In early January I joined a group of AI researchers from Microsoft and my fellow humanist Kathleen Fitzpatrick to talk at the Modern Language Association convention about the implications of artificial intelligence. Our panel was called Being Human, Seeming Human. Each participant came to this question of “seeming human” from a different angle. My own focus was on creativity. Here’s the text of my prepared remarks.
Today I want to talk to talk briefly about artificial intelligence and creativity. And not just creativity as it pertains to AI but human creativity as well. So, has anyone heard or played AI Dungeon yet?
AI Dungeon was released just a few weeks ago and it has gone absolutely viral. It’s an online text adventure you play in your browser or run as an app on your phone. Now, text adventure, that was a popular kind of game in the 1980s. A lot of people know Zork. In these games the player is offered textual descriptions of a house, a cave, a spaceship, dungeon, whatever, and the player types short sentences like go east, get lamp, or kill troll in order to solve puzzles, collect treasure, and win the game. There’s a parser that understands these simple commands and responds with canned interactions prewritten by the game developers. Text adventures are also known as interactive fiction and there’s a rabid fan base online that’s part geek nostalgia, part genuine fondness for these text-based games.
Interactive fiction often revolves around choice, where players have multiple ways to transverse the world and solve the puzzles. Following this generic convention, AI Dungeon opens up with major choice, literally which genre of text adventure you want. Fantasy, mystery, apocalyptic, and so on.
So here I picked fantasy and immediately I’m thrust into a procedurally generated story: a fantasy world entirely written by a natural language processing program.
Generating a static dungeon on the fly is one thing. But what’s amazing about AI Dungeon is that it’s not a scripted world so much as an improv stage. You can literally type anything, and AI Dungeon will roll with it, generating an on-the-fly response.
So here, we have a stock feature of fantasy text adventures, a dragon. And I eat it. The game doesn’t bat an eye. It runs with it and lets me eat the dragon, responding with a fairly sophisticated sentence that aside from its subject matter, sounds like something you’d read in a classic text adventure. “You quickly grab the dragon’s corpse and tear of a piece of its flesh.”
Let me be clear. No human wrote that sentence. No human preconceived a scenario where the player might eat the dragon. The AI generated this. Semantically and grammatically the AI nails language. It’s not as good at ontology. It lets me fly the dragon corpse to Seattle. The AI is a sponge that accepts all interactions. As you can imagine people go crazy with this. The amount of AI dungeon erotica out there is staggering—and disturbing.
Later I run into some people and I ask them about the MLA convention.
A man responds to my question about the MLA, “It’s a convention where all wizards use the same language. It’ll make things easier.”
Oh, that answer is both so right and so wrong.
So how does this all work? I obviously don’t have time to go into all the details. But it’s roughly this: AI Dungeon relies on GPT-2, a AI-powered natural language generator. The full GPT-2 set is trained on 1.5 billion parameters gleaned from over 40 gigabytes of text scrapped from the Internet. The training of GPT-2 took months on super-powered computers. It was developed by Open AI, a not-for-profit research company funded by a mix of private donors like Elon Musk and Microsoft, which donated $1 billion to Open AI in July.
One innovation of GPT-2 is that you can take the base language model and fine-tune it on more specific genres or discourse. For a while Open AI stalled on releasing the full GPT-2 set because of concerns it could be abused, say by extremist groups generating massive quantities of AI-written propaganda. In the more benign case of AI Dungeon, the AI is finetuned using text adventures scrapped from chooseyourstory.com.
There’s much more to be said about AI Dungeon, but I’ll leave you with just a few provocations.
Games are often defined by their rules. So is AI Dungeon a game if you can do anything?
Stories are often defined by their storytellers. Is AI Dungeon a story if no one is telling it?
And finally, a mantra I repeat often to my students when it comes to technology: everything comes from somewhere else. Everything comes from somewhere else. GPT-2 didn’t emerge whole-cloth out of nothing. It’s trained on the Internet, specifically, sources linked to from Reddit. There’s money involved, lots of it. Follow the money. Likewise, AI Dungeon itself comes from somewhere else. On one hand its creator is a Brigham Young University undergraduate student, Nick Walton. On the other hand, the vision behind AI Dungeon—computers telling stories—goes back decades, a history Noah Wardrip-Fruin explores in Expressive Processing. The genre fiction invoked by AI Dungeon has an even longer history.
All this adds up to the fact that AI Dungeon turns out to be a perfect object of study for so many disciplines in the humanities. Whether you think it’s a silly gimmick, an abomination of the creative spirit, the precursor to a new age of storytelling, whatever, I think humanists ignore AI storytelling at our own peril.
This summer I attended the first annual Institute for Liberal Arts Digital Scholarship (ILiADS) at Hamilton College. It was an inspiring conference, highlighting the importance of collaborative faculty/student digital work at small liberal arts colleges. My own school, Davidson College, had a team at ILiADS (Professor Suzanne Churchill, Instructional Technologist Kristen Eshleman, and undergraduate Andrew Rikard, working on a digital project about the modernist poet Mina Loy). Meanwhile I was at the institute to deliver the keynote address on the final day. Here is the text of my keynote, called “Your Mistake was a Vital Connection: Oblique Strategies for the Digital Humanities.”
