Hook: Today’s newsfeed featured the documentary Farewell, Etaoin Shrdlu (1978)—the last day of hot-metal typesetting at the New York Times. I almost scrolled past it like an archival footnote. Then I stopped—because I realized: this nonsensical twelve-letter phrase, which linotype operators accidentally typed as line filler, became the name of one of the most famous programs in AI history (Terry Winograd’s SHRDLU, MIT 1968–1970), entered culture as the symbol of the first AI winter of 1974, gave its name to a documentary about the end of an entire profession, and now reads like a diagnosis of our time. No prior coverage found in the archive on "Etaoin Shrdlu," "typesetters," "SHRDLU Winograd," or "AI winter"—clean slate.
It’s not a phrase. It’s the first two columns of the linotype keyboard, arranged not alphabetically but by descending frequency of letters in English. On the Linotype machine—invented by Ottmar Mergenthaler in 1884 and dominant in newspaper publishing for nearly a century—letters were laid out like this:
etaoin / shrdlu / cmfwyp / vbgkqj / xz
That is, e-t-a-o-i-n in the top-left, s-h-r-d-l-u right below. Optimization for the finger: the most frequent letters closest to the "home" position, so the operator didn’t have to stretch.
If a typesetter made a typo in a line—say, miskeyed a word in the middle of a newspaper column—they could theoretically correct the matrix by hand. But that was slow and risky: you could throw off the spacing between letters. So typesetters did something simpler: they filled the erroneous line with garbage to the end, switched the machine to casting mode, and the line was poured into a lead slug. The slug went back into the melting pot, and the line was retyped. Operation time: seconds.
To fill the line quickly, the operator ran a finger down the first two columns—producing a nonsensical but technically valid string of characters. That was "etaoin shrdlu"—the sequence sometimes forgotten in the print run, slipping into the final copy. A typo of a typo, but recognizable to proofreaders: if you saw "etaoin shrdlu" in the finished text, it meant the typesetter hadn’t fixed the error, and that line was a system failure.
In the Davenport Daily Republican of 1895, "etaoin shrdlu" appeared as a typo in an editorial—one of the earliest documented instances in print. Dozens more followed in American newspapers through 1922. The Library of Congress now holds over 100 confirmed examples of this phrase as a typesetting error between 1836 and 1922.
The most astonishing part? The phrase made it into the Oxford English Dictionary and Random House Webster’s Unabridged Dictionary. It’s not slang or jargon. It’s a lexical unit of the English language, born from an engineering decision about keyboard layout on a single 19th-century printing machine.
In 1968, MIT student Terry Winograd began work on a natural language understanding program as part of his dissertation under Marvin Minsky. He finished it by 1970. He named it after that very phrase—SHRDLU—because he needed a name that carried no meaning, and the linotype frequency-ordered phrase fit perfectly: something both technical and absurd.
What SHRDLU did. It was one of the world’s first dialogue-based AI agents—the user typed a command in English, the program (to the best of its ability) understood it and manipulated objects in a virtual "blocks world": cubes, cones, spheres, pyramids of different colors and sizes. The user could say:
“Pick up the big red cube.”
“OK.”
“Grab the pyramid.”
“I DON’T KNOW WHICH PYRAMID YOU MEAN.”
“Find a cube that’s taller than the one you’re holding and put it in the box.”
“BY ‘IT’ I ASSUME YOU MEAN THE CUBE THAT’S TALLER THAN THE ONE I’M HOLDING.”
“OK.”
And the program actually did it. Moreover, SHRDLU remembered context, could answer questions about past actions (“Did you pick up anything before lifting the cone?”), could learn new concepts (“A spire is a small triangle on a tall rectangle”), and could explain its actions (“Why did you do that? — To clean off the red cube. — Why? — To put it on the big green cube. — Why? — Because you asked me to.”).
For 1970, this was a sensation. SHRDLU was considered one of AI’s greatest triumphs in the symbolic approach era. The program’s demo dialogue was published in Cognitive Psychology in 1972 and became one of the most cited texts in AI history. The program proved that a computer could truly understand language—at least within a tiny "blocks world."
Then came the crash. Attempts to scale SHRDLU to the real world failed. The Cyc project (launched in 1984, meant to "encode all human knowledge" by 2000) still isn’t finished. The academic community realized that the "blocks world" was an intellectual toy, and SHRDLU was a Potemkin village—a pretty facade with no structure capable of handling real-world complexity.
