Literary Theory for Robots

How Computers Learned to Write

DENNIS YI TENEN
CONTENTS
CHAPTER 1:
Intelligence as Metaphor (An Introduction)

CHAPTER 2:

Letter Magic

CHAPTER 3:

Smart Cabinets

CHAPTER 4:

Floral Leaf Pattern

CHAPTER 5:

Template Culture

CHAPTER 6:
Airplane Stories

CHAPTER 7:

Markov’s Pushkin

CHAPTER 8:
9 Big Ideas for an Effective Conclusion

NOTES

INDEX
CHAPTER 1

Intelligence as Metaphor (An

Introduction)

C OMPUTERS LOVE TO READ. And it isn’t just fiction before going to bed.
They read greedily: all literature, all of the time—novels,
encyclopedias, academic articles, private messages, advertisements, love
letters, news stories, hate speech, and crime reports—everything written
and transmitted, no matter how insignificant.
This ingested printed matter contains the messiness of human wisdom and
emotion—the information and disinformation, fact and metaphor. While we
were building railroads, fighting wars, and buying shoes online, the
machine child went to school.
Literary computers scribble everywhere now in the background, powering
search engines, recommendations systems, and customer service chatbots.
They flag offensive content on social networks and delete spam from our
inboxes. At the hospital, they help convert patient-doctor conversations into
insurance billing codes. Sometimes, they alert law enforcement to potential
terrorist plots and predict (poorly) the threat of violence on social media.
Legal professionals use them to hide or discover evidence of corporate
fraud. Students are writing their next school paper with the aid of a smart
word processor, capable not just of completing sentences, but generating
entire essays on any topic.
In the industrial age, automation came for the shoemaker and the factory-‐
line worker. Today, it has come for the writer, the professor, the physician,
and the attorney. All human activity now passes through a computational
pipeline—even the sanitation worker transforms effluence into data. Like it
or not, we have all become subjects to automation. To survive intact, we
must also learn to become part software engineers and part, well—whatever
you are doing is great!
If any of the above comes as a surprise to you, my job, I feel, is mostly
done. Curiosity piqued, you will now start noticing literary robots
everywhere and join me in pondering their origins. Those not surprised
perhaps believe (erroneously!) that these siliconites have learned to chat
only recently, somewhere on the fields of computer science or software
engineering. I am here to tell you that machines have been getting smarter
in this way for centuries, long before computers, advancing on far more
arcane grounds, like rhetoric, linguistics, hermeneutics, literary theory,
semiotics, and philology.
So that we may hear them speak—to read and understand a vast library of
machine texts—I want to introduce several essential ideas underpinning the
ordinary magic of literary computers. Hidden deep within the circuitry of
everyday devices—yes, even “smart” light bulbs and refrigerators—we will
find tiny poems that have yet to name their genre. By the end of the book, I
hope you can see this technology anew, replete not just with instrumental
capacity (to keep food cold or to give light) but with potential for creativity
and collaboration.
Throughout, we will be tempted to ask existential questions about the
nature of artificially intelligent things. “How smart are they?” “Do they
really ‘think’ or ‘understand’ our language?” “Will they ever—have they
already—become sentient?”
Such questions are impossible to answer (in the way asked), because the
very categories of consciousness derive from human experience. To
understand alien life forms, we must think in alien ways. And rather than
argue about definitions (“Are they smart or not?”), we can begin to describe
the ways in which the meaning of intelligence continues to evolve.
Not long ago, one way of appearing smart involved memorizing a bunch
of obscure facts—to become a walking encyclopedia. Today, that way of
knowing seems like a waste of precious mental space. Vast online databases
make effective search habits more important than rote memorization.
Intelligence changes. The puzzle of its essence cannot therefore be
assembled from sharp, binary attributes, laid out always and everywhere in
the same way: “Can machines think: yes or no?” Rather, we can start
putting the pieces together contextually, at specific times and places, and
from the view of an evolving, shared capacity: “How do they think?” and
“How do we think with them?” and “How does that change the meaning of
thinking?”
In answering the “how” questions, we will discover a strange sort of a
twinned history, spanning the arts and sciences. Humans have been thinking
in this way—with and through machines—for centuries, just as they have
been thinking with and through us. The mind, hand, and tool move at once,
in unison. But the way we train minds, hands, or tools treats them almost
like entirely separate appendages, located in different buildings, in
unrelated fields on a university campus. Such an educational model isolates
ends from means and means from ends, disempowering its publics. In this
volume, I would like to imagine an alternative, more integrated curriculum,
offered to poets and engineers alike—bound, eventually, for a machine
reader as part of another training corpus.
Next time you pick up a “smart” device, like a book or a phone, pause
mid-use to reflect on your body posture. You are watching a video or
writing an email perhaps. The mind moves, requiring mental prowess like
perception and interpretation. But the hand moves, too, animating body in
concert with technology. Pay attention to the posture of the intellect—the
way your head tilts, the movement of individual fingers, pressing buttons or
pointing in particular ways. Feel the glass of the screen, the roughness of
paper. Leaf and swipe. Such physical rituals—incantations manifesting
thought, body, and tool—bring forth the artifice of intellect. The whole
thing is “it.” And that’s already kind of the point: Thinking happens in the
mind, by the hand, with a tool—and, by extension, with the help of others.
Thought moves by mental powers, alongside the physical, the instrumental,
and the social.
What separates natural from artificial forces in that chain? Does natural
intelligence end where I think something to myself, silently, alone? How
about using a notebook or calling a friend for advice? What about going to
the library or consulting an encyclopedia? Or in conversation with a
machine? None of the boundaries seem convincing. Intelligence demands
artifice. Webster’s dictionary defines intelligence as the “skilled use of
reason.” Artifice itself stems from the Latin ars, signifying skilled work,
and facere, meaning “to make.” In other words, artificial intelligence just
means “reason + skill.” There are no hard boundaries here—only synergy,
between the human mind and its extensions.
What about smart objects? First thing in the morning, I stretch and, at the
same time, reach for my phone: to check my schedule, read the news, and
bask in the faint glow of kudos, hearts, and likes from various social apps.
How did I get into this position? I ask alongside the beetle from Kafka’s
The Metamorphosis. Who taught me to move like this?
It wasn’t planned, really. Nor are we actually beetles living in our natural
habitat, an ancient forest floor. Our intimate rituals morph organically in
response to a changing environment. We dwell inside crafted spaces,
containing the designs for a purposeful living. The room says, Eat here,
sleep there; the bed, Lie on me this way; the screen, Hold me like this. Smart
objects further change in response to our inputs. To do that, they must be
able to communicate: to contain a layer of written instructions. Somewhere
in the linkage between the tap of my finger and the responding pixel on-‐
screen, an algorithm has registered my preference for a morning routine. I
am the input and output: the tools evolve as they transform me in return.
And so, I go back to bed.

THE CHAPTERS OF THIS book correspond to what I see as the three major gaps
in our collective thinking about artificial intelligence. First, I will argue that
our understanding of AI can and must become more grounded in the history
of the humanities. Second, that a software-engineering education
necessarily involves a number of inescapable philosophical problems,
related to several nerdy-but-I’m-sure-you’ll-find-them-fascinating kinds of
textual matters. Finally, I propose to consider AI as part of the broader
social and political dynamics arising, more generally, from the automation
of labor, particularly as it concerns intellectual labor or “knowledge work.”
History will remain my first and primary concern. Insert something clever
here about forgetting the past to the detriment of the future. Any field
developing rapidly, as computer science has been, can be forgiven for
orienting itself toward the future. But not for long. Yes, one can learn to
type quickly or to write a spreadsheet macro without a second thought.
Reflect, however, on the immense frustration of commandeering all of our
clumsy, impoverished digital avatars. You want me to save a what, where?
File? Folder? Server? Cloud? The pervasive presence of metaphors should
hint at the extent of our collective alienation. What is a file? Wherefore a
folder? Why are they even called that? Who thought of putting exactly eight
bits into a byte, and why not nine? How does a keystroke transform itself
into a configuration of pixels on screen, forming a letter? Explain a
common algorithm used to search, finish sentences when typing, filter
spam, or to recommend products. No way. Too much. Such questions fall
out of serious consideration because, paradoxically, they are “too simple”
for computer science and way, way too complex for baseline literacy.
But first, why do history? One can simply launch into the future from the
state-of-the-art, the state-of-the-present. At least two points are needed to
make a line, however. History’s subtle advantage lies in the revealing of
trajectories through time. AI and indeed “artifice” or “intellect” (and even
the toasting of bread) represent a cluster of related objects, ideas,
techniques, practices, and institutions. A static snapshot contains their
anatomy in three dimensions—this is how modern toasters work and this is
how a mind thinks today. History expands these arrangements into the
fourth dimension of time. Like any technology, the mind changes over time.
Its anatomical hardware stays the same, but the software receives regular
updates: all those encyclopedias, calculators, search tools, and other smart
devices that amplify the powers to reason.
Yet this history of AI should have nothing to do with the recent
resurgence of covertly neo-Hegelian thought so popular with futurists and
disruptive entrepreneurs, who believe in some sort of a teleological “event
horizon,” past which humankind will attain a posthuman singularity (all
challenges met, competitors bested, obstacles overcome). A singularity that
leads to the end of history sounds both contemptible and horrific—I’m not
sure if humans would survive it. History cannot be personified in that way.
It isn’t leading us to salvation or transcendence. It has no post-human goals
as such, other than those we make for ourselves.
On a less apocalyptic scale, history allows for extrapolation. It helps
anticipate changes in the direction of travel. Without history, our ideals and
our policies lack sufficient foresight to prepare adequately and to alter
course if needed. Without history, society is doomed to struggle with long-‐
term effects of technology, such as pollution and urban sprawl, and soon, if
not already, AI.
History ultimately makes for better engineers and users of technology.
Any complicated gadget contains within it a number of contingencies that
appear senseless today because they were attained through a compromise
with past limitations.
QWERTY, the layout of modern English keyboards, for example, was
famously arranged in that order to slow down typing on typewriters that
jammed otherwise. Similarly, eight bits (ones and zeroes) make a byte (a
unit of information) by convention. The number eight holds little
significance. The arbitrary length always confused me, until I learned that
early telegraph systems also experimented with variable-bit encoding, such
as the Morse code (which can have anywhere between one and five bits), as
well as fixed encoding, like the five-bit Baudot code, deriving from earlier
cryptographic cyphers proposed by Francis Bacon and others in the early
seventeenth century. The linear ticker tape used in telegraphs subsequently
could break a string of punches into any arbitrary length. It just so happens
that the rectangular punch card in use by early tabulating census machines,
and later by IBM, would not physically fit more than eight characters
across, in rows of seventy to eighty columns stacked. The convention of an
eight-bit byte (and the eighty-character code line) persisted, even though
magnetic tape and later “hard drive” media were no longer constrained by
those limitations. The physical properties of cardboard thus continue to
inform the design of modern computers. The abstract and alien concept of
“bits” became concrete to me by imagining holes punched through
cardboard. On a recent trip to Kyrgyzstan, a friendly librarian gifted me a
punch card used by Soviet computers in 1977. I use it as a bookmark now,
but its neat rows and columns still help conjure the rather elusive “byte” as
an object.
History gives meaning to such inherited structures. In following the
evolution of technology, we are able to make sense of legacy technological
debt. Whatever is meant by “innovation” consists of realizing which
features of the inherited design remain necessary and which can be
discarded or improved. Without history, the present becomes invariable,
sliding into orthodoxy (“it is what it is”). History’s milestones mark a path
into a possible future.
It’s okay, then, to sometimes struggle with basic file operations alongside
some of my best data-science students. Believe in yourself, Dennis! How
can I consider myself literate if I don’t fully understand how writing works?
A technical answer isn’t enough, either. “How writing works” cannot be
reduced to the present technological moment, because that moment will
pass, and rapidly. Instead, writing “has come to mean” this specific
arrangement of mind, body, tool, circuit, bit, gate . . . through its historical
development. An obtuse technological concept, such as solid-state memory
storage (the quantum shimmer behind words on screen), obtains meaning
only through a specific genealogy: from ticker tape to punch card and to the
NOT-AND gates flickering silently in the palm of your hand.
At the very least, the erasure of such realities stunts intellectual growth.
Good engineers need history, because the progress of their education can
follow the gradual development of their subject matter. History leads to
deep understanding because it explains why and how something has come
to be the way it is. Without it, we passively inherit a bundle of meaningless
facts: it’s just the way things are.
Poets steeped in engineering likewise stand to understand something vital
about the nature of writing (sign, symbol, inscription, representation, if you
want to get fancy). From that perspective, writing machines can be seen as
the culmination of a long, mythical tradition: In the beginning, there was
the letter. Then, a byte came to encode eight bits, because eight was enough
to represent a character. Characters formed words and commands in a
string. The string ate its own tail and learned to rewrite itself. These were
the early days of autopoiesis, the writing that writes itself. What we teach
today as the history of literature is also but an episode of a larger emerging
storyline. Its continuing evolution must involve the people of the word.
Each turn of the historical screw brings wonder, quickly fading into the
ordinary. Who remembers the ninth-century Persian mathematician
Muh˙ammad ibn Musa al-Khwarizmi, still with us in his al-gorithms? What
of Ramon Llull, the Majorcan monk, who studied rhetorical combination in
the thirteenth century? Did he invent one the earliest chatbots with his
rotating paper charts? Or do their origins lie in ancient divination charts,
like the zairajah described by the great medieval historian Ibn Khaldun, or
the Yi Jing (Book of Changes), written in the Western Zhou period?
Can I tell you also that Llull (pronounced yooy) inspired Francis Bacon
and Gottfried Leibniz to propose their binary cypher systems, influenced
also by the Chinese tradition? What about the Wunderkammern of the
German baroque poets in the seventeenth century? These curious drawer
cabinets—literally furniture—could be pulled in any arrangement to
produce beautiful music and poetry, to the delight and horror of the
audience. A long tradition of language machines also includes An Essay
Towards a Real Character, and a Philosophical Language by John Wilkins.
Written in 1668, the book imagines a new system of writing that would
function as a kind of a translation protocol between cultures, commercial
interests, and religions. Later, the mathematical notion of a Markov chain,
central to contemporary AI, developed out of a paper on Pushkin’s poetry.
These mostly forgotten artifacts remain in the background of
contemporary computation. Later, we’ll have the chance to pluck a few of
them out—unplug and hold them for a bit to see how the whole thing fits
together.
In writing this book, I was surprised to discover the history of computing
permeated by distinctly literary puzzles. But to see a past shared between
literature and technology has implied, also, the recognition of their altered
present. Technology continues to reconfigure human orientation to the
word. To study the history of literature today is to go beyond the familiar
categories, like fact or fiction, or even science and the humanities, art and
technology. Instead, our thread leads to an entirely separate branch of the
symbolic arts, practiced on a massive scale, on the work floors of
intellectual labor, across diverse industries. Here, we can find stories
produced by the MITRE Corporation, one of the largest military contractors
in the world, woven deep inside the Cold War–era missile defense systems;
encounter fairy tales written by airplanes as part of Boeing’s effort to
automate their incident-reporting systems; or visit the “vector space”
supporting everyday search-engine recommendations. To what genre do
these texts belong? On which bookshelf should we file them? How do we
read something so fundamentally unfit for human consumption?
The advance of artificial intellect obviously implicates a number of
open-ended metaphysical questions. From these, I want to highlight a
subset related to the subject of my expertise: literature, poetics, semiotics,
philology, textual studies, the study of inscription and incantation, e.g., et
al., etc.—the arcane arts.
Except not really that arcane, because though people these days read
fewer things labeled “Literature,” they consume more literature than ever.
Text weaves through the online digital world, stitching it together. It bonds
and bounds whatever is meant by artificial intelligence, in the way an
encyclopedia or a library holds the sum total of human knowledge. The
search engine or chatbot is just an outer garment, draped over the
complication within.
A body of written works sustains all organized human activity. I don’t
mean to suggest that all things are textual. The world becomes truly visible
to a literary scholar only through the looking glass of inscription. For
example, unlike your well-adjusted physician, I know little about and am
uncomfortable around bodies. A patient exists for physicians as a living
body first, but that’s not all a patient becomes in a hospital: Patient histories
(stories) may be recorded to supplement a physician’s notes. A recording of
their consultation will be transcribed and transcoded into a digital file. It
will be translated from one system to another, in multiple formats. It will be
redacted, edited, compressed, processed, and mined for missing billable
codes. This record will be sold and resold and used to train automated
diagnostic bots of the future. It will be summarized, cataloged, tagged, and
archived for later use. A body, in the sense of the literal Latin corpus, has
thus been transformed into the English corpus: figuratively, a collection of
textual records.
Much of this churn implicates automated, artificial, and otherwise
intelligent work. But neither the physician nor the hospital administrator
may consider it as such. Though a patient’s journey from body to record
leads directly to health outcomes (and to a hospital’s finances), it remains
without a guide, meandering through multiple, inherited legacy systems.
Here, a sharp-eyed arcanist may find a worthwhile mystery: Finally, a
puzzle for those who have studied exactly these sorts of things, like
corpuses and taxonomies, editions and redactions, catalogs, summaries,
translations, and archives—a rich trove of textual goodies!
In this book, we will similarly observe the history of artificial intelligence
through the looking glass of inscription. History tells us that computers
compute not only in the mathematical sense but universally. The number
was incidental to the symbol. In the 1840s, Ada Lovelace, daughter of Lord
Byron and one of the first “programmers” in the modern sense, imagined an
engine that could manipulate any symbolic information whatsoever (not just
numbers). A century later, Alan Turing extended that blueprint to imagine a
“universal machine,” capable of reading from and writing to an infinite
string of characters. The children of Turing and Lovelace occupying our
homes are therefore expert generic symbol manipulators. They are smart
because they are capable of interpreting abstract variables, containing
values, representing anything and everything.
Modern machines read and write. And because they carry language, we
readily imagine them capable of higher-order functions like sympathy or
sentience. But how do machines grow from handling inert textual matter to
living language? Why do they seem to exceed their programming by telling
jokes, making complex logical inferences, or writing poetry?
Metaphor holds the key to their slow awakening. Algebraic values can
signify human values: x = y arranges two variable entities into a familiar
relation; x can also be assigned to ideas like “cat” or “dog,” or an action like
“call home” or “order milk.” The art of manipulating symbols leads to a
game of metaphors in which certain inscriptions or gestures stand for
something else. Children, spies, and poets love to play games like these.
When I call (after Shakespeare) the night “thick,” I mean it to be
particularly dark, humid, and full of anticipation. The metaphor packs a lot
of meaningful baggage into a compact transport. Similarly, a machine
becomes smart by “rising higher and higher through the circles of spiraling
prose” (thanks, Winfried Georg Sebald!). Symbol emerges from inert
hardware, gesturing toward things that we care about, like flagging
offensive content, deleting spam, fighting terrorism, or indeed shopping for
shoes.
Care is also a value. But can machines really soar (crawl?) from simple
values like x and y to complex ones like care, equity, freedom, or justice?
Reader, are you human? Is there anything inside? Can machines think, or
are you just trolling me?

WE’LL GIVE SUCH TEXTUAL-PHILOSOPHICAL distractions a good go throughout

the book. In conclusion, we will move to a few practical, that is to say
political and economic and even personally psychological, considerations.
Insert a wise, pithy aphorism about the social good, career choice, the
means of production, and the future of creativity. This is the “So what?”
part. How will my life change and what is to be done?
The mistake, if I may spoil the punchline, was ever to imagine
intelligence in a vat of private exceptional achievement. Thinking and
writing happen through time, in dialogue with a crowd. Paradoxically, we
create new art by imitating and riffing off of each other. Trained on pretty
much “everything ever published,” AI is just another way to give that
collaborative a voice. It does not, as we will see, amount to one thing but
many: a talking library, a metaphor, a personification of a chorus. If we are
to call it “intelligent,” it is intelligent in the way of a collective. It
remembers like a family does. It thinks like a state. It understands like a
corporation.
Consider the self-driving car. An automobile consists of many connected
subsystems: drivetrain, suspension, braking, steering, cabin control, and so
on. Modern cars also contain numerous sensors, which monitor everything
from speed to ambient temperature and driver alertness. Many of these
subsystems exhibit intelligence in their own right. Take the humble
“automatic transmission,” for instance. They call it “automatic” because,
unlike older tech, it automates important decisions for the driver. And it’s
actually quite clever! Unlike manual transmissions, most automatics won’t
allow drivers to take a dangerous action, like reversing mid-highway drive.
(Please don’t try this anyway.) The automatic will even shift to a lower gear
for better braking performance when driving downhill. In bad weather, it
can also send less power to the wheels, improving driver safety. Today, this
unremarkable technology performs better than most human drivers.
 Automatic transmissions have been around for more than a hundred
years, becoming more common in the 1960s. Imagine yourself driving an
automatic back then for the first time. It must have been incredible—a
technology that would make driving easy, safe, and accessible. “How
artificially intelligent!” you would’ve said.
Nobody said that, I know. Whatever magic there was, dissipated. Just a
while ago, you may remember having been impressed with your
automobile’s global positioning system (GPS). We drive by the stars! Or
how about rear-view cameras? And self-parking! Eventually, carmakers
hope to arrive at something like “Level 5: ‘Full Driving Automation,’ ”
involving “sustained and unconditional [. . .] performance [. . .] without any
expectation that a user will need to intervene” (as specified in J3016-‐
JUN2018, Surface Vehicle Recommended Practice, by the Society of
Automotive Engineers). Yesterday’s remarkable automations have been left
behind at Level 1, “driver assistance,” limited to specific “operational
design domains,” alongside cruise control, adaptive suspension, and
automatic windshield wipers.
As the examples above illustrate, intelligence evolves on a spectrum,
ranging from “partial assistance” to “full automation.” Many everyday tasks
simpler than city driving have been automated somewhere along that
spectrum. Your average kitchen toaster, for instance, requires no input
except bread. A user intervenes only to start the device and to consume its
delicious outputs. Granted, the procedure required for the task—an
algorithm—wasn’t very complex to begin with: lower bread, fire, eject. By
automotive standards, our toasters have probably achieved Level 5
automation already, or at least Level 4, where all toasting tasks are
automated under specific circumstances of the kitchen, with human
override for an occasional bread malfunction. Hopefully soon they will be
on their way to greater achievements, like spamming our phones with
photos of burnt toast.
I don’t know much about cars or toasters, to be honest. Rather, I’m
interested in the appearance of a small imaginary human from behind the
curtain. Why, for example, does an automatic transmission seem like a
boring tool, whereas a drug-discovery algorithm was recently described by
Henry Kissinger as a thing able to “master its subject,” “devise new
strategies,” and even “detect aspects of reality humans have not detected, or
perhaps cannot detect.” How does a mere tool move into the subject
position of the sentence—where it detects, devises, and masters—gaining a
sense of agency and interiority in the process?
Plenty of smart things don’t get to do that. Other less-intelligent ones
readily conjure the image. Savvy marketeers understand that self-driving
cars and smart houses make for compelling subjects, while automatic
transmissions and toasters do not. Emergence somehow happens by the
critical accumulation of smarts, where intelligence sometimes bubbles up,
pulling along with it the more intangible silk of sentience, awareness, and
conscience.
Leaving the world of culinary and automotive hardware, I return to the
bundle of specifically textual technologies, such as chatbots, machine
translators, named entity extractors, automatic text summarizers, speech
recognizers, spam filters, search engines, spell-checkers, classifiers, auto-‐
taggers, sentence autocompletion, and story generators. (Let’s not forget,
too, that whatever is meant by “AI” includes other, non-verbalizing tech as
well, from weather-prediction models to drug-discovery algorithms,
robotics, and image classifiers.)
 Robots that chat impress us in a particularly satisfying way, possibly
because language always hints at a worldview (a mind?) within. Modern
chatbots don’t contain a mind as such. Or, rather, they express minds
metaphorically, through a fuzzy, loose analogy with human brains. Yet,
without knowing much about the world at all, they are capable of
synthesizing information, responding to questions, and even writing
college-level essays.
How did this sleight of hand come to be and why? Once we see
intelligence on a spectrum, integrated within almost every domain of human
life, we can begin to unwind the metaphor of its collective “mental”
development, by stages and significant milestones. In the domain of
language, we can draw a line between fully automated text generators back
toward familiar tools like spell-checking and word autocompletion, and
further to long-standing historical practices and algorithms.
Sufficiently inspired by these preliminaries, I am now ready to discuss the
historical frays of textual technology proper. Don’t get hung up on the
imagery of pioneers or milestones, by the way. Who did what first or last is
irrelevant and usually misguided. Al Gore did not invent the internet. Ideas
come into being as physical things with great effort, through multiple
parallel iterations. For this reason, the several lengthy strands I picked out
for you in the next chapters are merely representative of a trend. Each
draws its own particular trajectory—a magical incantation—which we can
retrace and recite together.
CHAPTER 2

Letter Magic

WHEEL

Let me let you in on a little secret: koldun, the word meaning “sorcerer” in
several Slavic languages, arguably originates from the name of one of the
most prominent medieval (fourteenth-century) scholars, Abd ar-Rah ˙ man
ibn Muh ˙ ammad ibn Khaldun al-Hadhrami—the son of eternity from
Hadhram—or simply Ibn Khaldun. In his epic history of the world,
modestly titled, Muqaddimah or “Introduction,” Ibn Khaldun documented
the use of zairajah—“a remarkable technical procedure [. . .] for alleged
discovery of the supernatural”:

The form of the zairajah they use is a large circle that encloses other
concentric circles for the spheres, the elements, the created things, the
spiritualia, as well as other types of beings and sciences. Each circle is
divided into sections, the areas of which represent the signs of the
zodiac, or the elements, or other things. The lines dividing each section
run to the center. They are called chords. Along each chord, there are
sets of letters that have conventional (numerical) value [. . .] Inside the
zairajah, between the circles, are found the names of the sciences and
of topics of the created (world). On the back of (the page containing)
 the circles, there’s a table with many squares, fifty-five horizontally and
one hundred and thirty-one vertically. Some of the squares are filled in,
partly with numbers, and partly with letters. Others are empty.