Forty years ago, the musician Brian Eno and painter Peter Schmidt published the first edition of what they called Oblique Strategies.Oblique Strategies resembled a deck of playing cards, each card black on one side, and white on the other, with a short aphoristic suggestion on the white side.
These suggestions were the strategies—the oblique strategies—for overcoming creative blocks or artistic challenges. The instructions that came with the cards described their use: “They can be used as a pack…or by drawing a single card from the shuffled pack when a dilemma occurs in a working situation. In this case, the card is trusted even if its appropriateness is quite unclear.”
When we look at some of the strategies from the original deck of 55 cards, we can see why their appropriateness might appear unclear:
And other strategies:
Make sure nobody wants it.
Cut a vital connection
Make a blank valuable by putting it in an exquisite frame
Do something boring
Honor thy error as a hidden intention
And one of my favorites:
Repetition is a form of change.
Brian Eno explained the origins of the cards in an interview on KPFA radio in San Francisco in 1980: The cards were a system designed to, as Eno put it, “foil the critic” in himself and to “encourage the child.” They were strategies for catching our internal critics off-guard. Eno elaborated:
The Oblique Strategies evolved from me being in a number of working situations when the panic of the situation—particularly in studios—tended to make me quickly forget that there were others ways of working and that there were tangential ways of attacking problems that were in many senses more interesting than the direct head-on approach.
If you’re in a panic, you tend to take the head-on approach because it seems to be the one that’s going to yield the best results. Of course, that often isn’t the case—it’s just the most obvious and—apparently—reliable method. The function of the Oblique Strategies was, initially, to serve as a series of prompts which said, “Don’t forget that you could adopt *this* attitude,” or “Don’t forget you could adopt *that* attitude.”
Other ways of working. There are other ways of working. That’s what the Oblique Strategies remind us. Eno and Schmidt released a second edition in 1978 and a third edition in 1979, the year before Schmidt suddenly died. Each edition varied slightly. New strategies appeared, others were removed or revised.
For example, the 2nd edition saw the addition of “Go outside. Shut the door.” A 5th edition in 2001 added the strategy “Make something implied more definite (reinforce, duplicate).” For a complete history of the various editions, check out Gregory Taylor’s indispensable Obliquely Stratigraphic Record. The cards—though issued in limited, numbered, editions—were legendary, and even more to the point, they were actually used.
David Bowie famously kept a deck of the cards on hand when he recorded his Berlin albums of the late seventies. His producer for these experimental albums was none other than Brian Eno. I’m embarrassed to say that I didn’t know about Bowie’s use of the Oblique Strategies
I knew about Tristan Tzara’s suggestion in the 1920s to write poetry by pulling words out of a bag. I knew about Brion Gysin’s cut-up method, which profoundly influenced William Burroughs. I knew about John Cage’s experimental compositions, such as his Motor Vehicle Sundown, a piece orchestrated by “any number of vehicles arranged outdoors.” Or Cage’s use of chance operations, in which lists of random numbers from Bell Labs determined musical elements like pitch, amplitude, and duration. I knew how Jackson Mac Low similarly used random numbers to generate his poetry, in particular relying on a book called A Million Random Digits with 100,000 Normal Deviates to supply him with the random numbers (Zweig 85).
I knew about the poet Alison Knowles’ “The House of Dust,” which is a kind of computer-generated cut-up written in Fortran in 1967. I even knew that Thom Yorke composed many of the lyrics of Radiohead’s Kid A using Tristan Tzara’s method, substituting a top hat for a bag.
But I hadn’t heard encountered Eno and Schmidt’s Oblique Strategies. Which just goes to show, however much history you think you know—about art, about DH, about pedagogy, about literature, about whatever—you don’t know the half of it. And I suppose the ahistorical methodology of chance operations is part of their appeal. Every roll of the dice, every shuffle of the cards, every random number starts anew. In his magisterial—and quite frankly, seemingly random—Arcades Project, Benjamin devotes an entire section to gambling, where his collection of extracts circles around the essence of gambling. “The basic principle…of gambling…consists in this,” says Alain in one of Benjamin’s extracts, “…each round is independent of the one preceding…. Gambling strenuously denies all acquired conditions, all antecedents…pointing to previous actions…. Gambling rejects the weighty past” (Benjamin 512). Every game is cordoned off from the next. Every game begins from the beginning. Every game requires that history disappear.
That’s the goal of the Oblique Strategies—to clear a space where your own creative history doesn’t stand in the way of you moving forward in new directions. Now in art, chance operations may be all well and good, even revered. But what does something like the Oblique Strategies have to do with the reason we’re here this week: research, scholarship, the production of knowledge? After all, isn’t rejecting “the weighty past” an anathema to the liberal arts?
Well, I think one answer goes back to Eno’s characterization of the Oblique Strategies: there are other ways of working. We can approach the research questions that animate us indirectly, at an angle. Forget the head-on approach for a while.