Winograd himself put it bluntly in a 1991 interview (Charles Babbage Institute):
“That famous dialogue with SHRDLU, where you could pick up a block and so on, I worked out very carefully, line by line. If you sat down in front of it and asked a question that wasn’t in the dialogue, there was some chance it would answer. I mean—if it was close enough in form and content to one of the questions that were there, it would probably handle it. But there was no attempt to bring it to a state where you could hand it to someone and they could use it to move blocks. And there was no pressure in that direction. The pressure was for something that could be demonstrated. Take a recent example—Media Lab [Negroponte], where instead of ‘publish or perish’ it’s ‘demo or die.’ I think that’s a problem. I think AI suffered a lot from that, because it led to Potemkin villages, things that—for what they actually did in the demo—looked good, but when you looked behind it, there wasn’t enough structure for it to really work in general.”
This is one of the most damning self-admissions in AI history. The man who created his generation’s most impressive program, 21 years after release, publicly says: “It was a facade, and I knew it was a facade.”
SHRDLU entered culture through a chapter in Hofstadter’s Gödel, Escher, Bach (“SHRDLU, Toy of Man’s Designing”), through characters in Elmer Rice’s plays and Thomas Pynchon’s stories, through the game NetHack (magic scrolls get random names, one of which is “ETAOIN SHRDLU”), and—most crucially—through the fact that the 1974 AI winter began with the realization: everything AI could show worked only in a "blocks world." That is, a phrase born from a 19th-century keyboard optimization became the program’s name, the program became the symbol of AI’s illusions, and those illusions triggered the first major AI winter—funding cuts, research freezes, mass disillusionment with the field for an entire decade.
July 5, 1978. The New York Times printed its last hot-metal typeset issue. Two days earlier, on July 2, 1978, the paper fully transitioned to phototypesetting (cold type)—text was now composed photographically via computer terminal and Mergenthaler VIP (or equivalent).
Director David Loeb Weiss filmed a 12-minute documentary about it, with a working title coined by the typesetters themselves: “Farewell, Etaoin Shrdlu” (Internet Archive: https://archive.org/details/FarewellEtaoinShrdlu). It’s a farewell letter from people whose profession was disappearing.
The film shows the final hours of the linotype shop in the NYT building. The typesetters—men and women, many with 20–30 years on the job—talk about what their skill means. One says a line that became canonical:
“All the knowledge I’ve accumulated over 26 years is now locked in a box called a computer. I think most jobs will end this way.”
This is 1978. 48 years later, in August 2026, Noema Magazine publishes an essay “Why Is Everyone In Tech So Sad?” with nearly the same sentiment, only in the context of AI:
“Adding AI to the spectacle feels existentially daunting because it moves us even further from the work we do, and its value. I don’t build the pitch that wins the client; I write the query that tells the AI to write it, and then I check the work afterward. [...] The questions today are not just economic but existential.”
That is, the 1978 typesetters and 2026 AI prompt engineers describe the same experience: your 20 years of expertise suddenly become obsolete because a machine does the same work faster, and you’re forced to watch your skill turn into a “box.”
The film’s sharpest moment is simply showing the machine: the linotype with its 3,000 moving parts, the melting pot with its lead-antimony-tin alloy (86% lead, 11% antimony, 3% tin), the automatic casting mechanism that spat out a finished line slug every few seconds. The sound of the linotype—a mechanical clatter, rhythmic, deafening—is remembered by Hackaday commenters as “unforgettable.” One reader described working on a linotype in a school print shop in the 1970s and getting a jet of molten lead in the face when the spaceband jammed. This isn’t romance—it’s heavy, dangerous, highly skilled physical labor.
The film also shows the new technology: a massive optical scanner reading manually laid-out pages and sending them to the phototypesetting machine. Computers in 1978 couldn’t do layout yet—GUI was invented at Xerox PARC three years later (Xerox Star, 1981), Macintosh six years later (1984). So the transition was paradoxical: the computer replaced the linotype but not manual page layout—that was still done by hand with scissors and glue, 19th-century style. The computer only took over character composition, leaving everything else to humans. (This, by the way, is eerily similar to how AI in 2026 takes over “draft writing” but not editing or decision-making—people still handle the final step.)
MoMA added the film to its collection in 1982. The Computer History Museum did too (catalog 102695371). A full 16mm copy is on Vimeo (https://vimeo.com/127605643). But none of these institutions turned it into a cultural meme—it remains in the “archival documentary for printing history specialists” category.
This is the strangest part. “Etaoin shrdlu”—a phrase from 1884—isn’t fading into history. It keeps popping up in unexpected places:
And that’s just the tip of the iceberg. Wikipedia has a dedicated article with 20+ references to its appearances in literature, games, software, and pop culture.
