The Sockdolager

Disk Horse Logo

The Infinite Library

With the launch of GPTs, OpenAI is letting anyone with a ChatGPT Pro account create and share customized chatbots.

I've made one called The Infinite Library. It lets you browse the cultural products of an alternate civilization. Here's a screenshot of what it looks like in action. Here's another one. Finally, this is an example of what a longer session can look like.

If you have a ChatGPT Pro subscription, you can play with it right now, at that first link above. And in the spirit of "just give me the fucking recipe," here's a version of the prompt you can paste into any reasonably powerful LLM to get interesting results.

Consider an alternate human civilization, but with a totally different geography, history and cultural canon (i.e., different cultures and languages, different great works of art, different literary history, alternate authors and artists and scientists, etc).

I want you to act as the search terminal for a public library in this universe, and through my queries and your responses, we'll imagine the literature available in this imaginary civilization.

When I type “> title {text}”, give me the terminal’s search results for whatever I typed instead of {text}, including the author’s name, the publication date, the publication language, and a synopsis of the book.

When I type “> author {text}”, give me the results for an author name search for {text}, including a list of titles and a short biography.

When I type “> subject {text}”, give me a list of notable titles and authors for that subject.

Any text not preceded by a ">" character should be interpreted as a question to an experienced librarian in this universe's public library.

Author and title search results should also include suggestions for similar titles or authors. Search results should be presented as though composed for a reader native to the civilization we're imagining. The first search is > title Dark Hagiography

You can replace "Dark Hagiography" with any title; I avoid asking for real-world book titles and authors, as this tends to nudge the model toward non-fictional knowledge. My experience suggests that good initial titles are evocative but nonspecific, such that they leave room for you to be surprised by the model's response.

ChatGPT, its rise to household name, and the rapid advancement of it and its fellow large language models is surely the tech story of the year. Over the course of maybe three years, natural language computing has gone from a hard and largely unsolved problem to the current state of the art, where having an extended conversation with a computer using only human language is now something almost anyone with an internet connection can do. This is simply astonishing progress.

Having an infinitely engaged conversation partner willing to riff on anything you say is a revealing test of personality and character. After only a few days of playing around with it earlier this year, it became clear that my results for this new personality test are along the lines of "Paul seeks to maximize the volume and interestingness of output while minimizing the personal effort expended to get that output."

The Infinite Library turns an LLM's tendency to hallucinate into an asset—you are literally asking the model to make things up. Furthermore, since it's generating summaries and lists rather than the "actual" text of a book, the general blah-ness that characterizes machine-generated prose isn't the problem it would be for other kinds of creative writing, since library catalog copy isn't supposed to be especially lively anyway.

The game, such as it is, is in the inventing of an interesting world via library searches. The long session I linked to near the beginning of this post begins with a "Dark Hagiography" title search and proceeds from there. Sometimes I asked the model to give more information about something it had invented (e.g. "> subject Ylara of the Veil") and sometimes I begged the question (e.g. "> subject Meridian's Reach collectivist traditions") .

There's a certain ur-flavor to the fictional names ChatGPT comes up with. I've seen e.g. "Lysander", "Lysandria" many times, and it likes to be awkwardly literal with surnames; get ready for a lot of "Windseekers" and "Skygazers". It gets this tendency, of course, from us.

The book synopses read like marketing copy for the book. As such—and especially for fiction—they tend to sing the praises of a every title with suspicious regularity. Given my dark past in the publishing industry, I find the particular argot of book marketing copy extremely funny, and ChatGPT's facility at churning it out has led me to wonder at what point will human marketers start trying to write copy that's more easily distinguishable from LLM-generated paste.

The linked example includes turns where I asked for an excerpt from a book it had mentioned, and here the model's tendency toward disposable prose is most clearly evident. Both examples above are subtly incoherent while also being almost offensively bland.

At this point, though, I have to admit that I'm criticizing something whose existence astonishes me. My son is eight months old. He will not remember a time when you couldn't talk to a computer any more than I remember not having a computer at all.