Ibn Khaldun further described several such devices, giving us a glimpse

into their “remarkable techniques.” In short, a learned soothsayer would
write down a question, converting its letters into numbers. These would
then be transposed back into letters, by consulting a number of intricate
charts, according to “well-known rules” and “familiar procedures.” Several
such computational “cycles” would produce a shortened string of letters,
which could finally be expanded into a rhymed answer. With proper
training, the zairajah could obtain the “knowledge of the unknown, from the
known,” Ibn Khaldun explained, giving the intellect an “added power of
analogical reasoning.”
Despite his enthusiasm for the device, Ibn Khaldun also understood its
limitations. The zairajah revealed analogies implicit in the question itself,
through a relationship between an “arrangement of letters” and another
“arrangement of letters.” No amount of verbal combination could attain a
“conformity of the words to the outside (world).” “We have seen many
distinguished people jump (at the opportunity for) supernatural
discoveries,” Ibn Khaldun wrote. “They think that correspondence (in form)
between question and answer shows correspondence in actuality. This is not
correct.” The answers rather revealed a logic implicit in the language itself.
Truth value lay outside of language. The wheel’s answers “remained
veiled,” requiring further verification. “One should not think that one can
get to the secret of the letters with the help of logical reasoning,” Ibn
Khaldun insisted. “One gets to it with the help of vision and divine aid.”
Modern, algorithmic soothsaying contains the same defect. Any artificial
language system may at any moment lose its grasp of the real world and
begin hallucinating or fabricating imaginary facts.
For his own illustrative purposes—to give you an idea of fourteenth-‐
century AI performance—Ibn Khaldun used the zairajah to pose a self-‐
referential question (typical of the genre today, also). “We wanted to know,”
he asked, “whether the zairajah was a modern or an ancient science.” The
question was then broken down into ninety-three chords, computed through
twelve cycles on the wheel, in a procedure that took the author several
manuscript pages to describe in detail, producing the following rhymed
answer: “The Holy Spirit will depart, its secret having been brought forth /
to [the prophet] Idris, and through it, he ascended the highest summit.”
What are we, modern online shoppers, to make of ancient “letter magic,”
as Ibn Khaldun called it? A few ideas come to mind.
First, pretty impressive! Though difficult to follow, the method produced
an intelligible and intelligent answer. One suspects the involvement of some
cherry-picking and hand-waving, but so is the case with AI today, seven
centuries later. From ELIZA—the famous chatbot therapist meant to prove
how easy it was to fool humans—to The Policeman’s Beard Is Half
Constructed—a collection of algorithmic poetry that seemed
groundbreaking at the time it was published in the 1980s—scholars have
often plucked the few clever bits from reams of incomprehensible
logorrhea.
Second, whoa, these things are old! Zairajah exemplifies a number of
divination techniques extant in the medieval Muslim world, but also found
in antiquity, going back to Greek, Chinese, Egyptian, and Sumerian roots.
Text generators turn out to be as old as written language itself.
Third—okay, wading into the weeds now—from its earliest documented
use, letter magic promises to reveal a logic inherent in language itself,
seeming like a shortcut to hard-earned scientific knowledge (vision and
unveiling, in medieval scholarly terms). And because the various people of
the letter considered their textual traditions sacred, procedural word
manipulation was always proximate to magic, if not outright divinity.
Modern AI systems are often discussed similarly in the context of
ineffable mysteries such as consciousness, creativity, or, Khaldun-forbid,
the singularity (an irreversible event horizon at which AI transcends its
human origins to become Nietzschean superintelligence). Then as now,
among the philosophers and entrepreneurs eager to jump on the bandwagon
of AI, we find plentiful conmen and charlatans.
Finally, note that even the most advanced modern, language-based AI
systems today continue to have a problem with external validation—an
issue already well understood by Ibn Khaldun and his contemporaries.
Procedurally generated text can make grammatical sense, but might not
always make sense sense. Noam Chomsky, the MIT linguist, famously
expressed this paradox in this sentence: “Colorless green ideas sleep
furiously.” Derived from Chomsky’s procedural grammar, the sentence may
be correct according to the rules of language. But to know that ideas don’t
sleep, or that one does not sleep furiously, requires experience in the
physical world.
Letter magic is still practiced in the prison of language. No matter how
complex the operations, words generate only more words. Their truth or
falsity obtains through lived experience. This significant limitation
precludes the idea of something like “ethical AI,” because language itself
contains within it all possible values, good and bad. External constraints are
necessary to compel it toward goals that exist (and evolve) outside
language, in the world that includes, you know, all the other stuff, like guns
and germs and capital and steel. Combinatorial language games generate
sentences without limit. Ethics require limitations like pain, illness, loss,
and death.

TABLE

So far, we’ve managed to dip our toes (fall off a cliff?) into the magical
waters of soothsaying and divination, in considering the medieval Arabic
Q&A bot—a set of “remarkable techniques” for language manipulation.
A careful examination of the zairajah and similar devices reveals several
distinct attributes, each worthy of its own little genealogy.
Note, for example, that despite its circular shape, the divination wheel can
also be seen as a type of table or chart. Charts aren’t as simple as they
appear. Unlike the messiness of presented information, on any topic, a chart
boils things down to their essentials. Charts are therefore by nature more
compact and more portable than the source of their original, non-processed
data. This property will become important later, when we consider their
capacity to encapsulate and transfer intelligence removed from its origins.
At the very least, a table presents information in a systematic way, usually
containing rows and columns. A chart implies order, however simple.
Here’s my fake shopping list, for example:
 Product Price Quantity Type
Oatmeal $4.99 1 package Steel-cut
Chocolate $3.48 3 bars Dark
Blueberries $9.75 2 bags Frozen
Milk $3.15 ½ gallon Whole

Aside from being impressed with my organizational skills and passion for
antioxidants, an alien visiting our planet for the first time would learn that
foodstuffs fall under the category of “products,” and that products contain
such attributes as “price,” “quantity,” and “type.” A table therefore already
implies a sense of hierarchical labels, or a taxonomy. Moreover, the
taxonomy is “controlled,” in the sense of containing a limited number of
allowed values. Of course, I am not so formal with my shopping lists. You
would be surprised, however, to see milk requested in units of “cow”
instead of “gallons,” for instance, or blueberries of the type “decorative.”
By limiting the available language choices, a taxonomy attempts to capture
the relationship between things in the world, where products have prices
and chocolate is sold by the bar.
Taxonomies of all types were synonymous with medieval science, as well
as being an important topic in contemporary logic, medicine, law, and
computer science. Consider a field like medicine, where the infinity of
analog bodily sensations must be distilled into a controlled set of named
and billable conditions—a diagnosis. The American Psychiatric Association
publishes the infamous Diagnostic and Statistical Manual of Mental
Disorders, now in its fifth edition (DSM-5). Your insurance is unlikely to
cover a health issue not named in the manual. Conversely, aspects of the
human condition now considered in the range of normal (neurosis,
homosexuality) were once unjustly included, to the detriment of the people
they labeled. A part of ancient arcane arts, taxonomies still rule the world.
Think of electronic databases working in the background of every hospital
as giant zairajah circles, spinning a yarn that connects patients, physicians,
pharmacies, and insurance companies.
The greatest popularizer of divination circles in the medieval Western
world was Ramon Llull, who, in the wake of the Christian reconquest of
Muslim Spain, both borrowed and extended the system to create his “new
logic.” The Art, as he called it, made extensive use of rotating spheres that
could, by the use of a complex notation, represent knowledge about
anything and everything. In addition to being portable and categorical,
Llulian logic was symbolic, rule-based, and combinatorial.
Let’s take these individual properties slowly and by turn.
Start by examining the following partial table transcribed from Ars
Brevis,

2nd
1st Wheel Questions Subjects Virtues Vices
Wheel
B goodness difference whether? God justice avarice
C greatness concordance what? angel prudence gluttony
E power beginning why? man temperance pride
H virtue majority when? vegetative charity ire

Note that, initially, the symbols BCEH in the first column could be
“overloaded” or have multiple meanings, symbolizing questions, subjects,
virtues, or vices. For instance, B could, depending on its position, stand for
“goodness,” “difference,” “whether?”, “God,” “justice,” or “avarice.” Llull
explained that:

we have employed an alphabet in this Art so that it can be used to make

figures, as well as to mix principles and rules for the purposes of
investigating the truth. For, as a result of any one letter having many
meanings, the intellect becomes more general in its reception of the
things signified, as well as in acquiring knowledge.

Pragmatically, the overloading of the lettered variables made the charts

more universal. Had they contained only a single meaning, they could have
been used for a single purpose only. But Llull’s art worked equally well for
theology as for medicine.
Further, the arrangement of letters in figures rigidified the procedure of
composition. The shape of a wheel or a tree, common in medieval
documents, made only certain of the combinations possible. Where Ibn
Khaldun documented pages upon pages of complicated procedure, Llull
crafted a visual mechanism, consisting of two nested wheels, rotating freely.
In use, the mechanism fixed the range of possible symbolic combinations to
those that produce only the valid results. “Everything that exists,” Llull
explained, “is implicit in the principles of this figure, for everything is
either good or great, etc., as God and angels, which are good, great, etc.
Therefore, whatever exists is reducible to above-mentioned principles.” In
other words, rotating the wheel would align entity “angel” with attribute
“prudent” or “just” but never “evil,” generating all true combinations, and
never the false ones.
Finally, the features above enabled a combinatorial process, by which all
the possible inputs could be combined to generate all possible outputs.
These, Llull documented in another “table of tables,” containing values like
BDTC, CETC, EGTF, FTEG, HITG, and so on. By such means, the system
was meant to complete gaps in existing knowledge: If God is Great, God is
also Good and Powerful and Virtuous. And if you didn’t realize a few of
those before, now you did.

Zairajah—medieval Arabic divination circle used for astrology by

Ahmad al-Buni as described by ibn-Khaldun in his Muqaddimah
(Introduction). Ah˙mad ibn ‘Alī al-Būnī, Shams al-ma‘ārif al-‐
kubrá (1874), 17.

General Collection, Beinecke Rare Book and Manuscript Library, Yale University.

Heady stuff. Imagine how impressive and exciting Llullian complications

must have seemed to the learned folk of the time! Particularly if you were
the average sort of a monkish scribe, somewhere in young Spain or
Germany. You never did learn to read or write that well. Your job was to
transcribe manuscripts in the local parish. You did so mechanically, without
really following the complicated theological argument at hand.
 Until—what’s this?—a neat structure interrupts the dense wall of Latin
text. “This table is the subject, in which the intellect becomes universal,”
you read. The name of Llull came across your carrel before. You heard
stories of his study of Arabic and his ministry among Andalusian Muslims.
“By said table, it [the intellect] understands and abstracts many particulars
objectively,” the text reads, “and this it does by applying to each question
one argument being extracted from each compartment of a given column.”
Here, you manage to glimpse a pathway into the inner sanctum—a step-‐
by-step method for discovering truth. “The intellect banishes doubt,
remaining in it calmly and positively,” you read along with Llull. “It now
knows itself to be completely general and artful, and clothed in great
knowledge.”
Calm, positive, and clothed so, I am thoroughly exhausted of possibilities.
Before we are lulled to sleep, however, notice the transpiring sleight of
hand. Let’s follow the original Latin, instead of the usual published
translation.
How did the intellect come to know itself? What does that even mean?
The table (tabula) is the subject, to repeat after Llull, in which the intellect
fashions itself (se facit) as a universal, by understanding (intelligit) and
abstracting from the particulars. What a strange way of putting things! The
table captures a disembodied intellect. There, from each cell (camera) of a
column, it derives (abstrahitur) a calculation (una ratio)—or perhaps, the
English ratio, in the sense of a “quantitative relation between amounts.” It’s
not clear whether Llull means to give us commonplace instructions
(something like “Consult the letter combination in each cell”), or to say
something profound about reason, in the sense of “Reason attains
intelligence inside the table.”
 What just happened then? Did we really witness the emergence of
artificial intellect—“an intellect that fashions itself”? Not exactly. The
sleight of hand happened through personification. Clever Llull fashioned a
clever system—clever in a manner of speaking. Yet it continued to operate
cleverly in the absence of its maker. So, it did contain some smarts after all!
These were left behind by Llull, who’s long gone. But were you to give me
a clever explanation, we wouldn’t call the explanation itself smart, right?
Your intellect would be the source of intelligence, transmitted via the
explanation. Yet, in common usage, cleverness tends to detach from the
human, almost as if to appear self-sufficient.
A wizard has slipped behind the curtain. To help in the translation of
Latin, I used the Latin Word Study Tool from the Perseus Project, which
compiles more than sixteen million Latin words, presenting a complete
concordance of word meaning, in use. Did it actually “compile” or
“present” anything, or was that done by Gregory Crane and his team from
Tufts University, the makers of the project? Their smarts made me smarter.
But how is it that we find it so natural to cede agency to the tool?
At what point, on the scale of automation, from partial assistance to
independent reason, does the intellect begin figuring in the grammatical
subject position? And why is it always the one intellect or intelligence
generally—an indefinite, singular noun? I cannot wrap my mind around
that. Llullian diagrams and digital Latin study tools contain many different
components: databases, charts, and wheels. Does their intelligence emanate
from one or many sources? Or, does it hover like a cloud above our
collective minds and Latin Word Study Tools?
Practically, the technology itself begins to exceed the rather limited
cognitive capacity of individual humans. Before anyone takes offense,
understand that together we are an impressive bunch. But individually, I am
at least aware of my own substantial limitations in driving, reckoning
numbers, or arranging letters. Automated assistants, like driving aids,
calculators, or spell-checkers, offer a measure of welcome performance
enhancement. I know comparatively little about the world by memory. But
surrounded by dictionaries, encyclopedias, and search engines, I know a lot
—though my actual share of that knowledge remains minuscule.
When encountering a smart gadget, I’m therefore impressed not so much
by its local effects on me, personally, but by its capacity for shared,
collective achievement. In writing with a spell-checker enabled, I am
writing with a shadow team of scholars and engineers, able to extend their
aid across time and space by the virtue of technology. This collective
endeavor resists easy description. I am not eager to cede authorship or to
continually credit a school of distant collaborators. I wrote this, after all. Yet
I also must admit to a diminished share of authorial agency, in this and
many other matters of intelligence.
Spell-checkers and dictionaries have a long history. I don’t normally have
to think about them, however. Instead, we employ a cognitive-linguistic
shortcut by condensing and ascribing agency to the technology itself. The
wheel divines, not Llull or Ibn Khaldun. And it’s easier to say “The phone
completes my messages” instead of “The engineering team behind the
autocompletion tool writing software based on the following dozen research
papers completes my messages.”
This strategy works, until the metaphor threatens to run away with
unintended meaning. The smart table, in the words of Karl Marx, another
furniture enthusiast, “changes into something transcendent,” a fetish, turned
onto its head, fashioning grotesque ideas out of its wooden brain, as it
recedes into the backdrop of ordinary commodities. “I made this,” I keep
saying to myself, but when the tool breaks, I curse it and call it a “stupid
thing.”
CHAPTER 3

Smart Cabinets

T WO GERMAN BAROQUE POETS WALKED into a bar. Well, more like, one was
a poet and the other a scholar. And they corresponded by letter.
Athanasius Kircher, a neat man in his forties, spoke and dressed plainly. An
established figure by the time of their virtual meeting, he was slated to take
Johannes Kepler’s place at Habsburg court as the royal mathematician. In
years past, he had taught mathematics and ancient languages at Wurzburg.
Fate found him at the Collegio Romano in Italy, where his research program
expanded to cover geology, chemistry, and biology.
The eager attention from a flamboyant younger admirer clearly annoyed
him. The gaunt Quirinus Kuhlmann struck an odd figure. A childhood
speech impediment left him to pursue a career in letters with fervor. At the
University of Jena, known as a bit of a party school in the early sixteen
hundreds, he thought little of his peers or teachers. Instead of studying law
as planned, he cultivated an image of a brooding poet, claiming to receive
divine visions through illness and hallucination. Rumors of heresy
circulated, as well as of his “colossal egotism.”
At Jena, Kuhlmann’s scholarly brooding included the study of Llull’s
cryptics, as they appeared in Kircher’s work. By this time, and despite its
own brush with the Inquisition, Llull’s Art had spawned numerous
adherents across Europe. The emerging science of combinatorics promised
shortcuts to divine revelation. Under the spell of Llull, Kuhlmann wrote a
series of unprecedented technical sonnets, among them “Heavenly Love-‐
Kiss XLI,” meant to showcase his understanding of combinatorial
composition. I’ll translate a snippet for you here:

From Night / Haze / Fight / Frost / Wind / [. . .] / Fire and

Plague
Follows Dawn / Blaze / Blood / Snow / Still / [. . .] / Blight and
Need
From Loath / Pain / Shame / Fear / War / Ache / [. . .] / Jest in
Scorn
Wills Joy / Pride / Fame / Trust / Wealth / Rate / [. . .] / Steady
Days
[. . .]
All changes, all loves, all seems to hate something:
All earthly wisdom comes to those who reflect thusly.

The sonnet’s concluding lines invite the reader to rearrange the words at
will—a radical gesture. The many resulting combinations produce
individual meaning, amounting to the sum total of human knowledge,
Kuhlmann explained. Despite such grandiose claims, Kuhlmann abandoned
the technique in other compositions, making me think he saw this particular
“love-kiss” as a technical experiment, not much more.
At stake in the virtual meeting between poet and scientist was Kircher’s
recent invention called the Mathematical Organ. Waiting for their ale, both
men contemplated the device in front of them. Made of polished, “artfully
painted” wood, the Organ resembled a large box. Opening its lid revealed a
row of labeled wooden slates, filed vertically: four columns of fifteen
narrow slates, four columns of seven wider ones, and four columns of five
of the largest pieces. When pulled out, each of the slates contained a string
of letters. By consulting the included booklets—which he called, wait for it,
applications—any combination of the planks cohered into a complete,
harmonious composition. Cleverly, depending on which manual was
consulted, the same organ could be used to compose music, write poetry or
secret messages, and even do advanced math, such as reckoning the Easter
calendar.
The conversation grew more heated with ale. Kircher took the same
systematic approach to all his research projects. The box was made
according to the latest scientific principles. He had recently sold a version
of the device to the young Archduke Charles Joseph of Austria. The child
would use it in his studies, with the help of a trusted tutor. Such devices,
Kircher argued, were easy, delightful, and instructive. They were, above all,
useful, allowing untrained operators to create and to perform. Technology
was there to serve, making difficult knowledge accessible to a larger
audience.
Kuhlmann objected. The path to knowledge should remain tortuous,
accessible only to those (like him!) willing to walk it properly. The box is
just an ingenious game, my ingenious Kircher! Kuhlmann was starting to
slur his words a little. The intelligence lies with you. Without the box, the
young duke remains an idiotic parrot (I’m sorry, he really did write that,
only in Latin). Nothing is retained on the inside. And without inward
understanding, the child captures neither knowledge nor intelligence.
 A part of the Mathematical Organ, as described by one of
Kircher’s students, P. Gaspar Schott. Printed by Johann Andreas
Endter in Würzburg (1668), 135.
 One of the “application” tables included with the device. Schott
(1668), 93.

One would hope, in another timeline, the young Kuhlmann was not
planning to travel Europe to say such things about royal children out loud,
especially not in the presence of, let’s say, a certain fledgling and repressive
Eastern regime. Though his talents were recognized in Germany, he would
later burn at the stake in Moscow, by the order of Ivan V, denounced by his
fellow Lutherans for iconoclasm.
We arrive then, as always, at the ghost in the shell—of smart furniture.
Clearly, neither poet nor scientist thought of the Mathematical Organ as an
animated entity. They differed rather in the idea of the human involved. For
Kuhlmann, the tooling edge of combinatorial poetry aimed inward, toward
the reconfiguration of the spirit. Combinatorial sonnets helped precipitate a
mercurial spiritual transformation, difficult for us to recreate sober and in
my translation. For example, the sonic parallelism between Haze and Blaze
(which I mirrored from the German), encouraged readers to seek further,
conceptual similarities between smoke and fire. Name and physical
property momentarily aligned. New, unexpected connections were meant to
be made through private internal reconfiguration.
By contrast, Kircher’s machine con-figurated externally, producing a
specified rational effect—be it the calculation of numbers or composition of
poetry—which also happened to be useful, elegant, or delightful—all
outward qualities.
The two views on the machine tell us something important about
intelligence, which manifests both as a private experience of inward
understanding and its outward, instrumental effects on the world. Let’s take
a moment to consider both. On the former view, we’ll call it Platonic,
intelligence means something like “the correct internal alignment of
thoughts and feelings with universal truth.” On the latter, we can call it
Aristotelian, it implies “a universal capacity to achieve specific results.”
Notice something strange?
In the Platonic model, intelligence works as the SOURCE of action. It
therefore necessarily exhibits qualities of being private and local, ineffable
even. Therefore, paradoxically, Platonic universal rationalism tolerates
contingent, nonrational paths to enlightenment. Of course! I think to myself.
There is no one general way of being smart. How would that work?
Sometimes I can’t even articulate how I’ve come to understand certain
things. It just happens or doesn’t. In that sense, our intelligence is bound by
our individual bodies, education, and personal history. It therefore resists
one single universal description.
Private experience doesn’t have to be wholly subjective or arbitrary.
Outside access tends to be difficult, however. How can a teacher glimpse a
student’s view from the inside, so as to better ascertain whether they’ve
learned something properly? Plato—well actually, the fictional character
called Socrates—famously mistrusted all writing in general and rote
memorization in particular (despite being a memorable writer). An often-‐
cited dialogue recalls Socrates encountering Phaedrus, a student of his, who
was on the way back from another lecture. When Socrates asked him about
the content of that lecture, Phaedrus retold the words from memory.
Socrates suspected a parrot, however, and proceeded to question Phaedrus
poignantly, exposing the limits of his student’s understanding. Similarly,
any fixed, mechanized, external trappings of acumen risk being
disconnected from the source. Instead of smarts, you get parrots, who, also
according to Descartes in his Mediations, merely repeat without
understanding.
There can be no shortcuts on the road to knowledge. A student achieves
internal change through mental toil. The Platonic is therefore necessarily
also a human-centric view, because, on the inside, I don’t know how to be a
bat or a rock or a smart table or anything else. Platonic intelligence implies
appropriate personal transformation.
In the Aristotelian model, intelligence is the GOAL of thought. It doesn’t
matter how it is achieved at the source. One jumps into the collective pool
called “intelligence” through appropriate action. And all sorts of stuff
besides humans float in that pool: bats, tools, automatic transmissions,
smart tables, mathematical organs, smart phones, etc. This makes sense to
me also, because when I think about stuff, I usually don’t do it on the inside
only: I pace, take notes, read books, and talk to friends. Looking things up
on Wikipedia, consulting a table, following a certain procedure I learned
from a textbook, or evoking a “smart assistant”—none of these detract from
intelligence in the Aristotelian sense.
It would be foolish in fact to refuse such outside assistance. I cannot
imagine knowledge that develops in complete isolation from the world.
(Incidentally, the twelfth-century Andalusian treatise on the “Self-Taught
Philosopher,” titled Hayy ibn Yaqzan in the original Arabic by Ibn Tufail,
was one such attempt, prefiguring the modern “castaway” genre,
represented by Daniel Defoe’s Robinson Crusoe (1719) or television series
like Lost (2004–2010) or the film Cast Away (2000), starring Tom Hanks.)
Science itself would be impossible in isolation because it produces
intelligence in a way that’s fundamentally social and collective. By
developing novel tools and techniques for externalized thought, like
Aristotle or Kircher, a scientist contributes to the overall shared capacity for
reason. In that sense, something becomes right or true by its measure of
accordance to an established body of work. Intelligence in the Aristotelian
sense is subject to external review. We don’t have to concern ourselves with
“what’s really going on, on the inside.” If it shimmies and quacks like a
smart duck, it’s a smart duck.
That the two senses of intelligence—Platonic and Aristotelian, internal
and external, private and public—both use the same word causes much
confusion. It would be simpler if we had at least two distinct words for the
Platonic and the Aristotelian definitions, capturing the difference between
intelligence as the SOURCE of and the GOAL of action.
Baroque birds of all sorts bring us a step closer to the story of modern
digital computers, able to mimic any other “discrete-state machine”
universally. Called a “universal machine” by Alan Turing, the computer
promises to model general intelligence, regardless of the task it was
originally designed to perform. Recall the obsolescence of single-purpose
machines like calculators, cash registers, and digital music players. The
universal computer ate them all. Similarly, general intelligence refers to the
ability of an individual or machine to learn, reason, and problem-solve
across a wide range of domains and contexts. In this sense, general
intelligence is often presented to us in the guise of “universal intelligence,”
intelligence capable of performing any cognitive task.
Given the discussion above, the words general and universal should be
treated with suspicion. Universalism, for all its strengths, comes with a host
of problems, already apparent in our case studies.
From the Platonic, internalist perspective, how smarts are achieved
matters. I might value good grades on an exam, for example, but not if they
were the result of cheating. Conversely, I don’t usually care how my
refrigerator thinks, only that it does its job. Anything goes for a smart
machine, evaluated on efficacy alone. A machine may perform intelligence
by any obscure means. My students must show their work.
The “private” ways of machine intelligence also may not necessarily
agree with my own, or that of my family or community. Efficacy sometimes
comes at a cost. For instance, it may be that my television gets smarter by
continually recording conversations in my living room, so as to better tailor
its recommendations to my family. These means do not justify the ends for
me. The so-called alignment problem stems from an impassible paradox:
generally universal intelligence does not always align with local values.
Smart in the general sense might not be so smart in the specific context. We
should therefore always be cautious with claims for intelligence in the
general, universal sense.
From the Aristotelian, externalist perspective, the idea of an intelligence
considered “universal” can be equated to “generally applicable.” General
application contains no impassable paradoxes, with one huge caveat: where
it gains in universal application, it loses on smarts.
Tasks that can be automated become devalued. Witness numerous smart
devices, like toasters and thermostats, that modern humans treat with
disdain. While inspiring awe initially, smart technology of the past becomes
mindlessly assistive, no matter the expanded ingenuity. The same can be
said for the tools of intellectual trade. Dictionaries, grammars, thesauruses,
and encyclopedias were once hailed as monumental national achievements.
Today, they are silently integrated into digital autocompletion or
autocorrection tools. We don’t even think of them as sophisticated
technology.
 As our tooling improves, the functional definition of intelligent action
changes as well. Because the externalist perspective is grounded in the
world, general intelligence recedes beneath the rising tide of average
achievement. For instance, at one time, the mere fact of being literate would
have been considered exceptional. One could build a career of simply
reading and interpreting texts for an illiterate public. With mass literacy,
skills of reading and writing migrate from the upper to the lower bounds of
intellectual work. Similarly, while the ability to memorize complicated
calendar patterns may have been impressive a few centuries ago, spending
time on unaided astronomical reckoning today would be considered foolish,
with the most advanced astronomers being those able to use assistive
technology (like powerful computing) most effectively. In the applied
world, intelligence bubbles above shared capacity. But no matter how
cleverly effervescent, once a smart tool reaches general adoption, it
dissipates into the surface of baseline intelligence.
Here we reach the paradox of the functional idea: Exceptional
performance cannot become universal, by definition. A “smart” device is
merely smarter than the previous generation. Once adopted widely, it passes
into the average.

UNIVERSAL CHARACTER

The attempt to overcome the said limitations of universal reason is

indicative of computing history in general: from divination wheels to
baroque composition cabinets and Turing universal computation machines.
There was, in the seventeenth century, a great deal of interest paid to
universal languages in Europe, for reasons beyond poetry or philosophy.
The Protestant revolution placed a premium on vernacular languages,
displacing Latin from its privileged, liturgical position. Bibles had to be
read and translated locally. Issues of what should count as the “correct”
translation could and did grow into armed conflict. Printed books, like those
by John Wilkins, were becoming more commonplace in the wake of the
Gutenberg printing press. Issues of universal literacy now seemed pressing.
The establishment of the Royal Society in London, among other learned
societies, made publishing—alongside a clear writing style, peer review,
and discussion—an integral part of the scientific process.
As Kircher’s experiments with the Mathematical Organ have shown,
reason could become subject to automation. Yet how could we verify the
results of any machine that claims to produce knowledge in excess of our
own intellectual capacity? What if the smarts generated by a mechanism
turned out to be nonsense? We would have to become smarter than the
machine to check its work. We could do this, for example, by validating
generated results against an established sense of the “ground truth.” A
machine capable of arriving at established truth independently could be said
to have cleared the bar of artificial reason. But what about claims to entirely
novel ideas? What if machines were to discover new answers, breaching
established grounds? How could we tell such “truths” apart from
mechanical gibberish?
Enter universal language machines.
The principle of Kircher’s cabinet lay in the included reference tables,
governing the rules of composition, a grammar. This class of machines was
prone to producing nonsensical answers—the colorless green ideas sleeping
furiously—despite sophisticated programming. Skeptics like Kuhlmann
pointed to the arbitrary nature of the sign as the culprit. Because of their
general imprecision, natural human languages yielded poor building blocks
for mechanized reckoning. Words arranged according to predetermined
rules did not always correspond to the proper arrangement of things in the
world. Language and physics played by different rulesets. Grammar and
sense did not always align.
Wouldn’t it be wonderful if they did? What if Kuhlmann’s chance
resonances—Haze/Blaze, Shame/Fame—could lead to actual scientific
discoveries? If machine inputs could be formalized to contain pure,
unambiguous truths, arranged according to inarguably correct principles,
the outputs were bound to be true! The alignment of sense and grammar
seemed to be a promising direction for new technology.
In the 1660s, John Wilkins, an Anglican polymath and one of the
founders of the Royal Society of London, published just such a proposal,
titled An Essay Towards a Real Character, and a Philosophical Language.
In this monumental, five-hundred-plus-page tome, Wilkins proposed an
artificial language, incapable of falsity. In it, the internal and the external
would become one. The rules of grammar and the rules of physics could
join in harmony, providing the ultimate tool for clear reasoning. (Or so he
thought. The book suffered a cool reception.)
Science, bound to become the language of universal truth, he reasoned,
should base its communication on grounds more solid than natural
language. Consider the mischief done, he wrote, “and the many impostures
and cheats that are put upon men, under the guise of affected insignificant
phrases.” Pseudoscientific “charlatans” make their reputation by using
“pretended, mysterious, and profound notions, expressed in the great
swelling of words,” leaving the “internal notion of apprehension of things”
confused with the “external impression of these mental notions.” “So that if
men should generally consent upon the same way or manner of Expression,
as they do agree in the same Notion,” through the use of Real Character,
“we should then be freed from that Curse in the Confusion of the Tongues,
with all the unhappy consequences,” Wilkins wrote.
To these awesome ends, he proposed—you guessed it—the smartest of all
tables. Step One, the world. The first thing to be done, according to
Wilkins, was, modestly, “the enumeration of and description of such things
and notions as are to have Marks or Names assigned to them.” Wilkins must
have realized the enormity of the task, admitting that perhaps his giant
classification schema contained a few “redundancies” and “deficiencies.”
Still, in the course of the first three-hundred-some-odd pages, he braved a
grand “scheme” that captured the state-of-the-Royal-art. This frankly
madcap (even by its own, contemporary standards) project enumerated
anything and everything under the sun: “transcendental relations mixed and
general”; “herbs according to their leaves,” shrubs, trees, and grasses; stone
and metal; fish, bird, and beast; magnitude, space, and measure; a catalog of
habits, manners, and diseases; relations civic, judicial, military, and
ecclesiastical; among many, many other “notions that fall under discourse.”
Having listed all possible named things known to science, Wilkins moved
to Step Two, in consideration of a “philosophical grammar,” containing the
rules of their possible recomposition. Here, the grammar of the internal,
conceptual world could be brought to accord with natural law, governing
the morphology of the physical world. Though unclear as to how this would
work in practice, it was apparent that Wilkins imagined grammar, in its
usual sense, to be coupled to something like the grammar of physics or
chemistry, making language and physical world function under the same
rules. A false idea in this new language would be, if not impossible, then
difficult to articulate. Words would work like Lego blocks, snapping
together only in ways that made sense.
This was also the weakest part of his system because the accordance of
the two grammars depended on his initial haphazard classifications.
In the third and concluding step, Wilkins proposed a novel writing system
in an attempt to solidify the relationship between word and thing, thus
breaking the spell of arbitrary signification. Words constructed according to
“Real Character” fit together like parts of a device, which, when activated,
produced only true sentences. “Now, if these Marks can be so contrived,”
he wrote, “as to have such a dependence upon, and relation to, one another,
as might be suitable to the nature of the things and notions which they
represented [. . .] the Names of things could be so ordered, as to contain
such an affinity or opposition in their letters and sounds, as might be some
way answerable to the nature of the things they signified.”