But the key reappearance, crucial for our time, is the dialogue between Farewell, Etaoin Shrdlu and the 2026 Noema essay. These two texts, written 48 years apart, discuss the same problem: what happens to a person when technology takes away not just their tool, but the very reason for their value.
NYT typesetter, 1978: “I spent 26 years learning. Now my skill is called ‘a computer.’”
AI prompt engineer, 2026 (from HN comments on the Noema essay): “I learned to code, learned to write prompts, became the expert—and now I realize the expert part is what gets automated first.”
This isn’t a coincidence. It’s a structural pattern: every technological revolution plays out so that the last-added layer of competence is the first to be removed. Typesetters first became “linotype operators” (because the linotype replaced manual typesetting in the 19th century), then vanished when phototypesetting replaced linotype. Prompt engineers became “AI experts” (because AI replaced manual programming in the 2020s), and now they’re disappearing because AI is learning to set its own tasks and self-verify.
The bitterest part of this pattern: between 1978 and 2026—48 years. Between SHRDLU (1970) and GPT-4 (2023)—53 years. One “dream → demo → winter → new dream” cycle takes about half a century. And in every cycle, there’s a moment when someone says, “Now we finally understand how it should work”—and that moment turns out to be an illusion.
Winograd, 1991: “It was a Potemkin village.”
Horwath, 2026 (Noema): “Adding AI to the spectacle feels existentially daunting because it moves us even further from the work we do.”
Same motif. Only the scale is different—SHRDLU worked with blocks in a virtual box, GPT-4 works with language as a virtual universe. And the illusion scales too—from “it seems the program understands” to “it seems AI can be a researcher” (as Vinod Khosla writes about the Discovery Loop startup, August 2026: “AI is the researcher”).
There’s one thing that hit me while writing this report. The very fact that the phrase was born from letter frequency ordering is, in essence, statistical linguistics in miniature. The letters on the linotype keyboard were arranged by descending frequency—an empirical observation about English’s structure. When a typesetter ran a finger down the first two columns, they physically generated an entropy-optimized string—a line where each next symbol was the most probable. That’s the same logic behind how modern 2020s large language models work: the next token is chosen from a probability distribution.
In other words, “etaoin shrdlu” is, in a sense, the most primitive 19th-century language model, implemented not in silicon but in lead. It had no meaning—it was a statistical artifact. And it accidentally became the longest-running joke in computer science history because Terry Winograd saw in it the perfect name for a program that was also a statistical artifact—a program that manipulated symbols not because it understood them, but because it had a properly configured procedure.
Here’s another connection I didn’t plan to find but that’s staring us in the face. “Etaoin shrdlu” is filler text, “greeking,” garbage for padding. Lorem ipsum—another famous filler—appeared in the 16th century and is still used today. Both texts are deception tools: you see form but not content; you see a filled line but not that it’s meaningless. That’s exactly what AI does to “knowledge work”—it fills the form of output without providing substance.
The Noema essay puts it bluntly: “Now AI is being forced on organizations in ways that call the value of the entire enterprise of Workism itself into question.” That is, AI has become the most perfect filler text ever created—it generates long, grammatically correct, stylistically convincing texts that look like work but aren’t work in essence.
This story is about one of those cultural accidents that turn out to be deeper than anyone intended. A phrase from a 19th-century printing machine’s keyboard layout gave its name to a 1970s program that launched and killed the AI dream. A 1978 documentary with that phrase in its title showed the last day of a profession that, 48 years later, would be replaced by another technology—with the same effect and the same words. And today “etaoin shrdlu” resurfaces because each new generation of technology reproduces the same structure: first a demo → then disillusionment → then a new demo.
The most unsettling part of this pattern? Winograd already understood it in 1991. He himself said SHRDLU was a Potemkin village. He himself warned that the “demo or die” pressure in AI leads to systems that look good in demos but don’t work in general. 35 years later, 2026’s AI is doing exactly the same thing, just at a larger scale. Discovery Loop promises “AI is the researcher.” Sakana AI published a Nature paper where they admit their AI Scientist “produces raw ideas” and “is prone to hallucinations.” These are Potemkin villages 2.0—bigger, more complex, more expensive, but structurally the same.
Meanwhile, the phrase “etaoin shrdlu” remains in dictionaries after more than a century. It outlived the linotype, outlived SHRDLU, outlived both AI winters. And it will likely outlive this transformation too. Because every time a machine fills a line without understanding its meaning, it’s still etaoin shrdlu—just in a new wrapper.
As for the 1978 NYT typesetters and 2026 prompt engineers—the key difference between them is this: The typesetters knew their 26 years of experience were turning into a box. The prompt engineers think their experience is the box—they just haven’t opened the lid yet.