My company Nimble Books today published the first book written using OpenAI's GPT-3 generative text engine. This was a big milestone for me personally as I have been exploring algorithmic publishing for most of the last decade. For more info, see:

Eccentric Dictionaries: An Experiment in AI-Enhanced Human Creativity (Nimble, 2021)

Write a book for Nimble using its OpenAI templates

Submit your AI-powered manuscript to Nimble, an AI-friendly publisher

Deep Dive

The following is an excerpt from the Introduction to Eccentric Dictionaries No.1.

I’ve always loved what I call “eccentric dictionaries.” Under that heading I include both standard dictionaries with serious definitions that are salted with a hefty dose of personality, like Samuel Johnson’s A Dictionary of the English Language, and humorous dictionaries with definitions not intended to be taken literally, like Ambrose Bierce’s A Devil’s Dictionary. It’s been a long-standing ambition of mine to write and publish eccentric dictionaries of my own, but there wasn’t an obvious way forward that didn’t involve spending months tediously creating thousands of database entries one at a time— until I got my hands on OpenAI’s GPT-3.

My goal was to create and publish a new eccentric dictionary in the spirit of Ambrose Bierce. I experimented with GPT-3’s davinci and instruct-davinci engines until I was able to “engineer” (write) a prompt that reliably creates random dictionary entries in the style of Bierce. I wrote a script that ran the prompt repeatedly overnight and created a big text file of results. Then I began to edit and write.

It took me about two days to go from about 1000 randomly created entries to 641 that I liked, presented here as 114 pages in a 5 x 8” chapbook format. I discarded about 30% of the initial batch for redundancy or incomprehensibility (often due to my coding errors, not GPT-3’s fault); I brought about 40% up to my standards with aggressive editing and writing; and about 30% required only light editing. A surprising portion of the total output—maybe 5-10%--struck me as brilliant and insightful. GPT-3 is an enjoyable collaborator that brings to the table a range of ideas and connections so wide as to be outside any individual human’s scope.

Because the style exemplars by Bierce were written over a period from 1881 to 1911, the GPT-3 output contains some dated and potentially offensive concepts. I removed many but left some so as to retain the period flavor. Before proceeding further, readers should realize that not everything will be compatible with modern sensibilities. Where possible, I flagged problematic entries with comments like “Obsolete” or “Dated.” I emphasize that responsibility for every word in this book is mine and mine alone.

GPT-3’s generative capabilities include the capacity to create new words that have no actual equivalent in any known language. I noted these cases as “Neologisms”. They are in fact sort of Super-Neologisms, more Neo than any ordinary neologism arising out of the normal processes of human language formation, and thus arguably even more interesting to word-lovers than regular neologisms.

GPT-3 also sometimes creates entirely plausible descriptions of facts that are not tethered to reality. Whenever there was a question in my mind as to whether what GPT-3 was saying was real, I checked a very well known search engine and added a note along the lines something along the lines of “Sadly, untrue”.

As the subtitle of the book indicates, I consider this very much an experiment. I already have thought of ways to considerably improve the diversity and quality of future prompts, outputs, and final entries. In a world of continuous software delivery, the quality of algorithmically enhanced publishing should always be improving.

As the title of the book suggests, I see this as just the first in a series, and expect to publish additional Eccentric Dictionaries on a variety of topics as tech, popular culture, and politics. There will be templates for other types of books, too.

GPT-3 is remarkable technology and there is a temptation to present the generated text output without editing as the evidence of “Artificial Intelligence.” But as a lifelong book-lover, author, and publisher, my initial encounter with the technology has led me in the opposite direction. I may change my mind in the future as capabilities evolve, but for now, this is most exciting to me as a partner for enhancing human creativity.


Fred Zimmerman
Publisher, Nimble Books LLC