At one point in the distant future, long after our sun has faded and its planets have passed, Pantography will produce the text, " in a village in la mancha, the name of which i cannot quite recalk". 1 One hour later, right on schedule, it will come out with a line that is indistinguishable from the opening words of the Penguin edition 2. If, upon seeing it, you were to hurry off to your local Waterstone’s, you would get nothing new from it other than a few capital letters.
Pantography takes brute force to the alphabet and generates one text after another. Although each one is born in the most banal way possible, it will eventually have amassed billions upon billions of texts that will be impossible to tell apart 3 from those written by human authors. It would also be possible, of course, to modify Pantography’s algorithm to generate only texts which are grammatically, syntactically, lexicographically and orthographically coherent by current understanding. 4 The entire body of texts thus produced could then be mistaken for one produced by someone with an extraordinary amount of time on their hands, some sort of prisoner of eternity.
There are other ways of getting a machine to produce texts, without using the indiscriminate strategy of Pantography’s incremental method. Quill, for instance, is a software program developed by Narrative Science which analyses a set of data and writes up a nifty little paragraph about its most significant points. 5 So far it is mostly used to report on regional sports events, real estate trends and stock market prospects. It comes up with little gems such as:
Wall Street is expecting higher profit for Celgene when the company reports its second quarter results on Thursday, July 25, 2013. The consensus estimate is calling for profit of $1.44 a share, a rise from $1.22 per share a year ago. For the fiscal year, analysts are expecting earnings of $5.77 per share. 6
Perfunctory, yes. But surely no worse than the following excerpt from Reuters, written, one presumes, by a human being:
Celgene Corp announced that for fiscal 2013, it raises adjusted diluted EPS guidance to a range of $5.80-$5.90 from the previous range of $5.55 to $5.65, an increase of approximately 19% over fiscal 2012 adjusted diluted EPS. GAAP diluted EPS is expected to be in the range of $4.17-$4.31 which includes upfront payment expense for collaborations. 7
Quill operates within a narrowly defined set of data and, so far, produces texts in a genre where standardised formulations are the norm. But it is already possible to see how its specialisations and applications will become ever broader. It’s not impossible to imagine it filing a perfectly serviceable analysis of an election result some years from now. Kristian Hammond, CTO at Narrative Science, has no lack of ambition for his child, predicting that it will win a Pulitzer by 2017 and that by 2027, “more than 90%" of news will be written by computers. 8
Hammond probably has in mind journalism with a more or less straightforward methodology: a piece of text that begins with a body of data—both foreground and background—identifies its most relevant aspects, contextualises it with some historical comparisons and then strings together a story. But it would be foolish to think that the same method cannot be applied to work that is normally assumed requires more imagination. As it is, Quill can probably write a perfectly readable chosistic novel in the style of Alain Robbe-Grillet. Or maybe the kind of coldly descriptive prose that Robert Musil was going for in the opening paragraph of The Man Without Qualities. Raymond Roussell’s procédé is an algorithm waiting to be automated. All writing begins with a body of data, at the very least the author’s own experience of the world, and text is a way of generalising and summarising the data. 9 Metaphor, irony, pathos, humour—these can all be “programmed in” 10. After all, they are rhetorical methods whose mechanics have been studied at least as far back as the Greeks.
In his essay Cybernetics and Ghosts 11, Italo Calvino, himself a collector and editor of Italian folk tales, welcomed the structuralists’ decipherment of the constitutive elements of myth that would eventually give us a sort of alphabet to work from. A distillation of narrative into its most basic buildings blocks, out of which any other story could be generated. 12 This sounds a lot like the “Good Old-Fashioned Artificial Intelligence” of symbolic manipulation that Hubert Dreyfus and others railed against in the 1970s, and the fashion has since moved on to a more heuristic approach that, rather than following an a priori procedure, learns to draw approximate inferences from given data. It is not impossible to imagine such a heuristic program using an enormous body of existing texts to learn how language is used and then have a go itself like a bright adolescent spending an uneventful summer omnivorously going through the local library and then, with the first rains of September, sitting down to write her first short story.
At this point, many would take consolation in re-categorisation, in finding ever more remote trap-doors and priest-holes where the Human Spirit can hide. They will console themselves with the notion there will always be something quintessentially human that machine writing cannot reproduce. This used to be the received wisdom about computer chess programs. It was believed that a grandmaster’s wit and ingenuity, his humanity, will always ensure that a computer will never ultimately beat the best human opponents 13. And then the Human Era of chess ended on 11 May 1997 when IBM’s Deep Blue won the second six-game match against Garry Kasparov. In the predictable acrimony that ensued, Kasparov accused IBM of cheating, on the grounds that he had detected creativity in the machine’s play—surely, he declared, a sign of human intervention! 14 That Wired article on Narrative Science, in its final paragraphs, sought to reassure panicking humanists that any Pulitzer on offer would be given not to the algorithm itself, of course, but to its coders. While one can certainly imagine an application that would be used by human authors to write fiction algorithmically, just as some musicians use SuperCollider to create algorithmic music 15 there is no reason to believe that there will never be a successor of Quill that is so autonomous that its output will be as attributable to its coders as The Handmaid’s Tale is to the programmers of Microsoft Word, or The Charterhouse of Parma to a duck’s arse.