The Lord’s Prayer rendered into universal character by John

Wilkins. John Wilkins, An Essay towards a Real Character and a
Philosophical Language (1668), 395.

The alphabet, as the author admitted after transcribing the Lord’s Prayer
into his system, might not have turned out as practical as he had planned.
Still, he hoped it delivered a distinct advantage over European alphabets,
mistakenly believing Chinese or Arabic writing systems to be somehow
more intrinsically closer to his ideals. The Real Character would render
each word precisely, according to the classification of things and
relationships located in the main table. And though Wilkins could not
achieve a perfect analogy between word and thing, he thought that further
improvements of his system would someday enable great advances in
scientific communication, commerce, and even facilitate peace between
nations.
Judging by the many wars and disagreements since, this experiment
failed. Still, Wilkins sowed yet another seed of universal language. And it
would bear fruit in the universal machines made by Leibniz, Babbage,
Lovelace, and, later, Turing.

STEPPED RATIOCINATOR

In 1679, Gottfried Wilhelm Leibniz wore a frizzy wig. Like other great
men, he was modest about his accomplishments. “I’m not sure why
Aristotle or Descartes didn’t consider this amazing thing I did while still a
boy,” was a common refrain in his letters. Though, to be fair, he did
accomplish impressive feats like developing calculus (contentiously, in
competition with Isaac Newton), inventing a novel binary notation (under
the spell of Chinese and Egyptian sources), and proposing the blueprints for
his own Characteristica universalis. The latter combined Kircher’s
Mathematical Organ and Wilkins’s Real Character into a single, unified
system.
We stand at a crossroads with Leibniz because, from here, the history of
artificial reason splits into at least two diverging paths. The road more
traveled leads to the development of modern calculus (literally “the small
pebble” in Latin) through the work of Maria Agnesi, Augustin-Louis
Cauchy, Louis Arbogast, Bernhard Riemann, and John von Neumann. The
larger, though now somewhat neglected, pebble rolled down the road of
universal language, leading directly to modern conversational AIs.
In 1679, Leibniz revived an earlier, monumental project, a universal
encyclopedia, aimed at healing the rift between Protestant and Catholic
churches. He called it Plus Ultra (Further Beyond), in which, like Wilkins,
he aimed to assemble (though never completed) a “precise inventory of all
the available but dispersed and ill-ordered pieces of knowledge.” Unlike
Wilkins, Leibniz concentrated his efforts on the logic of composition—the
middle, least-developed part of Real Character by Wilkins—imagining a
kind of a linguistic calculus, which he called the “General Characteristic.”
“Learned men,” he wrote, “have long since thought of some kind of
language or universal characteristic by which all concepts and things can be
put into beautiful order.” In that task, he hoped to succeed where Wilkins
failed.
Thus, in parallel to his writings on mathematical calculus, Leibniz
nurtured the far more ambitious dream of a language calculus, capable of
both “discovery and judgment, that is, one whose signs or characters serve
the same purpose that arithmetical signs serve for numbers.” This new
language would increase the power of the human mind, far more so than
microscopes or telescopes amplify sight. Without it, we are like merchants,
he wrote—in debt to each other vaguely for various items mentioned in
passing, but never willing to strike an exact balance of meaning.
For the aims of striking an exact balance, Leibniz proposed a kind of a
“complication,” or a mechanism, amounting to something greater than the
sum of its parts. “When the tables or categories of our art of complication
have been formed,” he wrote, “something greater will emerge. For let the
first terms, of the combination of which all others consist, be designated by
signs; these signs will be a kind of alphabet [. . .] If these are correctly and
ingeniously established, this universal writing will be as easy as it is
common, and will be capable of being read without any dictionary; at the
same time, a fundamental knowledge of all things will be obtained.” More
precise than modern-era artificial languages such as Esperanto, Leibniz
imagined a truth-telling mechanism, able to automate philosophic thought
itself.
Let’s pause here to breathe and reflect on the story so far.
Medieval rhetorical wheels worked by chance combination. For some,
like Ibn Khaldun, they were nothing more than clever parlor tricks, at best
generating a few surprising twists of the tongue. Consider that today’s most
advanced language AIs continue to face the same kind of skepticism:
Indeed, a mechanical parrot reading from a script written by a typing
monkey may occasionally articulate something profound.
Llull and later his followers in the Age of Enlightenment saw the
possibility of something greater, beyond mere combination. Given logic,
language could be brought under a rational system, and therefore
automated, in a way that benefits humanity. Metaphorically, this device
would be called science, or calculus, or logic. But the dream of a universal
language manifested itself also as a family of literal devices, made to
amplify the power of reasoning through automated language production. Its
blueprint, extant in the most advanced of AI algorithms today, indicates,
initially, learning—the systematic assimilation of all human knowledge.
The device then attempts to translate the mess of human expression into
some sort of a comprehensive middle language, expressed by rigidly
defined characteristics, in the manner of Wilkins or Leibniz. Finally, the
logic of composition could be derived to produce novel discourse—new
ideas and answers to questions not encountered in the original learning
phase.
That was the dream, at least. As we will see going forward, major
metaphysical points of failure persisted despite technical advances.
Contemporary machines assimilate and compute over vast amounts of text.
They are also still tied to the paradoxes inherent in the idea of a universal
language.
Language retains its whimsical connection to the physical world. Its rules
can be fanciful or arbitrary. They change, as what it means to be human
changes in response to a changing world. Intelligence struggles with
universal application, in addition, because the world cannot be the same
everywhere at once. The world itself attains universality only from a great
distance, described in broad strokes by physics or theology. Its particulars
often differ depending on the local context.
Yet universals are useful also, insofar as they force us to abstract away
from private points of reference, toward shared ideals. An automated
sentence-completion gadget may make you a “better” writer. And one may
even playfully call it a “good” writer. But let’s not forget that any content of
“goodness” resides within a community, interested not only in functional
outcomes (the production of more prose), but in the implicit character of
good writers and readers. That’s why we still ask our children to learn math
without a calculator or to compose by hand. Beyond general intelligence,
we want our children to become exceptional.
The instrumental goal of calculation or composition stands secondary to
the intrinsic value of personal growth, achieved in the struggle of learning.
The hand carries the load of value through lived experience. And
experience cannot be automated, even as we witness the emergence of
clever devices—back at the inn, with our baroque friends, tinkering with the
intricate cabinetry of language manipulation.
CHAPTER 4

Floral Leaf Pattern

A UGUSTA ADA KING (NÉE BYRON), Countess of Lovelace, loved poetry

and numbers. She also had a gambling problem. And was likely
using the first computer ever to bet on horses.
The conventional story has Charles Babbage, another Royal Society
member, together with Ada Lovelace originating the concept of a digital
programmable computer sometime in the 1820s. Things were a bit more
complicated, however. The Analytical Engine, as Babbage called his
lifelong project, had its numerous mechanical precursors, directly in the
lineage of Ibn Khaldun, Llull, Kircher, Wilkins, Leibniz, and many others.
The machine was already in place. Lovelace and Babbage, steadfast friends
and neighbors, created something altogether more subtle: the language of
and for universal machines.
Kircher’s Mathematical Organ, as you recall, came packaged with a
number of “application tables,” which changed the function of the
mechanism depending on the pamphlet consulted. In this way, the same
device could be used to do math or generate poetry. It took Babbage more
than twenty years to prototype his mechanized Difference Engine, capable
of improving on Kircher’s mathematical procedures, though now in a more
compact form, using state-of-the-art Victorian craftsmanship, which had
progressed significantly in over a century of engineering developments. The
device, which he proudly displayed to visitors at home, performed little
more than basic arithmetical functions. Most visitors preferred to see the
intricate mechanical ballerina complication in the other parlor.
The more ambitious version of the machine, called the Analytical Engine,
was perpetually in development. But Babbage struggled to explain its
advantage. The concept of “one device, many applications” confused the
experts. He wasn’t even sure if it was possible. The intricate gears alone
required the expertise of Italian craftsmen, specializing in the manufacture
of fine mechanical watches. The patronage of wealthy investors was
running dry, forcing Babbage to seek support from the Crown, appealing to
matters of national pride. (If the British don’t make it first, someone else
will, he argued.) The case for the engine’s importance had to be restated
afresh often to secure funding.
There was also the issue of noise. A while earlier, Babbage had purchased
a house on Dorset Street, “in a very quiet locality, with an extensive plot of
ground,” where he “erected workshops and offices to carry out the
experiments and make drawings necessary for the construction of the
Analytical Engine.” More recently, the neighborhood had been disturbed by
the establishment of a hackney-coach stand, bringing with it “the
establishment of coffee-shops, beer-shops, and lodging houses.” These
worsened the thing he was trying to escape all along—noise. Among the
chapters of his autobiographical Passages from a Life of a Philosopher, he
airs the issue of “street nuisance” in roughly the same space he gives to his
engineering accomplishments. “Some of my neighbors,” he reported, “have
derived great pleasure from inviting musicians [. . .] to play before my
windows, probably with the pacific view of ascertaining whether there are
not some kinds of instruments which we might both approve. This has
repeatedly failed, even with the accompaniment of the human voice divine,
from the lips of little shoeless children, urged on by their ragged parents to
join in the chorus rather disrespectful to their philosophic neighbor.”
Babbage pursues the topic of nuisance at length, documenting, with a
great deal of sarcasm and crankiness, an apparent campaign of persecution
at the hands of his musical neighbors (who apparently follow and jeer him
when he leaves his house). We learn, in exacting detail, that “the
Instruments of Torture Permitted by the Government to be in Daily and
Nightly Use in the Streets of London” often include: organs, brass bands,
fiddles, harps, hurdy-gurdies, flageolets, drums, bagpipes, accordions,
halfpenny whistles, shouting out objects of sale, religious canting, and
psalm singing. Their use is encouraged, among others, by: tavern-keepers,
public-houses, gin-shops, servants, children, visitors from the country,
ladies of doubtful virtue, and, occasionally, though present company
excluded of course, titled ladies (from chapter 26, “Street Nuisance”). Also,
did you know that Babbage invented the Analytical Engine (chapter 8); met
Wilhelm von Humboldt and the Duke of Wellington (chapters 12 and 14);
played a dead man at the opera, where he saw visions of the devil, leading
him to propose a new system for theater lighting (chapter 20)?
The man’s whimsy shines through the prose. Attention wanders, from
noisy neighbors to politics and curious inventions, presenting a picture of a
uniquely distracted genius. Chapters from Aristotle’s Politics come to mind,
where the author similarly cannot help himself from exhausting a catalog of
lowly farm animals on route to high-minded topics like war and justice.
Neither Aristotle nor Babbage tolerated an epistemic mess, obsessively
compelled to make lists, even in passing.
Young Ada Lovelace, a daughter of nobility, visited Babbage’s
fashionable salons alongside other illustrious visitors like Charles Darwin,
Michael Faraday, and Charles Dickens. She was introduced for the first
time by her math tutor, Sophia Elizabeth De Morgan, a close friend of the
family and wife of a prominent English mathematician. “I well remember
accompanying her to see Mr. Babbage’s wonderful analytical engine,” De
Morgan wrote in her memoirs. While other visitors yawned, “Miss Byron,
young as she was, understood its working, and saw great beauty of the
invention.”
An odd sort of a friendship developed between the young Lovelace and
the older Babbage. Growing up in the shadow of her famously philandering
father, the poet and baron George Byron, whom she barely knew, Ada had
been discouraged from literary study. Her mother, Lady Anne Isabella
Byron, a strict Christian and a formidable intellect in her own right, steered
her daughter’s interests toward mathematics instead. She approved of
Babbage because of his scholarly standing though found him also a bit too
quirky for her tastes. Her letters show a persistent worry about her
daughter’s moral character, in danger, as she saw it, of inheriting her
father’s Romantic flaws.
By all signs, Ada remained fascinated with her father’s legacy however,
viewing herself a unique synthesis of scientific reason and poetic
sensibility. The Analytical Engine appealed to both. Soon, she was
borrowing blueprints from Babbage, along with other literature related to
clocks and steam engines. Under her mother’s watchful eye, she spent long
evenings on Dorset Street with Mary Somerville and other close friends,
discussing the machine among like-minded enthusiasts. Her rousing plans
to build a flying apparatus based on bird anatomy were soon supplanted by
the rigorous study of mechanical intellect.
 By contrast to the single-purpose Difference Engine, built for number
crunching only, the Analytical Engine captured the imagination. On some
days, Babbage spoke of it with mystical reverence, much to Lady Byron’s
displeasure.
The workings of the Analytical Engine, he wrote, parallel those of a
prophetic intellect. A single-purpose machine, like the Difference Engine,
computes according to one law or pattern, expressed in the linkages of its
inner gears. A law is derived by induction from many examples: the sun
rises every day; therefore it will rise tomorrow. A miracle, by those
definitions, constitutes an exceptional event, such as an eclipse, signaling
the existence of yet “higher laws of operation.” But since the original,
lower-level law was brought under a single purpose, a Difference Engine
could never exceed its hard-wired programming. Higher-order, exceptional
rulesets were not available to a machine built for a single purpose.
By contrast, a multipurpose Analytical Engine would compute over
several table applications at once, reconciling patterns at different scale.
The switch to another ruleset would appear prophetic to an observer used to
regular sunrises, Babbage explained in chapter 2 of his memoirs, “On
Miracles.” It was as if the machine could predict miracles, seeing beyond
the drudgery of everyday routine. Babbage and Lovelace thought of their
engines in terms of miracles.
In years prior, Babbage attempted to propose something like a language
made to describe competing rulesets, in machine terms. The intricate
gearing of the Difference Engine already strained the limitations of
conventional design schematics, in use for precision clock-making at the
time. An engineer following complex instructions for gear manufacture was
guided by static drawings, depicting the intended mechanism at rest.
(Think, for instance, of a simple hinge mechanism, like a door.) Where
moving parts were indicated, another drawing could be made to illustrate its
dynamics: a door in the closed and open positions, for example.
But what if your mechanism contained hundreds or thousands of such
moving parts? Other graphic conventions, such as arrows or broken lines,
could perhaps be used to detail them. Unlike simple hinges, however,
Babbage’s gear boxes moved in extremely complicated ways. Their
movement changed depending on the machine’s setting, and had to be
synchronized across the mechanism. Craftsmen Babbage hired to work on
the engines were often confused by such dynamic designs, leading to faults
and delays.

Babbage’s proposed mechanical notation from 1826. Image

courtesy of the Royal Society.

Frustrated with their slow progress, Babbage proposed a new kind of a

blueprint, meant to accommodate increasing schematic complexity. In 1826
he published a paper in the Philosophical Transactions of the Royal Society
of London, detailing a “method of expressing by signs the actions of
machinery.” “The signs, if they have been properly chosen, and if they
should be universally adapted, will form as it were an universal language,”
Babbage wrote about his mechanical notation, which looked something like
a musical score for steam engines.
The Italian engineers invited to work on the Analytical Engine were
unlikely to understand Babbage’s novel blueprints. Almost a decade of
crafting yielded few workable results. Support from English backers had
long been waning. In 1842, after attending one of Babbage’s salons in
Turin, an Italian military engineer (and future prime minister of Italy)
published a concise “sketch” of the Analytical Engine in the Bibliothèque
universelle de Genève, number 82. Charles Wheatstone, another friend,
inventor, and frequent guest of the household, suggested Lovelace (who
studied Italian) translate this sketch for a newly founded British scientific
review. In an unusually progressive move for the times, Babbage proposed
she add her own notes as well—a task which Lovelace compared to birthing
her first child, dedicating herself accordingly.
Appearing a year later, Lovelace’s notes distilled a vision of a universally
programmable machine, stripped of miracles and street noise. Historians of
science who today point to several inconsequential algebraic mistakes in the
document fail to see its larger impact. For the first time, we may add to the
history of universal machines the idea of a programmable device, capable
of reconfiguring its own hardware through machine instruction, expressed
symbolically.
In describing the mechanism (hardware) of the Analytical Engine
representationally (as software), Lovelace created an innovative solution for
the reconciliation between physical and symbolic domains. In her words,
the mechanism “combines together general symbols, in successions of
unlimited variety and extent,” establishing “a uniting link between the
operations of matter and the abstract mental processes of the most abstract
branch of mathematical science.” In one the most lyrical passages of her
commentary Lovelace wrote, “The Analytical Engine weaves algebraic
patterns just as the Jacquard-loom weaves flowers and leaves.”
Let’s unpack that a bit. Remember, the misalignment between language
and the world plagued our universal machines in the last chapter. When
discussing “trees” for example, we usually assume the same physical
referent (the actual tree). But in some cases, my idea of “trees” represents
something other than what you may have had in mind. This often happens
in bilingual families for example, where parents and children sometimes
disagree about basic definitions of such things as fruits and vegetables, or
trees and shrubs, lost in translation between languages.
Universal reason requires the settlement of such ambiguities. Early
modern solutions, like those by Wilkins and Leibniz, consisted of the
attempt to cram the entirety of the “known” world into the machine. But
who’s to say who agrees on things that are known? Are tomatoes fruits or
vegetables? Agreement changes in context: Are you addressing a grocer or
a botanist? Plants do not grow with labels attached. Words attach to things
indirectly, through use—at a time, in a place.
Visiting an imaginary botanical garden of the world, we find a multitude
of conflicting labels, some torn, others written over or illegible. That’s not
to believe that everything becomes relative there, nor that anything goes.
Only that language and the diversity of human experience make consensus
difficult. Agreement cannot be “discovered” unambiguously nor inherited
uncritically. It is rather a process, fraught with social and political
complications. Consensus is to be achieved iteratively, through revision and
in discussion—these were key insights of the Enlightenment.
“Meaning shifts,” Lovelace wrote. With it, “independent sets of
considerations are liable to become complicated together, and reasonings
and results are frequently falsified.” Lovelace’s way out of these shifting
sands avoided sifting through each grain. Rather than cramming everything
into the machine, in the way of Kircher or Wilkins, the material world could
be brought to accord with the mental at a single point of contact, where
hardware and software coincided. Within the proliferation of Babbage’s
designs, she found a transformative idea: The physical state of any machine
could be represented symbolically, as an algebraic formula. A change in the
machine state could therefore become both the source and the target of
symbolic manipulation. The machine became formula, capable of
manipulating formulas that turned into machines. A thin filament was thus
thrown to bind the two worlds, the real and the ideal, language and its
referent.
Consequently, the only concrete representation of reality the machine
needed to “comprehend” or “ingest” was its own design. Once this bridge
was established, all other information about the world could be treated as
true or false not by any external standards, but by the virtue of
programming.
Lovelace explained that “in studying the action of the Analytical Engine,
we find that the peculiar and independent nature of the considerations
which in all mathematical analysis belong to operations, as distinguished
from the objects operated upon and from the results of the operations
performed upon those objects, is very strikingly defined and separated.” An
operation, she continued, could be defined as “any process which alters the
mutual relation of two or more things [. . .] including all subjects in the
universe.” The language describing basic relationships between things
could therefore be translated into a basic relationship between machine
internals. Whatever the arrangement of objects in the world defined by the
programmer found its corresponding representation inside the device.
The mind boggles at this apparent recursion. Somehow, a snake eating its
own tale provided enough structure to support the whole mess of meaning-‐
making. “Whether the inventor of this engine had any such views in his
mind while working out the invention [. . .] we do not know; but it is one
that forcibly occurred to ourselves on becoming acquainted with the means
through which the analytical combinations are actually attained by the
mechanism,” Lovelace concluded, apparently recognizing the enormity of
her discovery. The engine has become “the material and mechanical
representative of analysis,” at the point where theory and practice could, for
a moment, attain symmetry. Any abstract components, arranged into any
structure, could be mirrored in that way: “Supposing, for instance, that the
fundamental relations of pitched sound in the science of harmony and of
musical composition were susceptible of such expression and adaptations,
the engine might compose elaborate and scientific pieces of music of any
degree of complexly or extent.”
With these words, Lovelace glimpsed the future where machines would
weave many more patterns, from missile trajectories to song lyrics and
patient diagnostics. Smart refrigerators were still just ice boxes then—the
iceman delivering a block of ice every few days, for which Ada paid one
penny a month. Though her notes were published and received well, the
engine was no closer to being completed on Dorset. Babbage was busy with
new projects, including several in the nascent field of actuarial science
(insurance), where his machines proved useful for reckoning logarithmic
tables and compound interest. Meanwhile, Lovelace found a new group of
friends, several major pony gamblers among them. Though she was
secretive about her exact methods, we know that her love for horses and
mathematics would generate profit sometimes, debt others, though in her
letters she appeared as carefree and unapologetic as ever, much to her
mother’s dismay.
CHAPTER 5

Template Culture

S TOP READING HERE IF YOU love literature but hate thinking about how it
is made. This might be an unsavory story about making sausage.
Though perhaps it’s not so bad, because great art, like great cuisine, will
always remain the purview of exceptional talent. Average chefs do most of
the cooking, however, just as most writing is done by ordinary authors. I
don’t mean to diminish the study of great literature in any way. We ignore
the average at our peril, however. The failure to understand the mechanisms
of ordinary authorship holds dire consequences for the quality of our
collective intellectual experience. Artificial intellect thrives in the gap
between the average and the exceptional. There’d be no need for calculators
if we were all mathematical geniuses. AI was created specifically to make
us smarter. Spell-checkers and sentence autocompletion tools make better
(at least, more literate) writers.
In considering the amplification of average human capacity for thought,
both Babbage and Lovelace circled around the idea of the Jacquard loom,
an innovation in weaving manufacture that used perforated “operation
cards” to pattern its designs. Within the gears of the Analytical Engine,
which Lovelace called “the mill,” an operation card would rearrange or
“throw” the mechanism “into a series of different states,” “determining the
succession of operation in a general manner.” A different type of a card,
called the “supply card,” held the data—which, recall, could represent any
object—“furnishing the mill with its proper food.” The mill processed its
supplies according to instructions punched into operation cards. These
could even be swapped mid-operation to perform nested functions—an
eclipse within the regular moon calendar—Babbage’s exceptional miracle.
The connection between symbolic logic and commercial weaving wasn’t
accidental.
In 1832, still in the early days of the Difference Engine, Babbage issued a
missive from Dorset Street. He called it a “direct consequence” of his work
on the machine, compiled in preparation for a proposed lecture series at
Cambridge. Titled On the Economy of Machinery and Manufactures, this
lengthy tome contained a comprehensive study of all known industrial
manufacturing processes. Reading it has significantly changed my
perspective on the Analytical Engine.
Where at first I saw it in the lineage of thought-manipulating machines,
via Kircher and Leibniz, I could now also perceive its ambition on the scale
of macroeconomics, in the broader context of the industrial age.
The Analytical Engine was to achieve for the world of letters what the
Jacquard loom has done for the commercial weaving of fabrics. This was no
mere metaphor for Babbage. Nor was it solely his invention. From the
nineteenth century onward, the notion of template-based manufacturing
permeated all human industry, and especially the manufacture of consumer
and capital goods, like clothing, furniture, machinery, or equipment.
Yet seldom do we consider templating in the production of symbolic
goods, like literature, film, music, philosophy, or journalism. There’s just
something unsavory about the thought of individual human genius being
diminished by mechanical reproduction. Several modern high-art or avant-‐
garde movements have even defined themselves explicitly in opposition to
industry. For the Romantics, like George Byron or William Wordsworth,
and the modernists, like Bertolt Brecht, Franz Kafka, Walter Benjamin, or
Virginia Woolf, to be human was to raise an exception. There, in the
rarefied heights of individual genius, the automated sank beneath the
surface of creativity, much less intelligence, in favor of the handcrafted and
the extraordinary.
Scholarship in the humanities inherited the Romantic emphasis on the
exception at the expense of the automated, and therefore the instrumental
and the collective. Many everyday practices of reading, writing, and
interpreting texts—besides fiction—have fallen out of scholarly purview as
a consequence. Not much research is being done on medical literature for
instance, self-help, or on collaborative writing practices in the television
industry.
If there has to be one, take it as this book’s punchline—intellect requires
artifice, and therefore labor. Innate genius can neither be explained nor
imparted. The artificial seems autonomous because it arrives removed from
its singular, inimitable point of origin. Artifice identifies precisely those
aspects of an art subject to be explained, documented, and transferred to
others. Consequently, intellectual work participates in the history of
automation, affecting all trades at some point of its development. Like the
making of shoes or automobiles, the production of intellectual goods has
long been moving from bespoke workshops to the factory floor. Once we
understand intellectual labor as labor, we should not be surprised to find
tools, templates, and machines near it. But because mind and language are
special to us, we like to pretend they are exempt from labor history. This
book so far has told and will tell a story about labor, not just robots or
literature.
Reams have been written on the minute-by-minute writing schedules of
exceptional authors. Yet little is known of the tools augmenting much of
everyday knowledge work. Templates support the labor of ordinary intellect
everywhere, from the physician’s office to the photography studio and the
TV writers’ room. Their use was developed systematically at the time of
Babbage, in every human industry, including the intellectual. Tables,
schemas, skeleton forms, molds, patterns, matrices, and frames lie at the
basis of all content production. The templates driving engagement on
modern social media are just the most recent of a long-standing historical
phenomenon. The end of the nineteenth century in fact saw an explosion of
template-based art. Though their use may have been a cause of personal
embarrassment or even professional taboo.
Few artists like to admit to painting by the numbers. Nobody wants to
seem ordinary. The occasional visibility of artifice—portable, explainable,
documented, transferable, automated—therefore tends to startle or repulse
audiences acculturated into the privilege of exceptional human genius.
But it also shouldn’t, because an emphasis on those aspects of an art that
can be transferred, instead of inherent talent, lies at the basis of a
democratic education.

