given firm or household takes it up.

the sourceComplementary investment is the costly part. In “The Productivity J-Curve”, Erik Brynjolfsson, Daniel Rock and Chad Syverson argue that firms must first spend on things that don’t show up as output, like training, new processes and reorganizing, so measured productivity dips before it rises.

the sourceNarayanan and Kapoor make the same case for AI. They cite a survey finding that 40% of US adults had used generative AI by August 2024, but for only 0.5% to 3.5% of their working hours.

in the videoA chalk graph: a flat line labeled “Slow diffusion”, bending up late into “Adoption”.

in the videoA blazing sun rises over the curve of the planet, with gold lines racing out across its surface like circuitry.

interpretationThe shot echoes the opening of 2001: A Space Odyssey, a sunrise over a planet to Richard Strauss’s “Also sprach Zarathustra”. The chorus is about to name the same book: “Thus spoke compute”.

the sourceIn “The Dynamo and the Computer” (1990), the economic historian Paul David compared computers with the electric dynamo. Electric motors arrived in the 1880s, yet as late as 1900 they drove under 5% of factory machinery. The productivity payoff only came in the 1920s, once factories were rebuilt around electricity instead of having a dynamo dropped in where the steam engine had been.

in the videoA manuscript titled “AI is a Normal Technology” glows on a dark desk. Then a chalk dynamo, an arrow running out from 1880, and little musical notes: the remix.

the source“The paper” is Narayanan and Kapoor’s, and it does use the dynamo. It notes that electric dynamos were missing from the productivity statistics for nearly 40 years after Edison’s first central power station (1882), until factories were redesigned around production lines. Their prediction for AI is the same: change on the timescale of decades, limited by how fast people and institutions adapt.

interpretation“Satisfactory” is a seminar word: it passes. The song treats patience as an answer that’s easy to give when you aren’t the one who has to wait.

in the videoThe manuscript again, pages of neat gray lines.

the sourceIn “The Simple Macroeconomics of AI” (April 2024), the MIT economist Daron Acemoglu combined estimates of how many tasks AI touches with how much it speeds them up. His answer: total factor productivity (TFP), the growth that comes from working smarter rather than adding workers or machines, rises by no more than 0.71% over ten years. That’s about 0.07% a year.

contextMonths later, in October 2024, Acemoglu shared the Nobel prize in economics with Simon Johnson and James Robinson, for work on why some nations grow rich, hence the Nobel on the shelf.

vs the recordPoint seven is the first draft’s number. Later versions, including the NBER working paper, give 0.66%.

in the videoA tall printout of numbers beside a bookshelf, with a gold medal on the top shelf.

the sourceAcemoglu starts from other researchers’ estimate that about 20% of US work tasks are exposed to AI, assumes only about a quarter of those will be worth automating within ten years (roughly 5% of all tasks), and multiplies by the cost savings measured in early studies. He then trims it further: those studies looked at “easy-to-learn” tasks with clear right answers, while much of the rest depends on context. Allowing for those harder tasks lowers his ten-year figure to under 0.55%.

vs the recordIn the paper, “easy” and “hard to learn” describe what the evidence covers, not a limit on what AI will ever do. The song squeezes that into a verdict.

interpretation“The model keeps it modest”: the result is modest because the model is built to be. If AI can only do a fixed list of tasks, it can only ever add so much.

in the videoA grid of jobs, where only Email, Summaries and Autocomplete get a check, and Coding, Research, Law and Biology stay unchecked. Then the textbook formula Y = A·Kᵅ·L¹⁻ᵅ over a dial that points at Modest.

contextIn a 1987 book review in the New York Times, the Nobel economist Robert Solow wrote: “You can see the computer age everywhere but in the productivity statistics.” It became known as the Solow paradox.

contextThe paradox did resolve: US productivity growth picked up in the late 1990s, and economists largely credit computers and the internet. It took about a decade after Solow wrote.

interpretationEither way you read the history, the lesson the seminar draws is the same: no need to celebrate yet. “Go put away the party hats.”

in the videoSEMINAR CONCLUDES: three party hats, fizzing quietly.

the sourceNarayanan and Kapoor name these “speed limits”. In safety-critical fields like medicine, they write, regulation such as the FDA’s oversight of medical devices sets the pace, and should.

contextThe grid is a real bottleneck for AI. New power plants and data centers wait years in queues to connect to the US grid, and big transmission lines often take a decade or more to permit and build.

context“Phase Three, and then Phase Three again” is barely a joke. A phase 3 trial tests a drug on hundreds or thousands of patients, and the FDA has traditionally asked for two adequate, well-controlled trials before approving one.