One way of working that I’ve increasingly become convinced is a legitimate—and much-needed form of scholarship—is deformance. Lisa Samuels and Jerry McGann coined this word, a portmanteau of deform and performance. It’s an interpretative concept premised upon deliberately misreading a text. For example, reading a poem backwards line-by-line. As Samuels and McGann put it, reading backwards “short circuits” our usual way of reading a text and “reinstalls the text—any text, prose or verse—as a performative event, a made thing” (Samuels & McGann 30). Reading backwards revitalizes a text, revealing its constructedness, its seams, edges, and working parts.
Let me give you an example of deformance. Mary Lee and Katharine are two social media stars, with tens of thousands of followers on Twitter each. They’re also great white sharks in the Atlantic Ocean, tagged with geotrackers by the non-profit research group OCEARCH. Whenever either of the sharks—or any of the dozens of other sharks that OCEARCH has tagged—surfaces longer than 90 seconds, the tags ping geo-orbiting satellites three times in order to triangulate a position. That data is then shared in real-time on OCEARCH’s site or app.
The sharks’ Twitter accounts, I should say, are operated by humans. They’ll interact with followers, post updates about the sharks, tweet shark facts, and so on. But these official Mary Lee and Katharine accounts don’t automatically tweet the sharks’ location updates.
Sometime this summer—well, actually, it was during shark week—I thought wouldn’t it be cool to create a little bot, a little autonomous program, that automatically tweeted Mary Lee and Katharine’s location updates. But first I needed to get the data itself. I was able to reverse engineer OCEARCH’s website to find an undocumented API, a kind of programming interface that allows computer programs to talk to each other and share data with each other. OCEARCH’s database gives you raw JSON datathat looks like this to a human reader:
But to a computer reader, it looks like this:
Structured data is a thing of beauty.
Reverse engineering the OCEARCH API is not the deformance I’m talking about here. What I found when the bot started tweeting location updates of these two famous sharks was, it was kind of boring. Every few days one of the sharks would surface long enough to get a position, it would post to Twitter, and that was that.
Something was missing. I wanted to give this Twitter account texture, a sense of personality. I decided to make Mary Lee and Katharine writers. And they would share their writing with the world on this Twitter account. The only problem is, I don’t have time to be a ghost writer for two great white sharks.
So I’ll let a computer do that.
I asked some friends for ideas of source material to use as deformance pieces for the sharks. These would be texts that I could mangle or remix in order to come up with original work that I would attribute to the sharks. A friend suggested American Psycho—an intriguing choice for a pair of sharks, but not quite the vibe I was after. Mary Lee and Katharine are female sharks. I wanted to use women writers. Then Amanda French suggested Virginia Woolf’s novel Night and Day, which just happens to feature two characters named Katharine and Mary. It was perfect, and the results are magical.
Now, Katharine tweets odd mashed-up fragments from Night and Day, each one paired with historical location data from OCEARCH’s database. On December 9, 2014, Katharine was very close to the shore near Rhode Island, and she “wrote” this:
Katharine: Down all luxuriance and plenty to the verge of decency; and in the night, bereft of life (09-Dec-2014) pic.twitter.com/SEEsv3FBKm
In every case, the journal part of the tweet—the writing—is extracted randomly from complete text of Night and Day and then mangled by a Python program. These fragments, paired with the location and the character of a shark, stand on their own, and become new literary material. But they also expose the seams of the original source.
Whereas Katharine writes in prose fragments, I wanted Mary Lee to write poetry:
The line of heroes stands, godlike: Though we wander about, the tangled thread falls slack.
Mary Lee follows the cut-up method described by Brion Gysin decades ago. I’ve made a database of 1,288 lines of H.D.’s most anthologized poetry. Every tweet from Mary Lee is some combination of three of those 1,288 lines, along with slight typographic formatting. All in all, there are potentially 2 billion, 136 million and 719 thousand original poems that Mary Lee could write.
The snow is melted, we have always known you wanted us. My mind is reft.
What kind of project is @shark_girls? Is it a public humanities project—sharing actual data—dates, locations, maps—that helps people to think differently about wildlife, about the environment, about the interactions between humans and nature? Is it an art project, generating new, standalone postmodern poetry and prose? Is it a literary project, that lets us see Virginia Woolf and H.D. in a new light? Is it all three?
I’ll let other people decide. We can’t get too hung up on labels. What’s important to me is that whatever @shark_girls is about, it’s also about something else. As Susan Sontag wrote about literature: “whatever is happening, something else is always going on.” And the oblique nature of deformance will often point toward that something else. Deformance is a kind of oblique strategy for reading a poem. If the Oblique Strategies deck had a card for deformance it might read:
Or maybe, simply,
Another kind of deformance—another oblique strategy for reading—are Markov Chains. Markov chains are statistical models of texts or numbers, based on probability. Let’s say we have the text of Moby Dick.
Just eyeballing the first page we can see that certain words are more likely to be followed by some words than other words. For example, the pronoun “I” is likely to be followed by the verb “find” but not the word “the.” A two-gram Markov Chain looks at the way one pair of words is likely to be followed by a second pair of words. So the pair “I find” is likely to be followed by “myself growing” but not the pair of words “me Ishmael.” A three-gram Markov parses the source text into word triplets. The chain part of a Markov Chain happens when one of these triplets is followed by another triplet, but not necessarily the same triplet that appears in the source text. And then another triplet. And another. It’s a deterministic way to create texts, with each new block of the chain independent of the preceding blocks. Talk about rejecting the weighty past. If you work with a big enough source text, the algorithm generates sentences that are grammatically correct but often nonsensical.
