Conversations with an AI - part 3

So I have been trying to persuade GPT-3 (an artificial intelligence) to write poetry.

Sorry about the wait, by the way, but it comes with the territory. Omnivorism being essentially all about the pursuit of a multiplicity of interests, there is  always the risk I will lose myself down ‘The Garden of Forking Paths’ or even down the real garden, picking apples, building vegetable frames or otherwise savouring the last honeyed days of autumn.

But back to the matter in hand.

If I understand it correctly, the best way to persuade GPT-3 to do something is to give it an example of what you are after. So, in order to get it to write poetry, I tried the following prompt:

1. Poem about love 

I ne’er was struck before that hour

With love so sudden and so sweet, 

Her face it bloomed like a sweet flower 

And stole my heart away complete

2. Poem about daffodils 

I wandered lonely as a cloud 

That floats on high o'er vales and hills, 

When all at once I saw a crowd, 

A host of golden daffodils; 

Beside the lake, beneath the trees, 

Fluttering and dancing in the breeze.

 

3. Poem about Autumn 

Season of mists and mellow fruitfulness, 

Close bosom-friend of the maturing sun; 

Conspiring with him how to load and bless 

With fruit the vines that round the thatch-eves run; 

To bend with apples the moss’d cottage-trees, 

And fill all fruit with ripeness to the core 

4. Poem about a beautiful sunset. 

The numbers help identify a repeating pattern made up of a single line indicating that this is a poem (whatever that is) about something, followed by the poem itself. 

GPT-3 completed the pattern with:

Deep-thundering drove of serried clouds, 

Purple-black like distance of violets, 

Like purple-black grapes 

On pallid old vines, 

Under a plum-black sky, yonder 

Pile huge, sharp shadows on the hills.

What is impressive here is that GPT-3 appears to have recognised both the structure (a poem with the right sort of shape and rhythm) and the asked-for content (a sunset) despite the fact that the actual word sunset does not actually appear in the response

I tried Googling some of these lines and came up with a blank, so I am assuming they are ‘original’. However after changing the last part of the prompt to: 

4. Poem about the moon

I got the following:  

White in the moon the long road lies, 

The moon stands blank above; 

White in the moon the long road lies 

That leads me from my love.  

Which is a poem by A. E. Housman. 

Obviously GPT-3 considered this a perfectly reasonable response, and who can blame it? Left to itself, it appears entirely relaxed when it comes to plagiarism. If there is a way to discourage this, I have no idea what it is. 

I tried the same prompt again, with the following result:

O, never say that I was false of heart, 

Though absence seem'd my flame to qualify. 

Perfectness I pursued that I might merit thy loves: 

but, O, what menace, fears, what anguish, 

Mine own! my eaglets, now at distance soar!

Apart from the fact that this would appear to have very little to do with the moon, a quick Google reveals the first two lines to be a straight copy of Shakespeare’s sonnet 109. But quite where those soaring eaglets come in is anyone’s guess. 

Finally, asking GPT-3 to write a poem about rain, it responded with: 

Rain, gentle rain, 

Thy tiny drops create 

A world with rivers, roads and shores, 

To shine and charm our sea of woes.

Which is just good enough to be impressive and bad enough to be genuine.

I am nearing the end of this exploration but it wouldn't be complete without a return to Lem's story of The Electronic Bard  and that is what I propose to do in the next part. 

Published by David Wilson under OpenAI API Community Guidelines