Open a new file in any writing application and it will offer you some templates to choose from -- resume? term paper? newsletter? sales pitch? Here are the molds you can pour your work into. Who could object? Given that you're making something that is expected to work in a particular way, variation from the standard isn't welcome. You make today's like yesterday's, from the template that works for them all. An eccentric version would just gum up the works, like a deformed part on an assembly line.
But now we have ChatGPT, Claude, Gemini and other Large Language Models. Using them, you can "templatize" a lot more than typography and layout.
“Artificial’’ assistance for churning out words has been around for centuries
You can, for example, have a LLM write that recommendation letter someone asked for. Given that the model has been trained on data that includes thousands of letters, it could well be more effective than your version (after you check it for made-up facts and figures).
Yet isn't the point of a letter of recommendation to show that the subject impressed a living person enough to take the trouble to write? In the absence of human effort, it becomes meaningless, doesn’t it? Much of the angst these models cause stems from the feeling that they're intruding into the realms where we express our uniqueness as creative, communal, ethical beings.
As Dennis Yi Tenen puts it in Literary Theory for Robots: How Computers Learned to Read, sometimes we don't care if a piece of writing was human-made or machine-made. If it quacks like a duck it's a duck. But sometimes the psychology matters, which means the biology matters. A human brain wrestling with phrases, human fingers typing. We want the thing to taste like a duck too. So, when ChatGPT tastes like a duck without being a duck, some of us freak out.
Until I read Yi Tenen's brief and brilliant book, I thought this was a wholly 21st-century quandary, created by new digital technology. From him, I learned that it is not.
Written work has been "automated" -- made from standardized templates, in which each instance will be just like the last -- for nearly two centuries. “Tables, schemas, skeleton forms, molds, patterns, matrices, and frames lie at the basis of all content production,” Yi Tenen writes. That includes not just sales brochures and government reports, but novels, plays and movie scripts. ChatGPT and its ilk aren't a sudden disruption of writing practice. They're the continuation of a centuries-long process that has brought automation to almost everything people do.
From ‘Plotto’ and other aides to ChatGPT
And, in the book's most mind-blowing passages, Yi Tenen, once a Microsoft engineer and now a professor of English at Columbia, describes remarkable finds amid archives of forgotten books and magazines. He shows that our supposedly novel digital tools to assist writers are the direct descendants of 19th century guides and gizmos designed to do the same thing.
One consequence of industrialization was that literacy rates increased, as did incomes and the physical infrastructure to make books, newspapers and magazines. There followed an "expansion of literary markets, with the corresponding emergence of popular genres--from self-help to guidebooks and travelogues to books on homemaking and home improvement, farmers' almanacs, how-to manuals, children's literature, pamphlets, true crime, detective fiction, and pornography." Rising demand for the product meant supply had to increase. And it did.
"The total number of titles published in the United Kingdom grew from almost 3,000 in the 1840s to more than 10,000 in the 1900s,” Yi Tenen writes. Similarly, “the total number of titles printed in the United States (including translations and reprints) grew tenfold, from around 2,000 in 1876 to 17,000 in the 1930s. " What was considered respectable output for a writer in the 17th century was paltry by the more demanding standards of the 19th. The market needed a lot more words.
Did writers make a killing? Foolish reader, writers never make a killing. Instead of paying writers more, society found a means to make each one produce more words. It was the same method used in other industries: Examine what people do with practice, intuition and muscle memory unique to the craftsperson — and turn that into explicit instructions that any person, or maybe even a machine, can execute.
On a nearby shelf as I write this is a book called "20 Master Plots," which I picked up somewhere, somehow, during one of my fiction-writing phases. Books like that, Yi Tenen explains, multiplied beginning in the 19th century as writers looked to master the procedures for churning out their works at a relentless pace. Books like “James Knapp Reeves's Practical Authorship (1900); The Technique of the Novel (1908) by Charles Horne; Writing the Short-Story: A Practical Handbook on the Rise, Structure, Writing, and Sale of the Modern Short-Story (1909) by Joseph Esenwein; Harriott Fansler's Types of Prose Narratives: A Text-Book for the Story Writer (1911); Henry Phillips's The Plot of the Short Story (1912); The Technique of the Mystery Story (1913) by Carolyn Wells; The Technique of Play Writing (1915) by Charlton Andrews; The Technique of Fiction Writing (1918) by Robert Saunders Dowst; Plots and Personalities (1922) by Edwin Slosson and June Downey; and William Cook's Plotto: A New Method of Plot Suggestion for Writers of Creative Fiction (1928), among many other examples." So did publications like Author, started in Boston in 1889 as "a monthly magazine to interest and help all literary workers."
There was also a boom in what we'd now call "knowledge management," as people wrestled with the problem of gathering material for their writing labor. Even Mark Twain cashed in with a patented "self-gumming" scrapbook for would-be authors to collect their clippings.
Yi Tenen quotes a writer of the 1930s describing this industrial writing approach to the craft, saying he followed a particular school, creating stories that "were all like Fords. They were of limited horsepower, neat, trim, and shiny, taking up very little road space, structurally correct and all following the blueprint without the slightest deviation." What Yi Tenen calls "template culture" was everywhere.
If your eyes glazed over reading the list of titles I just quoted, imagine reading all that stuff. But Yi Tenen’s patience and persistence were repaid with a mind-boggling discovery.