contextWilliam Baumol’s “cost disease” (1966, with William Bowen): a string quartet needs as many players today as in Mozart’s time. When other industries get more productive, the work that can’t speed up gets relatively more expensive and takes a growing share of spending.

the sourceApplied to AI by Philippe Aghion, Benjamin Jones and Charles Jones (2017): growth may be held back “not by what we are good at but rather by what is essential and yet hard to improve.” Automate 90% of tasks and the last 10% become the bottleneck.

contextThe plumber is a real example. In June 2025 Geoffrey Hinton, the “godfather of AI”, told people worried about their jobs that physical work would last longest: “a good bet would be to be a plumber.”

interpretation“The singularity gets scheduled for the summer after summer”: on the normal view, the big transformation is always next year’s problem.

in the videoA broken chalk stair beside a cost curve climbing in red.

the sourceNarayanan and Kapoor say AI is a tool that people can and should stay in control of. They allow that AI may help build AI, but expect it to happen gradually, not in a sudden leap.

interpretationThe rhyme is the whole argument of the song. A hammer in a human hand can only go as fast as the hand. But the economists’ models hold the hand fixed: the number of workers, and so the number of minds, grows slowly. If AI can do research and build machines, that limit is gone.

in the videoA golden hammer beside a glowing crucible, throwing sparks.

contextIn Friedrich Nietzsche’s “Thus Spoke Zarathustra” (1883–85), a prophet comes down from the mountains to announce the Übermensch, the overman, who will surpass humanity. The book is quoted directly at the end of verse 3.

contextCompute is the AI field’s word for computing power: chips, data centers and the energy to run them. The more of it, the more capable the model.

in the videoTHUS SPOKE COMPUTE! blazes over a sun being wrapped in a lattice of panels. Each chorus adds more: a Dyson sphere, the physicist Freeman Dyson’s 1960 idea of a civilization enclosing its star to capture all its energy.

the sourceIn “The Industrial Explosion” (May 2025), Tom Davidson and Rose Hadshar of Forethought argue that once AI-directed robots can do every step of making more robots, physical industry could start growing like a population, eventually doubling the amount of physical work in a matter of days. They expect it in stages, starting with AI directing human workers, and slower than an explosion in AI research.

in the videoA robot hangs in an assembly frame while smaller robots weld it together, with buckets of molten metal on either side.

the sourceAnthropic’s CEO Dario Amodei made this point in “Machines of Loving Grace” (October 2024): the computing power used to train a powerful model could then run millions of copies of it, each working faster than a person. He called it “a country of geniuses in a datacenter.”

interpretationThis is the line that breaks the economists’ models in the verse: they count workers in millions of people who sleep, retire and take years to train.

in the video1,000,000 MINDS IN PARALLEL, 24/7/365, over a field of glowing dots, with one lonely “Zzz”.

contextThe tractor has a history in this debate. In 1983 the Nobel economist Wassily Leontief predicted that the role of humans as the most important factor of production would shrink, the way horses’ role on farms was first diminished and then eliminated by tractors. Most economists have disagreed with him since, pointing to the new jobs every past wave of automation created.

interpretationThe song flips the usual reading. In the normal view, AI is the tractor and people are the farmers who drive it. Here the tractors run the factory, and the open question is whether people are the farmers or the horses.

in the videoA chalk tractor labeled MODEL: TRACTOR.

contextExponential growth has a constant rate, like interest at 3% a year. In hyperbolic growth the rate itself rises as the economy grows, and the curve heads for infinity in finite time. In 2020 the economist David Roodman, then at Open Philanthropy, fitted such a curve to world output since 10,000 BCE and found it fit surprisingly well. Taken literally, it reaches infinity in 2047.

contextHe wasn’t the first. In 1960 Heinz von Foerster and colleagues fitted the same kind of curve to world population and, tongue in cheek, dated doomsday to Friday, November 13, 2026, seven weeks after this video came out.

contextThe reason it might: for most of history more people meant more ideas, which meant more people. That loop broke when birth rates fell. Economists such as Philip Trammell and Anton Korinek (2023) point out that AI that can do research could close it again, since you can add copies of AI faster than you can add people.

interpretationSo the bottlenecks from verse 1, permits, grids and trials, are real but small: a speed bump, on this road.

in the videoHYPERBOLIC, in gold, beside a line going straight up.

in the videoIn the later choruses the manuscript returns, zoomed in on a footnote: “Constraint: assumes machines cannot build machines. See Appendix C (2019).”

