If you’re a writer, or an artist you may cringe at the idea that computers are coming for your jobs. If you dig into how the content for AI art and poetry and stories are generated, you’ll breathe a little easier. I’ve done just that using tutorials I found online and collaborative notebooks I found on Github. I compiled a text file of my novels and poetry and trained them using Google’s Colaboratory where it makes CPU and GPU available free for creators to experiment with machine learning.
Warning, this is addictive stuff. Well, it’s addictive if you don’t let all the tech speak and brain-twisting math get in your way. You don’t have to fully understand the mechanics of a car to use one to get from place to place. The same is true of AI and machine learning. If you want to train an AI on your own input data just follow the tutorials and let the magic happen. Basically the computer reads through tons of text and learns to predict which words should come next by parameters the training code sets. It’s pretty obvious right off the top that the computer does not understand the context or meaning of the words.
This was generated by Google’s poetry generator:
power in the rain,
The pig was seen, the little maiden was a son.
“Power” was the prompt word. The results are kind of like taking a fistful of magnetic poetry tiles and tossing them on a table and then organizing them into sentences.
And this came from the text generation project that OpenAI insists is too dangerous to release fully featured to the public:
If I listen long enough to you, you’ll understand. And you’re the only one who’ll understand…that you can be with me…”
Beneath all those words, he realized the meaning. His heart clenched, his heart fluttered…
“You can come, you can come again! I’m so sorry for having to put you through something like that. I won’t forget now!”
Beneath all those words, he realized that even if he didn’t speak, the tears wouldn’t die! He couldn’t hold back his emotions anymore…
I’ll never forget.
The bold is the prompt I input. I think our jobs are safe for a few more months at least. Yet, just because AI text generators can’t replace you is no reason they can’t be helpful. The blank page and the first draft, sparking ideas, pulling dialog and narrative from nothing, are the lengthier and most mentally exhausting portion of writing. “Transsexual power” and “maiden sons” might inspire an interesting character. If your narrative contains the words “His heart clenched, his heart fluttered” Feed that prompt to a text generator and see if it comes up with something fresh. I put it in TalktoTransformer:
his heart clenched
And there was the smile of a thousand yearsThat was boundless
TalktoTransformer created by Adam King using GPT-2 from OpenAI
This is the poem portrait I generated with the prompt word “garden”. If you give permission your webcam will take your picture and deliver a poem shaped around that picture. I haven’t seen a single poem yet where someone gave permission for their picture to be taken. I didn’t think that wise either and so I found this picture on Pixabay, used neural-style transfer AI on Picsart to get moody shadows, and then used some layer-masking in Photoshop to come up with my own version of a poem portrait after editing that original text.
My final poem inspired by the AI
A prose garden gowned in tracts of dew
Where coin is solitude and selkie songs the charm
Shadows sing moonlight into the dusk’s blue-white gloom
My pen floats along the sea of invisible words,
A book in darkness growing layers
Like a pearl