Publishing One of the World’s First A.I. Written Books — The Poetic Prototype
And it it now also the World’s First A.I. created audiobook! As in both written and spoken by an A.I. (more on that coming soon!). See the demo:
As we all know, A.I. has long been prophesized to replace us all. One of the few domains of human society assumed to be safe from this encroachment was the Arts. Most assuredly — we all assumed — A.I. could not replicate the artistic side of humanity.
As those following the field already know, AI generated imagery has been a thing for a few years now. Sites like ArtBreeder and DeepArt let anyone make AI images. Or you can create images using nothing but code and an input image, as I did here (some coding required for that demo). Some have even sold at auction. Basically we could say that the domain is quite saturated.
But one thing that hasn’t been as well explored is AI-generated literature. Sure, in the AI community we have explored the efficacy of generating technical text using AI and modern NLP (Natural Language Processing). Famous models like GPT-3 might even be familiar to many outside of the AI community. They work fine on generating technically-proficient and grammatically correct texts. But again, those lack an artistic orientation.
Where’s the style? In the aforementioned domain of AI painting we have styles ranging from van Gogh to Renaissance Portraiture to the trippy output of DeepDream. All of varying qualities. You could run one of those algorithms 1000 times and get a different result every time. With this AI writer I wanted to explore that kind of variation, that kind of style. More technically, I wanted to explore style transfer as applied to natural language generation algorithms, but the technicalities of that are a discussion for another day.
In such a task we can take two major approaches: the high-entropy approach & the low-entropy approach. The high-entropy approach produces more statistical randomness, whereas the low-entropy approach is far more orderly. The low-entropy approach produces a corpus that is far more grammatically and thematically coherent. To put it in a relatable context, the low-entropy approach is kind of like taking a photograph and putting an Instagram filter over it. It might be slightly different from a real photograph, but chances are a lot of people will look at it and not even know that it’s edited. To the contrary, the high-entropy approach is like taking a photograph and turning it into a full painting. It might be something relatively subtle like a Renaissance-style portrait or it might be a fully impressionist van Gogh style portrait, but nobody is going to confuse it for a photograph.
With The Poetic Prototype I took the high-entropy approach, making a more creative AI.
The inspiration (aka the training dataset) for this AI was 18th & 19th century naturalist poetry by a variety of authors. Some other similar styles of the period were thrown in to give the AI a bit more to work with. It should be noted that spelling and grammatical conventions were different at this time. This style of poetry also tends to take more artistic liberties than both the technical writings and longer-form literature of its time. The AI is more familiar with the writing conventions of these poems than it is with the conventions of modern times.
I felt relatively confident that a vanishingly minute number of readers would enjoy reading 500 pages of rather matter-of-fact descriptions of trees, the sky, birds, and the like. That’s what the low-entropy models wrote. So I decided to just turn up the high-entropy dial all the way. Personally I enjoyed this writing much more. Sure, it gave plenty of grammatical errors and even totally nonsensical descriptions. But I would be willing to wager that few modern readers of this style of poetry read it to learn basic grammar or even for any sort of narrative structure. Sure, some might read it for a supposed message. But lets have a look at well-known author Samuel Taylor Coleridge’s poem Sonnet: To the River Otter (this work and many others by Coleridge were part of the dataset used to create the AI):
Dear native Brook! wild Streamlet of the West!
How many various-fated years have past,
What happy and what mournful hours, since last
I skimm’d the smooth thin stone along thy breast,
Numbering its light leaps! Yet so deep imprest
Sink the sweet scenes of childhood, that mine eyes
I never shut amid the sunny ray,
But straight with all their tints thy waters rise,
Thy crossing plank, thy marge with willows grey,
And bedded sand that vein’d with various dyes
Gleam’d through thy bright transparence! On my way,
Visions of childhood! oft have ye beguil’d
Lone manhood’s cares, yet waking fondest sighs:
Ah! that once more I were a careless child!
Did you get a message? Maybe an appreciation for otters? Maybe you just enjoyed the rhythm or the relaxing nature of the imagery? That’s what I wanted to do with this AI: create a heavily stylistic writing that is more enjoyable than it is technically-proficient or thematically benevolent. I wanted to create the van Gogh portrait rather than the photo filter, except with writing style.
Don’t worry I’ll publish some of the more technical ones soon. But I think most readers, particularly avid readers of similar styles of poetry, will enjoy the novel nature of the present AI.
When reading the book you’ll most certainly notice some of these alluded-to grammatical errors and even some non-sensical statements. Lets look at the following from the opening lines of Chapter 3:
With dreams of sorrow come,
And so clouds not fears his land of bright so spirit place,
The shadows so have beautiful the sight
Fall in the sense of stone
Then he gaint of the wiser and the ship
Is trouble the links of all most beauty seems the light
Of all the last saddess the more strike a spectering more
And what clouds and I steep in sleep.
Most assuredly it is quite the novel scene. I would most certainly recommend reading this aloud to get a full appreciation for the AI’s focus on rhythmic structure, alliteration around the letter ‘s’ and pacing. Further try reading it exactly as you think 18th-19th century poetry should sound. Then try reading it aloud in that way to someone who has no clue that it was written by an AI. Can they tell that something is up? Do they enjoy the sound of it? My bet would be no and yes, respectively.
One interesting thing to note is the manner in which it makes spelling mistakes or mentions. Some are artifacts of the training dataset. For instance, we can see that Samuel Coleridge adds ‘t’ and ‘d’ to certain words to enhance the pronunciation and expression of this words. The AI can be seen doing the same, turning ‘gain’ into ‘gaint’ in the 5th line. Yet the misspelling ‘saddess’ (which isn’t a word in this period’s English) seems to be made to play up the local ‘s’ alliteration. Moreover saddess rhymes more fluidly with the preceding word last (try saying ‘last saddest’ and ‘last saddess’ aloud). The AI very much seems to prefer rhythmic structure and rhythmic flow to technical proficiency, a style not dissimilar to that of the authors that inspired the AI.
At the very least, it is certainly interesting to read something highly entropic yet still strongly focused on poetic and rhythmic structure. It is certainly something you won’t find anywhere else, which is kind of the point. I’ll publish more technically-focused A.I. writings in the future, but certainly do let me know if you enjoy this style of this one!