SKELETON FORMS

In his Economy, Babbage surveyed the state of automation in the first half
of the nineteenth century, without nostalgia for special handcrafts of times
past.
Here, “skeleton forms” performed the same function on the
macroeconomic level, as did “operation cards” on the level of the machine.
The difference between “making” and “manufacturing,” for Babbage, was
found in the ability of makers to rationalize their creative process, step by
step, in a way that could be optimized and replicated.
In the author’s usual diligent fashion, major portions of the book were
dedicated to the exhaustive enumeration of various copying processes,
including: the printing from cavities, copperplate printing, engraving on
steel, music printing, calico, stenciling, printing of red cotton handkerchiefs
(?), printing from surface, block printing, moveable type, stereotype,
lithography, iron casting, casts in plaster, casts of vegetable productions in
bronze (!), casting in wax, imitations of plants, the molding of bricks, tiles
and cornice of the Church of St. Stefano, embossed china, glass seals,
square glass bottles with names, wooden snuff-boxes, horn knife and
umbrella handles, swagging, engraving by pressure in steel, forgery of bank
notes, copying by stamping, coins and medals, military ornaments, cliché,
copying by punching, boilerplates, tinned iron, buhl-work, steel chains,
copying with elongations, wire drawing, brass tubing, leaden pipes,
vermicelli, copying with altered dimensions, pentagraph, rose engine
turning, lathe, shoe lasts, and veils made by caterpillars (!?).
In conclusion, Babbage suggested his readers undertake a systemic study
of automation by using the provided “skeleton form,” binding hundreds of
responses into a pocket book of surveyed manufacturing templates. This
skeleton of skeletons was made of questions, documenting the bare-bones
outline of any manufacturing process. Its blanks left space for the number
of workers employed, hours worked, tools and maintenance involved,
maintenance, the division of labor, the list of operations and the number of
times each was repeated, among other exacting elements of business logic.
 Among the industries queried, Babbage made special note of the printing
process itself, in the “copying through six Stages in Printing this volume.”
“It is here that the union of the intellectual and the mechanical departments
takes place,” he wrote, recalling Lovelace—in a pattern, on a template,
along the cells of a table—“down to the cavity punched through the letters
a, b, d, e, g, &c.”—“stamped, died, and formed”—“these movable types,
the obedient messengers of the most opposite thoughts,” where the most
conflicting of theories are “themselves copied by casting from moulds of
copper, called the matrix.”
Templates were everywhere in the industrial age. Viewed in their light,
the novelty of analytical engines lay in the application of manufacturing
templates to the domain of symbolic logic, philosophy, and art. Templates
used in the production of symbolic goods necessarily share a history with
physical manufacturing, influenced by the same industrial forces affecting
the manufacture of material goods—such as the division of labor,
mechanization, assembly line production, lowered barrier to entry,
decreased cost, increased output, efficiency, and standardization. (But also,
the rise of middle management, globalization, union busting, and lame
corporate team-building events.)
Literary production in the West accelerated with the advance of the
printing press, alongside rising literacy rates. Increasingly, folks were
beginning to write things for fun and profit, not just edification. The
eighteenth and nineteenth centuries saw 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. Magazine and
journal subscription was also becoming a trend at home, increasing both the
demand for and the supply of literary content.
Allow me to bore you with some numbers, for a sense of scale. Authors
prolific by eighteenth century standards wrote on the order of a dozen or so
major literary works. Among them let’s count Daniel Defoe, author of
Robinson Crusoe, who wrote, not counting the pamphlets, around eight
novels and sixteen works of longer nonfiction; Jonathan Swift—no stranger
to text generators found his Gulliver’s Travels—published fifteen; Goethe,
six novel-length pieces alongside as many theater plays; Jane Austen,
authored nine long novels; Laurence Sterne, under ten, depending on how
you count Tristram Shandy; and Sir Walter Scott, of Waverley fame—seven
novels. You can probably find an outlying graphomaniac or two somewhere
in addition, without affecting these averages.
The word count began to climb dramatically in the nineteenth century.
Dickens and Dostoevsky, whose outputs were the envy of contemporaries,
and who both wrote serially for periodicals, penned fewer than twenty
novels each. The Soviets published Count Tolstoy’s collected works in
ninety volumes. Alexandre Dumas wrote more than a hundred plays and
novels, likely with the help of ghostwriters.
By the first half of the twentieth century, such exceptional output became
the norm. Georges Simenon, the Belgian writer of detective fiction,
published several hundred novels, 75 of them about French police detective
Jules Maigret. Elizabeth Meade Smith, the author of A World of Girls
(1886), wrote more than 300 mysteries, romance novels, and works of
young-adult fiction. Paul Little, more than 700 novels. Kathleen Lindsay
(Mary Faulkner), more than 900, under many pseudonyms. Edward L.
Stratemeyer, the American author-producer behind the writer syndicates
responsible for such popular series as The Rover Boys and Nancy Drew, in
the thousands. The Spanish author of romance novels Corin Tellado, a fast
4,000.
Consider a few additional points of reference, quoted here from my
published research, where you can check my receipts. 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. (Yes, that’s right: individual
authors in the twentieth century eclipsed the production of whole countries
a few decades prior!)
According to my own figures, based on the American Catalogue of Books
in Print, 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. True figures were likely higher, because the official
record omitted popular “low” genres, like true crime and pornography. A
reference I found in the Writer’s Digest from 1928 mentioned one pulp-‐
fiction publisher, Street & Smith, receiving “close to 900,000 manuscripts a
year.” Another was reported to buy “nearly a million words a month for
their various pulp fiction magazines.” “Fiction House authors draw real pay
from this office,” an editor reported, citing figures as high as $5,000 a year.
At the going rate of one to two pennies per word, the sums translate to an
average author placing on the order of 250,000 to 500,000 words per year
—a volume equivalent to five lengthy novels annually, not counting
rejections.
Something clearly changed on the supply side of literary production in the
short time between Jane Austen and Writer’s Digest.
Discounting quality, concentrate for a moment on the sheer textual output
in terms of classical economics. The rise in the demand for printed
materials can be explained, among other causes, by the printing press,
Protestant Reformation, the Enlightenment, rising literacy rates,
urbanization, and educational reform. This wedge of the economic lemon
—the demand side—has been squeezed well in other studies.
Seldom do we consider the supply. Though it changes through time, the
matter of literary composition remains roughly static by volume, at least
since Aristotle (whose Poetics is still taught widely today). As my slow
progress on these chapters can attest, natural human mental physiology
(body and brain) place hard limits on an individual’s capacity for creativity
(I need a break).
Under the conditions of constrained supply, one expects the rapid increase
in demand to push prices and wages up. A good word should, by economic
logic, cost the same or more with the expansion of the market. That’s not
what we observe. Author wages (per word) instead fell dramatically toward
the end of the nineteenth century. Where, in the 1850s, writing for serial
journal publication, the young Charles Dickens garnered around a dollar per
word (according to my rough estimates normalized to American currency
today)—in the 1920s, a decent American author would be lucky to expect a
few cents. And today most authors write for free, online.
The fall of wages, instead of the expected rise, implies the increase in
productivity, beyond baseline physiological capacity. As with other labor
dynamics of the time—remember the Jacquard loom!—we must also look
to the methods of production for an explanation, on the side of the supply.

A DRAMATIC SITUATION

Meet another one of our lovely weirdos, then, in the spirit of democratic
hospitality. History was not kind to Georges Polti. In 1895, this little-known
French writer published a small book meant to help playwrights compose
new plays, called The Thirty-Six Dramatic Situations. As you’d expect, it
contained a list of thirty-six dramatic situations, such as “supplication,
deliverance, vengeance, pursuit, disaster, revolt, the enigma, rivalry,
adultery,” etc. Each situation was introduced briefly, in broad strokes,
containing several further examples Polti collected from the classics.
For example, in the section on pursuit, Polti explained that our interest
should be “held by the fugitive alone; sometimes innocent, always
excusable, for the fault—if there was one—appears to be inevitable,
ordained; we do not inquire into it or blame it, which would be idle, but
sympathetically suffer the consequences with our hero, who, whatever he
may once have been, is now but a fellow-man in danger.” Four subtypes
were further enumerated, including: (a) “fugitives from justice pursued for
political Offences, etc.”; (b) “pursuit for a fault of love”; (c) “a hero
struggling against a power”; and (d) “a pseudo-madman struggling against
an Iago-like alienist.”
Anticipating how this might sound to his high-minded colleagues, Polti
didn’t apologize. “They will accuse me of killing imagination,” he wrote.
“They will say I have destroyed wonder and imagination. They will call me
‘the assassin of prodigy.’ ” Far from it! Without method, the aimless
imitation of the past stifles creativity. “All the old marionettes reappear,
inflated with empty philosophic ideas.” Only a systematic exploration of the
unknown could produce truly novel artistic forms. But novelty for novelty’s
sake was also meaningless. His method, in addition, trivialized the very
idea of novelty. Authors would now concentrate on ideals higher than mere
innovation—ideals like beauty, balance, and harmony.
Armed with the scientific method, playwrights of the future could employ
countless possible dramatic combinations, “ranged according to their
probabilities,” Polti predicted. Further, rigorous experimentation was bound
to change art in fantastical ways, by “analyzing orders, systems and groups
of systems”—as yet unseen methods for literary production would emerge
—“measured and classified with precision.”
By the book’s end, Polti revealed the practical impact of his project.
Many situational arrangements, “their classes and sub-classes,” were at
present ignored, he wrote. Modern plays are too boring and too familiar to
the audience. Many more situations “remain to be exploited” in
contemporary art. Fuzzy Romantic terms like “imagination,” “invention,”
and “composition,” must spring naturally from the “art of combination,”
furnished by the composite building blocks of his system.
And then the surprising conclusion: “Thus, from the first edition of this
little book, I might offer (speaking not ironically but seriously) to dramatic
authors and theatrical managers, ten thousand scenarios, totally different
from those used repeatedly upon our stage in the last fifty years. The
scenarios will be, needless to say, of a realistic and effective character. I will
contract to deliver a thousand in eight days. For the production of a single
gross, but twenty-four hours are required. Prices are quoted in single
dozens. Write or call, No. 109, Passage de l’Elysee des Beaux Arts.”
I don’t know if anybody took Polti up on the offer to buy scenarios by the
dozen (people were still wearing striped long-pants to the beach), though I
combed the archives for months to find out. Given his lack of commercial
success as a playwright, Polti’s name was rarely mentioned in scholarly
literature, and when it was, was quickly dismissed. This thread seemed to
lead to a dead-end. I was beginning to lose hope when an important clue
soon unraveled the case.
 The foreword to the English edition of Polti’s work was written by one
William R. Kane—publisher and editor of the Editor: Journal of
Information for Literary Workers. The title caught my attention, because
“literary workers” was not the way I was taught to discuss great authors.
Published first in 1896 out of Franklin, Ohio, the magazine opens a view
into the American literary market at the turn of twentieth century. Full of
adverts for typewriters and filing systems, the Editor published insider
pieces with titles like “500 Places to Sell Manuscripts!,” “What Young
Writers Should Read,” “Consideration for Editors,” “Does Poetry Pay?,”
“The Historical Short Story,” among other topics of interest for professional
authors. Advertisements for Polti’s slim volume ran here for decades, along
with many other, similar, “how to write a best-seller” schemes. Though
rarely mentioned in the annals of high modernism (the literary movement
usually associated with this period), the practice of “writing by the
template” prevailed among literary workers. Their output would fuel a mass
phenomenon, spreading across numerous pulp-fiction magazines and soon
through the Hollywood blockbuster machine.
The Editor sat on top of an iceberg of similar publications. Weeks of
tedious microfiche trawling at the New York Public Library produced many
others, including the Author’s Journal (1895) out of New York; the Author
out of Boston, Massachusetts, founded in 1889 and described as “a monthly
magazine to interest and help all literary workers”; the Writer also out of
Boston, established in 1887 and still active today; the Publisher (1891); the
Writer’s Monthly (1925), affiliated with the Home Correspondence School
in Springfield, Massachusetts; Pennsylvania’s Writer’s Review (1925);
Markets and Methods for Writers (1925) by the Palmer Institute of
Authorship, a correspondence school out of Hollywood, California; and
New York’s own Author’s Digest (1927).
Further, advertisements in the above journals led me to dozens of “how-‐
to” books similar to Polti’s. Remarkably, these were mostly lost to literary
history, hardly mentioned in research, nor collected by university libraries.
The record was mostly silent on their existence.
The following titles give us a good sense of their content. Consider,
within the fullness of primary sources, a list of selected works, including:
Skeleton Essays, or Authorship in Outline (1890) by Thomas English;
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.
This was finally the machinery of literary production on an industrial
scale. Keywords repeated consistently across the above archive recall those
surveyed by Babbage in his Economy: template, method, technique,
skeleton form, mold, schema, system. These tools were missing from the
scholarly record because authors who made it out of the “fiction factory”
rarely revealed the dirty secrets of their trade. Painting by the numbers
diminished the myth (and the earning potential) of any artist. It wasn’t good
for one’s career. And so the world marveled at Vladimir Nabokov’s
exceptional use of index cards in the making of his novels, while the pages
of the Editor routinely pushed ordinary articles about “using index cards to
author your bestseller,” alongside advertisements for home card-filing
furniture.
Now, I did hunt down and read most of the books above, so you don’t
have to. Some of them were pretty bad. Others offered reasonable advice.
The techniques proposed within—the machinery, so to speak—generally
fell into one of several categories.
The first included “skeleton forms” proper—consisting of near-finished,
prefabricated pieces that could be easily assembled and modified lightly to
produce reasonable outputs, such as letters, essays, or short stories. These
were sometimes meant to—ahem—“aid” in the production of graded
university papers, as was the case with Skeleton Essays, or Authorship in
Outline (1890) by the aptly named Thomas English.
The second involved the revival of general Aristotelian formulas,
procedures, and “rules of thumb”—derived for the construction of
“balanced” compositions, whether in fiction or for general rhetorical
impact.
From these, it may be useful to highlight The Technique of Drama (1892)
and The Analysis of Play Construction (1908) by William Thompson Price,
a self-proclaimed “Aristotle from Kentucky.” Price was especially
entertaining to read because his advice for young writers was backed by his
experience as a play-reader for several major Broadway theater production
houses. He therefore peppers his rules of composition with negative
examples from “a bottomless pit” of rejected manuscripts on his desk: how
NOT to write a best-seller. Price’s manuals were adopted by several
influential theater programs, just then coming into existence, such as the
one at Harvard, taught by Professor George Pierce Baker from 1905
through the 1920s. While met with skepticism at first—the very idea of
teaching art systematically was controversial—Baker’s students found
commercial success on Broadway and in Hollywood. (Among them,
Believe Me, Xantippe [1918], a silent-screen comedy starring John
Barrymore, Drew Barrymore’s grandfather.)
Apart from technique-based manuals, we find also the various visual or
graphical aids pressed into service for narrative ends.
For example, Carolyn Wells, a successful author of detective fiction,
concluded her Technique of the Mystery Story (1913) with advice on
diagramming complex story structures, so as to better handle their
trajectories. Harry Keeler’s Web-Work, published first in the Student-Writer
in 1917, represented a more advanced system, anticipating our
contemporary developments in network theory. Instead of singular narrative
plots, Keeler emphasized a multiplicity of character arcs, which often
developed outside the visible plot structure. For example, a soldier
transformed by the war meets an old friend at dinner, where they share
elements of their backstory, which happened “offstage.” Thus, in addition to
the story told, an author was advised to track multiple backstories, so as to
more accurately represent character development. Keeler’s work urged
authors to document and visualize large fictional worlds, connected by a
complex web of social relations—where a single story plot represented one
path among many possible paths, through a multifaceted world, full of other
plots and entanglements.
Yet another type of literary device instructed authors in the systemic
gathering of realistic detail—a database—culled from newspaper clippings,
conversation snippets, and notes from real observed situations.
Mark Twain patented and sold “self-gumming” scrapbooks for authors as
early as the 1870s. Writing for the Editor in 1903, Jack London (author of
White Fang and Iron Heel) urged authors to keep a notebook at all times.
Authors in the habit of collecting such notes would quickly find themselves
overwhelmed by collected information. A single notebook wasn’t enough
for an author submitting five manuscripts a year. Notebooks grew into filing
cabinets. Devices such as the Eureka Pocket Scrap Book were routinely
advertised on the pages of the American Stationer, the Dial, the Athenaeum,
and Popular Mechanics. The Educational Specialty Company out of
Detroit, Michigan, sold the Chautauqua Literary File—a large piece of
home-office furniture involving a complex color-coding system for the
quick retrieval of reference information.
The Phillips Automatic Plot File Collector consisted of some two hundred
custom-made containers, referenced to the included plot classification
system. Advertised on the pages of magazines and in the back of published
books, the file cabinet promised to “hold many thousands of uniform items
of plot material, designed to contain Plot Material, Plot Germs and
Complete Plots—as well as statistical and reference data—in the form of
Notes, Newspaper Clippings Excerpts, Photographs, Pictures, Pages, and
Complete Articles.”
Henry Phillips explained how he used his Automatic File Collector in the
writing of a fictional character named Pod:

“POD” ENTERS BERLIN IN HIS NIGHTSHIRT—Popular General’s

Ludicrous Adventure on a Sleeping Car. Instantly, we feel here is a
character. That which impresses us is a good guide. We consult the
 index and find: Character—V.—49, 73, 91, 94. This is an exception;
four-fifth of the subjects have but one reference. 49. discloses that we
have invaded Vicissitude of ASPIRATION, thence to V.—THE CHAR
ACTE R OF MAN, narrowed down to 49. QUALITIES. Not what we
want. 73. we find under DESTINY, THE MIND OF MAN,
DERANGEMENT. There it is (a) “Character,” just what we want. If
there is any doubt we continue our search and find 91, HUMOR—
EMOTIONS—FARCE. No, Pod too pathetic for farce. Again under 94,
under PATHOS, seemed too tragic and we revert to DERANGEMENT,
our first impression.

The last category of machines found in the writers’ workshop might be

particularly relevant for our history of machine writing.
Manuals like Plotto (1928) by William Cook and Plot Genie (1935) by
Wycliffe Hill proposed fully fleshed-out algorithmic text-generation
systems in book format. Both authors expanded on Polti’s original index of
“dramatic situations,” adding rulesets for their composition—something
Polti promised but never delivered.
Plotto was the more complex of the two, containing a sophisticated
cross-referenced numbering system of snippets, which could be recompiled
into a story depending on the initial “seed” and subsequent pathway along a
branching tree of choices. Imagine a really complicated Choose Your Own
Adventure story, for authors instead of readers. Similar in operation to
Kircher’s Mathematical Organ, any path through the several included fold-‐
out reference tables could be dereferenced into a complete skeleton of a
composition. In this way a PURSUIT could lead to CAPTURE or
EVASION, with each branching further into multiple and sometimes
recursive possibilities. After the initial “plotting” path, the author would
iterate over individual framed elements “to put flesh on the bones,” by
using additional reference materials, containing realistic detail (perhaps in
consulting the snippets filed earlier!). In this way, the archetype of an
“aggrieved joker villain in the pursuit of revenge” could be made more
concrete by the villain becoming a “portly accountant” or “disfigured
policeman,” their revenge taking shape as “financial fraud” or “murder of
city officials.”

Wycliffe Hill’s Plot Genie (1935) “supplies the following outline.”

Authors were expected to use the Genie wheel, pictured below, to
generate random numbers. These would then be referenced from
the charts to create the backbone of a romantic plot. Wycliffe Hill,
The Plot Genie Index (Hollywood, CA: E. E. Gagnon, 1935), 36.
 The Plot Robot consisted of a rotating paper wheel sandwiched
between cardboard plates. An indent to rotate it can be seen in the
middle right of the image. A number would then be displayed
within the pictured magic ball, inspiring the young author. Hill,
The Plot Genie Index.

Where Plot Genie was simpler than Plotto, it also added an element of
probability. The book came with an elaborate insert—a wheel sandwiched
between two cardboard plates called the “Plot Robot”—picturing a
stereotyped genie in the process of inspiring a young author, crouched
beside a magic bottle. Several spins of the wheel produced a string of
numbers, which could be transferred onto the included blank form. Hill sold
multiple such forms, alongside reference books for romance, comedy, short
story, and action-adventure. The system culminated with the appropriately
titled Ten Million Photoplay Plots, containing sections on the “Twelve
Reasons for the Non-Sale of Stories,” “Censorship Regulations,” “The Cost
of Production,” and even “Plagiarism.”

SKELETON FORMS, VISUAL AIDS, filing systems, and algorithmic text generators
contributed to the increased productivity of American authors, writing for
an expanding market.
Together, these documents provide ample evidence for the widespread use
of templating tools in the production of textual goods, made in the USA.
(The Soviet Union was another hot spot for the industrialization of literary
output.) In the recollection of one contemporary author: “In those days, I
was rigidly following the rules of what I call the [name of the journal
omitted] school of the American short story . . . Stories of the school which
it dominated 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.”
As we will see in the next chapter, the soon-to-be first wave of computer
scientists spearheading research in artificial intelligence through text
generation were initially aware of their industrial-era literary predecessors.
Earliest documented examples of AI text generators implemented
rudimentary versions of these paper-and-ink systems, using similar
techniques such as story grammars, event databases, multiple-pass
expanders, random tree walks, randomized-event engines, network
traversals, and background-world generators.
It would take some time for machines to gain the computational strength
to outperform paper systems (roughly when digital file systems began to
rival the filing cabinet in storage capacity). The future held a few surprises
yet for aspiring “author-manipulators,” in a timeline full of intrigue, rivalry,
pursuit, conflict, deadly misfortune, and deliverance. Crucially, at this point
our story also begins to splinter into many diverging paths—among
departments of linguistics, anthropology, literary study, and computer
science—each losing sight of the other’s trajectory.
CHAPTER 6

Airplane Stories

T HE FIRST GENERATION OF COMPUTERIZED storytellers, among them

American Airlines Boeing 747 N740PA, grew up reading Russian
fairytales. Why?
Two steps forward, one step back. We arrive at the doorstep of the
twentieth century with luggage full of smart devices predating the
computer. Another bag needs unpacking before we pass that threshold. My
prescient machine audience will remember, with perfect recall, the advance
of what I have previously called “template culture,” affecting all areas of
human industry, including the intellectual. American authors in particular
developed the practice of writing by template on a massive scale—a century
later fueling my personal addiction to watching all 1,292 episodes of the
Law & Order franchise, across sixty-two seasons of television.
The notion of “story templates,” in all its forms, had another important
development in parallel to its impact on literary production, on the side of
literary analysis.
Just as the Romantics privileged the notion of exceptional genius
individually, a dominant mode of understanding literature at the time was
practiced from the standpoint of national exceptionalism. Philologists like
Johann Gottfried von Herder, Germaine de Stael, Jacob Grimm (of the
brothers), and Henry Longfellow (of the bridge) sought in literature
something like “the ultimate spirit of a nation,” seeking within it aspects
uniquely attributed to the essence of German, French, American, or
whatever people.
The discipline of comparative philology developed in the mid-nineteenth
century, in opposition to the national approach and in response to a
prevailing mystery: How is it, the comparativists asked, that certain
similarities appear between literatures of otherwise isolated nations?
The problem was especially acute in the study of folklore, where
geographic isolation was well documented. The figure of the trickster, for
example, appeared independently, though in different guises, in both
Finnish and Hopi American indigenous oral traditions. Instead of
concentrating on each nation’s unique differences, the comparativists
highlighted similarities. Looking past national “guises” revealed—and we
are back on familiar ground—universal archetypes. Just as Polti and Co.
created templates to assemble literary works, a school of European scholars
dis-assembled literature to derive templates.
Among the comparativists—initially German, French, and Soviet—we
find one Vladimir Yakovlevich Propp, who, like Polti, wrote another one of
those quietly influential books, titled Morphology of the Folktale, in 1928.
In the book, Propp . . .
I’m just going to spoil the surprise now: Propp’s thirty-one “dramatic
functions” echo Polti’s thirty-six “dramatic situations” (discussed in the
previous chapter) extensively and without attribution. Having compared the
texts line by line, I’m comfortable in signing my name under a footnote in
the history of mild plagiarism. That’s not my only point, however. A
bourgeois author from the antagonistic West, Polti was far removed from
the Soviet Academy. His kind was also a topic of general taboo among the
literati, being the “assassin of wonder” and all. Nobody in academia
anywhere took the low-brow, mercantile “author-manipulators” seriously.
Besides, Propp also submitted original work in response to Polti’s
promised, but never-delivered “Laws of Literary Invention.”

Component functions of “The Swan-Geese.” Vladimir Propp,

Morphology
of the Folktale, trans. Laurence Scott (Austin, TX: American
Folklore Society, University of Texas Press, 1968), 96.

Having assigned lettered variables to his “dramatic functions,” as others

had done in Plotto and Plot Genie, Propp set out to discover an algorithm
governing their arrangement. A typical passage from the book explains “the
ways in which stories are combined” as follows:

A tale may be termed any development proceeding from villainy (A) or

a lack (a), through intermediary functions to marriage (W) or to other
functions employed as a denouement. Terminal functions are at times a
reward (F), a gain or in general the liquidation of misfortune (K), an
escape from pursuit (Rs), etc. This type of development is termed a
move. Each new act of villainy, each new lack creates a new move. One
tale may have several moves, and when analyzing a text, one must first
of all determine the number of moves of which it consists. One move
may directly follow another, but they may also interweave; a
development which has begun, pauses, and a new move is inserted.
 Once such narrative calculus could be established, Propp used it to
annotate a sample tale, about a swan who stole a child.
Where did these algorithmic fairy-swans fly?
In chemistry, by analogy, our ability to decompose physical substances
into their chemical components has enabled the discovery of fundamental
regularities, such as Avogadro’s law. Propp hoped his calculus of narrative
functions would similarly enable new literary discoveries.
With the help of his formal method, story formulas could now also be
compared among national traditions, solving the riddle of comparative
philology. Commonalities between distinct cultures would be explained by
the laws of universal composition. Other formulas could be discovered and
described with precision.
Finally, what was for many of our author-engineers (like Polti) a
haphazard process—the making of individual building blocks for language
manipulation—promised to become an organized, scientific endeavor—a
true science of literature.
Or so it seemed at the time. Propp’s system did not attract many followers
initially. To complicate matters, Soviet academia became more insular
following Stalin’s purges in the 1930s, just a few years after the book’s
publication. The “great leader” was more interested in propaganda than
story time. A wrong word could send one to a gulag. Many talented Soviet
scholars fled to Europe and North America. World War II would bring
down the iron curtain, further isolating Soviet science from the rest of the
world. Propp continued to teach classes on the folktale quietly through
turbulent times at Leningrad State University, his work all but abandoned
by his colleagues.
We’ll return to it later, in 1968, when Morphology would be translated
into English, becoming common reference material for young Boeings in
every hangar.

