The Text Singularity

How artificial intelligence will shape the future of writing


There are several versions of the same cartoon: A 10th grader sits at her desk at 9 p.m. on the night before a big essay is due. She puts her hand on her dictionary and says, "Okay, all the words are right there. I just need to put them in the right order."

It's funny because putting the words in order, right or not, is "writing." And writing is hard, and important. Or at least it was. It may no longer be, after the text singularity.

Scientifically, a singularity is a boundary beyond which the known physical rules no longer apply, or where the curvature of space-time is no longer defined. If such boundaries exist, we cannot imagine or understand them.

A new kind of singularity—the "technological singularity" described by Verner Vinge—is easy to imagine, though hard to understand. One example would be "the text singularity," using the fourth generation of generative pre-trained transformer (GPT4) technology that will likely drop this year.

GPT4 are artificial intelligence (A.I.) software that produce grammatically ordered text. Such chatbots are often mocked because they write stilted prose. But what if chatbots wrote a lot of things? In fact, what if chatbots wrote everything: all possible sequences of words, in all languages, in all formats, from short haikus to Remembrance of Things Past? It's not a conceptually challenging problem, though it would take substantial developments in text generation and storage to produce that many sentences.

By the end of this year, GPT4 chatbots will be able to produce, in less time than it takes to read this sentence, millions of texts on all the topics that you can think of, as well as those no one has ever thought of. Most of those "texts" would be nonsense, of course, but the advantage of GPT4 technology is that it could create all the text, of every length. That means that along with all the nonsense, we would have all the best texts, also.

The problem would be equivalent of the 10th grader's dictionary, just one step further along. "All the words are right there," but we would need some way of choosing among the library of trillions and trillions of texts to find the one that serves our needs.

That is not a very conceptually challenging problem, either. We have a substantial corpus of texts, dating from antiquity to five minutes ago, that we consider good, useful, or entertaining, and thus worth publishing. Those texts give us an obvious training set for a selection process, enabling an enormous profusion of A.I. entities operating in parallel to prune the set of all possible texts to the (much smaller, but still enormous) set of possibly useful texts. Let's call those "cullbots," because they cull or prune the set of all possible texts to a much smaller set of possibly useful texts, based on the features of existing texts that humans have decided are worth keeping around.

These first two steps—creating and storing the set of all possible texts and then culling that set using "learned" features of existing text—are conceptually simple. Though computationally intensive, both are finite, well-defined tasks. The resulting corpus will not be all possible word sequences, most of which would be nonsense, but a much smaller set of texts where word sequences form sentences and coherent "thoughts," as judged by cullbots.

If the process of creation can be replicated indefinitely, with a selection filter, we will have arrived at J.L. Borges' idea of a Library of Babel, except that the library will contain all texts of all lengths and styles, stored on servers instead of a physical library. 

The feedback loop would then be closed by repeatedly rating the texts, first at a gross level, and then at a decentralized personal level. The texts that attract the most citations and views, from the following generation, and the one after that, get higher status in searches that return the "best" texts, as the selection process iteratively culls the dross.

With many GPT4 chatbots producing text constantly, and cullbots pruning the corpus of text constantly, there are no humans involved at all—except eventually as readers. A "generation" in this process might be a day at first, then a few seconds, and then a small fraction of a blink of an eye. Learning and updating becomes faster, and more text becomes available as the training set. There is no reason to wait for anything to be published—and "published" doesn't mean printed in paper anyway; it means posted on the internet. The process would spin off on its own, dynamically updating itself with only high-level human supervision.

That's when we hit the singularity. Remember, a singularity in this context means passing through an event horizon the other side of which suspends the rules as humans know them. Worse, humans cannot imagine, by definition, what the new rules will be, or if there are any "rules" at all. Finally, the event horizon is one-way: once crossed, it closes, at least from the perspective of those who have crossed it.

The write-publish-cite/write-publish cycle is already accelerating. All that needs to happen is for the cycle to become independent, relying only on A.I. entities, and the singularity will spin up. In a short period of time, by historical standards, all the things that have not yet been written will be written. All the things that never would have been written at all, at least not by human authors, will be written. And cullbots will suggest which of that corpus might interest you on a Sunday afternoon.

It's a singularity because writing is finished, forever. There cannot possibly be human writers, because we will have stored all the texts that are possible: Nothing will remain unwritten. One qualification might be that writing will survive as a boutique skill, like a home-cooked meal: The food would have been better in a restaurant, but look how much I love you, dear! Handwriting and extemporaneous words in a thank-you note would then be doubly retro. A.I.s have already noticed, producing what looks like handwriting and personalized notes.

One might object that cullbots can't possibly judge good writing. And that's right, but all the cullbots will do at first is prune text that is meaningless. (Do we include Finnegans Wake in the training set?)

What we do with the host of texts that survive the first rounds of cullbot pruning is up to us. This may seem like an impossibly complex problem, but we are used to this kind of selection process. It's just that until now the selection of the "canon" or "popular writing" has operated on a human time scale. The great theater of Greece: Were playwrights just better? No, we selected—unintentionally, by preserving some of them—the "best." Why are the greatest hits from the 1960s, '70s, and '80s so popular on the radio? A selected sample from the best of any decade excels any random sample of current music.

The personal curation portion of the process to come should already be familiar to you. There has been an explosion in the quantity and variety of musical choices available, so much that you couldn't possibly choose. But Spotify suggests some songs or albums, and over a surprisingly short time you can train the A.I. to "know" your tastes. If you have streamed a video, Netflix is ready to suggest another video that that A.I. "thinks" you will enjoy. Perhaps most interestingly, the TikTok A.I. trains itself by showing you essentially random videos, without you making any choices at all, and then "learns" from what you watch which other videos might be desirable. In all three cases, a gigantic unorganized mass of material is ordered and curated, with no human agency at all.

So what will our 10th grader do instead of writing an essay that night? There's no way of knowing, because what things will look like on the other side of the singularity can't be predicted, or even understood, until we go through it. But she won't be writing, because there will be nothing left to write.