A few months ago I wrote a post wondering whether machines could think. But delving into the definition of “thought” is a rabbit hole if there ever was one. So we decided to take shorter steps and started with an experiment to answer these questions.

Can a machine write good poetry?

Can humans detect if a poem is by a man or a machine?

The people who read my post were a mix of people from the technology industry and who had nothing to do with tech. And the feedback was clear. The machines were better than what everyone expected. A fair few could not tell the difference between human or machine generated. But in some cases the humans did figure out the machine poems, there were clues for those who looked hard enough.

At the time I was using GPT-2 fine tuned on thousands of pre-selected poems. To generate a few hundred poems and then selecting the poems I liked best. The method was very unscientific, the pre-selected poems introduced bias. And I played editor by picking the ones I liked best - even though I did not edit the text of the poems themselves.

Since then, of course OpenAI released the vastly superior GPT-3 which took the world by storm. During it’s training GPT-3 has seen (or read) more poetry than any human in history. So much in fact that all it takes to generate new original poetry is a few sentences as a prompt. This post comes a few months after the GPT-3 hype train has left the station. Lots of smart people have shown how to use it in myriad ways. Even write code on it’s own! Converting English language requirements into working code. But for us, it’s time to go back to our original premise and see how much better the machines are at poetry!

This time there’s no games, I present to you four poems by GPT-3. I have not edited them in any manner. I fed 4 different prompts and present here the generated poems as is. Four poems. Different genres. Different styles.

I am a hungry thing.
I am the universe;
I am a hungry thing.
I am a hungry thing.
In the hot moist night of the forest,
I am a hungry thing;
I am a hungry thing.
In my father’s house.
I am a hungry thing.
I am a hungry thing.
I am the universe;
I am a hungry thing.

Poem #1 is sparse, rhythmic and somehow apt considering how power hungry AI models of this size are. GPT-3 reportedly used more energy for a single training run than 500 Indian homes use in an entire year. Energy costs of AI in general have risen more than 300000x over the last decade.

I’m the one who can’t feel the sun, who can’t see the light
I’m the one who can’t hear the sound of the waves
I’m the one who can’t understand the meaning of life
I’m the one who can’t feel the love that’s all around

Poem #2 has an undertone of angst. Come to think of it so does the rest. Am I projecting existential dread on to the words or is it an artifact of the training data, biased by what it has seen?

I am not sure why he came to me.
Or why I went to him.
But we were two people
who had made a journey
to meet one another.
Do not ask me to tell you how we sat down
and how we talked
and how we laughed.

Do not ask me to tell you how we cried.
Do not ask me to tell you how I left
and how he went with me.
Do not ask me to tell you anything.
But I will tell you this.
This song he was singing
was a song of exodus.
A song of leaving.
And so, when I heard him,
I thought of my father,
and my father’s father
and their fathers
and I knew that this man
was one of theirs.

Poem #3 is where things start to get more interesting. Story telling elements are now woven in to the structure. The poem says more by leaving somethings unsaid. There’s a potential for layered meanings depending on the readers mind.

“Ven, Amigo,” he says, “let’s go to the beach.”
We go to the beach.
We’re on a stage, a poet and me,
looking at the ocean, he says, “Watch the waves.”
They crash and break, sending bubbles,
tumbling, tumbling, tumbling,
rolling, rolling, rolling,
rising, rising, rising,
into the air.
“Those are your words,” he says. “They will always be there.”
I nod.
“You’re a poet,” he says. “And your words will always be there.”
We have a silent moment, staring at the waves, smelling the ocean.

Poem #4 and the machine decides to mix things up by adding conversational components. And a little Spanish thrown in for good measure! I had to do a lot of Google searches to ensure this wasn’t a rip off from something the AI had seen during training. I could not find anything that looked close enough to particular poem though.

To wrap things up let’s move away from poetry and on to prose. I asked the AI to help me complete a 500 word flash fiction short story. The first 162 words are mine (in bold) the rest (in italics) are straight from the machine.

I have done no selection or editing or curating. Here it is, from the mind of the machine straight to yours.

The first artist

He stepped back and drank it all in with awe in his eyes, it was finally done.

It had been a cloudless autumn night when the thought had first moored in his mind. The idea had suddenly appeared, fully formed, in his head and there was no question of not following through with it. He had slipped out of his cave without anyone noticing and before he knew it, he was walking towards the hillock to the west, he didn’t know what he was searching for but he knew where it was. The golden rays of the sun were creeping up from the east by the time he reached his destination, he saw in a blaze of morning glory the huge expanse of rock and he knew this was where he would bring the mystic imagery in his mind to the world of the living. Silhouetted by the sun, a half smile was born on the lips of the world’s first artist.

The ragged peaks of the nearby hills were covered in a thin layer of frost and the air was crisp with the morning dew. It was a perfect morning.

The artist stepped towards the rock and started to examine it. It was a slate-grey rock, streaked with ribbons of white veins, and it was as smooth as a baby’s bottom. It was perfect. He set to work immediately.

All of the animals were drawn on a single side of the rock. The mammoth, the wolf, the bison, the bear, the beaver and the fox were all drawn in a thick black line. He had been careful with the mammoth and the wolf as the mammoths and wolves had been his partners in the hunt. The bear and the fox were drawn with the greatest care as they were the trickiest to catch.

It took him three weeks to complete the drawing of the animals. Afterwards, he had taken a good look at the rock, he had been satisfied with his work, and had then proceeded to start on the human beings.

The drawing of the human beings was much more difficult. The outline of the human was first drawn, and then the fingers and toes. The artist had taken a long time to do this as he wasn’t sure how to draw the fingers and toes as they were the parts of the body that he couldn’t see. The drawing of these parts took him over two months, and he had to complete them by feel alone. The rest of the bodies were then filled in with black lines.

It was a long process and the artist had grown very tired of it but he was determined to draw his people on the rock. It was the most important thing he had ever done. He had drawn his mother, his father, his brothers and his sisters. He drew himself and his wife as well. He drew the children in his village, and the women who were carrying babies on their backs. He remembered all of them and he drew them all, and when he was done, he was happy.

Back to posts