Let’s generate some Markov chains of Moby Dick on the spot. Here’s a little script I made. If it takes a few seconds to load, that’s because every time it runs, it’s reading the entire text of Moby Dick and calculating all the statistical models on the fly. Then spitting out a 3-, 4-, or 5-gram Markov chain. The script tells you what kind of Markov n-gram it is. The script is fun to play around with, and I’ve used it to teach what I call deep textual hacks. When I show literature folks this deformance and teach them how to replace Moby Dick with a text from their own field or time period, they’re invariably delighted. When I show history folks this deformance and teach them how to replace Moby Dick with a primary source from their own field or time period, they’re invariably horrified. History stresses attentiveness to the nuances of a primary source document, not the way you can mangle that very same primary source. Yet, also invariably, my history colleagues realize what Samuels and McGann write about literary deformance is true of historical deformance as well: deformance revitalizes the original text and lets us see it fresh.
All of this suggests what ought to be another one of Brian Eno and Peter Schmidt’s Oblique Strategies:
Misreading is a form of reading.
And to go further, misreading can be a form of critical reading.
Now, let me get to the heart of the matter. I’ve been talking chance operations, deterministic algorithms, and other oblique strategies as a way to explore cultural texts and artifacts. But how do these oblique strategies fit in with the digital humanities? How might oblique strategies not only be another way to work in general, but specifically, another way to work with the digital scholarship and pedagogy we might otherwise more comfortably approach head-on, as Brian Eno put it.
Think about how we value—or say we value—serendipity in higher education. We often praise serendipity as a tool for discovery. When faculty hear that books are going into off-site storage, their first reaction is, how are we going to stumble upon new books when browsing the shelves?
A recent piece by the digital humanities Victorianist Paul Fyfe argues that serendipity has been operationalized, built right into the tools we use to discover new works and new connections between works (Fyfe 262). Serendipomatic, for example, is an online tool that came out of the Roy Rosenzweig Center for History and New Media. You can dump in your entire Zotero library, or even just a selection of text, and Serendipomatic will find sources from the Digital Public Library of America or Europeana that are connected—even if obliquely—to your citations. Let your sources surprise you, the tagline goes.
Tim Sherratt has created a number of bots that tweet out random finds from the National Library of Australia. I’ve done the same with the Digital Public Library of America, creating a bot that posts random items from the DPLA.Similarly, there’s @BookImages, which tweets random cat-pics images from the 3.3 million public domain images from pre-1922 books that the Internet Archive uploaded to Flickr.
Fyfe suggests that “these machines of serendipity sometimes offer simple shifts of perspective” (263)—and he’s totally right. And simple shifts of perspective are powerful experiences, highlighting the contingency and plurality of subjectivity.
But in all these cases, serendipity is a tool for discovery, not a mode of work itself. We think of serendipity as a way to discover knowledge, but not as a way to generate knowledge. This is where the oblique strategies come into play. They’re not strategies for discovery, they’re practices for creativity.
Let me state it simply: what if we did the exact opposite of what many of you have spent the entire week doing. Many of you have been here a whole week, thinking hard and working hard—which are not necessarily the same thing—trying to fulfill a vision, or at the very least, sketch out a vision. That is good. That is fine, necessary work. But what if we surrendered our vision and approached our digital work obliquely—even, blindly.
I’m imagining a kind of dada DH. A gonzo DH. Weird DH. Which is in fact the name of a panel I put together for the 2016 MLA in Austin in January. As I wrote in the CFP, “What would an avant-garde digital humanities look like? What might weird DH reveal that mainstream DH leaves out? Is DH weird enough already? Can we weird it more?”
My own answer to that last question is, yes. Yes, we can. Weird it more. The folks on the panel: Micki Kaufman, Shane Denson, Kim Knight, Jeremy Justus will all be sharing work that challenges our expectations about the research process, about the final product of scholarship, and even what counts as knowledge itself, as opposed to simply data, or information.
So many of the methodologies we use in the digital humanities come from the social sciences—network analysis, data visualization, GIS and mapping, computational linguistics. And that’s all good and I am 100 percent supportive of borrowing methodological approaches. But why do we only borrow from the sciences? What if—and this is maybe my broader point today—what if we look for inspiration and even answers from art? From artists. From musicians and poets, sculptors and quilters.
And this takes me back to my earlier question: what might a set of oblique strategies—originally formulated by a musician and an artist—look like for the digital humanities?
Well, we could simply take the existing oblique strategies and apply them to our own work.
Do something boring.
Maybe that’s something we already do. But I think we need a set of DH-specific oblique strategies. My first thought was to subject the original Oblique Strategies to the same kind of deterministic algorithm that I showed you with Moby-Dick, that is, Markov chains.