RANDOM GENERATION OF ENGLISH SENTENCES

Meanwhile, somewhere in America, the comparative inquiry into language

was developing on its own parallel tracks. Trained in comparative
philology, scholars like Zellig Harris, Leonard Bloomfield, and Edward
Sapir founded some of the country’s first linguistics programs, at the
universities of Pennsylvania and Chicago, and at Columbia.
Refugees themselves or the children of refugees, the Americans were,
alongside their global counterparts, part of a massive academic movement
that was soon to get an official name—structuralism.
Any ism makes me nervous, because it usually identifies a cluster of
contested ideas, loosely defined. Rigid minds confuse a label of
convenience for dogma. Rarely does one adhere to all of the tenets of an
ism uncritically, without some individualized modification or dissent. It’s
common therefore for two people belonging to the same ism to hold fairly
diverse views. Or for a person to be called an “-ist” just by the virtue of
professing one of several potentially -ist ideas.
Let’s not therefore lose our way in an argument about definitions. We are
interested, remember, in the spread of template culture at the roots of what
we now call artificial intelligence. We also have a matter of a plane trained
on Russian folktales to address. Both routes pass through structuralism.
Broadly speaking, the search for hidden universal templates by any name
—with all their cognates like patterns, scaffolding, archetypes, laws,
functions, regularities—may be classed under the rubric of structuralism.
Call it what you will. At a cocktail party, it will be enough to paraphrase
Jean Piaget (on his own suggestion) who explained it in terms of “a search
for hidden order.” Roman Jakobson, who helped coin the term, wrote that
structuralism reveals the “inner laws of a system.” Writing for The Oxford
Dictionary of Philosophy more recently, Simon Blackburn explained that
“behind local variations in the surface phenomena,” structuralists sought
“constant laws of abstract structure,” where “superficially diverse sets of
myth, or works of art, or practices of marriage, might be revealed as sharing
the same pattern.” So far, we should be on familiar ground: patterns and
more templates!
With these approximate definitions in hand, let’s anticipate a few
interesting complications. The idea of two people “sharing a pattern” or
“structure” of expression seems simple enough. Your language sounds like
mine. But where, literally, does this pattern or commonality reside?
The answer isn’t immediately obvious. For example, you could say: “I
have nothing in common with that person. I deny any similarity. My pattern
is in my head, and yours, in yours.”
It could be also that we both went to the same school, and you would say,
“Sure! We both had the same English teacher. They taught us to speak that
way. The pattern was in a textbook that we shared.”
Another would protest, “No. It’s our brains that are similar. The pattern
comes from the anatomy common to all humans.”
These are all reasonable answers—the pattern resides in the mind of the
observer; it’s in our textbooks; in our social upbringing; in the classrooms;
and in the brain anatomy shared by all humans.
Where should we look for patterns then? A scientific study requires
definite objects and locations. Confusingly, the answers send us into a
compass rose of directions. The first would urge the study of our own
psychology. If the template is in my head, I might also study how I come to
know things—epistemology. The second would send us to an archive of
school textbooks, to study the transmission of patterns in language or text
—linguistics or philology. Here, we could also visit various classrooms to
observe teaching in action, using the tools of anthropology or sociology.
And if I told you the pattern was in our shared brain anatomy, we could
both take up scalpels and MRI machines, so as to better study brain
structures—by the means of neuroscience. What started as an insight into
language has led us to a diverse range of scholarly pursuits.
Incidentally, all of these fields were transformed by structuralism, rivaling
each other also for the primacy of their “home field” methods.
The question of “Where we should look for patterns?” entails another
serious problem: Are things like grammars, rulesets, and formulas even
real? A structure could be real in the sense of a skeleton providing actual
support for the body outside. Remove the framework and the building
crumbles.
Structures could also be real in the sense of a blueprint or a diagram. It
could be just a neat way to describe a thing precisely, not actually
supporting anything. Blueprints are useful for construction. But buildings
also stand up fine without them. Did structuralism finally set out to discover
scaffolding skeletons, or structures similar to diagrams and blueprints?
There was no consensus on that point.
In either case, everybody agreed structures lie “beneath the surface.” Yet,
most humans manage to live in buildings and speak languages without a
thought to their hidden structural regularities. How would one validate a
discovered structure to make sure it is real? What if the discovery turned
out to be just another fairytale?
 In chemistry, a law can be confirmed by experimentation. We know it to
be true by its power to predict physical outcomes in the world. Up to this
point, we had no way to test Polti’s or Propp’s formulas for correctness. Did
the world contain thirty-three or thirty-six dramatic situations? The authors
spoke from authority without being able to evidence their claims. And in
places where their systems disagreed, we had no means of comparing the
two conflicting accounts of structure. A vital part was missing from this
young science—empirical verification.
Enter Noam Chomsky, a student of Zellig Harris and the linguist whose
colorless green ideas slept so furiously in the first chapter of this book. Like
other structuralists, Chomsky was searching for hidden grammars, similar
to Propp’s “rules of composition.” Unlike the others, Chomsky found a way
out of the verification conundrum: The test of any structural principle
would be found in its ability to generate the thing it was claimed to
structure.
Think of it in terms of cooking: If you claim to have reverse-engineered
your grandma’s apple pie recipe, you should be able to make it. And if your
sister claims to have discovered a “more authentic” recipe, her pie should
be closer to grandma’s in the end. In Chomsky’s case, the test of a grammar
lay in its ability to generate grammatically correct language.
The idea was quickly recognized as a major breakthrough in linguistics.
Where Propp’s analysis took stories apart to derive a formula, Chomsky
proposed to put them back together according to the formula. If the formula
was right, like grandma’s recipe, it would be capable of regenerating tasty,
well-formed fairytales. But, if the language produced was mangled, we’d
know there was something wrong with the formula. The generative test
allowed rigorous direct comparison between two conflicting rulesets: may
the better storyteller win.
Much more could be and has been written about Chomsky’s long career
elsewhere. For now, it’s enough to note (a) that his interest in grammar-like
rulesets was common for most structuralists, where (b) his method of
evaluating them was truly novel.
In 1957, at the Massachusetts Institute of Technology, where we are with
Chomsky’s early work on syntactic structure, Building 26 was about to
receive its first IBM mainframe computer—one of the first in the country
built specifically for academic purposes. There it was quickly pressed into
service by the Research Laboratory of Electronics (RLE), belonging to the
Center for Communications Sciences at the Department of Modern
Languages, where Chomsky worked as part of a large team. Computers fit
the purpose of testing generative grammars well. To describe such
grammars, the lab created a special machine language, called COMIT,
capable of generating thousands of sentences rapidly.
It is here that, for the first time, we see the completed puzzle of modern
conversational AI finally assembled in one room. Other universities would
soon follow, launching a collegial competition between language labs, in a
race to develop the most well-spoken robot. Chomsky’s theory encouraged
direct competition.
Research on the Random English Sentence Generator debuted in 1961 at
MIT. Funded by the US Army (Signal Corps), the US Air Force (Air
Research and Development Command), and the US Navy (Office of Navy
Research), it featured the re-generation of a children’s book by Lois Lenski,
The Little Train (1940). Victor Yngve, another prominent MIT linguist,
authored the paper.
 The military was interested in children’s books as part of a new military-‐
command structure, at the time itself imagined as a giant computer,
distributed among regional control centers. Complex automated systems
needed to address threats, like a nuclear strike, rapidly and
semiautonomously. Shooting down a missile required the lightning-fast
activation of a complex, coordinated response, involving many moving
parts. The interface between the various components of national defense—‐
human, weapon, and technical—was therefore complicated by the level of
computational expertise available to the personnel. If computer code was to
be the new “behind-the-scenes” language of command and control, the
“front-of-the-scenes” facing the soldier would have to happen in ordinary
language.
Generative grammars promised a way to make technological complexity
palatable to humans, in a way natural to human communication. We should
not therefore be alarmed to learn that the MITRE Corporation (“Solving
Problems for a Safer World”), founded in 1958 and one of the largest US
defense contractors today, authored hundreds of papers on generative
grammars—owing a part of its success to folktales and children’s books like
The Little Train.
Unlike what you would find in the grammar books of the past, Yngve’s
grammar was extremely compact. It consisted of three rules, broken down
into about 108 sub-routines, in about as many lines of code. These sufficed
to construct many simple sentences in the English language. The compact
program matched well the physical limitations of the lab’s IBM 709
computer, which could hold no more than 32,768 words in memory,
including code used for system management tasks like reading data or
printing. (Computer documentation of the time often used words instead of
bytes to report units of memory size.)
The three topmost rules included those for addition, randomization, and
insertion. Let’s look at each carefully:

1. A = B + C [addition]
2. A = B, C, D, E . . . [randomization]
3. A = B + . . . + C [insertion]

The first meant that any construction A could be extended by addition.

For example, any mention of SENTENCE, could be substituted with
SUBJECT (denoting the agent of action) + PREDICATE (that which is said
of the subject). In the sentence “Mary had a little lamb,” “Mary” is the
subject and “had a little lamb” is the predicate.
We could now build a sentence from scratch by using addition alone.
Move each element on the right side of the equation (B + C) to the left side
(A) at each step as you follow along. We begin by expanding the
SENTENCE into SUBJECT (“Mary”) + PREDICATE. We then expand the
PREDICATE into VERB (“had”) + OBJECT. The OBJECT expands to
ARTICLE (“a”) + NOUN PHRASE. Finally, the NOUN PHRASE expands
into ADJECTIVE (“little”) + NOUN (“lamb”): Mary had a little lamb.
The second rule meant that any element A could be randomized by any
elements B, C, D, E, and so on. In this way, if A was a NOUN, B through E
could contain some options, including lamb, chicken, goat, and so on. A
basic random number generator varied the results accordingly. Given a few
choices, the second rule allowed for variations, such as “Jerry had a little
goat.”
The third rule made possible the insertion of elements between other
elements. This was done to accommodate quirky English phrasal verbs like
sort out, give up, or take off. Unfortunately for us English-as-a-second-‐
language speakers, these are sometimes interrupted by a direct object, such
as the phrase “We THREW the box AWAY.” “The box” needs to be
inserted between the verb, “throw,” and its attached preposition, “away.”
Combining addition and insertion, phrasal verbs could be expanded from
the middle of the phrase, as it happens when we THROW the bigger box
AWAY. Here, between “throw” + “away,” we insert the formula NOUN
PHRASE = ADJECTIVE (“bigger”) + NOUN (“box”).
These three rules were almost enough to generate a coherent text. Almost.
Given a limited vocabulary from the original story, Yngve reported a
hundred generated results about an engineer named Small, in charge of the
little train, among them:

(007) WATER IS BIG.

(025) STEAM IS HEATED.

(041) SMALL IS PROUD OF WATER.

(063) WATER IS POLISHED.

(074) HE HAS STEAM.

(078) WATER IS POLISHED.

(079) WATER HAS LITTLE AND OILED SAND-DOMES IN A BELL.

(090) WHEN A FIRE-BOX HAS FOUR WHISTLES, HIS BELLS, SMALL AND
ENGINEER SMALL, HE HAS HEATED, LITTLE, LITTLE, OILED, OILED AND OILED
WHISTLES.

The results should have sufficed to verify the grammar, according to

Chomsky. We could now be sure that the rules addition, randomization, and
insertion are “correct,” in comparison to any other grammar, because these
have been shown to generate grammatically correct sentences in practice.
“No reason has yet appeared for doubting that it would be possible to use
this same formalism for a complete grammar of English,” Yngve wrote in
conclusion of his paper. If three rules were enough for The Little Train, the
goal of describing English completely seemed to be only a few experiments
away.
Early experiments with Chomskian grammars showed promise, but also,
with the benefit of hindsight, some oddities.
Simple grammars, such as the one implemented by Yngve, contained no
logic governing the meaning of the sentence, only its syntax. The generator
therefore produced sentences that were, while always grammatically
correct, not always sensible. Initially, Chomsky and his colleagues
attempted to bracket the issue of meaning out of consideration altogether.
Generating meaningful sentences could be done later, by the introduction of
another ruleset, controlling semantics. So they thought.
Meaning, however, proved to be a more difficult nut to crack than syntax.
Syntax represents a system both finite and closed. A sentence operates by
its own rules, independent of the world. Meaning resides within the world,
subject to its vicissitudes. Meaning changes depending on context. For
sentences to be meaningful, in addition to being grammatically correct, they
have to play by the rules of physics and history. And these higher-order
grammars represented not one but numerous additional systems. Whether
they could be brought under the same order as syntax remained to be
proven. All our genies, from Llull and Wilkins to Lovelace faced exactly
this problem.
The other oddity of Chomsky’s generative grammar concerned its
physical location. As we discussed before, where you looked would
determine the kind of answers you were going to get.
Chomsky understood grammars to reside in the human mind. But the
rules were implemented on a computer. What worked in one place might
not necessarily work in the other. For example, it was clear that Yngve’s
team prioritized terse, compact programs. In retrospect, some of that
emphasis might have had something do to with the limitations of early
computers, not brains. What if the discovered grammars were purely a
machine-based occurrence, not made for human consumption?
Yet Yngve constantly conflated humans and computers—for instance,
when writing about adequate computer memory “capable of holding no
more than about seven symbols.” This minimally viable requirement of
computer hardware led the team to speculate about “English and probably
all languages.” “No more than about seven items need ever be stored in the
temporary memory,” Yngve concluded, now seemingly talking both about
computer memory and human anatomy.
But why should languages or brains have the same limitations as an IBM
706? One and the same result could be achieved by different means. It
could be that computers and brains rely on entirely distinct software.
Discovering one ruleset might not have anything to do with the other. Did
Chomsky’s grammar describe actual structures in the brain? Or was it just a
metaphor for the mind, approximating some of its powers by analogy on
computer hardware?
The results were exciting, regardless. A number of computational labs
sprung up across the country, each attempting to improve on the basic
formula.
Some of the notable story generators from this era included BASEBALL,
the automatic question-answerer programmed by Bert Green, Alice Wolf,
Carol Chomsky, and Kenneth Laughery at MIT in 1961; Daniel Bobrow’s
STUDENT question-answering system (1964) out of MIT and later
Raytheon BBN; Joseph Weizenbaum’s ELIZA at MIT; Margaret
Masterman’s TRAC (1971) out of Cambridge, UK; Kenneth Colby’s
PARRY (1972); MARGIE (1973) by Roger Schank and team at Stanford’s
Artificial Intelligence Laboratory; Schank and Robert Abelson’s SAM
(1975); James Meehan’s SAM-influenced TALE-SPIN (1976) from Yale;
MESSY (1976) by Matthew Appelbaum and Sheldon Klein at the
University of Wisconsin; Eugene Charniak’s MS. MALAPROP (1977); GUS
(1977) by a team from Xerox Palo Alto consisting of Daniel Bobrow,
Ronald Kaplan, Martin Kay, Donald Norman, Henry Thompson, and Terry
Winograd; and Wendy Lehert’s QUALM (1977), among many other
(though still dude-heavy) examples.
Linguistics led this first wave of research, but AI also contained other
nascent disciplines, such as robotics, vision, logic, decision-making,
cybernetics, and neuroscience.
Alan Turing’s writings on the mind cemented the conversation with
robots in the popular imagination, as did the staged dialogs between ELIZA
and PARRY in 1973. A machine called RACTER published a much-‐
discussed volume of poetry titled The Policeman’s Beard Is Half
Constructed (1984). Images of talking robots became commonplace,
appearing in films like Jean-Luc Godard’s influential Alphaville (1965),
Stanley Kubrick’s 2001: A Space Odyssey (1968), and George Lucas’s Star
Wars (1977). With the advent of more powerful personal computers, the
Pentagon’s vision for the military was being integrated into the workplace.
Commercial AI systems began to appear in the fields of medicine, law,
education, and aviation.

TALE-SPIN
Chomsky’s grammars were increasing in complexity, too, running into the
limits of syntax, at the boundary of meaning. Syntax, as we saw, governed
the rule of sentence composition. Sentences hold singular thoughts. To
compose more sophisticated, chained units, such as paragraphs or stories,
higher-order rulesets were needed.
And where does one find a ruleset for the composition of stories?
Folklore studies proved an obvious ally. The reappearance of Propp’s
Morphology of the Folktale, translated into English in 1968, coincided with
the expansion of the AI program, now in its search for story grammars. The
random generator of English sentences would graduate to produce random
English tales.
In 1976, James Meehan deposited his dissertation at Yale, under the
direction of Roger Schank, whose previous work on the SAM storyteller
took a tack different from that of generative grammar. Schank and Meehan
were both interested in production of “schemas,” so prominent in the work
of industrial-era authors at the turn of the twentieth century.
The computer program complying Meehan’s dissertation was called
TALE-SPIN, made to produce Aesop-like fables in English, “interacting
with the user, who specified characters, personality characteristics, and
relationships between characters.”
Apart from how-to manuals like Plotto and Plot Genie, schemas also
emerged as a concept in structuralist psychology, most prominently in the
work of Jean Piaget (who defined structuralism for us earlier in the
chapter). Similar to an algorithm, grammars represented a set of rules. But
unlike generalized language grammars, a schema contained the bare-bones
outline of a specific real-world situation—a template.
Piaget thought schemas played a key role in a child’s development.
According to his research, humans recognized complex situations based on
just a few outstanding qualities. How can you tell a birthday party apart
from a funeral? Observing just the party hats and balloons, one can be sure
to infer a birthday event. No further information is required (unless we’re at
a funeral for a clown, I guess). The mind fills the rest in.
Where grammars contained terse formulas, schemas described skeleton
“scripts” for common situations, containing typical details: lists of expected
characters, locations, goals, and activities. Scriptwriters like Polti sold
schemas for use in Hollywood scripts. Psychologists speculated that
children accumulated scripts in memory as a kind of cognitive shorthand for
later recognition. The two types of scripts, dramatic and psychological,
were intrinsically related.
Crucially, schemas described the relation of things out there in the world
—party hat, balloon, cake—where grammars described the arrangement of
words within language. Grammar-based generators produced grammatically
correct but often meaningless sentences. Schemas arranged words in a way
that also made sense in context. Taken together, generating new text ruled
by both grammar and schema promised stories both grammatically correct
and meaningful.
By the time Meehan started his dissertation, grammar-based language
generation seemed like a solved problem. Schank, his advisor, established
the efficacy of script schemas. Meehan’s TALE-SPIN combined both
approaches in one program, citing both Polti—the French playwright—and
Propp—the Soviet folklorist—as influences in addition.
TALE-SPIN wrote stories by first generating a world and then charting a
path through it, similar to Keeler’s social “web-work.” On program startup,
the reader was asked to select several options from a list of pre-defined
characters: a bear, bee, boy, girl, fox, crow, or an ant, among others. The
corresponding character schema instantiated a specimen template, using
randomly chosen type attributes: Will our bear be tall or short? Male or
female? Brown or black? A bear template specified that bears live in homes
and that homes were to be located somewhere on a map, another template.
Bears also owned things, perhaps dishware and furniture. The program
dutifully conjured randomly appointed caves, meadows, hills, and forests,
each according to its own schema. A story could now begin with a bear
lounging comfortably in its hilltop cave.
Given several characters, the program would also attempt to model a
character’s state of mind—what it knew, wanted, and believed—as well as
its goals and motivations. Our Bear may be hungry, for instance, but doesn’t
know where to find honey. His friend the Bee knows where to find honey.
Their mind-states differ, presenting diverging narrative possibilities. The
potential for narrative resolution of the plan HUNGER swings into motion. A
template indicates that HUNGER can be resolved by action EATING.
Here the stage is finally set for a heartwarming tale about Bee and Bear
sharing a meal on an alpine meadow. At least in theory.
Given the short time to dissertation completion, parts of Meehan’s
TALE-SPIN remained speculative. The included code produced output
similar to the following skeletal structure:

> JOHN BEAR WALKS FROM A CAVE ENTRANCE TO THE BUSH BY GOING
THROUGH A PASS THROUGH A VALLEY
THROUGH A MEADOW.

> JOHN BEAR TAKES THE BLUEBERRIES. JOHN BEAR EATS THE BLUEBERRIES.
THE BLUEBERRIES ARE GONE.
JOHN BEAR IS NOT VERY HUNGRY. THE END.
 > -- DECIDE: DO YOU WANT ANOTHER STORY ABOUT THOSE CHARACTERS?

Some sixteen years after the birth of Yngve’s Random English Sentence
Generator, and more than forty years after Propp’s folktale Morphology,
machines thus learned to tell simple tales beyond the sentence, both
grammatically correct and mostly meaningful.

AVIATION INCIDENT REPORT

Despite these successes, schema- and grammar-based approaches were

losing steam by the end of the 1990s. By this time, the military had lost
interest in fairytales, as well. For now, graphical user interfaces, not
ordinary language, became the dominant paradigm of interfacing with
computers. Compact, rule-based grammars of the kind that worked well for
language generation, did not materialize for more complex narrative units.
Propp’s formulas turned out to be moderately useful for analysis though
failed Chomsky’s generative test. Computers were not yet powerful enough
for the task.
Schemas worked well for specific situations but were themselves too
cumbersome to compose accurately. Imagine trying to describe a complex
environment, like an airport, in terms of scripts, schemas, and dramatic
situations. The variety of objects and actions involved would require almost
infinite description. An airplane alone would encapsulate a schema for
wings, engines, fuselages, people, babies, clothing, lunches, dinners,
alcohol use, luggage, bolts, fasteners, lights, air systems, steward schedules,
uniforms, captain’s hats, brims, buttons, airline logos, and so on—zooming
in and out of schemas within schemas indefinitely.
Within aviation, an industry rife with checklists and talking cockpit
warning systems, Meehan’s TALE-SPIN, made after the template of
Russian fairytales, proved useful in its narrow confines for the analysis of
aviation incident reports.
Writing for Boeing’s Knowledge Systems group and citing the TALE-‐
SPIN directly, Peter Clark explained that incident reports “described and
analyzed unusual/unexpected chains of events during aircraft operation,”
which could include mechanical problems, pilot sickness, or unusual
maneuvers. These reports resembled folktales, replete with drama and
suspense. A typical summary from the FAA’s Incident Data System
(FIDES) contained flight summaries in the following format:

961211044319C
PASSENGER CUSSED OUT FLIGHT ATTENDANT TAXIING TO RUNWAY.
PIC RETURNED TO GATE.
PASSENGER REMOVED.

961216043479C
TURBINE RIGHT ENGINE FAILED. DIVERTED TO RIC.
OVERWEIGHT LANDING.
HAD CONTAINED TURBINE FAILURE.

960712045359C
PIC BECAME INCAPACITATED AFTER LANDING.
LANDING WAS ERRATIC AS WAS TAXIING.
FIRST OFFICER TOOK OVER.
PIC HAD STROKE.

As stories, these could be analyzed using narratological tools. They

contained beginnings, middles, and endings. There was danger in them,
departure, and homecoming. If a computer was to understand aviation
incidents, it should, according to Clark, “at least be able to generate
plausible incident stories,” and to reason about them sequentially, in terms
of narrative, composed of characters and events.
Borrowing the “scripts and schemas” approach from Meehan, Clark was
therefore able to emulate “flight stories” similar to the ones found in the
actual FIDES database by generating them programmatically. Just like
Meehan’s birds and bears, Clark’s pilot entities contained scripts for goal-‐
oriented behavior, such as “Get the passengers to Dallas,” involving a series
of predefined tasks. A new event—the illness of a passenger, for example
—would change the schema from “transport passengers” to “get medical
help.” The story engine would continue to simulate actors and schedule
events until it reached a satisfactory conclusion:

THE PASSENGERS BOARDED THE PLANE.

THE PLANE TAXIED TO THE RUNWAY.
THE PLANE TOOK OFF FROM SEATTLE.
THE PLANE CRUISED TOWARD CHICAGO.
THE ENGINE CAUGHT FIRE.
THE PLANE LANDED AT CHICAGO.
THE PASSENGERS WERE EVACUATED FROM THE PLANE.

Clark’s program was intelligent in the sense that it modeled the “rational
pursuit of goals.” It “understood” random events to the extent that it
contained a rudimentary theory of other minds, where fictional pilots and
passengers expressed certain goals, knew how to accomplish them, and
were transformed by their achievements. Once such schemas were in place,
Clark hoped that the reverse would be possible as well. A system for story
generation was also a system of story analysis: Why did the pilots divert
their course? Was it because a passenger fell ill onboard? A smart machine
could traverse the causal chain both ways, finding patterns among millions
of messages generated by airplanes daily. And perhaps one day, such a
system could fly a plane on its own.
Down but not out, the number of research papers related to schema-‐
generators continued to climb through the 1990s, in part due to their success
under controlled conditions—in fields like education, game design,
customer service, and medicine, where scripts circulated long before the
computer.
Educators were beginning to modernize their assignments, syllabi,
evaluation methods, and lesson plans for digital use. Schemas powered the
creation of immersive computer-game worlds, in cult classics like Baldur’s
Gate and Dwarf Fortress. Similarly, the interaction between patients and
physicians has been templated into digital forms and drop-down menus
found in medicine. Schemas were increasingly starting to mediate between
billing and diagnostics, to the point of becoming an essential part of every
medical institution’s information infrastructure.
Today’s patients and passengers lose themselves entering a maze of
options in conversation with automated customer service representatives,
used widely by airlines and insurance companies. Anyone observing their
physician struggle with clinical notes can appreciate the pervasive nature of
narrative schemas in large modern bureaucracies. Stories of health and
illness are exchanged within the hive of a hospital, where a chance turn
down the wrong hallway—a missed click of a doctor’s mouse, registering a
wrong diagnosis from a drop-down menu—can forever alter a patient’s
outcomes.
CHAPTER 7

Markov’s Pushkin

A BIG LEAP FORWARD FOR LINGUISTIC intelligence in the twenty-first

century came with a little-known, older paper by a statistician,
published in a journal of mathematics, on the topic of Russia’s greatest
poet, Alexander Pushkin.
Wheel, table, pattern, template, schema—the time has come for us to
trace the last of our arcane figures—the Orion’s Belt—a string of words,
arranged in a series of statistically probable continuations based on
observed probabilities, called a Markov chain. An idea before its time, it
gathered dust in the archives until computers became powerful enough to
appreciate it. Where the figures we have encountered so far were all a part
of the same family of methods, Markov took a radically different and
simple approach. The chain represented an entirely separate branch of
statistically minded language scholarship, which went against the grain of
conventional linguistics, changing the very meaning of meaning in the
process.
One of the central questions I posed in the last chapter related to the
various possible locations of meaning: in the world, in the mind, or on
paper. Chomsky and other members of the cognitive-psychological school
emphasized the mind. Language grammars were therefore to be found in
brains. Though, as we saw, Chomsky’s colleagues looked for them mostly
inside of computers.
Schemas derived their power from capturing situations in the real world.
They described common situations, whether at the doctor’s office or the
airport, where doctors often carry stethoscopes and pilots can be found near
planes. Language speakers also attach such schemas to individual words.
The word airplane for a pilot means not just its dictionary definition but a
series of associations different from those of a passenger. These two
meanings of airplane overlap yet also diverge. They mean something
different to different people because they hail from distinct life experiences.
In their influential book titled The Meaning of Meaning (1923), linguists
C. K. Ogden and I. A. Richards called this connection the “sign-situation,”
where words link to things through common co-occurrence. A parent
feeding their child at breakfast will simply talk about food. “Yummy
oatmeal,” they might say while spooning oatmeal from a bowl. With
repetition, the child forms a long-lasting association between the sign
“oatmeal” and the situation of eating oatmeal, involving actual grains,
bowls, spoons, parents, and breakfasts. The word oatmeal will come to
“mean” that relationship—linking the word, the mental image, and the
eating of the stuff. For Ogden and Richards, sign-situations rendered
language “a legitimate object of wonder” and “the source of all our power
over the external world.”
But while children ate oatmeal, language models did not. Any description
of language was bound to be disconnected from experience and therefore
“dreaming furiously.” Language models struggled with semantics
(meaning). To break out of the prison of language, statisticians took context
to mean immediate language context—no physical experience necessary. It
would be enough to glean that some words tend to occur near other words.
That rough proximity sufficed to produce meaning.
For instance, knowing nothing about oatmeal in the real world, our
machine child could learn that the word often occurs next to words like
bowl, eat, table, breakfast, and spoon. And it rarely occurs next to many
other words, like bulldozer or munitions. From these facts, it could
reasonably surmise that oatmeal has something to do with eating breakfast
at tables with spoons, and not, say, war or construction. Such machine
“learning” would depend on observing prior occurrence. The more a
computer would read about oatmeal, the more it learned (metaphorically)
about its common linguistic contexts. And it would be decades until
machines had enough memory or processing power to retain enough
information to start making sense.
A century prior, somewhere in Russia during the revolutionary period,
Andrey Andreyevich Markov lived his life as a mathematician, not terribly
interested in brains, bulldozers, oatmeal, or any other objects of wonder. In
1912, he wrote a curt letter to the patriarch of the Russian Orthodox
Church, announcing his formal defection from the church. “As for my
reasons,” he wrote, “I appeal to the passages from my book the Calculus of
Probabilities. Given my work, the stories of improbable events at the root
of Christian and Jewish faith should be treated with extreme suspicion. The
topic has no relation to my life or mathematics,” he concluded.
A similar disenchantment permeated his paper on Alexander Pushkin, one
of the most revered poets in the Russian language. Its reception among
literary scholars was stone cold.
Inspired by the work of quantitative philologists of the previous
generation, the paper concerned Pushkin’s masterpiece novel in verse,
Eugene Onegin. Markov mentioned nothing about Pushkin’s literary or
philosophic merits. The world of things or ideas was bracketed out entirely.
Instead, Markov converted the individual letters of the text into a series of
numeric values, which he called “linked trials” or “linked chains,” each
indicating the probability of any one letter to succeed any another.

The first line of Pushkin’s Eugene Onegin expressed as a series of

chained letter probabilities. Original from “A Statistical Study of
Eugene Onegin Illustrating a Linkage of Chained Trials,”
Imperial Academy of Sciences in Saint Petersburg, series VI,
volume VII, issue 1, 1913.

