A few poems

I recently had the opportunity to play around with OpenAI’s API for their GPT-3 language model. Given some natural language prompt, GPT-3 will try to create text that answers a question, continues a story, creates a list, or whatever seems to make sense in context. The flexibility and power of the model is pretty incredible.

One thing you can do with the API is create a fine-tuned version of GPT-3 based on your own dataset of example text, so that the model’s output more specifically targets a particular format or style. So after using the stock models to make the AI write me the plot of a movie about a witch who goes to prom, and after I got a few suggestions of campaign slogans for a fish who is running for president, I decided I wanted to try building a fine-tuned model for generating poetry.

Shel Silverstein seemed like a good poet to model: his style is conversational and simple, so I thought he would have good odds of producing coherent output (compared to say, E. E. Cummings). And just as importantly, his style is playful. I want to have fun with this!

After a few hours of scraping text from the internet and fiddling with the API, I had a model that I could prompt with my own made up poem title and get back a brand new poem in Shel Silverstein’s style. I’d like to share some of my favorites so far:

“A Walk In the Wind”

A walk in the wind,
A gossamer thin thread,
A flutter of a butterfly’s wings,
And years can’t mend my broken heart.
For you and I, we once were one,
But now we’re two, and I can’t go on,
‘Cause we’ve grown so far apart.
And I’ll walk in the wind
And you’ll walk in the sun.

“The Kicking Gentleman”

He was a kicking man,
They called him Kick.
He kicked the men who sold him bad oranges,
He kicked the men who laughed at him,
He kicked the men who passed him on the street,
And the men who drove him mad.
He kicked the men till finally he had enough
And then he went and kicked the whole darn town,
And everybody ran and hid.
He kicked the men till finally he had enough
And then he went and kicked himself instead—
That’s how I heard it happened.
And now here’s Kick with his story to tell—
Me, of course, being a kicking man.

“A Seal In an Overcoat”

There was a seal in an overcoat
Who walked around all tuckered out.
He was tired of swimming on dry land,
So he went to live in a wet coat.

“Anna Is So Cool”

Anna is so cool, she gets ice cream all year long.
She wears shorts in the snow and tucks her t-shirt in.
She rides bareback on the roller-coaster,
And if that scares you, she says ‘Scat!’
She walks into the school unbuttoned, unzipped,
With her dukes flapping open wide.
The boys all roar, ‘Hey! Hey! Hey!’
…And the girls all scream…’Oh!’
…And the boys all shout…’Hurray!’
…And the girls all howl…’Scat!’
Anna is so cool, she gets ice cream all year long.

“The Wizard of Loneliness”

Said the Wizard of Loneliness to the Witch of Greed
‘I’m so lonesome, I could die.’
Said the Witch of Greed, ‘Put a spell on it!’
And the Wizard of Loneliness did just that.
He put a spell on it that made him lonelier than ever,
And he says, ‘It won’t work. I’m still lonelier than ever.’
Said the Witch of Greed, ‘It’s not fair. Why is everyone else happy?’
Said the Wizard with a sneer, ‘Oh, everyone’s happy all right…’
Except him… except him… he’s the only one who isn’t!
Said the Witch with a screech, ‘That’s not true at all!’
And she put a spell on it that made him happier than ever,
Said the Wizard with surprise, ‘How did you do that?’
Said the Witch with glee, ‘I learned it from my cat.’

“No Parking”

There is no parking on the divine.
God has no need of wheels.
He has two perfectly good legs
And four perfectly good feet.
He can walk down the road
And knock on your door,
If He wants to look like that.
He can park in a minute or two,
If He wants to look like that.

“A Room Full of Elbows”

If you were the only person in the world
And you had to choose between right and left
Which would you choose?
The right hand or the left?
If you’re clever, you’ll say both.
If you’re really clever, you’ll say neither.
For if you’re the sort of person that’s always right,
You’ll soon exhaust the supply of left.
But if you’re the sort of person that’s always wrong,
You’ll find there’s no one to your right at all.

“Treasure Map”

You’ll find it at the bottom of the sea,
At the bottom of the lake,
At the bottom of the hill,
You’ll find it if you dig hard enough.
You’ll find it in a cave with bats,
You’ll find it in a mine.
You’ll find it in an empty lot,
You’ll find it on the moon.
But you can’t find it just sitting there,
It must be SEEN!

A few thoughts

There’s something uncanny to me about reading these poems. At points, they really seem to evoke some deeper meaning; poetry has that feeling that there is a transcendent truth just behind the words that prose couldn’t express. But do these poems have deeper meaning?

High school English class left me with the impression that reading poetry is about decoding some hidden message left by the author. By performing a close enough analysis of what the author is trying to say, you get to solve the puzzle of what the poem actually means. In this case, seeing as we’re just looking at draws from a conditional probability distribution, the lens of authorial intent is obviously problematic. The metaphors are, in a sense, not real. There is nothing cached behind the text, nothing waiting to be found. (Or is the original author Shel Silverstein still communicating with us through these poems, his voice and values distilled into latent variables?)

The text seems to challenge me: if this were a metaphor, what do you think it would mean? Knowing that it was created by an AI changes my experience of the poems because it shifts my interpretation to be about creating meaning, instead of analyzing meaning. I like this mode! Reading these AI generated poems, I don’t feel constrained by what they are supposed to be, because there is no supposed to be. My imagination floats between all sorts of moods and symbolism, and they are all equally true. By reading the poem, I create it.

I don’t know much about literary criticism, but I know that theorists have spent a long time debating whether the locus of meaning is in the author, the text, the reader, or society writ large. AI poetry seems like an interesting test of whatever beliefs you have here: if you saw these poems without knowing they were AI generated, how would that change your interpretation of them? What does that imply about where meaning resides?

I recently came across a review by the composer and music critic Jan Swafford of “Beethoven X”, which is an AI generated completion of Beethoven’s unfinished tenth symphony. Jan is very skeptical of the aesthetic value of AI. He pans this particular musical work as flat, but there is also a philosophical angle. He says: “Artificial intelligence can mimic art, but it can’t be expressive at it because, other than the definition of the word, it doesn’t know what expressive is… The only true, meaningful intelligence is in a body, and likewise the only true and meaningful creativity.” For Jan, the human struggle that goes into producing a work of art is an essential part of its value—a very Romanticist point of view.

I’m sympathetic to the idea that the artist’s process has aesthetic importance. But perhaps the idea of artistic process can be expanded?

The artist Sol LeWitt is known for his large scale wall drawings. Here’s an example of one of his pieces, “Wall Drawing 419”:

The wall is bordered and divided horizontally and vertically into four equal parts with a 6-inch (15 cm) black ink band. Each quarter has alternating parallel 6-inch (15 cm) bands of white and color ink bands. Upper left: gray; upper right: yellow; lower left: red; lower right: blue.

That’s it. That’s the work of art.

In a museum setting, LeWitt’s piece may be instantiated by following these instructions to produce something that looks like this:

But LeWitt’s “actual” art is the above instructions. LeWitt was a conceptual artist. He explained it: “When an artist uses a conceptual form of art, it means that all of the planning and decisions are made beforehand and the execution is a perfunctory affair. The idea becomes a machine that makes the art.”

To build a generative language model like GPT-3, you specify a model architecture, optimization functions, and training data, which combine to produce parameters, which in turn produce language. It’s like writing instructions for writing instructions for writing poetry—conceptual art squared. While there might be a tad bit more complexity involved, is a fine tuned GPT-3 really so different from LeWitt’s art?