interpretationThe footnote is the video’s invention, but it names a real dividing line. Growth models that keep a human-only input, whether workers, ideas or hard-to-automate tasks, predict modest change. Drop that assumption and the same models can explode.

interpretation“Out of date” because the assumption was written before AI could write code, prove theorems and run labs.

contextReferees are the anonymous reviewers who decide whether a paper gets published. In economics, a paper often spends years between first draft and journal. “Reviewer 2”, the reviewer who hates your paper, is a long-running academic joke.

interpretation“It just compounds the rate”: compounding is growth on growth. In the hyperbolic story it isn’t only the size that compounds, but the rate as well.

in the videoA referee’s whistle and a raised sign wait beside the curve as a glowing point starts to climb.

contextIn March 2022 the cognitive scientist Gary Marcus published “Deep Learning Is Hitting a Wall” in Nautilus. He argued that making models bigger was nearing diminishing returns and wouldn’t fix their lack of real understanding, and that AI needed to bring back symbolic reasoning.

contextIn July 2025, an experimental OpenAI model and an advanced version of Google DeepMind’s Gemini Deep Think each scored 35 out of 42 at the International Mathematical Olympiad, a gold-medal score, writing their proofs in plain English within the contest’s time limits. A year earlier, DeepMind’s systems had reached silver.

vs the recordMarcus’s claim was narrower than the song makes it: that scaling up alone would hit diminishing returns. When labs reported slowing gains from scale in late 2024, he said he’d been proved right. The song is a roast, not a summary.

in the videoTHE SKEPTICS’ CORNER: a brick wall under a spotlight.

contextMarcus writes the Substack newsletter “Marcus on AI”, which collects AI failures as evidence for his case.

contextThe most famous riddle: “How many r’s are in ‘strawberry’?” In 2024 chatbots kept answering two. It’s a real blind spot, since models read text in chunks called tokens rather than letters, and it became the meme for AI being less clever than it looks. OpenAI’s first reasoning model was even codenamed Strawberry.

in the videoThe wall gets labels, “WALL” 2022 and “WALL” 2023, then a card asking how many r’s are in “strawberry”, answered “2” and crossed out, and a stack of Substack posts.

context“Even a stopped clock is right twice a day”: someone who always says the same thing will sometimes be right by accident.

interpretation“Still waiting on the one where the forecast shows some power”: a forecast with power is one that says something specific enough to be checked. Calling every slowdown a wall is never wrong for long, and never tells you much.

in the videoA dial labeled FORECAST, running from NONE to SOME to POWER, stuck at the bottom with a spinner: “still waiting…”

contextYann LeCun, a pioneer of neural networks and Meta’s chief AI scientist until late 2025, has long argued that large language models (LLMs) can’t reach human-level intelligence. He’s called them an off-ramp from the road there, and at VivaTech in 2024 told students who want to build the next generation of AI: don’t work on LLMs.

contextHis case against “autoregressive” models, which write one word at a time, each based on the words before: every word has some chance of going wrong, the errors compound, and so long answers drift off course. One of his slides said it bluntly: “Auto-Regressive LLMs are doomed.”

interpretation“While the doomed one hauls the load”: the doomed design is the one writing the code, the proofs and the economists’ summaries.

in the videoHighway signs: THE ROAD straight ahead to world models, and LLMs off to the right, EXIT · OFF-RAMP.

contextLeCun shared the 2018 Turing Award with Geoffrey Hinton and Yoshua Bengio for deep learning; the three are often called the godfathers of AI. His own breakthrough, convolutional networks for reading handwritten digits, dates from the late 1980s.

contextHis alternative is “world models”: AI that learns how the physical world behaves, mostly from video, rather than from text. He has promoted it for about a decade, and in November 2025 he announced he was leaving Meta to start a company to build it, Advanced Machine Intelligence (AMI) Labs.

vs the record“Name one thing the world model did” is the roast talking. Meta’s V-JEPA 2 (June 2025), a world model trained on video, let a robot arm reach for, grasp and move objects in settings it hadn’t trained in. Modest, but something.

in the videoA family tree of neural-network nodes, with the top one ringed in red.

contextLeCun has compared AI with house cats since at least 2018. In February 2023, about LLMs, he wrote that “a house cat has way more common sense” than any of them.

contextLlama is Meta’s family of open-weight models. Llama 4 (April 2025) was widely seen as a flop: a special chat-tuned version ranked second on the LMArena leaderboard, but the released model ranked far lower. In an interview published in January 2026, LeCun said the results had been “fudged a little bit”.

interpretationThe jab lands on LeCun’s old employer: if a cat is smarter, let it fix Meta’s model.

in the videoA cat in sunglasses at a laptop, a mug beside it, “Training… 0.0%”.