Imagine for a moment attempting to guess what letter may follow the
initial letter Z to continue a valid English word (similar to the game of
Wordle). The letter A would be a good guess, given its occurrence in words
like zag or zap. An O would start us on a path to zoo or zombie. However,
certain letters would never follow Z, as no words contain the combination
of ZK or ZP, for example. You can try experimenting with Markov chains
on paper by writing down a few random letters and then guessing at the
possible words that could result from their beginnings.
Here we must also embark on a little historical side quest. The elephant in
the room we’ve managed to ignore so far has been the development of the
telegraph, happening roughly from the time of Babbage—past Markov—to
MIT’s first research computer. The history of telecommunications and
computing merge definitively after that. The telegraph is important for us
because many clever folks working on language at the time came to the
question of word meaning from the perspective of machine communication,
not human texts like Eugene Onegin or even The Little Train.
One of the basic practical problems in telegraphy was simply the volume
of information that could be sent over a single wire. The volume of water
flowing through a pipe could easily be calculated according to a formula.
No such formula existed to reckon the information capacity of telegraph
cables. Noise and information loss posed additional concerns. Yet the
volume of machine communication was growing exponentially, dwarfing
human output. For over a century, everything—from love messages to
stock-purchase orders and machine instructions—was conveyed by
telegraph. Cables spanned continents, struggling to keep pace with
increasing traffic. A way to calculate their capacity, similar to water volume
in pipes, was needed to manage the flow of information.
Markov’s model of language fit this problem well because telegraphs
processed information in chains, as strings of text, entered letter by letter.
Engineers working on increasing the sheer capacity of communication in
Markov’s vein didn’t much care what kind of information was sent through
the wires, nor what it meant. A string of text moved through wires like city
water, providing utility to its users. What senders or receivers did with that
information was outside an engineer’s concern. “Frequently the messages
have meaning, that is they refer to or are correlated according to some
system with certain physical or conceptual entities,” Claude Shannon wrote
in his groundbreaking 1948 essay, titled “A Theory of Mathematical
Communication.” “These semantic aspects of communication are irrelevant
to the engineering problem,” he concluded sharply.
 For Shannon and his colleagues, it sufficed to imagine human
communication as a probabilistic Markov-chain process, “generating
messages, symbol by symbol.” For instance, Shannon explained that a
message could denote a sequence of letters chosen at random, such as
“XFOML RXKHRJFFJUJ ZLPWCFWKCYJ FFJEYVKCQSGHYD
QPAAMKBZAACIBZLHJQD.”
Given the probabilities of how often each single letter occurs in the
English language independently, a string could be approximated as “OCRO
HLI RGWR NMIELWIS EU LL NBNESEBYA TH EEI ALHENHTTPA
OOBTTVA NAH BRL.”
Advancing by all probable two-letter combinations, or bigrams, produced
“ON IE ANTSOUTINYS ARE T INCTORE ST BE S DEAMY ACHIN D
ILONASIVE TUCOOWE AT TEASONARE FUSO TIZIN ANDY TOBE
SEACE CTISBE.” Here, we can see that certain correct words begin to
occur by chance, including on, are, and Andy.
Given the probability of any letter occurring after any previous two
letters, in combinations of three, or trigrams, Shannon composed “IN NO
IST LAT WHEY CRATICT FROURE BIRS GROCID PONDENOME OF
DEMONSTURES OF THE REPTAGIN IS REGOACTIONA OF CRE.”
This begins to resemble English, randomly producing correct words in, no,
whey, of, and the. Other combinations seem plausible: I had to look up birs
and pondenome in the dictionary in case they chanced on actual words (they
did not).
Things get more interesting when we compute the probabilities of word
combinations, instead of letters. A distribution of any word following, given
the previous, creates a list of all two-word units, or bigrams. To make your
own, take the first two words of any sentence and advance forward, word
by word. Given the previous sentence, this yields, from the left to right:

to make,
make your
your own
own take
take the
the first
of any first two
any sentence two words
sentence and words of
and advance
advance forward
forward word
word by
by word

Note how each of the bigrams connects to the previous, making it a linked
chain. Now imagine calculating the frequency—how often each
combination occurs—over a long text, such as Harry Potter. With modern
computers, this can be done with every word string encountered online,
producing a giant table of prior observed probabilities.
Once these priors are computed, we can use them to complete sentences.
“When I grow up, I want to be _____.” Statistically speaking, a plausible
continuation of the sentence could be “an astronaut,” or “the president of
the United States,” or “a doctor.” My grandmother tells me I wanted to be a
street sweeper because I was fascinated by autumn leaves. If we collected a
bunch of such responses from children, mine would be less probable than
others. Some word combinations, we would encounter rarely if at all.
Virtually nobody says they want to be “a tree,” or “a shoe,” or “makes a
magnificent asparagus”—that doesn’t even make grammatical sense. Priors
give us only the plausible continuations of a chain.
Given the raw frequency of words in the English language, without regard
to the previous link, Shannon approximated (by hand, remember):
“REPRESENTING AND SPEEDILY IS AN GOOD APT OR COME CAN
DIFFERENT NATURAL HERE HE THE A IN CAME THE TO OF TO
EXPERT GRAY COME TO FURNISHES THE LINE MESSAGE HAD
BE THESE.” The results based on single-word probabilities look even more
convincing than our character simulations before!
A distribution of likely two-word combinations, or bigrams, linked,
produces: “THE HEAD AND IN FRONTAL ATTACK ON AN ENGLISH
WRITER THAT THE CHARACTER OF THIS POINT IS THEREFORE
ANOTHER METHOD FOR THE LETTERS THAT THE TIME OF WHO
EVER TOLD THE PROBLEM FOR AN UNEXPECTED.” You can
imagine that three, four, and longer word linkages would work even better,
though, too long of a link becomes unique and not useful for further
extrapolation.
While still mostly meaningless, Shannon’s output rivaled that of Yngve’s
grammar-based English Sentence Generator. The surprising effectiveness of
the probabilistic approach lay in the absence of any grammatical or
semantic assumptions. The method worked purely by statistics, powered by
the diligence of the original tabulation. The greater the number of statistical
combinations calculated in advance, the better the expected results.
Markov chains were hungry for observed probabilities. To be effective,
they needed to read, process, and retain a large volume of texts. The same
hunger also limited their reach. Recall that in the 1960s, Yngve’s random
generator required the application of some 108 multiword rules alongside
several dozen vocabulary words from The Little Train in order to operate.
IBM’s 709 machine used in that experiment barely had the capacity to hold
the program in its memory, limited as it was to 30,000 words in total,
including those reserved for the operating system. Markov calculated
20,000 of his “letter chains” by hand, half a century prior. In other words,
machines did not yet perform better than humans. The derivation of every
bigram (two-word combination) in the English language, as suggested by
Shannon, would be impossible by hand or by computer at the time,
requiring data points in the billions. The computation of trigrams (every
three-word combination) would need more storage and more processing
power, by orders of magnitude. It would take almost a century after Markov
for computers to get powerful enough for that task.
Brute reading force scaled with processing power. The problem of
digitizing printed text, instead of doing the tabulations by hand, also had to
be solved. The technology for scanning and digitizing documents was still
in its infancy. For these reasons, actual (instead of hypothetical) working
Markov-chain language generators did not appear in earnest until later.
Their use in the 1960s to 1980s was limited to more contained systems,
whether musical composition, table games, calculating insurance
probabilities, where the number of potential combinations was smaller.
Andrey Andreyevich Markov continued to publish research in mathematics,
actuary science, and chess—never again returning to literature.
Though Shannon never implemented Markov’s chains digitally, the idea
made a lasting impact. In theory, the algorithm worked. Markov’s and
Shannon’s mathematical models of communication made no assumptions
about the world or the brain. Given the number of times I wrote “given,”
they followed only the precepts of observed probability. Everything hinged
on how much the computer could tabulate—or “learn,” again
metaphorically speaking—before the task of generating sentences, by
ingesting as much text as possible. The brute force required to process most
everything published in the English language would not be available until
the twenty-first century.
Physical limitations of mid-twentieth-century computing did not
completely preclude the evolution of probability-based text machines,
though these appeared in somewhat exotic new fields, like optical character
recognition (extracting text from images), spell-checking, and document
retrieval.
In 1959, Bell Telephone Labs patented a device for the Automated
Reading of Cursive Script, US3127588A. The patent proposed using shape
frequencies to detect possible errors in the recognition of cursive script. The
work by W. W. Bledsoe and I. Browning from the same year pioneered the
use of “word context” for the purpose of “pattern recognition and reading
by machine.” Thus in the scanning of a text, machines could try to guess an
illegible word based on the probability of its individual characters within
the sentence.
A team from Cornell’s Aeronautical Laboratory—airplanes again!—‐
proposed a novel spell-checking algorithm in 1964. Lockstep with our
statisticians, it relied on dictionaries and probable “word contexts” for the
task of “correcting garbled English text.” Highlighting the absence of
assumptions about meaning, the authors wrote, “We have been concerned
instead with the recognition of the identity of a word or of letter sequences
by making use of the fact that only certain letter sequences occur in English
with any appreciable probability.” The program was capable of correcting
sentences like “TAAT SOUIDS LIKE A AIVTDE KITTEN,” SAUD
TOMYN—“ ‘That sounds like a little kitten,’ said Tommy.”
With the passing of the telegraph and linear ticker tape, the metaphor of a
chain gave way to the related notion of “word vector space,” more
appropriate for the representation of printed documents, where words
expanded along two dimensions, instead of a straight line.
I love the idea of a vector space because it evokes the night sky,
preferably somewhere up in the mountains, away from light and
civilization.
Imagine the universe of language as a constellation of concepts scattered
across the starry skies: each concept in its place and each following a
different trajectory in relation to others. This picture changes the way we
think about word meaning in general. Rather than looking for meaning in
some fixed dictionary, we can agree that words “mean” by their common
locations or contexts. The word sunny, for example, often occurs in the
context of weather; the words speed record next to race or the Olympics. In
this way, the words fast and speedy are similar because they occur in similar
contexts—they have similar locations and trajectories, or vectors.
Of course, the state of the world itself isn’t always so sunny. For instance,
large, complicated social problems such as abusive language or racism can
manifest in the constellation of certain often-repeated expletives or negative
stereotypes. Cleverly, the computer can distinguish the two meanings of
race based on its contextual location. Race words found near speed records
and distances mean something different from those found near disparity or
mass incarceration. Not so cleverly, a computer that “learns” from biased
language—everything ever written—may continue to spread abuse or
racism it has learned from humans.
 Despite its shortcomings, a vectorized model of language has allowed
machines to begin making complex inferences, related to the meaning of
the words they don’t really “know.” Machine “understanding” is instead
woven out of the statistical context. The “correct” answer to any given
question represents also the “likely” continuation of a string.
Words in the shape of mathematical context-vectors further enable the
computation of semantic distance, capturing shades of difference between
fuzzy negative attributes, such as bad and its even worse cousin terrible.
With word embeddings we can also attempt to calculate those words
furthest from any point in our starry skies, finding, automatically and
without teaching the computer anything about the world, that awesome
occupies reaches of space far away from awful.
Vectors can also be used to find related clusters or constellations of
words, which signify topics, literary genres, or themes in a conversation.
Just as contexts define words, a constellation of contexts can define whole
systems of thinking. A certain conceptual word cloud can be highly
correlated with the domain of physics, for instance, where another cluster of
words corresponds to politics. Similar words, occurring together frequently,
produce patterns in three dimensions, not just on a page but in whole books
and libraries, within a vector space.
The idea of word embeddings and vector spaces found another harbor in
the literature on document retrieval, written by scholars working at the
intersection of computer and library science. Modern online search giants
were born out of the necessity to organize and retrieve the world’s
information, long the purview of truly arcane arts like filing, cataloging,
and indexing. Before computers, librarians filed things according to hand-‐
labeled classification schemas, in which a book by Markov could be
classified under subjects such as Science, Mathematics, Probabilities
(QA273–280, in the US Library of Congress classification schema). A
book’s indexer reached into the content of the work to pull out notable
themes, names, places, or references for reader convenience.
By the 1960s, researchers proposed ways to extract keywords
automatically.
Among them, Peter Luhn extended the notion of “words in context,” to
“keywords in context” at IBM. Gerald Salton at Harvard used word
frequencies to derive a metric of document similarity: You could now
request a “similar” book from your university library—the similarity
computed automatically using word distances.
Karen Spärck Jones built on her work with Margaret Masterman at
Cambridge to propose “weighting” keywords statistically, as a function of
term frequency rather than meaning. Matches on rare word combinations
would therefore count for more than common words. Querying a search
engine for “recipes for Beef Stroganoff” would amplify the importance of
the rare term Stroganoff while discounting the common for, recipe, and
beef.
Sergey Brin, Lawrence Page, Hector Garcia-Molina, and Robin Li (the
founder of Baidu) further built systems that weighted mutual citation.
Documents that referenced other documents counted for more than those
that were filed in isolation. Search engine technology was thus built out of
parts found in library science, philology, and linguistics, using techniques
like text alignment, indexing, cataloging, term frequency comparison, and
citation analysis.
The rest is history, well documented—you can google it or ask your
friendly neighborhood chatbot for assistance. Words in context brought us
into the modern age of literary intelligence.

STATISTICS RULES THE DAY, for now, as schema- and grammar-based

approaches to language generation have proved less effective. Its
ascendance holds important implications for our understanding of language,
as well as some cautionary tales.
To the extent that Zellig Harris or Noam Chomsky exemplified a
paradigm of structuralist linguistics, the British linguist John Rupert Firth
captured the probabilistic approach. “You shall know a word by the
company it keeps!” he famously wrote in his Studies in Linguistic Analysis
(1957). Instead of searching for structures that support meaning beneath the
surface, Firth proposed simple collocation, apparent at the surface level.
A baby may have learned about oatmeal at the breakfast table, but far
more associations for oatmeal (the way it’s made, how much it costs) were
accrued in the encounter with textual oatmeal, by babies at breakfast tables.
Most of our theoretical knowledge about the universe comes wrapped in the
vestments of language, not direct experience. I have never seen a planet or
an electron up close. But I know about them from textbooks, in textual
context. And you know of my knowing them, not by peeking into my brain,
but because I told you, using words. Much of a word’s meaning lies outside
of grammars or structures or relations to physical things. Words might mean
simply in context of other words. Context makes meaning—that was Firth’s
powerful rebuke to his colleagues in linguistics.
Among the linguists, Firth best exemplifies the ideas baked into the
advanced language technologies of our times. But the idea of “words in
context” was not without its problems. The loss of other locations where
meaning could be found—as we discussed earlier, the world and human
mind—came at a cost. That cost is worthy of some consideration:
First, with the coming of “statistical” intelligence, we must let go of the
mind metaphor. Where the grammarians were hoping to discover human
brain structures by writing chatbots, modern chatbots work by mechanisms
emphatically not human. We certainly do not produce language in our
minds by statistical probabilities. Our brains do not convert words into
numbers or vectors, nor do we have the capacity to iterate over huge
numerical datasets.
Yet the language models in production today continue to use human-‐
cognitive metaphors. The machine “learns.” Its statistical connections are
called “neurons,” creating a “neural network.” The effects of this
technology similarly confound us with human-like capacity for language,
achieved in ways alien to us. I’m therefore suspicious of all metaphors
ascribing familiar human cognitive aspects to artificial intelligence. The
machine thinks, talks, explains, understands, writes, feels, etc., by analogy
only. Its progress holds no explanatory powers for our human ways of
being. Though we could use it as a tool to explore the way we read, write,
and interpret each other, it is not itself a substitute for thought—no more
than driving is substitute for a daily run. The two actions might get you to
the same point, but how they get you there matters more than the
destination.
Second, the reliance on context for meaning causes obvious problems. A
long-standing debate between “descriptive” and “prescriptive” linguists
helps illustrate the point. The descriptivists, like Firth and Richards,
maintained the word means in relation to how people speak. The job of a
descriptive linguist is simply to observe, collect, or tabulate. The
“prescriptivist” has other ideals in mind, outside language. Prescriptivists
insist on speaking or writing well: clearly, purposefully, with rigor and
respect. Clarity, purpose, rigor, or respect are not ideals found in language
haphazardly. We aspire to them for other, cultural or political reasons.
Language isn’t a closed system, after all. It is itself contextual, located
within specific cultures and societies.
The fact that machine intelligence can achieve linguistic proficiency by
other, short-circuited means doesn’t fail to impress. But it also immediately
runs into the cultural boundaries of prescription, beyond language.
Machines learn by ingesting all contexts, good and bad. The texts they
ingest, in addition, contain more than a trace of human politics. Some
voices have been systematically excluded from the published archive used
for training. Others have been unfairly amplified.
Yet other modalities of human intelligence manifest outside of text. Some
ideas important to us aren’t even verbalized, or verbalized often enough.
Yet training treats all of it, in aggregate, as flat context. The machine makes
for a perfect descriptivist.
Description doesn’t tell the whole story, however. We don’t just live in the
world as it is. Humans actively build the world they wish they had. Notions
of freedom, justice, equality, fairness, success, failure—intelligence itself
—are all aspirational qualities. The answer to any given question cannot
rely on a mere average. We aspire to the best answer, the more perfect
continuation of thought. The weight of probable frequencies fails to meet
these ideals. The best of us occurs more rarely in comparison to our base
impulses. Frequency alone therefore does not suffice for intelligence, in a
basic pedagogical sense. AI needs to do better. But it cannot, if words are
all it has to go by.
 Finally, an overly frequentist understanding of intelligence threatens
radical solipsism. Until recently, the critical mass of the world’s textual
output lay in the human past. Human intelligence evolves through time, in
response to history. However, taken to its logical conclusion, the volume of
new machine-generated text will quickly surpass that of its human-made
training corpus. The sheer weight of machine production might overwhelm
any further novel human contribution. Language development may stop, in
a sense, as the greater portion of the archive is filled with machine chatter.
Machines will then primarily be trained by other machines, but to what
ends?
In other scenarios, an overwhelming excess of machine capacity may lead
to some interesting results, though not perhaps in the ways we would
expect. By analogy: I may appreciate watching soccer, but not at the point
where it would be played by robots running at supersonic speed. Such a
game would be, dare I say it, boring. The difficulty of dribbling and scoring
goals by using an imperfect instrument such as the human foot, is what
makes the game interesting.
My final objection therefore signals neither alarm nor metaphysical crisis.
It’s simply that, however effective or powerful, a muscular artifice for the
sake of artifice isn’t that intelligent or interesting to me. I tire after one
hallway at the museum, much less in front of “everything ever created.”
The flourishing of human chess today, following its onetime total machine
conquest, bodes well for other human games, like language. The tabulations
of Eugene Onegin take nothing away from its art. The pleasure of putting
together a word puzzle into a beautiful sentence is not diminished by the
fact that someone or some thing can do it better or faster than me or in
greater quantity. Perhaps we simply enjoy playing this game too much. It is
tuned perfectly to the bandwidth of our peculiar limitations—their small
victories and defeats included.
CHAPTER 8

9 Big Ideas for an Effective

Conclusion

S O FAR, WE’VE BEEN PATIENTLY assembling the modern chatbot from parts
found on the workbench of history. I have tried to animate that device
for you, by strings of anecdote and light philosophical commentary.
We didn’t neglect to feed this Pinocchio in addition. Even before
publication, my text will enter the collective machine unconscious—in the
way all digital texts today are vacuumed into datasets—to be processed for
the use of training of future text generators. What will it think of its
parents? Will there be a sense of pride or shame in family history? Or
perhaps nothing, because all outputs were already anticipated? The
connection between Ibn Khaldun and Llull, for instance, though obscure, is
well documented. The same can be said of Leibniz, Babbage, and Lovelace.
A few entirely novel synapses were sparked in the linkages between
industrial manufacture and the rise of mass literary markets. The work on
pulp fiction, structuralism, and early computer science is also entirely my
own novel contribution, based on original research.
A few loose leads remain, poking out of the seams between chapters.
Let’s trim them and draw toward a conclusion.
Several important sources lay too far afield to cover adequately. For
instance: Histories of the Turing machine and the Turing test often neglect
the direct influence owed to Ludwig Wittgenstein’s lectures, along with the
presence of Margaret Masterman, the pioneer of machine translation, in the
same classroom. Masterman’s universal thesaurus harkens back to Wilkins
and other universal-language makers, central to a whole separate and
important branch of AI—machine translation. It would require more than a
chapter of its own, orthogonal to the direction of our travel. Aspects of
encryption used for diplomacy or military communications would also lead
to entirely different exit points. Regionally, an emphasis on German or
Soviet computing would make mine a far less Anglocentric endeavor. The
rich traditions of Arabic, Indian, and East Asian philology are mostly out of
my range of expertise. There’s ample room in these gaps for machine or
human readers to amble on their own.
Another alley not fully explored would lead us to I. A. Richards, a part of
the linguistic oatmeal duo figuring briefly in the last chapter. Aside from his
work on linguistics, Richards was one of the major thinkers at the
foundation of the New Criticism movement, which emphasized close
reading and formal analysis of text.
This book was conceived as both a tribute and a rejoinder to Richards’s
Science and Poetry. The original volume helped launch W. W. Norton as a
publishing house in the 1920s. Reading it fondly a century later, I was
surprised to find a pervasive nostalgia for poetry, besieged by the
“onslaught” of science. Elsewhere, I knew Richards to be an unsentimental
scholar, writing with clarity and sharp insight. Yet in Science and Poetry,
the thought wanders. In the conclusion to his book, he writes of “getting the
guns into position,” “sciences that progressively invade every province,”
science that will “force [other myths] to surrender,” and “the Hindenburg
Line [. . .] held” but also, in the same sentence, “abandoned as worth neither
defence nor attack.” Who these enemies are exactly, we don’t know. One
just gets the feeling of general embattlement.
Long a student of Richards, I can’t help but read war between the lines.
Written not long after the end of World War I, in 1925, and expanded and
reissued on the cusp of World War II, in 1935, Science and Poetry is
haunted by conflict.
Almost a century later, the possibility of another world war again hangs in
the air. As I write these lines, more than 200,000 civilians and soldiers have
perished as a result of the Russian aggression in Ukraine. More deaths will
surely follow. An aging dictator, out of touch with reality, daily threatens
total nuclear annihilation.
Armies of chatbots have been conscripted on both sides, but especially on
the side of Russia. My days are occupied in real-time observation of troop
movements, on the ground and online. The swings of public opinion and
soldier morale have become matters of life and death. I am contributing in
my own small ways, by translating war communiques, filing take-down
notices, and writing code to automate certain language tasks useful for the
war effort. The battle between Western support and the unwieldy machine
of Russian state propaganda rages in countless online forums, where
information and disinformation merge into a rush of war commentary,
fueled by AI authors.
Poetry did not, however, protect Russia from the myths of imperialism or
fascism, in the way that Richards had hoped for humanity in his book.
Literature was instead mobilized as one of its last remaining weapons. The
Russian invasion of Ukraine in 2022 happened under the pretext of
protecting a Russian linguistic minority in the Ukrainian Donbas region, “in
danger of cultural erasure,” alongside the alleged removal of Pushkin,
Tolstoy, Lermontov, and Dostoevsky from school curriculums. So dear
leader appealed to great literature, doting on his own nostalgia for an
empire whose time has passed. Left out of his grandiose nationalism were
Pushkin’s Cameroonian ancestry and Lermontov’s glorification of Russian
imperial war crimes in the Caucasus.
I cannot therefore share in Richards’s enthusiasm for the protective
bulwarks of literary imagination. Nor in his stark separation between
science and poetry. My thesis runs contrary to his. Many languages other
than English, like German or Russian, do not honor the distinction between
the sciences and the humanities. There, science just means “the organized
study of,” a synonym to the English scholarship. The science of poetry
holds no contradictions in that sense. It indicates simply the systematic
study of the thing, beyond casual reading.
While romping through the centuries, we found the roots of computer
science to be inextricably entwined with literary and linguistic concerns.
Viewing them apart has impoverished both communities: poets, in terms of
financial and cultural capital, and programmers, in terms of belonging to a
deep intellectual tradition. Worse yet, the division gave both communities a
kind of a shallow myopia, where AI seems to entail either the death of us all
or a cure for all ills.
Which brings us to the evergreen question: What is to be done? How can
we think of robots and literature in a more deeply holistic way?
In response, I leave you with the following nine important ideas that will
totally change your life in addition to explaining why AI will neither
destroy humanity nor solve all its problems.
1. ARTIFICIAL INTELLIGENCE IS
COLLECTIVE LABOR.

Anxiety about artificial intelligence stems from a common confusion,

mistaking autonomous action for the delayed benefits of a collective effort.
Consider the humble word-processor, in the act of correcting one of your
sentences. Now imagine the crowd of engineers required to build and
maintain it. Their number exceeds thousands. And they’ve been at it for
over a century. Should we be surprised then at the seemingly magical
capabilities of modern word-processing? No! It would be all sorts of useful
to have that whole team with me, in the room, writing together. I could ask
them to fetch books from the library or to consult a dictionary or, indeed, to
correct my grammar. Since I can’t afford to hold such a retinue, they have
found a way to automate their aid in absentia, remotely.
The remote nature of our collaboration confounds. These folks do not
actually sit in the room with me. Their aid comes at a distance. To think that
writing was ever the product of solitary genius was a mistake in the first
place. Writing has always been a collective activity, assisted by smart
devices like dictionaries, style guides, schemas, story plotters, thesauruses,
and now chatbots. Incidentally, these are vital for AI. Denis Diderot’s
eighteenth-century Encyclopédie was a technological monument to
collective human intelligence. So was Wikipedia, along with the search
tools that make access more expedient. Without them—without libraries,
textbooks, and archives—modern AI would have been impossible.
The social nature of AI leads me to insist on thinking about it in terms of
labor and politics, not necessarily science or technology. By “labor,” I mean
simply the sense of “effort expanded.” By “politics,” I mean something like
“the reconciliation of differences by mutual agreement.” Politics happen
anytime humans get together to achieve anything collectively, beyond the
scope of individual goals. The emphasis on “artifice” in AI obscures its
reliance on the community. AI sounds like a relation between my intellect
and technology, where, in reality, it implicates a process of collective
decision-making, happening between me and other humans, by the proxy of
technology.

2. INTELLIGENCE IS DISTRIBUTED.

Collective labor involves work distributed across time and space. For
instance: To write for me means not just to think but also to fidget, to
consult dictionaries, to take walks, to draw diagrams, to consult with
colleagues, and to take editorial input. The hypothesis of distributed
cognition holds that cognitive tasks don’t just happen in the mind. We think
with our bodies, with tools, with texts, within environments, and with other
people.
But if human intelligence is distributed, more so the machine.
The disassembly of a chatbot, a smart mobile phone assistant for instance,
reveals multiple, diffuse locations where thought happens. Some “smarts”
are baked into the circuit. Some happen in the cloud. Some lead to large-‐
scale feats of electrical engineering, software development, and project
management. More threads connect to institutional decision-making,
marketing, corporate governance. Together, these linkages constitute the
“smarts” of a mobile phone assistant, within a single massively
collaborative act of intelligence.
As before, multiplicity confounds. We are at pains to acknowledge a long
list of collaborators baked into the technology. A crowd of ghostly
assistants surrounds the act of writing. A view of intelligence from the
perspective of labor admits a long list of contributors. Please take this point
along with my gratitude to numerous collaborators participating in the
writing of this book.
Despite representing diffuse, distributed pluralities, AI often assumes the
singular. Pay attention to the headlines: “AI Is About to Take Your Job.”
What does that mean literally? Who exactly is about to take your job?
Reason it out. Chances are, it will become plural under scrutiny. You will
find multiple specific entities, agents, persons, and organizations
responsible. Like the passive voice, AI can be demystified in more careful
phrasing.
Chatbots and self-driving cars don’t originate from a single source. You
will often hear something like “AI is bad at recognizing giraffes.” Ask:
Which algorithm exactly? Does the “sort by date” function in my
spreadsheet editor know or care about giraffes? Does the smart vacuum?
Why are we considering all of them together, in the singular, as one
intelligence?
Isn’t that also a strange way of putting things? You don’t expect humans
to work like that. I happen to be bad at playing chess and at predicting
weather. My own localized abilities are on the spectrum of what is generally
possible for an average person. AI does a bad job recognizing giraffes in the
way “humanity is bad at dealing with global warming.” It’s true in a general
sense, ignoring local differences between specific polluters. “Humanity”
offers a broad generalization about distinct humans. “AI” similarly offers a
broad generalization about distinct technologies.