Among these guides for writers that Tenen found in the archives is a book by a Frnech playwright named Georges Polti, published in 1895. It's called The 36 Dramatic Situations, and it's yet another of these books offering templates for writers to assemble their works.
But Tenen discovered that Polti's book is unusual. It is an unacknowledged source for a book published 30 years later, Morphology of the Folktale, by a Soviet scholar named Vladimir Propp. "Propp's thirty-one ‘dramatic functions’ echo Polti's thirty-six ‘dramatic situations’ […] extensively and without attribution," Yi Tenen writes.
Why is this significant? Because Propp was an influence on later thinkers who set out to reverse the template-to-story operation. If a writer could follow a simple schema to construct stories that were superficially different, why couldn’t a scholar study superficially different stories to find the simple schema that underlies them all?
This quest for "hidden universal templates" came to be known as structuralism. One of structuralism's most profound impacts was on linguistics. There, beginning in the 1950s, Noam Chomsky began his search for the deep underlying rules of language -- the regularities that underlie all of its variations. Once a theorist had a grammar of such rules worked out, it could be tested by reversing the process. The rules of the grammar, when followed, should be able to generate the language it described.
These "generative grammar" tests were soon being done by computers. MIT had a computer doing that by 1957. It was not long before other computers were dedicated to producing stories. One of the more practical descendants of Propp's book, Yi Tenen writes, is a system with which airplanes produced incident reports about in-flight problems.
Large Language Models aren't an intrusion of the mechanical and industrial into a sheltered creative space. They're the latest form of an industrialization that began long ago.
In short, Large Language Models aren't an intrusion of the mechanical and industrial into a sheltered creative space. They're the extension of an industrialization process that began more than a century ago.
Which means, on the one hand, that their further development is likely to mirror what happened with other tools and technologies -- what now seems miraculous will become part of life's background, little noticed except when it screws up.
On the other hand, when a technology and its use are products of history -- that is, products of events that need not have happened, which depended on other events that need not have happened -- then we can be sure: This wasn't inevitable. It could have been otherwise. We have the kind of AI we have today because people made decisions in times and places where their decisions had an effect on others. Neither their discoveries nor their products were inevitable. And if they could have decided otherwise, that means we, today, can decide otherwise.
I don't mean that we should decide to chuck AI in some sort of Butlerian jihad (though I suppose we could). I mean that we could choose to use AI differently, or to have different uses for AI than we currently have. The message of this eye-opening book is that robots, AI and other tech are neither inevitable nor alien. They are things people made. So the right question to ask about AI isn’t how to cope with this alien thing. It’s what to do with this thing we’ve made.
This Week’s Briefs
1 Four-legged robot flamethrowers
You know that scene in the thriller where the idealistic scientist realizes that her perpetual motion machine or breakthrough DNA isn't going to be used to end hunger or make the blind see? Instead the soundtrack blares in alarm as she learns it's going into fighter planes or junk food or concentration camps.
I wonder if something like it might be playing out at Boston Dynamics, makers of the Spot quadraped, now that the Thermonator has gone one sale.
The Thermonator is a "robot dog" mounted with a flamethrower. According to its maker, Throwflame, its use cases include "wildfire control and prevention" (sic), agricultural management, ecological conservation, snow and ice removal, and entertainment. I'm not sure how serious the company is about all this, given its product video, where a voice intones that famous Oppenheimer quote about death, destroyer of worlds, right before the robot's 30-foot jet of flame blasts forth. It's not hard to think of how things just might possibly go wrong here, and I'll leave you to it.
The flamethrower is not mounted on a Spot, perhaps because Boston Dynamics has a strict prohibition on attaching weapons to its robots and no doubt because a basic Spot costs $75,000. But now that "robot dogs" have been around a while, other companies can make them more cheaply and with fewer restraints on their uses. As several people have noted, the Thermonator appears to be a flamethrower mounted on a the GO2 quadraped made by Unitree, a Chinese company. A GO2 costs $1,600, according to the Unitree website. Throwflame is selling its Thermonator for $9,420.
It's a reminder that we shouldn't put too much faith in the assurances of innovators that their product can't be used for bad ends. Because if their product works, it will have competitors who aren't on the same wavelength.
2 Police robot takes a bullet in the line of duty
A man with a rifle was holed up in a house in Barnstable, Massachusetts a few weeks ago, with a SWAT team surrounding him. State Police troopers sent in three robots, one of which was a Spot equipped with a camera, to locate the gunman inside the house. When the Spot found the suspect, he knocked it over and ran into another room. The Spot got back up and found the suspect who -- "with apparent surprise" that the robot could right itself -- knocked it down again and shot it.
BD is giving state troopers a new Spot and keeping this one "for research," according to the report in Police Magazine.
3 A one-philosopher chatbot
A generation ago you knew you were on the map when you got your own Wikipedia page. Perhaps in the future, the mark of success will be a chatbot dedicated to you, and you alone? In any event, there is now such a bot dedicated to the works of Luciano Floridi, the philosopher of information and AI ethics. You can try the "luflotbot" here.
4 Another day, another humanoid
Last week I accidentally cut out mention of another humanoid robot company -- Mentee Robotics, makers of the Menteebot. Mea culpa.
We would be so lucky to have a Butlerian Jihald.