Calvino, on the other hand, cheerfully welcomed the prospect of machines taking over the task of writing—precisely because then, " the decisive moment of literary life will be that of reading". But Calvino was flattering us, thinking us as humble and sophisticated as himself. 16 He describes us all as if we have long been impatient to be rid of this distasteful burden of “self-expression”, to devote ourselves fully, like consecrated monks, to the one role in literature that has humility and dignity, that of the reader. That we have longed to abandon the kitchen for the dining room, gracious guests at the magnificent table of literature.
As things stand, we are still stoically peeling onions among the gurgling pots in the kitchen. Several million books are published every year. These are books that real live human beings conceived of, researched, planned, discussed, and then typed and re-typed with their own hands, often for little or no reward. Quill, and all the other machine-writing programs that will succeed it, are solutions for a problem that does not exist. Whatever the motivation, the production of texts is clearly a task that we are very happy to do for ourselves. And even if we ever do get fully serviceable writing machines, one suspects that after a brief fad of automation, we will want to return to hand-making our texts. Just as we did with mashed potatoes. 17
If this interest in artisanal writing continues to grow at its current pace 18 it will become ever harder for writers to find readers prepared to read their texts, or even to find any literate person who is not so occupied with their own writing that they have any time for reading. 19 More and more texts will die the unconsummated death of the slush-pile. We don’t need a machine for writing—what we need is a machine for reading.
When Friedrich Nietzsche first obtained a typewriter, he was overwhelmed by the speed with which he could now produce texts. Or rather, he found that his friends, his initial readers, were overwhelmed: “What I really need is a machine that can read what I produce.” 20
It’s hard to imagine what Nietzsche could have thought such a machine would look like. He had just taken delivery of an upright Danish typewriter, made of clunky keys and springs. 21 Surely he couldn’t have thought that by a different, more ingenious arrangement of these mechanical parts, a machine for reading could be devised?
In fact, even with our vastly greater familiarity with machines that seem to perform uncanny tasks, that can place us anywhere in the world and carry our voice and image dutifully across oceans, it remains almost impossible, even on the level of a thought experiment, to imagine what a reading machine would look like.
We can imagine a writing machine simply by extrapolating from current technologies, by imagining machines that become ever subtler, ever more complex until their methods are impossible to distinguish from the normal artifice of authorship. Writing can be faked; because once the text is produced we are able to take it at face value. It is often irresistible to infer the mind of the author from what we read but that is all post hoc. Language can be produced even accidentally, and carried naively. 22 After all, a whole realm of meaning is produced by misspelling, mispronunciation and malapropism.
But it is impossible to imagine a reading machine. After all, we already have simple writing machines but we don’t have equivalent machines that can read these simple texts. Except in the Chinese Room sense that machines read data from each other, that our computer reads data from a memory stick.
Or, rather, it is impossible to imagine a reading machine without imagining a fully conscious machine. We can perhaps imagine a sort of network whose “reading” feeds back into life, through which we can influence the world. Or one that, having read our texts, goes on to produce further texts under their influence. But can we imagine a machine that gives us authorial fulfilment when it reads our texts?
In fact, the reason why we don’t consider Quill’s texts to have been “authored” but rather “produced” or, at best, “written,” is because these machines, and their foreseeable successors, can’t read what they’re writing. 23 It’s not primarily their shortcomings as writers that disqualify them, but their shortcomings as readers, their lack of subjectivity as readers. Maybe, when we do have writing machines, we will start treating the provenance of the texts to be as mystical, as incorporeal, as the fruits of the muses once were. 24 Indeed, if we assume the machines producing the texts to be merely naive carriers then it would be very hard to not infer, however wrongly, a deeper source, a truer origin, an “intelligent designer.”