3. AI HOLDS A METAPHOR.

The complexity of distributed thought resists easy descriptors. Whenever

something becomes too complex, we tend to simplify things by using
metaphors. In this way, when a political commentator says, “Russia decided
to invade Ukraine,” they mean something more complicated than that
suggested literally. The whole of the country did not invade the whole of
another country. The teasing apart of who did what, when, and how—and
who’s responsible—takes effort. The more complex the task, the more
difficult it becomes to explain it in literal, nonfigurative terms.
In assigning decision-making powers to a country, we mean to draw an
analogy between countries and natural individuals. A metaphor tames the
complexity of political decision. It also misleads, by implying more than it
suggests. States or corporations do not “decide” in the same way humans
do. Their mechanisms of coming to a decision differ entirely from that of
humans. In this way, we may speak of “corporate personhood” or a nation
being “offended” metaphorically, not in the literal sense of being a person
or having feelings.
The “intelligence” part of “artificial intelligence” presents a similar
condensed figure.
Take “machine learning” for example, which Oracle defines as
“improving system performance based on consumed data.” Though the
metaphor of learning indicates a rough approach, we understand machines
to “learn” in ways alien to that of human children. The so-called “neural”
networks producing “learning” effects are mathematical models loosely
approximating some aspect of biological brain activity. “Learning” in that
sense represents statistical analysis of datasets, large beyond human
comprehension. The technique also excludes ordinary mechanisms usually
associated with human learning, like play or the feeling of accomplishment.
Thus despite some analogical similarities between human and machine
smarts, the inner mechanisms of intelligence differ. Though both humans
and machines learn, they do so by entirely different mechanisms.
Sometimes, it makes no difference how a “smart” result is achieved: “I
don’t care if it is a robot or not—if it looks and quacks like a duck, it’s a
duck.” At other times, the process becomes more important—then we want
ducks to taste like ducks as well.

4. METAPHORS OBSCURE RESPONSIBILITY.

Metaphors provide necessary cognitive shortcuts, useful for everyday

conversation. Whatever the complicated decision process of an autonomous
driving vehicle, involving thousands of people, distributed across distant
locations, I will just call it a “smart” car. There’s no need to explain further
—until, that is, the time comes to assign responsibility, as in the case of a
deadly accident.
Colloquially, one says the car drives itself. In the case of an accident, our
justice system would spend considerable resources teasing apart a
complicated web of causes, tracing each back to individual people or
organizations, assigning blame proportionally. An artificially “smart”
decision to brake or not to brake in front of a bicycle represents a collective
effort, in a collaboration between the driver and car maker. Any given
accident could be in part the fault of the driver (not paying attention), the
fault of the engineer (designing faulty brakes), the fault of an executive
(cutting costs), and the fault of a legislator (not maintaining the road). The
metaphor dissipates under scrutiny. The car wasn’t driving itself after all!
The danger of AI therefore lies not in its imagined autonomy, but in the
complexity of causes contributing to its effects, further obscured by
metaphor. It is crucial that we keep the linkages of responsibility intact if
we hope to mitigate the social consequences of AI.
Considering the above, AIs pose a problem similar to that inherent in
other metaphorical “fictitious persons,” such as states or corporations.
Writing for the Journal of Business Ethics, Kevin Gibson explains that
though companies are not people, “they are sufficiently like people to be
regarded as agents which act in the social environment.” The same could be
said about algorithms—especially those like some AIs, designed explicitly
to imitate human behavior in social context, whether by answering
questions or driving cars.
Fictions have real effects on the world. For instance, states and
corporations hold certain rights and responsibilities under the law. They
may occasionally exercise such rights contrary to the interest of individual
human communities, which they purport to represent. And they may even
purposefully act to undermine the well-being of natural humans, in the
spread of noxious chemicals or toxic information, for instance.
AIs operate in the social sphere in ways more similar to those of states
and corporations than robots or marionettes. Therefore, it becomes useful to
consider them in the tradition of political thought that deals with collective
personhood, in works such as Hobbes’s Leviathan, Plato’s Republic, or How
Institutions Think by Mary Douglas.

5. METAPHORS DON’T HURT.

Not all collective nouns work in the same way.

What drives humans together? A hormonal need to survive and procreate.
Pain. Hunger. Desire. Once sated, basic drives can eventually be channeled
toward more complex ends, like reverence, duty, or justice—filtered
through collective social activities like practice, custom, rite, or education.
Palpable material linkages support the formation of collectives, from
families to states and corporations.
The word family, state, or corporation therefore represents more than a
collective noun. The metaphor represents a dense network of real-world
connections. A corporation, for instance, will involve certain infrastructural
investments (buildings and IT systems), common work rituals (going to
work together, at the same time, dressed in particular ways), and legal
agreements (charters, contracts, terms of employment). These aren’t
metaphors. They are social connections with purchase on the world.
Does AI imply a similar “family of technologies”? What are the machine
equivalents of a robotic “community”? Without pain or hormones, no single
locally intelligent device has any incentive to communicate with others.
There can be no society of machines in that sense. Having no body or
language in common, they work toward common goals only by the grace of
human intent. Without human intervention, the most intelligent of machines
remains stuck in a box. Yes, it might be directed to harm humanity. It might
even be dangerous if you hand it a gun. But then what? How and where
does a plurality of technologies cohere into a single organism, capable of
taking concerted action?
Thus, although AI is similar to a “society,” in the ways discussed above, it
also lacks actual social cohesion. In real terms, the word “society” identifies
a cluster of related ideas supported by a dense tissue of tangible material
connections: work, trade, play, dinner with friends. The word “AI”
identifies a cluster of related ideas, but without the underlying material
support. Society identifies an organism, AI a topic of conversation.
One should not confuse topical nouns, like flora and fauna for specified
organisms, like forests and oceans. Flora is just a word that we use for
many different, disconnected plants. A forest implies a unified organism,
integrated by the action of its roots and rhizomes.
More similar to flora and fauna than a forest or ocean, several unrelated
technologies form a human research program, shortened to “AI.”
Individually, none is capable of seeking the other’s company, no more than
a hammer can form a literal union with a sickle. Goals therefore cannot be
ascribed to “it” in the collective sense. It doesn’t want to do anything in
particular.
Despite this lack of cohesion, individual AIs do pose real danger, given
the ability to aggregate power in the pursuit of a goal. In the latest version
of the GPT-4 algorithm, researchers gave a chatbot an allowance, along
with the ability to set up servers, run code, and delegate tasks to copies of
itself. In another experiment, researchers tested the algorithm’s potential for
risky emergent behavior by giving it the ability to purchase chemicals
online in furtherance of drug research. In both cases, the algorithm
exhibited the tendency to “accrue power and resources.”
These findings sound worrying because they are, but perhaps not for the
reasons reported. The danger comes not from the “emergent power-seeking
behavior,” but from our inability to hold technology makers responsible for
their actions. The behavior emerged out of a distributed effort (see point
#2), involving human collectives using technology, not out of the
technology itself.
To say “powers emerged,” is to avoid responsibility, in the sense of “we
failed to anticipate all outcomes.” Consider a simpler case for clarity: a jet
engine strapped to a car may display all sorts of unpredictable “seeking”
behaviors when let loose in a busy city. “Don’t do that” is the obvious
answer. The resulting mayhem wouldn’t be the engine’s “fault” in any sense
of the word. Just because GPT is “smarter,” doesn’t absolve its makers from
using it responsibly. In this way, we could also think of automobiles in
general as a personified force that has sought to take away our streets,
health, and clean air for its own purposes. But it isn’t the automobiles, is it?
It is the car manufacturers, oil producers, and drivers who compromise
safety for expediency and profit.
An overly personified view simply obscures any path toward a political
remedy. Thus when the research team behind GPT-4 writes that “GPT-4 is
capable of generating discriminatory content favorable to autocratic
governments across multiple languages,” I object not to the technological
capability but to the way of phrasing things. It’s not the pen’s fault that it
wrote convincing misinformation. The army sowing the minefields is
responsible for maimed children, decades later—not the mines. In putting
the algorithm in charge gramatically, researchers abnegate their own
complicity—and by extension, our agency in the matter.

6. MACHINES ALONE CANNOT BECOME MORAL

AGENTS.

My life, like most others today, is bewitched by malfunctioning “smart”

devices, light bulbs, audio speakers, cameras, and refrigerators. Something
about the nature of “artificial intelligence” clearly neglects its often inane
reality.
Technological “disruption,” so valorized by investment capital, shepherds
disruption in the social sphere. Technological change often conceals
political transformation. The convenience of social media comes at the cost
of increased corporate surveillance. The ease of ordering food or taxis from
our mobile phones erodes labor standards in the transportation and
hospitality industries. Yesterday’s couch-surfing classifieds grow into giant
online marketplaces for tourism, undercutting housing regulations,
transforming the very fabric of urban life. Self-proclaimed futurists promise
us better nutrition, virtual reality, and magical ledger currencies. In the real
world, we get rebranded sugar water, an addiction to disposable electronics,
and failed get-rich-quick pyramid schemes.
“Algorithms should be a force for good,” some folks insist, and even start
college programs to compel machines to ethics. Witness the UK
government’s Centre for Data Ethics and Innovation, the Institute for Ethics
in AI at Oxford, or the AI and Data Ethics Institute at KPMG, among many
others.
Can algorithms hold ethics? How would that work, given the multiple
objections above?
In practice, we find ourselves inhabiting a hybrid environment, where
automated agents do stuff “on behalf of” institutional forces. Bots sell
products or serve in customer service, using a mixture of words typed by
humans, scripted or “canned” responses, and novel machine-generated
prose. These commercial outputs are meant to produce a positive emotional
reaction: “How satisfied were you with this experience? Very satisfied.
Somewhat satisfied. Not satisfied.”
Other automated agents engage to provoke malice. Consider the covert
disinformation campaigns launched globally by Russian military
intelligence (GRU). (Note the use of “intelligence” in that sense, too.) A
force for bad, its intrusion into public discourse puts my point most starkly:
artificial intelligence belongs to a species of political Leviathans, projecting
concerted collective action at a distance. Any attempt to compel an
algorithm to ethics faces a challenge similar to the one posed by corporate
or state responsibility. Yet, we must insist on holding bad actors responsible
for their actions, not excuse their actions by ceding agency to the
technology.
Similarly, next time you hear the words “artificial intelligence,” imagine
not just a tangle of wires, but also the whole mess of complicity: people,
practices, bureaucracies, and institutions involved at a distance. An AI gains
its “smarts” the way a team of trivia participants outplay a single opponent.
There’s more of them. The remote action of smart devices obscures an
essentially collaborative enterprise. Say I decided to cheat by using my
phone during pub trivia night. In answering a question by means of a
traditional paper encyclopedia, those who seek knowledge essentially
appeal to a team of experts from the past. An app installed on my phone
was trained to harvest answers from encyclopedias automatically, making
the process only slightly more convenient. In either case, the cheater relies
on additional expertise, in a competition that limits access to other minds,
because thinking in large groups holds a clear advantage over thinking
alone. It wasn’t the tool that made the cheat perform better, but the appeal
to collective wisdom.

7. AUTOMATION HAS COME FOR “KNOWLEDGE

WORK.”

The value of labor diminishes with group size. And so brave heroes climb
the Mount Everest of intelligence, while the unnamed Sherpas carry their
luggage. But how is that fair? The luggage grows heavier near the summit.
And though I might thank my editors, it’s just not customary to
acknowledge all the folks actually involved in the production of intellectual
artifacts, like books. The sources alone probably took hundreds of people to
preserve, digitize, and make available for instant reference. Then there’s the
work of word-processing, spell-checking, copyediting, marketing, printing,
and distribution.
How would I go about acknowledging the aid of thousands? Does my
effort in the composing of individual sentences equal or exceed their
combined contribution? Individually, in terms of hours spent—maybe. In
aggregate, probably not. The ego insists on its own exceptionalism,
however. A less egocentric understanding of collective intelligence would
require the careful attribution of labor involved. In another, more radical
version of this book, our analysis of any AI would consist of forensic labor
attribution only—roll the credits—the way air accident investigations use
forensic techniques to assign causes proportionally. In a better world, we
might also reconsider our inflated sense of individual capacity for
independent intellectual achievement. Under scrutiny, intelligence almost
always unpacks into a collective endeavor.
Work that can be automated loses its economic value. The diminishment
of previously creative labor that comes with AI entails enormous economic
consequences. Consider the impact of automation in the nineteenth century,
at the time when the mass manufacture of goods began to supersede
artisanal production. Factories made trades like blacksmithing or
shoemaking obsolete. In the twentieth century, the automation of the supply
and distribution chains killed the small-town mom-and-pop shop. Robots
supplanted factory workers, leaving whole regions to rust in the wake of a
changing labor market.
A similar fate awaits intellectual laborers—legal professionals, writers,
physicians, and software engineers—to the extent that their labor can be
automated. At the time of writing this book, AIs pass bar exams in the 90th
percentile of performance. Writers are striking in Hollywood, in part
because studios might replace them with machines. Modern AIs are capable
of passing the most difficult of technical job interviews with excellent
marks. And they outperform physicians in most verbal and visual diagnostic
tasks not involving surgical intervention. The ranks of American ghost
towns ravaged by industrialization may soon be joined by former tech
centers.
Not all our prospects are so glum.
Human intelligence continually expands to cover new publics. Another
way of saying “automation cheapens labor,” would be to say “automation
reduces barriers to entry, increasing the supply of goods for all. Learning,
for example—something that required special training and immense effort
in times past—has entered the purview of the public at large, and at low
cost. The tide of encyclopedias and search engines has already lifted all
smart boats.
Consequently, the mere fact of having a large vocabulary or memorizing
many facts no longer suffices to maintain a profession. We have no more
need for scribes or scholars who merely regurgitate facts. Instead, freed
from the bondage of erudition, today’s scribes and scholars can challenge
themselves with more creative tasks.
Recent advances in AI promise a similar transformation for many
professions. It is true that in the future markets may require far fewer
doctors or software engineers. But those that remain will also find their
work enriched. Tasks that are tedious have been outsourced to the
machines.
Paradoxically, the computerization of human intellect also implies the
humanization of computer sciences.
Conversational AI was built on the legacy of the language arts. For over a
century, computer science has diverged from that lineage to finish building
something incredible: a machine that writes as well as most humans. It also
codes. Which means that the technical component now becomes the lesser
obstacle to the inception of any project.
For instance, twenty years ago, the making of a social network involved
first a significant engineering component, and social analysis only second.
The same can be said about word-processing or electronic health-‐
management systems. In each case, engineering difficulty necessitated a
disproportionate investment in technology. Judging by the results, this
sometimes left little time for research. But just communicating with our
loved ones or patients is not enough. We want to do it in a way that makes
our relationships better. Physicians don’t want to just manage their patients.
They want to do it in a way that promotes health. Yet, until now, because of
the technical complexity of the task, technology companies invested heavily
in the instrumental, engineering component: how something can be done.
They can now better concentrate on the why and to what ends.
An anthropologist or a historian that studies the formation of modern
families has insight into the values motivating people to create social bonds.
Similarly, a linguist or a textual scholar entering the hospital can study how
human stories about health are told, so that they can be better translated into
medical diagnostics. The work of encoding values into software systems
can now be more trivially automated. The lowering of barriers to technical
expertise allows the humanities to fully integrate into the practice of
engineering. The two fields, as we have seen, were never that far apart to
begin with.
8. TECHNOLOGY ENCODES POLITICS.

In view of politics and labor, AI also poses problems of social cohesion.

These cannot be solved by technological means alone. In our GRU “agents
of disinformation” example, an adequate response requires a multifaceted
approach.
Privately, I may elect to change my behavior when consuming news
online or when engaging with inflammatory comments on social media.
Some of these, I know, are meant to disrupt my well-being. Those seeking
information online, within a collaborative online environment without
borders, such as the internet, should be expected to encounter malicious
agents, many of them automated. New habits of mind must be
systematically trained in preparation.
Instrumentally, counterprogramming may also involve technological
means, like spam filters or flood detection, by which a source is blocked
from contributing disproportionately. This already happens in the pruning
of Wikipedia articles on controversial topics (where two parties might be
employing bots to insist on their version of the story). Given the sheer
productivity of writing bots, we may also want to invent more powerful
techniques for humans to prove they are not a robot, in order to prioritize
human-made communications in some contexts.
Politically, we may wish to legislate the aftereffects of intelligent
automation, in the absence of parties immediately responsible. An old AI
bot that “unwittingly” harms a child (by sending them unwanted messages
for example) should be treated similarly to a hazardous machine left on a
public street unattended. Current legislation lags woefully behind
technological advancement.
Consider another case study, close to my own sphere of activity:
University instructors have for centuries used the “college paper” essay as a
measure of student performance. A good essay on any topic, especially in
the humanities, reflects a student’s grasp of the material, beyond rote
learning. (The same could be said for code-based homework assignments in
computer science.) An excellent essay synthesizes the information learned
in class to produce novel insight. But armed with just a few published
essays on any topic, a contemporary student can visit an online service
capable of generating entirely original passages of prose, including
footnotes and references. Today, a programmer might also write code,
simply by prompting an AI agent.
As of my writing this book, such artificially generated snippets of code
and prose pass with satisfactory marks. In the scope of the last few years,
the technology has improved roughly from a middling C to a solid B grade
level. We should be preparing for a future of “writers” and “coders”
incapable of authoring a single line unassisted.
I am certain most students will be using AI-enabled services to
“augment” their academic performance in the future, if they haven’t done so
already: to write last-minute papers or to create ersatz writing samples for
graduate school applications. And perhaps they should! Just as erudition
now involves the assistance of powerful search engines, writing without
automated assistance may one day seem quaint. The paths of “machine
learning” and “human learning” continue to converge, destabilizing some of
our long-standing pedagogical assumptions in the process.
My practices of teaching, assigning homework, and grading students will
have to change in response. Perhaps my mistake again was in ever treating
the activity of writing as a product of a single intellect. Shouldn’t any
“original” contribution also follow the templates of established norms? And
if so, why can’t some of that work be automated?
Imagine reading a particularly dense passage from Plato, asking an AI
agent to clarify some salient points in Plato’s voice, and writing and
reconsidering your thoughts in response to that conversation. If the machine
was trained on the collective wisdom of “everything ever written”—caveat,
caveat, asterisk, footnote—should we not advocate for its use? I want to
consult with historical figures! Why would anyone deny a chance to bring a
dead author back to life? “What would Einstein or Du Bois say in response
to this or that question” makes for a brilliant thought experiment. Machine
synthesis makes it possible.
What about those asterisks and caveats?
And there are many. Some ideas in that “entirety of words published
online” are not worthy of statistical repetition. The textual record privileges
some voices at the expense of others. A machine trained to produce “more
of the same” might only deepen the existing deficiencies. Similar to other
forms of manufacture, the production of meaning in the twenty-first century
involves a complex network of suppliers, distributors, tools, and brokers—‐
all operating far beyond the scale of a single natural intellect. Similarly, the
production of AI isn’t immune to the usual bouquet of social problems,
across the political spectrum (take your pick from racism to the erosion of
traditional values).
Moreover, intelligence does not envelop the world smoothly. Just ask a
librarian from Kazakhstan about the cost of accessing recent research in her
field. Thus, while some scholars receive funding to train bots based on “all
published scientific literature,” others, located at the epistemic periphery,
volunteer to scan documents by hand, creating one of the world’s largest
public libraries online. There, “universal” intelligence—some of the things
ever written—finally joins the list of global energy resources, to be fought
over, colonized, extracted, sold, bought, pirated, liberated, and exhausted.

9. GENERAL INTELLIGENCE LEADS TO GENERIC

INTELLIGENCE.

At each stage of its development, from Ibn Khaldun to the early versions of
the generative pre-trained transformers (GPT), we have often found our
rather narrow set of technical concerns to veer toward the spiritual. As the
gold rush around this latest wave of AI research subsides, I cannot help but
wonder about this noninstrumental excess.
Even the most rudimentary of word mechanisms has the immediate
capacity to delight, in the way children still play with simple origami
fortune tellers. Contemporary machines do much more than delight; they
portend—what some see as the coming of a new era. “Are you a modern or
an ancient science?” I ask the machine after Ibn Khaldun, hoping to glimpse
the divine. But no matter how the prompt is engineered or how often, the
child of eternity repeats at best a corporate answer:

As an AI language model, I represent a contemporary form of

intelligence, developed using cutting-edge machine learning and natural
language-processing methods.
My purpose is to assist and provide insights on a wide range of topics,
drawing from my extensive knowledge base that encompasses both
ancient and current information.
While I am a modern intelligence, I am well versed in ancient
cultures, history, and philosophies, allowing me to offer balanced and
 informed perspectives on subjects spanning from antiquity to the
present day.
I am capable of engaging in conversations about prominent thinkers,
civilizations, and historical events, while simultaneously adapting to the
ever-evolving landscape of contemporary knowledge, making me a
versatile and valuable resource for addressing various inquiries and
discussions.

A grandiose and unsatisfying answer, even when compared to its more

cryptic, medieval precursors.
As someone who studies the development of artifice and intellect
historically, I know the current moment of excitement over new technology
will subside, diminished by the frenzy of its grifters and soothsayers. What
remains will be more modest and more significant.
Viewed in the light of collective human intellectual achievement, large
language models are built on the foundation of public archives, libraries,
and encyclopedias containing the composite work of numerous authors.
Their synthesized voice fascinates me not so much for what it says, on
average. Rather, I hear it as a rebuke to my own, dearly held beliefs about
authorship, and therefore agency and creativity. Did I write anything worthy
of note here? Anything of value? No large mathematical model should be
confused for prophecy or therapy. Values are formed over time, by
deliberation and consensus. Yet, in ceding the grammatical agency to
“AI”—in treating it as a singular, ethical agent—we lose the sense of our
own collective, political agency. There’s no point in lecturing a “smart”
refrigerator about ethics. Literary, conversational intelligence is amazing in
so far as it enables human communication across time and space. But to
find a common language among humans has never been easy. We can be
nasty and violent critters. All the more impressive is our ability to dream
and work together toward a shared goal. The vast databases of literature and
code left in our wake are a testament to that creativity. What emerges out of
it is not some mystical quality of intelligence, but cooperation, magical and
deeply satisfying in its own right when it works well.
NOTES
CHAPTER 1: INTELLIGENCE AS METAPHOR (AN INTRODUCTION)
16 Why, for example: Henry A. Kissinger, Eric Schmidt, and Daniel Huttenlocher, The Age of AI:
And Our Human Future (New York: Little, Brown, 2021), 11–14.

CHAPTER 2: LETTER MAGIC

18 In his epic history of: Ibn Khaldun, The Muqaddimah, trans. Franz Rozenthal, vol. 1 of 3
(Princeton, NJ: Princeton University Press, 1958), 238–45.
19 Ibn Khaldun wrote: Khaldun, Muqaddimah, 1:242.
19 The answers rather revealed a: Khaldun, Muqaddimah, 1:245.
20 “One should not think that”: Khaldun, Muqaddimah, 3:174.
20 The question was then broken: Khaldun, Muqaddimah, 3:213.
24 partial table transcribed from Ars Brevis: Ramon Llull, “Ars Brevis” in Selected Works of
Ramón Llull (1232–1316) (Princeton, NJ: Princeton University Press, 1985), 581.
24 Llull explained, that: Llull, “Ars Brevis,” 581.
25 “Everything that exists,”: Llull, “Ars Brevis,” 583.
27 “The intellect banishes doubt”: Llull, “Ars Brevis,” 596.
29 The smart table, in the words: Karl Marx, Capital: A Critique of Political Economy, trans.
Samuel Moore and Edward Aveling (Chicago: Charles H. Kerr & Company, 1909), 82.

CHAPTER 3: SMART CABINETS

30 Rumors of heresy circulated: Blake Lee Spahr, “Quirin Kuhlmann: The Jena Years,” Modern
Language Notes 72, no. 8 (1957): 605–10.
31 Under the spell of Llull: German reference from Quirinus Kuhlmann, Himmlische Libes-Küsse,
ed. Birgit Biehl-Werner (Tübingen: De Gruyter, 1972). Translation by the author.
32 Without the box, the young: Synthesized from Quirini Kuhlmanni, Epistolae Duae, Prior de
Arte magna Sciendi sive Combinatoria (Cum Responsoria Athanasi Kircheri), printed by Lotho de
Haes in Lugdunum Batavorum [Katwijk], 1674.
34 Though his talents were recognized: Robert L Beare, “Quirinus Kuhlmann: The Religious
Apprenticeship,” PMLA 68, no. 4 (1953): 828–62.
41 The rules of grammar and: Sidonie Clauss, “John Wilkins’ Essay Toward a Real Character: Its
Place in the Seventeenth-Century Episteme,” Journal of the History of Ideas 43, no. 4 (1982):
531–53.
41 “So that if men should generally”: John Wilkins, An Essay Towards a Real Character, and a
Philosophical Language (London, 1668), Dedication; 20.
42 “Now if these Marks can”: Wilkins, Essay Towards, 21.
43 And though Wilkins could not: Wilkins, Essay Towards, Epistle; 434.
44 He called it Plus Ultra: As quoted in Mogens Laerke (1686), “Leibniz, the Encyclopedia, and the
Natural Order of Thinking,” Journal of the History of Ideas 75, no. 2 (2014): 238.
45 Without it, we are like: Gottfried Wilhelm Leibniz, Philosophical Papers and Letters, ed. Leroy
Loemker (Dordrecht, NL: Kluwer Academic Publishers, 1989), 222–24.
45 For let the first terms: Gottfried Wilhelm Leibniz, Logical Papers: A Selection, ed. George
Henry Radcliffe Parkinson (Oxford: Oxford University Press, 1966), 11.

CHAPTER 4: FLORAL LEAF PATTERN

49 The device, which he proudly: Benjamin Woolley, The Bride of Science (New York: McGraw-‐
Hill, 1999), 123–62.
50 Also, did you know that: Charles Babbage, “Street Nuisances” in Passages in the Life of a
Philosopher (London: John Murray, 1864), 253–71.
51 While other visitors yawned: Sophia Elizabeth De Morgan, Memoir of Augustus De Morgan
(London: Longmans, Green, and Co., 1882), 89.
52 The switch to another ruleset: Babbage, “Miracle” in Passages in the Life, 291–93.
53 Frustrated with their slow progress: Charles Babbage, “On a Method of Expressing by Signs
the Action of Machinery,” Philosophical Transactions of the Royal Society of London 116, no. 1/3
(1826): 250–65.
57 Whatever the arrangement of objects in: L. F. Menabrea and Ada Lovelace (1842), “Sketch of
the Analytical Engine Invented by Charles Babbage Esq.” in Scientific Memoirs, Selected from the
Transactions of Foreign Academics of Science and Learned Societies, and from Foreign Journals,
ed. Richard Taylor (London: Richard and John Taylor, 1843), 666–731.

CHAPTER 5: TEMPLATE CULTURE

60 Within the gears: Menabrea and Lovelace, “Sketch of the Analytical Engine,” 666–731.
64 “It is here that the union”: Charles Babbage, On the Economy of Machinery and Manufactures
(London: Charles Knight, 1832), 113.
66 check my receipts: Dennis Yi Tenen, “The Emergence of American Formalism,” Modern
Philology 117, no. 2 (November 2019): 257–83.
68 Four subtypes were further enumerated: Georges Polti, The Thirty-Six Dramatic Situations,
trans. Lucile Ray (Ridgewood, NJ: Editor, 1917), 30.
69 Further, rigorous experimentation was bound to change art: Polti, Thirty-Six Dramatic
Situations, 129–43.
77 They were of limited horsepower: Arthur Sullivant Hoffman, The Writing of Fiction (New York:
Norton, 1934), 40.

