Sophocles: Lost & Built

Some time ago, when AI was still nascent and with only trial versions of GPT-3 available in the OpenAI Playground (this was a few months before the worldwide release of ChatGPT) a random thought struck me. Since LLMs were so powerful and efficient at reproducing authorial style, voice, tone, verbiage, syntax, and sentence structure - all patterns that are noticeable by the human brain with enough reading experience, and easily expressible via computational linguistics - why couldn't we use one to embody a past author?

It wasn't long after ChatGPT came out at the first flood of chatbots proclaiming to be reminiscent of previous historical and authorial figures accompanied them, some with intense amounts of detail and others mimicking their subjects in name only. But the thought I had before the ChatGPT release stuck with me. What if we could, in fact, reproduce not just the authorial style, but the text itself? What if we could turn that dream into a reality? What if, through prompt engineering and design, we could "recompose" or "reconstruct" a lost text? What if we could perform an act of literary necromancy, reviving a lost text to as close to its original state as we deemed possible?

The answer to all of those questions was "yes, we can do that."

Using a novel prompt engineering methodology I developed - the same one which has found a home in multiple Fortune 500s and local businesses alike - I was able to, using GPT-4 and a correspondingly designed agent, construct the "Sophobot," an agent that writes in the diction, style, and tone of the great Greek. This process was arduous, and included a few stops and starts (notably a full agent redesign upon the release of GPT-4 and a giant series of now-useless GPT-3.5 JSON docs for training purposes that were never implemented). Eventually, however, Sophobot came to literary life, and was ready to reconstruct a text.

But what text? Sophobot was equipped to mimic the style, tone, and linguistic variations of the playwright, but could it ever compose a plot or design characters that would be faithful to the original text?

The answer here is no - not unless I wanted a series of hallucinatory "almost characters" espousing dialogue crafted out of whole cloth in a vaguely Sophoclean style. The answer, then, was to turn to extant historical sources, and the text I identified for reconstruction was "The Searchers," a satyr play which is missing significant fragments, but remains one of the best-preserved examples of Sophic fragmentary text to date.

More importantly, there was contextual data to direct me toward the play's original plot structure. A dramatization of the Homeric Hymn to Hermes, "The Searchers" documents a humorous and sometimes satirical take on the Hymn, which involves Apollo searching for his lost herd of cattle and eventually finding them stolen by the neophyte Hermes. The two come to an agreement before Zeus, a lyre is exchanged, and everyone walks away happy.

With this context in mind, including an array of characters, roles, and important plot points that must be included, the stage was set, if you'll excuse the intentional pun. I worked my way through the historical documents for accuracy and plot structure, reconstructing various parts and speaking roles and stage direction using Sophobot. The emergent text, once I iterated the Sophobot agent enough times, was powerful and effective, and resulted in a fully-reimagined play, as close to what the original Sophocles may have written as I personally believe is possible with the current state of artificial intelligence.

The final step was checking my work, and after a series of computational linguistic tests to certify similarity to the original Sophic text (including reading level, vocabulary density, sentence, syllable, and word length) I was satisfied. This was, indeed, the closest I could get to reconstructing the original play.

On November 30th of 2023, the fully-reconstructed text was presented in a live reading at the University of Pittsburgh-Johnstown #UPJ by the generous efforts of John Teacher and faculty and staff of the Theatre Arts department (photo).

It was, not to put too fine a point on it, the first time anyone had heard this text in over 2000 years. It demonstrated that with enough imagination, AI and LLMs truly unlock amazing possibilities. It also demanded the necessary next question: what text is next? What, if anything, can any longer be deemed out of our reach? How can this assist scholarship in everything from literature to history?

I have to pause here to give thanks and credit to Josh Byers at the Tribune-Democrat, who's coverage of the presentation was amazing. Thanks to him, this achievement made it to news sources both local and global. You can find his original article about the presentation here: https://www.tribdem.com/news/ai-completed-sophoclean-play-presented-at-upj/article_fabc8d3a-8fd8-11ee-9b2b-1b8f63ccbc8b.html

I've included a brief synopsis of the project in my "projects" page under my bio, but it occurs to me now that I've never described the entire thing in detail as a post, so I thought I'd do that. AI is the key that opens many doors, and I'm hopeful that you'll get inspired to construct your own project or agent after seeing just what the software is capable of.

Personally, I'm hopeful that the full process, including my data, agent design, and reconstructive efforts find a home in the pages of an academic journal at some point. In the meantime, I can't wait to see what you create!