Humane Ingenuity 9: GPT-2 and You
by Dan Cohen
Carlotta Corpron (1901-1987) was an artist and photographer who used light to create abstract works from which figures would sometimes emerge. (“Strange Creature of Light,” 1948, via the International Center for Photography.)
This newsletter has not been written by a GPT-2 text generator, but you can now find a lot of artificially created text that has been. For those not familiar with GPT-2, it is, according to its creators OpenAI (a socially conscious artificial intelligence lab overseen by a nonprofit entity), “a large-scale unsupervised language model which generates coherent paragraphs of text.” Think of it as a computer that has consumed so much text that it’s very good at figuring out which words are likely to follow other words, and when strung together, these words create fairly coherent sentences and paragraphs that are plausible continuations of any initial (or “seed”) text.
If you are new to GPT-2, the best way to see it in action is on Adam King’s “Talk to Transformer” website, which just implemented the full-sized GPT-2 model, and allows you to prompt it with any text you wish. I put in the first two sentences of Jane Austen’s Sense and Sensibility (in bold) and the computer spat out the rest:
The family of Dashwood had long been settled in Sussex. Their estate was large, and their residence was at Norland Park, in the centre of their property, where, for many generations, they had lived in so respectable a manner as to engage the general good opinion of their surrounding acquaintance. They were also well known for their hospitality and philanthropy, especially for their charitable donations of meat and milk to the poor. The Dashwoods were not, in fact, known to have any money, but had an excellent credit with all classes, and were well looked upon by their neighbors. The Dashwoods were a most benevolent family. Their land was large, and their possessions were in good order, and the house well kept. They had four children, namely, John, Sarah, John and Mary. These four children were all educated, and all well known for their benevolence and philanthropy, so that their name is familiar in…
You get the picture. If a bit dull, it’s still rather impressive, and certainly a major advance over prior text generators of this sort. (Austen would have had a field day picking up this thread and writing about two Johns fighting over their common name and potential spouses.) The web is now filled with examples using GPT-2, including the New Yorker‘s computational autocomplete of paragraphs in its article on predictive text (how meta).
The most interesting examples have been the weird ones (cf. HI7), where the language model has been trained on narrower, more colorful sets of texts, and then sparked with creative prompts. Archaeologist Shawn Graham, who is working on a book I’d like to preorder right now, An Enchantment of Digital Archaeology: Raising the Dead with Agent Based Models, Archaeogaming, and Artificial Intelligence, fed GPT-2 the works of the English Egyptologist Flinders Petrie (1853-1942) and then resurrected him at the command line for a conversation about his work. Robin Sloan had similar good fun this summer with a focus on fantasy quests, and helpfully documented how he did it.
OpenAI worried earlier in this year that GPT-2 might become a troubling anarchy loosed upon the world, and while we surely would like to avoid that, it’s not what I want to focus on in this issue of the newsletter. (If you are concerned about GPT-2 and devious trickery, please note that other AI researchers are working on countervailing tools to identify “fake” text, by using GPT-2’s strength—its statistical reliance on the common words that follow other words—against it in a nifty jiu-jitsu move.)
I’m actually less interested in whether GPT-2 has achieved some kind of evil genius, and more interested in what is really happening in the interaction between its generated texts and the reader. Before we worry about GPT-2’s level of intelligence, we should remember what occurs during the act of reading, and why we appreciate fiction in the first place. And that has much more to do with the activity in our own minds than the mind of the author, however real or fake.
We bring to bear on any text all of our prior experience and emotions, as well as everything we have read and thought. We complete a text, no matter how coherent it is; we fill in any blanks with what we believe should be there, or through our imagination. We ourselves are a preprocessed, mammoth, unique corpus, a special composite lens that colors what our senses encounter.
From this perspective, GPT-2 says less about artificial intelligence and more about how human intelligence is constantly looking for, and accepting of, stereotypical narrative genres, and how our mind always wants to make sense of any text it encounters, no matter how odd. Reflecting on that process can be the source of helpful self-awareness—about our past and present views and inclinations—and also, some significant enjoyment as our minds spin stories well beyond the thrown-together words on a page or screen.
[Carlotta Corpron, “Light Creates Bird Symbols”]
My thanks to the HI subscribers who responded to my question about whether I should include some discussion of our library renovation in this space. The unanimous sentiment was yes, so I’ll drop some bits in here from time to time. (I have also discovered that two members of this list are also directors of large university libraries that are beginning renovations; helpful to compare notes!)
One complicated topic for us, and undoubtedly others, has been flexible space for focused study, collaboration, and creative production. Because the square footage of libraries is finite (with the obvious exception of Borgesian libraries), you often need to design spaces for multiple uses, especially in a library that (like ours) is open 24/7 and thus goes through different cycles of use as day turns into night and back again.
I have looked at a number of flex spaces in libraries, and I’m not sure that we have solved for this problem. Many approaches seem a bit immature, with an overreliance on movable furniture. Some of the more interesting conversations I’ve had with architects recently have focused not on physical elements like furniture but on more ethereal elements like shifting acoustic design and phased lighting. If you know of a space—in a library or anywhere else—that has worked well for different uses, I’d love to see it.
[Carlotta Carpron, “Patterns in a Glass Cube”]