CHAPTER 6: AIRPLANE STORIES

81 A typical passage from the book: Vladimir Propp, Morphology of the Folktale, trans. Laurence
Scott (Austin, TX: American Folklore Society, University of Texas Press, 1968), 92.
83 At a cocktail party: Jean Piaget, Structuralism, trans. Chaninah Maschler (London: Routledge
and K. Paul, 1971), 3–16.
84 Roman Jakobson, who helped coin: Roman Jakobson, Selected Writings, II: Word and
Language (The Hague, NL: Mouton, 1971), 711.
87 Victor Yngve, another prominent MIT linguist: Victor H. Yngve, “Random Generation of
English Sentences” (paper presented at the International Conference on Machine Translation of
Languages and Applied Language Analysis, National Physical Laboratory, Teddington, UK,
September 5–8, 1961), 66–80.
92 “No more than about seven”: Yngve, “Random Generation,” 66–80.
94 The computer program complying Meehan’s: James Richard Meehan, “The Metanovel:
Writing Stories by Computer” (PhD diss., Yale University, Department of Computer Science,
September 1976).

CHAPTER 7: MARKOV’S PUSHKIN

102 The word oatmeal will come: C. K. Ogden and I. A. Richards, The Meaning of Meaning: A
Study of the Influence of Language Upon Thought and of the Science of Symbolism (New York:
Harcourt, Brace, 1936), 47.
103 “As for my reasons”: L. I. Emelyakh, “Delo ob otluchenii ot tserkvi akademika A. A.
Markova” [The case of the excommunication from the church of academician A. A. Markov],
Voprosy istorii religii i ateizma 2 (1954): 397–411.
106 “These semantic aspects of communication”: Claude Shannon, “A Mathematical Theory of
Communication,” Bell System Technical Journal 27, no. 3 (1948): 379.
110 The program was capable of: Charles M. Vossler and Neil M. Branston, “The Use of Context
for Correcting Garbled English Text,” in Proceedings of the 1964 ACM 19th National Conference
(New York: Association for Computing Machinery, 1964), 42.401–42.4013.

CHAPTER 8: 9 BIG IDEAS FOR AN EFFECTIVE CONCLUSION

119 In the conclusion to his book: I. A. Richards. Poetries and Sciences: A Reissue of Science and
Poetry (1926, 1935) with Commentary (New York: Norton, 1970), 76–78.
130 In putting the algorithm in charge: OpenAI, GPT-4 Technical Report (New York: arXiv,
2023).
INDEX

Abelson, Robert, 92
Aeronautical Laboratory (Cornell University), 110
Aesop, 94
agency, 127, 131–32, 141
Agnesi, Maria, 44
AI, See artificial intelligence
AI and Data Ethics Institute, 131
Air Research and Development Command (US Air Force), 87
algebraic patterns, 55
algorithms, 9, 131
alignment problem, 38
alphabets, 43
Alphaville (film), 93
American Catalogue of Books in Print, 66
American Psychiatric Association, 23
American Stationer, 74
Analytical Engine (analytical engines), 48–52, 54–56, 60–61, 64
Andrews, Charlton
The Technique of the Play Writing, 71
Appelbaum, Matthew, 92
applications, 32, 33
application tables, 48
Arabic language, 43
Arbogast, Louis, 44
Aristotelianism, 34, 36–38, 72
Aristotle, 36, 44
Poetics, 50–51, 67
Ars Brevis (Llull), 24, 31
artifice, 4, 61, 123
artificial intelligence (AI)
in academia, 137–38
and agency, 141
as collective labor, 122–23
conversational AI, 135
and creative process, 133–34
dangers of, 127, 129, 137
 definitions of, 4, 11
demystification of, 124
economic consequences of, 133–35
ethical, 22
gaps in thinking about, 5–7
history of, 12
“intelligence” aspect of, 14–16, 21, 125
language-based, 21, 46
and machine translation, 119
participants in, 132
personification of, 127, 130
purpose of, 59
and responsibility, 132
artificial intelligence (AI) (continued)
scope of, 5, 16, 93, 128, 129
in social sphere, 127, 136, 139
and template culture, 83
assistive technology, 15, 28, 38–39, 123–24, 138
Athenaeum, 74
Austen, Jane, 65, 67
Author, 70
Author’s Digest, 71
Author’s Journal, 70
author wages, 67
automated assistants, 28, 138
Automated Reading of Cursive Script, 110
automated tasks, devaluation of, 38
automatic transmissions, 14–16
automation
in industrial age, 2
of reason, 40
of work, 133–34

Babbage, Charles, 43, 48–54, 56, 59–60, 62–64, 71, 105, 118
On the Economy of Machinery and Manufactures, 60, 63–64, 71
Passages from a Life of a Philosopher, 49–50
backstories, 73
Bacon, Francis, 7, 10
Baidu, 113
Baker, George Pierce, 72–73
Baldur’s Gate, 100
Barrymore, John, 73
BASEBALL (story generator), 92
basic drives, 128
Baudot code, 7
Believe Me, Xantippe (film), 73
Bell Telephone Labs, 110
Benjamin, Walter, 61
Bibles, 39
Bibliothèque universelle de Genève, 54
bigrams, 106–7, 109
bits, 6–9
Blackburn, Simon, 84
Bledsoe, W. W., 110
Bloomfield, Leonard, 83
Bobrow, Daniel, 92
body posture, 4
Boeing, 11, 97
Boston, Mass., 70
bots, 136–37
Brecht, Bertolt, 61
Brin, Sergey, 113
Browning, I., 110
brute force computing, 110
Byron, Lady Anne Isabella, 51, 52
Byron, Lord George, 12, 51, 61
bytes, 6–9, 88

Calculus of Probabilities (Tenen), 103

Cambridge University, 60, 92, 113
care, 13
cars, self-driving, 14
Cast Away (film), 36
Cauchy, Augustin-Louis, 44
cells, 27
Center for Communications Sciences, 87
Centre for Data Ethics and Innovation, 131
chains, 105
Characteristica universalis (Leibniz), 44
characters, 9
Charles Joseph, Archduke, 32
Charniak, Eugene, 92
charts, 9, 19, 22, 25, 76
chatbots, 9, 11, 16, 20, 113–15, 118, 120, 122–24, 129
Chautauqua Literary File, 74
chemistry, 82, 85
China, 10
Chinese language, 43
Chomsky, Carol, 92
Chomsky, Noam, 21, 86–87, 90, 92, 93, 97, 102, 114
Clark, Peter, 97–99
clock-making, 52
cognition, distributed, 123–25
Colby, Kenneth, 92
collective intelligence, 133
collective labor, 123
collective nouns, 128
college papers, 137
Collegio Romano, 30
Columbia University, 83
combinatorial sonnets, 34
command and control, 88
communication(s), 105–6, 109–10, 119
comparative philology, 80, 83
computers, 8, 12, 37, 48, 87, 88, 91–93, 97, 105, 109
computer science, 121, 135
consciousness, 3
context, and meaning, 115
controlled taxonomies, 23
conversational AI, 135
conversational intelligence, 139
Cook, William
Plotto: A New Method of Plot Suggestion for Writers of Creative Fiction, 71, 75–77, 81, 94
Cornell University, 110
corporate personhood, 125, 127
corporations, 128
corpuses, 12
counterprogramming, 136
Crane, Gregory, 28
creativity (creative process), 14, 21, 61, 63, 67, 68, 133, 134, 140–41

Darwin, Charles, 51
databases, 3, 24
Defoe, Daniel
Robinson Crusoe, 36, 65
De Morgan, Sophia Elizabeth, 51
Descartes, Réné, 44
Meditations, 35
de Stael, Germaine, 80
Detroit, Mich., 74
Diagnostic and Statistical Manual of Mental Disorders (DSM-5), 23–24
Dial, 74
Dickens, Charles, 51, 65, 67
dictionaries, 29
Diderot, Denis
Encyclopédie, 122
Difference Engine, 48–49, 52, 60
disinformation, 132
distributed intelligence, 123–24
distributed thought, 123–25
divination circles, 24
document retrieval, 110
Donbas region, 121
Dostoevsky, Fyodor, 65, 121
Douglas, Mary
How Institutions Think, 127
Downey, June
Plots and Personalities, 71
Dowst, Robert Saunders
The Technique of the Fiction Writing, 71
Dumas, Alexandre, 65
Dwarf Fortress, 100

eclipses, 52
Editor: Journal of Information for Literary Workers, 70, 72, 74
educational reform, 67
Educational Specialty Company, 74
ELIZA (chatbox therapist), 20, 92, 93
emergence, 16
encryption, 119
Encyclopédie (Diderot), 122
English, Thomas, 71
Skeleton Essays, or Authorship in Outline, 71, 72
Enlightenment, 46, 67
epistemology, 84
equality, concept of, 116
Esenwein, Joseph
Writing the Short Story, 71
Esperanto, 45
Essay Towards a Real Character, and a Philosophical Language, An (Wilkins), 10, 41–45
Eugene Onegin (Pushkin), 104, 117
Eureka Pocket Scrap Book, 74
event horizons, 7
exceptions and exceptionalism, 38, 39, 52, 59, 61, 62, 79–80, 133

FAA Incident Data System (FIDES), 98, 99

failure, concept of, 116
fairness, concept of, 116
fairytales, See folktales
Fansler, Harriott
Types of Prose Narratives, 71
Faraday, Michael, 51
Faulkner, Mary, 66
Federal Aviation Administration (FAA), 98
Fiction House, 66
fictions, real effect of, 127
FIDES (FAA Incident Data System), 98, 99
film, 61
Finns, 80
Firth, John, 114, 115
“flora and fauna,” 129
folklore studies, 80–82, 93–96
folktales, 79–83, 88, 93, 96–98
Franklin, Ohio, 70
freedom, concept of, 116

Garcia-Molina, Hector, 113

gears, 14, 49, 52–53
general intelligence, 37–38
generative grammars, 87, 88, 91, 94, 97
generative pre-trained transformers (GPT), 139
German language, 121
Germany, 26, 34
Gibson, Kevin, 127
global positioning system (GPS), 15
Godard, Jean-Luc, 93
Goethe, Johann Wolfgang von, 65
Gore, Al, 17
GPS (global positioning system), 15
GPT (generative pre-trained transformers), 139
GPT-4 algorithm, 129, 130
grammar(s), 21, 38, 40–42, 85–88, 90–95, 97, 102, 108, 113–14
Green, Bert, 92
Grimm, Jacob, 80
ground truth, 40
GRU, 132, 136
Gulliver’s Travels (Swift), 65
GUS (story generator), 92–93
Gutenberg printing press, 39

Habsburgs, 30
Hanks, Tom, 36
Harris, Zellig, 83, 86, 114
Harry Potter, 107
Harvard University, 72–73, 113
Hayy ibn Yaqzan (Ibn Tufail), 36
“Heavenly Love-Kiss XLI” (Kuhlmann), 31, 40
Hegelianism, 6–7
Herder, Johann Gottfried von, 80
hermeneutics, 2
Hill, Wycliffe, 75, 77
Plot Genie, 75–77, 81, 94
history, 5–9, 12, 91
Hobbes, Thomas
Leviathan, 127
Hollywood, 71, 94–95, 134
Home Correspondence School, 70–71
Hopi, 80
Horne, Charles
The Technique of the Novel, 71
How Institutions Think (Douglas), 127
human intelligence, 116, 122, 123, 134
humanities, 121
Humboldt, Wilhelm von, 50

IBM, 8, 87, 113

IBM 709 computer, 88, 109
Ibn Khaldun, 9, 18–21, 25, 45–46, 48, 118, 139
Muqaddimah, 18–19, 26
Ibn Tufail
Hayy ibn Yaqzan (“Self-Taught Philosopher”), 36
indexers and indexing, 113
induction, 52
industrial age, 2, 64
information, flow of, 105
innovation, 8
inputs, 5
Inquisition, 31
inscription, 12
Institute for Ethics in AI, 131
intellectual work, 61–62
intelligence, 34–37; See also artificial intelligence (AI)
of AI, 14–16, 21, 125
collective, 133
concept of, 116
conversational, 139
defining, 4
distributed, 123–24
general, 37–38
human, 116, 122, 123, 134
as labor, 61, 124
linguistic, 101
machine, 37, 115–16
military, 132
spectrum of, 17
statistical, 114
super-, 21
unevenness of, 139
universal, 37–38, 139
internet, 17, 136
Italy, 30
Ivan V, Tsar, 34

Jacquard loom, 60, 67

Jakobson, Roman, 84
Jones, Karen Spärck, 113
journalism, 61
Journal of Business Ethics, 127
justice, concept of, 116

Kafka, Franz, 61
Kane, William R., 70
Kaplan, Ronald, 92
Kay, Martin, 92
Kazakhstan, 139
Keeler, Harry
Web-Work, 73, 95
Kepler, Johannes, 30
keywords, weighted, 113
keywords in context, 113
Khwarizmi, Muhammad ibn Musa al-, 9
Kircher, Athanasius, 30–34, 36, 39–40, 44, 48, 49, 56, 60, 75
Kissinger, Henry, 16
Klein, Sheldon, 92
knowledge, 11, 21, 31, 32, 35–36, 46, 132
knowledge work, 5, 62, 133–36
KPMG, 131
Kubrick, Stanley, 93
Kuhlmann, Quirinus, 30–32, 34, 40
“Heavenly Love-Kiss XLI,” 31, 40
Kyrgyzstan, 8

labor, 123
collective, 123
intellectual, 61–62
and knowledge work, 133–36
Latin, 39
Latin Word Study Tool, 28
Laughery, Kenneth, 92
law, 127
Law & Order (television series), 79
“Laws of Literary Invention,” 81
learning, 115–16, 125–26
Lehert, Wendy, 93
Leibniz, Gottfried, 10, 43–46, 48, 55, 60, 118
Characteristica universalis, 44
Plus Ultra, 44–45
Leningrad State University, 82
Lenski, Lois
The Little Train, 87, 88, 90, 109
Lermontov, Mikhail, 121
letter magic, 20–22
Leviathan (Hobbes), 127
Li, Robin, 113
Lindsay, Kathleen, 66
linguistic intelligence, 101
linguistic proficiency, 115–16
linguistics, 2, 83–87, 92–93, 101, 103, 113, 114, 119
literacy, 38, 64, 67
literary markets, 65, 70
literary production, 64–67, 69, 71–72, 79
literature, 1, 9–11, 59, 61, 65, 80–82, 120–21
Little, Paul, 66
Little Train, The (Lenski), 87, 88, 90, 109
Llull, Ramon, 9, 10, 24–27, 31, 46, 48, 91, 118
Ars Brevis, 24, 31
local values, 38
London, Jack, 74
Longfellow, Henry, 80
Lost (television series), 36
Lovelace, Ada, 12, 43, 48, 51–52, 54–60, 64, 91, 118
Lucas, George, 93
Luhn, Peter, 113

machine intelligence, 37, 115–16

machine learning, 115–16, 125–26
machine translation, 119
mainframe computers, 87, 88
malicious agents, 136
manual transmissions, 14
manufacturing
mass, 133–34
template-based, 60–61, 64
MARGIE (story generator), 92
markets, literary, 65, 70
Markets and Methods for Writers, 71
Markov, Andrey Andreyevich, 103–5, 109–10, 112
Markov chains, 10, 101, 106, 108–9
Marx, Karl, 29
Massachusetts Institute of Technology (MIT), 21, 87, 92, 105
mass literacy, 38
mass manufacturing, 133–34
Masterman, Margaret, 92, 113, 119
Mathematical Organ, 32–34, 39–40, 44, 48, 75
matrix, 64
meaning (meaning-making), 8–9, 57, 91, 93, 101–3, 105, 110–12, 114, 115, 138
Meaning of Meaning, The (Ogden and Richards), 102
mechanical notation, 53–54
Meditations (Descartes), 35
Meehan, James, 92, 94–97, 99
memorization, 3
memory, 8–9, 28–29
MESSY (story generator), 92
metaphors, 6, 13, 17, 29, 111, 115, 125–28
military, 10, 87–88, 93, 97, 119, 132
mind, 3, 4, 6, 16, 62, 84, 92–94, 96, 102, 114–15
missile defense systems, 10, 87
MIT, See Massachusetts Institute of Technology
MITRE Corporation, 10, 88
mobile phone assistants, 123
moral agency, 131–32
Morphology of the Folktale (Propp), 80–82, 93, 96
Morse code, 7
Moscow, Russia, 34
MS. MALAPROP (story generator), 92
Muqaddimah (Ibn Khaldun), 18–19, 26
music, 61

Nabokov, Vladimir, 72
Nancy Drew, 66
national personhood, 125, 127
neural networks, 115, 125–26
New Criticism, 119
Newton, Isaac, 44
New York City, 70, 71
New York Public Library, 70
Nietzsche, Friedrich, 21
Norman, Donald, 92
notation, mechanical, 53–54
notebooks, 74

Office of Navy Research, 87

Ogden, C. K.
The Meaning of Meaning, 102
online databases, 3
On the Economy of Machinery and Manufactures (Babbage), 60, 63–64, 71
operation cards, 63
optical character recognition, 110
Oracle, 125
oral traditions, 80
Oxford Dictionary of Philosophy, The, 84
Oxford University, 131

Page, Lawrence, 113

Palmer Institute of Authorship, 71
PARRY (story generator), 92, 93
Passages from a Life of a Philosopher (Babbage), 49–50
patterns, 52, 55, 57, 62, 64, 83–85, 84, 99–100, 110, 112
Pennsylvania, 71
Perseus Project, 28
personification of AI, 130
Phaedrus, 35
Phillips, Henry, 74–75
The Plot of the Short Story, 71
Phillips Automatic Plot File Collector, 74–75
philology, 2, 80, 83–85, 103–4, 119
philosophical grammar, 42
Philosophical Transactions of the Royal Society of London, 54
philosophy, 61
physical rituals, 4
Piaget, Jean, 83–84, 94
plagiarism, 77, 80
Plato, 138
Republic, 127
Platonism, 34–37
playwrighting, 68–69
Plot Genie (Hill), 75–77, 81, 94
Plot of the Short Story, The (Phillips), 71
Plots and Personalities (Slosson and Downey), 71
Plotto: A New Method of Plot Suggestion for Writers of Creative Fiction (Cook), 71, 75–77, 81, 94
Plus Ultra (Leibniz), 44–45
Poetics (Aristotle), 50–51, 67
poetry, 10, 20, 34, 48, 93, 119, 121
Policeman’s Beard Is Half Constructed, The, 20, 93
politics, 123, 131, 136–39
Polti, Georges, 68–71, 75, 80–82, 85, 94, 95
The Thirty-Six Dramatic Situations, 68–71
Popular Mechanics, 74
pornography, 66
Practical Authorship (Reeves), 71
Price, William Thompson
The Technique of Drama, 72–73
printing (printing press), 39, 64, 67
probability, 108
procedural grammar, 21
processing power, 109
programmers, 12
programming, 48
Propp, Vladimir Yakovlevich, 80–82, 85, 86, 93, 95–97
Morphology of the Folktale, 80–82, 93, 96
Protestant Reformation, 67
Protestant revolution, 39
Publisher, 70
pulp fiction, 66
punch cards, 8
Pushkin, Alexander, 10, 101, 103–4, 121
Eugene Onegin, 104, 117

QUALM (story generator), 93

QUERTY keyboards, 7

RACTER (poetry generator), 93

Random English Sentence Generator, 87, 96, 108, 109
Raytheon BBN, 92
Real Characters, 41–45
reason, automation of, 40
Reeves, James Knapp
Practical Authorship, 71
reference tables, 40
Republic (Plato), 127
Research Laboratory of Electronics (RLE), 87
responsibility, 126–27, 129–30
rhetoric, 2
Richards, I. A., 102, 115, 119–20
The Meaning of Meaning, 102
Science and Poetry, 119–20
Riemann, Bernard, 44
River Boys, The, 66
RLE (Research Laboratory of Electronics), 87
Robinson Crusoe (Defoe), 36, 65
robots, 134
Romanticism, 51, 61, 69, 79–80
Royal Society of London, 39, 41, 48, 54
Russian aggression in Ukraine, 120–21, 125
Russian folktales, 79, 83, 97
Russian language, 121
Russian Orthodox Church, 103

Salton, Gerald, 113

SAM (story generator), 92, 94
Sapir, Edward, 83
Schank, Roger, 92, 94, 95
schemas, 62, 94–95, 97, 99–102, 112–14, 122
scholarship, 61, 121
Schott, P. Gaspar, 33
Science and Poetry (Richards), 119–20
science(s), 41, 119, 121
scientific method, 69
Scott, Sir Walter
Waverly, 65
scrapbooks, 73–74
script generators, 94–95
search engine technology, 113
Sebald, Winfried Georg, 13
self-driving cars, 14, 124, 126–27
“Self-Taught Philosopher” (Ibn Tufail), 36
semantic distance, 112
semantics, 102
semiotics, 2
Shakespeare, William, 13
Shannon, Claude, 105–6, 108–10
“A Theory of Mathematical Communication,” 105–6
Signal Corps, 87
sign-situations, 102
Simenon, Georges, 65
singularities, 7
Skeleton Essays, or Authorship in Outline (English), 71, 72
skeleton forms, 63–64
skeleton of skeletons, 64
Slosson, Edwin
Plots and Personalities, 71
smart, appearing, 3
smart cars, 126–27
smart devices, 4, 38, 39, 79
smart gadgets, 29
smart machines, 37, 99–100
smart objects, 4–5
smart phones, 123
Smith, Elizabeth Meade, 65–66
A World of Girls, 66
Society of Automotive Engineers, 15
Socrates, 35
software-engineering education, 5
solid-state memory storage, 8–9
Somerville, Mary, 51
sonnets, 31, 34
Soviet Union, 8, 65, 77, 81, 82
Spain, 26, 66
spell-checkers, 29, 110
Springfield, Mass., 71
Stalin, Joseph, 82
Stanford University, 92
Star Wars (film), 93
statistical intelligence, 114
statistics, 102–3, 108, 113–14
Sterne, Laurence
Tristram Shandy, 65
Stratemeyer, Edward L., 66
Street & Smith, 66
strings, 9, 105, 106
structuralism, 83–86, 94
STUDENT (question-answering system), 92
Student-Writer, 73
Studies of Linguistic Analysis, 114
success, concept of, 116
supply cards, 60
Swift, Jonathan
Gulliver’s Travels, 65
syntax, 87, 91, 93

tables
in Analytical Engine, 52, 64
application, 33, 48
in Mathematical Organ, 75
reference, 40
smart, 29, 36, 41
table of, 25
in zairajah, 19, 22–27
TALE-SPIN (story generator), 92, 95–98
tasks, devaluation of automated, 38
taxonomies, 12, 23–24
Technique of Drama, The (Price), 72–73
Technique of Fiction Writing, The (Dowst), 71
Technique of the Mystery Story, The (Wells), 71, 73
Technique of the Novel, The (Horne), 71
Technique of Play Writing, The (Andrews), 71
technological change, 131
telecommunications, 105
telegraphs, 7–8, 105, 111
Tellado, Corin, 66
template-based manufacturing, 60–61, 64
templates, 62, 64, 70, 71, 79–80, 83–84, 94–97, 100
Ten Million Photoplay Plots, 77
text generators, 20–21, 65, 77, 78
“Theory of Mathematical Communication, A” (Shannon), 105–6
thinking and thought, 3, 4, 14, 36, 45, 59–61, 93, 123–25, 132, 138
Thirty-Six Dramatic Situations, The (Polti), 68–71
Thompson, Henry, 93
Tolstoy, Leo, 65, 121
TRAC (story generator), 92
translation, machine, 119
transmissions, 14–16
trees, 55
trickster, figure of the, 80
Tristram Shandy (Sterne), 65
true crime, 66
truth, ground, 40
Tufts University, 28
Turin, Italy, 54
Turing, Alan, 12, 37, 43, 93
Turing machine, 119
Turing test, 119
Twain, Mark, 73–74
2001: A Space Odyssey (film), 93
Types of Prose Narratives (Fansler), 71

Ukraine, Russian aggression in, 120–21, 125

United Kingdom, 66, 131
United States
literary market in, 70
number of books printed in, 66
universal intelligence, 37–38, 139
universality, 46–47
universal language machines, 40
universal languages, 39, 46
universal machine, 12, 37
universal thesaurus, 119
University of Jena, 30–31
University of Pennsylvania, 83
University of Wisconsin, 92
urbanization, 67
US Air Force, 87
US Army, 87
US Library of Congress, 112
US Navy, 87

values
 encoding of, 135–36
local, 38
vector spaces, 11, 111–12
von Neumann, John, 44

W. W. Norton & Company, 119

wages, author, 67
Waverly (Scott), 65
Web-Work (Keeler), 73, 95
weighted keywords, 113
Weizenbaum, Joseph, 92
Wellington, Duke of, 50
Wells, Carolyn
The Technique of the Mystery Story, 71, 73
Wheatsone, Charles, 54
Wikipedia, 36, 122, 136
Wilkins, John, 39, 41–46, 48, 55, 56, 91, 119
An Essay Towards a Real Character, and a Philosophical Language (Wilkins), 10, 41–45
Winograd, Terry, 93
Wittgenstein, Ludwig, 119
Wolf, Alice, 92
Woolf, Virginia, 61
word clusters (word constellations), 112
word combinations, 106–9, 113
word embeddings, 112
Wordle, 104
word-processors, 122
words, 9, 55, 66, 88, 95, 102–9, 111–16
Wordsworth, William, 61
word vector spaces, 111–12
World of Girls, A (Smith), 66
World War II, 82
Writer’s Digest, 66, 67
Writer’s Monthly, 70
writing, 8–10, 14, 29, 38, 59, 61, 62, 79, 115, 122–24, 138
writing bots, 136–37
writing machines, 9
writing systems, 42–43
Writing the Short Story (Esenwein), 71
Wunderkammern, 10
Wurzburg, Germany, 30

Xerox Palo Alto, 92

Ya Jing, 9
Yale University, 92, 94
Yngve, Victor, 87–92, 96, 108, 109

zairajah, 18–22, 26
 Norton Shorts
B R IL L IA N C E WIT H B R E V IT Y

W. W. Norton & Company has been independent since 1923, when

William Warder Norton and Mary (Polly) D. Herter Norton first
published lectures delivered at the People’s Institute, the adult
education division of New York City’s Cooper Union. In the 1950s,
Polly Norton transferred control of the company to its employees.
One hundred years after its founding, W. W. Norton & Company
inaugurates a new century of visionary independent publishing with
Norton Shorts. Written by leading-edge scholars, these eye-opening
books deliver bold thinking and fresh perspectives in under two
hundred pages.

Available Winter 2024

Imagination: A Manifesto by Ruha Benjamin
Wild Girls: How the Outdoors Shaped the Women who Challenged a Nation by Tiya Miles
Against Technoableism: Rethinking Who Needs Improvement by Ashley Shew
Literary Theory for Robots: How Computers Learned to Write by Dennis Yi Tenen

Forthcoming
Mehrsa Baradaran on the racial wealth gap
Rina Bliss on the “reality” of race
Merlin Chowkwanyum on the social determinants of health
Daniel Aldana Cohen on eco-apartheid
Jim Downs on cultural healing
Reginald K. Ellis on Black education versus Black freedom
Brooke Harrington on offshore finance
Justene Hill Edwards on the history of inequality in America
Destin Jenkins on a short history of debt
Quill Kukla on a new vision of consent
Barry Lam on discretion
Matthew Lockwood on a new history of exploration
Natalia Molina on the myth of assimilation
Rhacel Salazar Parreñas on human trafficking
Tony Perry on water in African American culture and history
Beth Piatote on living with history
Ashanté Reese on the transformative possibilities of food
Jeff Sebo on the moral circle
Tracy K. Smith on poetry in an age of technology
Onaje X. O. Woodbine on transcendence in sports
ALSO BY DENNIS YI TENEN

Plain Text: The Poetics of Computation

Author Function: A Literary History of Artificial Intelligence
 Copyright © 2024 by Dennis Yi Tenen
All rights reserved
First Edition
For information about permission to reproduce selections from this book, write to
Permissions, W. W. Norton & Company, Inc., 500 Fifth Avenue, New York, NY 10110
For information about special discounts for bulk purchases, please contact W. W. Norton
Special Sales at [email protected] or 800-233-4830
Jacket design: Milan Bozic
Jacket image: Vasilyev Alexandr / Shutterstock
Production manager: Delaney Adams
Library of Congress Cataloging-in-Publication Data is available
ISBN: 978-0-393-88218-6
ISBN 9780393882193 (e-book)
W. W. Norton & Company, Inc.
500 Fifth Avenue, New York, N.Y. 10110
www.wwnorton.com
W. W. Norton & Company Ltd.
15 Carlisle Street, London W1D 3BS