Here are a few of the Markov Chained Oblique Strategies my algorithm generated:
Breathe more human. Where is the you do?
Make what’s perfect more than nothing for as ‘safe’ and continue consistently.
Your mistake was a vital connection.
I love the koan-like feeling of these statements. The last one made so much sense that I worked it into the title of my talk: your mistake was a vital connection. And I truly believe this: our mistakes, in our teaching, in our research, are in fact vital connections. Connections binding us to each other, connections between domains of knowledge, connections between different iterations of our own work.
But however much I like these mangled oblique strategies, they don’t really speak specifically about our work in the digital humanities. So in the past few weeks, I’ve been trying to create DH-specific oblique strategies, programmatically.
The great thing about Markov chains is that you can combine multiple source texts, and the algorithm will treat them equally. My Moby Dick Markov Chains came from the entire text of the novel, but there’s no reason I couldn’t also dump in the text of Sense and Sensibility, creating a procedurally-generated mashup that combines n-grams from both novels into something we might call Moby Sense.
So I’m going to try something for my conclusion. And I have no idea if this is going to work. This could be a complete and utter failure. Altogether, taking into account the different editions of the Oblique Strategies, there are 155 different strategies. I’m going to combine those lines with texts that have circulated through the digital humanities community in the past few years. This source material includes Digital_Humanities, The Digital Humanities Manifesto, and a new project on Critical DH and Social Justice, among other texts. (All the sources are linked below.) I’ve thrown all these texts together in order to algorithmically generate my new DH-focused oblique strategies.
[At this point in my keynote I started playing around with the Oblique DH Generator. The version I debuted is on a stand-alone site, but I’ve also embedded it below. My talk concluded—tapered off?—as I kept generating new strategies and riffing on them. We then moved to a lively Q&A period, where I elaborated on some of the more, um, oblique themes of my talk. As nicely as this format worked at ILiADS, it doesn’t make for a very satisfying conclusion here. So I’ll wrap up with a new, equally unsatisfying conclusion, and then you can play with the generator below. And draw your own conclusions.]
My conclusion is this, then. A series a oblique strategies for the digital humanities that are themselves obliquely generated. The generator below is what I call a webtoy. But I’ve also been thinking about it as what Ted Nelson calls a “thinkertoy”—a toy that helps you think and “envision complex alternatives” (Dream Machines 52). In this case, the thinkertoy suggests alternative approaches to the digital humanities, both as a practice and as a construct (See Kirschenbaum on the DH construct). And it’s also just plain fun. For, as one of the generated oblique strategies for DH says, Build bridges between the doom and the scholarship. And what better way to do that than playing around?
Benjamin, Walter. The Arcades Project. Edited by Rolf Tiedemann. Translated by Howard Eiland and Kevin McLaughlin. Cambridge, Massachusetts: Belknap-Harvard UP, 1999.
The Expressive Work of Spaces of Torture in Videogames
At the 2014 MLA conference in Chicago I appeared on a panel called “Torture and Popular Culture.” I used the occasion to revisit a topic I had written about several years earlier—representations of torture-interrogation in videogames. My comments are suggestive more than conclusive, and I am looking forward to developing these ideas further.
Today I want to talk about spaces of torture—dungeons, labs, prisons—in contemporary videogames and explore the way these spaces are not simply gruesome narrative backdrops but are key expressive features in popular culture’s ongoing reckoning with modern torture. Continue reading “Sites of Pain and Telling”→
Seeking to have a rich discussion period—which we did indeed have—we limited our talks to about 12 minutes each. My presentation was therefore more evocative than comprehensive, more open-ended than conclusive. There are primary sources I’m still searching for and technical details I’m still sorting out. I welcome feedback, criticism, and leads.
An Account of Randomness in Literary Computing
MLA 2013, Boston
There’s a very simple question I want to ask this evening:
Where does randomness come from?
Randomness has a rich history in arts and literature, which I don’t need to go into today. Suffice it to say that long before Tristan Tzara suggested writing a poem by pulling words out of a hat, artists, composers, and writers have used so-called “chance operations” to create unpredictable, provocative, and occasionally nonsensical work. John Cage famously used chance operations in his experimental compositions, relying on lists of random numbers from Bell Labs to determine elements like pitch, amplitude, and duration (Holmes 107–108). Jackson Mac Low similarly used random numbers to generate his poetry, in particular relying on a book called A Million Random Digits with 100,000 Normal Deviates to supply him with the random numbers (Zweig 85).
Published by the RAND Corporation in 1955 to supply Cold War scientists with random numbers to use in statistical modeling (Bennett 135), the book is still in print—and you should check out the parody reviews on Amazon.com. “With so many terrific random digits,” one reviewer jokes, “it’s a shame they didn’t sort them, to make it easier to find the one you’re looking for.”
This joke actually speaks to a key aspect of randomness: the need to reuse random numbers, so that, say you’re running a simulation of nuclear fission, you can repeat the simulation with the same random numbers—that is, the same probability—while testing some other variable. In fact, most of the early work on random number generation in the United States was funded by either the U.S. Atomic Commission or the U.S. Military (Montfort et al. 128). The RAND Corporation itself began as a research and development arm of the U.S. Air Force.