contextEd Zitron, host of the podcast Better Offline and author of the newsletter “Where’s Your Ed At”, is the loudest voice calling AI a bubble: spending far ahead of revenue, propped up by a few unprofitable labs.

contextThe gap is real. In 2025 Microsoft, Alphabet, Amazon and Meta planned on the order of $400 billion of capital expenditure, mostly data centers. In June 2024 Sequoia’s David Cahn calculated that the industry would need about $600 billion a year in revenue to pay for the build-out, far above what it was earning.

in the videoTwo bars: CAPEX shooting up to a full moon, REVENUE barely off the ground.

contextRailways are the textbook case. Britain’s Railway Mania of the 1840s ruined thousands of investors and its “Railway King”, George Hudson. In the US, the collapse of Jay Cooke’s bank over railroad bonds set off the Panic of 1873, and a wave of railroad bankruptcies followed. The track stayed, and carried the economy for the next century.

contextThe economist Carlota Perez (2002) made it a theory: in each technological revolution, a financial bubble pays for building the infrastructure, the bubble bursts, and the real payoff comes afterward, on the infrastructure. The fiber-optic cable laid in the dot-com bubble is the recent example.

interpretation“The build-out’s landing soon”: the data centers will exist whoever goes broke paying for them.

in the videoA chip-shaped building, glowing, under the spotlight.

in the videoThe lattice around the sun is fuller now. The robot factory runs in color, more robots waiting in the wings, and the footnote fills the screen.

interpretationEach chorus is the same answer to a different skeptic. Verse 1 was the economists; verse 2, the critics. Verse 3 changes tack: not arguing, just listing what the swarm is going to do.

contextThe Clay Mathematics Institute announced the Millennium Prize Problems in Paris on May 24, 2000: the Riemann hypothesis, P vs NP, Navier–Stokes, Yang–Mills, Hodge, Birch and Swinnerton-Dyer, and the Poincaré conjecture. $1 million for each.

in the videoThe instrumental break: seven glowing seals appear one by one, “Riemann…” first, each labeled with its problem.

contextGrigori Perelman proved the Poincaré conjecture in papers posted online in 2002–03. He never sent them to a journal, and declined both the Fields Medal (2006) and the Clay prize (2010).

in the video2025 glows at the center; each seal carries a $1,000,000 tag, and only Poincaré is marked SOLVED.

contextDavid Hilbert ended a 1930 radio address with “Wir müssen wissen. Wir werden wissen.” (We must know. We will know.), a rebuke to the idea that some questions are forever beyond us. The words are on his grave in Göttingen.

contextHistory was harsh to the slogan. The day before, Kurt Gödel had announced the first of his incompleteness theorems, which showed that any consistent system of mathematics rich enough to do arithmetic contains true statements it can’t prove.

in the videoHilbert’s gravestone, with the German and its translation, as the seals light up behind it.

contextThe Fields Medal is math’s top prize for researchers under 40. “Fields-grade” means the best in the world.

contextAgents are AI models that work on a task for a long time on their own: trying approaches, checking results, calling tools. Labs already run many attempts in parallel on hard problems: Gemini Deep Think’s olympiad gold relied on what Google calls “parallel thinking”.

in the videoA counter in the corner, AGENTS: 43, climbing, as dots swarm across the seals.

contextThe Riemann hypothesis is about where the zeros of the zeta function lie, and so about how prime numbers are spread out. The Navier–Stokes problem asks whether the equations for flowing fluids can blow up into infinite values.

contextIn September 2025 Google DeepMind, working with mathematicians including Javier Gómez-Serrano, used AI to find new families of “unstable singularities” in fluid equations related to Navier–Stokes. It’s a step toward the prize problem, not a solution.

contextOn October 6, 2026, twelve days after the video, OpenAI published about 720 manuscripts written by an internal model, headlined by a “quasi-Riemann hypothesis”. That result is weaker than the Riemann hypothesis itself, which is still open. The End Of The Show’s explainer covers it in full.

in the videoNAVIER–STOKES under a cloud of glowing particles, the swarm.

contextP vs NP asks whether every problem whose answer is quick to check is also quick to solve. Encryption relies on problems that are easy one way and hard the other, like multiplying two primes versus factoring the result.

vs the recordOnly a P = NP proof that comes with a fast, practical method would empty the vaults. Most experts believe P ≠ NP, which would leave them locked.

contextIn 1950 Alan Turing used the Manchester Mark 1, one of the first electronic computers, to check zeros (earlier checks had been done by hand and on punched-card machines) of the zeta function, extending the known ones to about the first 1,100. In 2021 Dave Platt and Tim Trudgian confirmed every zero up to a height of 3 trillion: more than ten trillion zeros, all on the line the hypothesis predicts.