But the compelling need for just such a reading machine will impel us to build one. First we will have to imagine it. Time has taught us that tasks that seem uniquely human tend to appear more mechanistic once we have learnt to automate them. Nietzsche spoke of his typewriter anthropomorphically, whereas to us it appears as a merely mechanical device. The feat of beating a Chess grandmaster was considered a uniquely human ability until we managed to reduce it to a heuristic process. The mysterious act of translation is carried out with increasing competence using probablistic methods. Maybe we will manage to create a reading machine, and it will appear as banal a task as geolocation. We continue to mechanise tasks that we once considered “intelligent,” and redefine intelligence as something ever more elusive. 25
As with most other artificial-intelligence projects, a Mechanical Turk could serve as an interim solution. Large buildings in our post-industrial towns and in developing economies could be given over to deskfuls of hired readers poring over texts written by more affluent people, elsewhere in the world, people with time, attractive psychological complexities and issues with their mother.
This Mechanical Turk and its fully automated successors could become a built-in part of our word-processors such that once we have written a text, it is also instantly read and its entire life-cycle—from conception to production to consummation—happens instantaneously. It could be plugged directly into Creative Writing departments. 26 And, of course, it would become irresistible to plug reading machines directly into writing machines.
But would the two separate applications still exist? 27 Quote article about robots reading robots 28 Not strictly parallel, of course. Those robots did not need the narrative sugar sprinkled on their data.
Reading machines can be connected to writing machines and it would short-circuit meaning. But would machines need this human detour of narrative? 29
Is narrative a human detour or could they be created - and is it only a human detour. 30 Is this the plane of our existence; machines connected to each other would have gnosis. 31 Our plane of existence is just that plane of tentativity, of makeshift approximations, of language. Narrative is the playing out of information over a human-scale timeline.
- 1An English language example would be simpler, though we would then lose the allusion to Pierre Menard.
- 2Unless, of course, in the meantime Penguin sees fit to commission another translation.
- 3Cf. Flusser in Into the Universe of Technical Images.
- 4Although, as far as Pantography itself goes, this would be against its spirit. See The Fallacy of Noise.
- 5 Quakebot is another example.
- 6 forbes.com
- 7 reuters.com
- 8See Can an Algorithm Write a Better News Story Than a Human Reporter? a -better-news-story-than- a -human-reporter
- 9And it is in this sense that Calvino, in his essay, concludes that we already have Literature Machines that process data and turn them into texts, they’re made of flesh and blood and neurons and synapses and we call them authors. A bit of a Humanist cop-out, I feel.
- 10Cf Joke-bot, via On the Media, “I like my notes like I like my fingers, sticky.”
- 11Originally a speech delivered in 1967. See Calvino 1997.
- 12Cf. Ramon Llull, Leibniz.
- 13Citation needed.
- 14Tangentially, in Zweig 2011, the protagonist is a dullard whose only interest is chess, and the money he makes from it. There is no wit to his game, just a rote implementation of memorised games, he brings no human experience to the table. Mention, here or elsewhere, that Deep Blue was a computer that relied on specialist hardware whereas contemporary chess programs use heuristic rather than procedural, algorithmic methods. Much like the current generation of machine translation.
- 15Do pseudo-code in novels.
- 16Cf. Borges being more proud of what he’s read than what he’s written. Benabou?
- 17Cf. Cadbury Smash commercial.
- 18Do graph, also use for The Reader. Find sweet-spot. In fact, this essay obviously has a lot in common with it. Don’t overlap too much.
- 19The secondary act of language becomes writing, rather than reading. If once we spoke and read, now we speak and write.
- 20TRACE quote, via Kittler.
- 21Kittler points out that, like other early typewriters, like the Remington I for instance, the typist could not see what they were writing. This was a brief period when reading and writing were divorced; though it remains the definition of a “touch typist” that one is able to type without looking at the text being produced.
- 22That, indeed, is the whole point of language, that thought and meaning can be stored in and carried through even inanimate objects.
- 23Cf. Mead and reflexive language. Aboulafia 2006, p. 11, and Kittler re early typewriters that hid the produced text.
- 24Wired’s attribution of a Pulitzer to the coders, is possibly part of its anti-Singularity agenda. Or just a nifty winding-up paragraph.
- 25Turing’s definition of intelligence seems extremely permissive to us now that we have very effective chatbots. But Turing’s point was precisely that we ascribe intelligence to anything that convincingly achieves a task we consider—or previously considered—to require intelligence.
- 26More applications.
- 27Machines reading machines. See article where it is suggested that this is exactly what’s happening with financial data.
- 28In financial world.
- 29Meaning would be shortcircuited in the way that automated trading stock makes a mockery of the market. This is relevant -the stock market represents human whims and desires, at a human speed.
- 30Cf: “Time is nature’s way to keep everything from happening at once.” See also: Pantography: The Fallacy of Further Brevity.
- 31This gedankenexperiment gives us a (rare) glimpse of a universe outside of language, we see electronic gnosis hurtling past the deliberate impediment of human language. Language exists to stop everything from being known at once. (Cf. also Foucault)