Now the thing with going down a list of random numbers in a book, or pulling words out of hat—a composition method, by the way, Thom Yorke used for Kid A after a frustrating bout of writer’s block—is that the process is visible. Randomness in these cases produces surprises, but the source itself of randomness is not a surprise. You can see how it’s done.
What I want to ask here today is, where does randomness come from when it’s invisible? What’s the digital equivalent of pulling words out of a hat? And what are the implications of chance operations performed by a machine?
To begin to answer these questions I am going to look at two early works of electronic literature that rely on chance operations. And when I say early works of electronic literature, I mean early, from fifty and sixty years ago. One of these works has been well studied and the other has been all but forgotten.
My first case study is the Strachey Love Letter Generator. Programmed by Christopher Strachey, a close friend of Alan Turing, the Love Letter Generator is likely—as Noah Wardrip-Fruin argues—the first work of electronic literature, which is to say a digital work that somehow makes us more aware of language and meaning-making. Strachey’s program “wrote” a series of purplish prose love letters on the Ferranti Mark I Computer—the first commercially available computer—at Manchester University in 1952 (Wardrip-Fruin “Digital Media” 302):
YOU ARE MY AVID FELLOW FEELING. MY AFFECTION CURIOUSLY CLINGS TO YOUR PASSIONATE WISH. MY LIKING YEARNS FOR YOUR HEART. YOU ARE MY WISTFUL SYMPATHY: MY TENDER LIKING.
M. U. C.
Affectionately known as M.U.C., the Manchester University Computer could produce these love letters at a pace of one per minute, for hours on end, without producing a duplicate.
The “trick,” as Strachey put it in a 1954 essay about the program (29-30), is its two template sentences (Myadjectivenounadverbverb your adjectivenoun and You are my adjectivenoun) in which the nouns, adjectives, and adverbs are randomly selected from a list of words Strachey had culled from a Roget’s thesaurus. Adverbs and adjectives randomly drop out of the sentence as well, and the computer randomly alternates the two sentences.
The Love Letter Generator has attracted—for a work of electronic literature—a great deal of scholarly attention. Using Strachey’s original notes and source code (see figure to the left), which are archived at the Bodleian Library at the University of Oxford, David Link has built an emulator that runs Strachey’s program, and Noah Wardrip-Fruin has written a masterful study of both the generator and its historical context.
As Wardrip-Fruin calculates, given that there are 31 possible adjectives after the first sentence’s opening possessive pronoun “My” and then 20 possible nouns that could that could occupy the following slot, the first three words of this sentence alone have 899 possibilities. And the entire sentence has over 424 million combinations (424,305,525 to be precise) (“Digital Media” 311).
On the whole, Strachey was publicly dismissive of his foray into the literary use of computers. In his 1954 essay, which appeared in the prestigious trans-Atlantic arts and culture journal Encounter (a journal, it would be revealed in the late 1960s, that was primarily funded by the CIA—see Berry, 1993), Strachey used the example of the love letters to illustrate his point that simple rules can generate diverse and unexpected results (Strachey 29-30). And indeed, the Love Letter Generator qualifies as an early example of what Wardrip-Fruin calls, referring to a different work entirely, the Tale-Spin effect: a surface illusion of simplicity which hides a much more complicated—and often more interesting—series of internal processes (Expressive Processing 122).
Wardrip-Fruin coined this term—the Tale-Spin effect—from Tale-Spin, an early story generation system designed by James Mehann at Yale University in 1976. Tale-Spin tended to produce flat, plodding narratives, though there was the occasional existential story:
Henry Ant was thirsty. He walked over to the river bank where his good friend Bill Bird was sitting. Henry slipped and fell in the river. He was unable to call for help. He drowned.
But even in these suggestive cases, the narratives give no sense of the process-intensive—to borrow from Chris Crawford—calculations and assumptions occurring behind the interface of Tale-Spin.
In a similar fashion, no single love letter reveals the combinatory procedures at work by the Mark I computer.
MY AFFECTION LUSTS FOR YOUR TENDERNESS. YOU ARE MY PASSIONATE DEVOTION: MY WISTFUL TENDERNESS. MY LIKING WOOS YOUR DEVOTION. MY APPETITE ARDENTLY TREASURES YOUR FERVENT HUNGER.
M. U. C.
This Tale-Spin effect—the underlying processes obscured by the seemingly simplistic, even comical surface text—are what draw Wardrip-Fruin to the work. But I want to go deeper than the algorithmic process that can produce hundreds of millions of possible love letters. I want to know, what is the source of randomness in the algorithm? We know Strachey’s program employs randomness, but where does that randomness come from? This is something the program—the source code itself—cannot tell us, because randomness operates at a different level, not at the level of code or software, but in the machine itself, at the level of hardware.
In the case of Strachey’s Love Letter Generator, we must consider the computer it was designed for, the Mark I. One of the remarkable features of this computer was that it had a hardware-based random number generator. The random number generator pulled a string of random numbers from what Turing called “resistance noise”—that is, electrical signals produced by the physical functioning of the machine itself—and put the twenty least significant digits of this number into the Mark I’s accumulator—its primary mathematical engine (Turing). Alan Turing himself specifically requested this feature, having theorized with his earlier Turing Machine that a purely logical machine could not produce randomness (Shiner). And Turing knew—like his Cold War counterparts in the United States—that random numbers were crucial for any kind of statistical modeling of nuclear fission.