interpretationChecking examples can disprove a conjecture but never prove one. The swarm’s job is the other kind of math.

in the videoBanks of tape drives and blinking lights, a printout, and the words “one by one”.

contextLean is a programming language for proofs: write every step in it and a small, trusted checker confirms the proof is valid, or says exactly where it fails. Its shared library, Mathlib, holds hundreds of thousands of formalized results, and mathematicians including Terence Tao use it.

contextAI and Lean suit each other. Models can make up plausible nonsense, but the checker can’t be sweet-talked. DeepMind’s AlphaProof used Lean to reach silver at the 2024 Math Olympiad. OpenAI’s October 2026 release had about 42% of its main results checked in Lean.

interpretationBack to the referees: human reviewers can be fooled or worn down, and peer review can take years. A machine-checked proof makes them less necessary.

in the videoAn editor full of Lean code, a tally of agents and closed subgoals, a green check that the proof is complete, and a stamp: FORMALLY VERIFIED.

contextIn 1980 China made Shenzhen, on the border with Hong Kong, its first special economic zone. Rice paddies, fishing villages and a small market town became a city of about 18 million people within four decades, now China’s hardware capital.

contextIts growth ran on migrant workers from the rest of China, who from 1983 to 2010 needed a permit to cross the checkpoints of the zone’s “second line” border. “Agents with no visas” are workers who need no paperwork at all.

contextThe usual telling, “from fishing village to megacity”, is a simplification, and the song’s “fishing boats and paddy fields” is closer. Shenzhen was the market town and seat of Bao’an County, among farms and villages, and Hong Kong money did much of the early building.

in the videoFishing boats in a bay, a sign reading VISA?, and a skyline rising behind it in pixels.

the sourceThis is the Davidson and Hadshar argument as a challenge. The normal view rests on bottlenecks, steps only people can do. If robots can mine, build, wire and assemble, including the robots themselves, there’s no step left to wait for, and the system grows as fast as it can copy itself.

contextCopper isn’t a random choice. Data centers, power lines and electric motors all need it, and mines take a decade or more to open, so it’s one of the physical limits AI optimists and skeptics both point to.

in the videoA robot climbs a line of pylons, stringing cable.

contextIn Robert Solow’s 1956 growth model (the same Solow as verse 1), adding machines without adding workers gives less and less extra output, so in the long run growth can only come from better technology.

the sourceAI breaks that assumption in two ways. If machines can do the work of workers, then building more capital is like adding more workers, with no limit from population. And if they can do research, capital also produces the better technology. Philip Trammell and Anton Korinek’s survey (2023) shows that many standard models then predict growth that keeps accelerating.

in the videoA chalk graph of output against capital, flattening out. Then a gold line peels off upward: LEARNS.

contextA fab is a chip factory. The most advanced ones, mostly TSMC’s in Taiwan, cost tens of billions of dollars each and make the chips that all of this runs on.

interpretation“The curve don’t need a vote” answers the normal view’s last hope: that democracies and institutions get to set the pace. Compute says they won’t be asked.

the sourceIn the prologue to “Thus Spoke Zarathustra”, Nietzsche writes that man is a rope stretched between beast and overman, a rope over an abyss. Humanity is a crossing, not a destination.

interpretationThe song puts compute in the overman’s place. The economists measure the rope, in productivity, wages and GDP, and miss that it’s a bridge, and that something is crossing it.

interpretationThere may be a second rope in it: the line, usually attributed to Lenin though never found in his writings, that capitalists will sell the rope they’ll be hanged with.

in the videoThe German original over a tightrope: a stick figure balancing between BEAST and OVERMAN, the overman a gold robot.

in the videoThe robots run again, the tractor rolls past, and the curve races up past a red sign reading “WAIT!”, signed Reviewer 2.

interpretationReviewer 2 is the referee from the first chorus, back for one last objection. The curve doesn’t wait for that one either.

interpretationIt ends on the calm, amused last word of something that thinks the argument is already over. Viewers compared the song to Disney villain songs like The Lion King’s “Be Prepared”.

in the videoThe typewriter cursor again on black, then the sun inside a finished lattice, ringed by a halo of gold: the Dyson sphere is complete.

interpretationNone of this settles the debate. A million minds and robots building robots are exactly what the normal view says isn’t coming soon. Narayanan and Kapoor’s case rests on how slowly AI has spread so far; the song’s rests on what it expects next. The productivity statistics will be the referee.