I have more to say about randomness in the Strachey Love Letter Generator, but before I do, I want to move to my second case study. This is an early, largely unheralded work called SAGA. SAGA was a script-writing program on the TX-0 computer. The TX-0 was the first computer to replace vacuum tubes with transistors and also the first to use interactive graphics—it even had a light pen.
The TX-0 was built at Lincoln Laboratory in 1956—a classified MIT facility in Bedford, Massachusetts chartered with the mission of designing the nation’s first air defense detection system. After TX-0 proved that transistors could out-perform and outlast vacuum tubes, the computer was transferred to MIT’s Research Laboratory of Electronics in 1958 (McKenzie), where it became a kind of playground for the first generation of hackers (Levy 29-30).
In 1960, CBS broadcast an hour-long special about computers called “The Thinking Machine.” For the show MIT engineers Douglas Ross and Harrison Morse wrote a 13,000 line program in six weeks that generated a climactic shoot-out scene from a Western.
Several computer-generated variations of the script were performed on the CBS program. As Ross told the story years later, “The CBS director said, ‘Gee, Westerns are so cut and dried couldn’t you write a program for one?’ And I was talked into it.”
The TX-0’s large—for the time period—magnetic core memory was used “to keep track of everything down to the actors’ hands.” As Ross explained it, “The logic choreographed the movement of each object, hands, guns, glasses, doors, etc.” (“Highlights from the Computer Museum Report”).
And here, is the actual output from the TX-0, printed on the lab’s Flexowriter printer, where you can get a sense of the way SAGA generated the play:
In the CBS broadcast, Ross explained the narrative sequence as a series of forking paths.
Each “run” of SAGA was defined by sixteen initial state variables, with each state having several weighted branches (Ross 2). For example, one of the initial settings is who sees whom first. Does the sheriff see the robber first or is it the other way around? This variable will influence who shoots first as well.
There’s also a variable the programmers called the “inebriation factor,” which increases a bit with every shot of whiskey, and doubles for every swig straight from the bottle. The more the robber drinks, the less logical he will be. In short, every possibility has its own likely consequence, measured in terms of probability.
The MIT engineers had a mathematical formula for this probability (Ross 2):
But more revealing to us is the procedure itself of writing one of these Western playlets.
First, a random number was set; this number determined the probability of the various weighted branches. The programmers did this simply by typing a number following the RUN command when they launched SAGA; you can see this in the second slide above, where the random number is 51455. Next a timing number established how long the robber is alone before the sheriff arrives (the longer the robber is alone, the more likely he’ll drink). Finally each state variable is read, and the outcome—or branch—of each step is determined.
What I want to call your attention to is how the random number is not generated by the machine. It is entered in “by hand” when one “runs” the program. In fact, launching SAGA with the same random number and the same switch settings will reproduce a play exactly (Ross 2).
In a foundational work in 1996 called The Virtual Muse Charles Hartman observed that randomness “has always been the main contribution that computers have made to the writing of poetry”—and one might be tempted to add, to electronic literature in general (Hartman 30). Yet the two case studies I have presented today complicate this notion. The Strachey Love Letter Generator would appear to exemplify the use of randomness in electronic literature. But—and I didn’t say this earlier—the random numbers generated by the Mark I’s method tended not to be reliably random enough; remember, random numbers often need to be reused, so that the programs that run them can be repeated. This is called pseudo-randomness. This is why books like the RAND Corporation’s A Million Random Digits is so valuable.
But the Mark I’s random numbers were so unreliable that they made debugging programs difficult, because errors never occurred the same way twice. The random number instruction eventually fell out of use on the machine (Campbell-Kelly 136). Skip ahead 8 years to the TX-0 and we find a computer that doesn’t even have a random number generator. The random numbers must be entered manually.
The examples of the Love Letters and SAGA suggest at least two things about the source of randomness in literary computing. One, there is a social-historical source; wherever you look at randomness in early computing, the Cold War is there. The impact of the Cold War upon computing and videogames has been well-documented (see, for example Edwards, 1996 and Crogan, 2011), but few have studied how deeply embedded the Cold War is in the software algorithms and hardware processes themselves of modern computing.
Second, randomness does not have a progressive timeline. The story of randomness in computing—and especially in literary computing—is neither straightforward nor self-evident. Its history is uneven, contested, and mostly invisible. So that even when we understand the concept of randomness in electronic literature—and new media in general—we often misapprehend its source.
Bennett, Deborah. Randomness. Cambridge, MA: Harvard University Press, 1998. Print.
Turing, A.M. “Programmers’ Handbook for the Manchester Electronic Computer Mark II.” Oct. 1952. Web. 23 Dec. 2012.
Wardrip-Fruin, Noah. “Digital Media Archaeology: Interpreting Computational Processes.” Media Archaeology: Approaches, Applications, and Implications. Ed by. Erkki Huhtamo & Jussi Parikka. Berkeley, California: University of California Press, 2011. Print.
—. Expressive Processing: Digital Fictions, Computer Games, and Software Sudies. MIT Press, 2009. Print.
Zweig, Ellen. “Jackson Mac Low: The Limits of Formalism.” Poetics Today 3.3 (1982): 79–86. Web. 1 Jan. 2013.
(This is the text of my five minute position statement on the role of computational literacy in computers and writing. I delivered this statement during a “town hall” meeting at the annual Computers and Writing Conference, hosted at North Carolina State University on May 19, 2012.)
I want to briefly run through five basic statements about computational literacy. These are literally 5 statements in BASIC, a programming language developed at Dartmouth in the 1960s. As some of you might know, BASIC is an acronym for Beginner’s All-Purpose Symbolic Instruction Code, and the language was designed in order to help all undergraduate students at Dartmouth—not just science and engineering students—use the college’s time-sharing computer system.
Each BASIC statement I present here is a fully functioning 1-line program. I want to use each as a kind of thesis—or a provocation of a thesis—about the role of computational literacy in computers and writing, and in the humanities more generally.
10 PRINT 2+3
I’m beginning with this statement because it’s a highly legible program that nonetheless highlights the mathematical, procedural nature of code. But this program is also a piece of history: it’s the first line of code in the user manual of the first commercially available version of BASIC, developed for the first commercially available home computer, the Altair 8800. The year was 1975 and this BASIC was developed by a young Bill Gates and Paul Allen. And of course, their BASIC would go on to be the foundation of Microsoft. It’s worth noting that although Microsoft BASIC was the official BASIC of the Altair 8800 (and many home computers to follow), an alternative version, called Tiny BASIC, was developed by a group of programmers in San Francisco. The 1976 release of Tiny BASIC included a “copyleft” software license, a kind of predecessor to contemporary open source software licenses. Copyleft emphasized sharing, an idea at the heart of the original Dartmouth BASIC.
10 PRINT “HELLO WORLD”
If BASIC itself was a program that invited collaboration, then this—customarily one of the first programs a beginner learns to write—highlights the way software looks outward. Hello, world. Computer code is writing in public, a social text. Or, what Jerry McGann calls a “social private text.” As McGann explains, “Texts are produced and reproduced under specific social and institutional conditions, and hence…every text, including those that may appear to be purely private, is a social text.”[1. McGann, Jerome. The Textual Condition. Princeton, NJ: Princeton University Press, 1991, p. 21.]
10 PRINT “GO TO STATEMENT CONSIDERED HARMFUL”: GOTO 10
My next program is a bit of an insider’s joke. It’s a reference to a famous 1968 diatribe by Edsger Dijkstra called “Go To Statement Considered Harmful.” Dijkstra argues against using the goto command, which leads to what critics call spaghetti code. I’m not interested in that specific debate, so much as I like how this famous injunction implies an evaluative audience, a set of norms, and even an aesthetic priority. Programming is a set of practices, with its own history and tensions. Any serious consideration of code—any serious consideration of computers—in the humanities must reckon with these social elements of code.
10 REM PRINT “GOODBYE CRUEL WORLD”
The late German media theorist Frederich Kittler has argued that, as Alexander Galloway put it, “code is the only language that does what it says.”[2. Galloway, Alexander R. Gaming: Essays on Algorithmic Culture. Minneapolis: University of Minnesota Press, 2006, p. 6] Yes, code does what it says. But it also says things it does not do. Like this one-line program which begins with REM, short for remark, meaning this is a comment left by a programmer, which the computer will not execute. Comments in code exemplify what Mark Marino has called the “extra-functional significance” of code, meaning-making that goes beyond the purely utilitarian commands in the code.[3. Marino, Mark C. “Critical Code Studies.” Electronic Book Review (2006). <http://www.electronicbookreview.com/thread/electropoetics/codology>.]
Without a doubt, there is much even non-programmers can learn not by studying what code does, but by studying what it says, and what it evokes.
10 PRINT CHR$(205.5+RND(1));:GOTO 10
Finally, here’s a program that highlights exactly how illegible code can be. Very few people could look at this program for the Commodore 64 and figure out what it does. This example suggests there’s a limit to the usefulness of the concept of literacy when talking about code. And yet, when we run the program, it’s revealed to be quite simple, though endlessly changing, as it creates a random maze across the screen.
So I’ll end with a caution about relying on the word literacy. It’s a word I’m deeply troubled by, loaded with historical and social baggage and it’s often misused as a gatekeeping concept, an either/or state; one is either literate or illiterate.
In my own teaching and research I’ve replaced my use of literacy with the idea of competency. I’m influenced here by the way teachers of a foreign language want their students to use language when they study abroad. They don’t use terms like literacy or fluency, they talk about competency. Because the thing with competency is, it’s highly contextualized, situated, and fluid. Competency means knowing the things that are required in order to do the other things you need to do. It’s not the same for everyone, and it varies by place, time, and circumstance.
Translating this experience to computers and writing, competency means reckoning with computation at the level appropriate for what you want to get out of it—or put into it.