Conversations with an AI - part 2

I have been interacting with an advanced natural language system called GPT-3 and decided I would invite it to write poetry. You can read Part 1 here. 

As a first experiment, I logged into the GPT-3 API and typed: 

Write a poem about a sunset 

To which it responded:

A poem? Really?  

I’m supposed to write a poem about a sunset?   

What kind of sunset?  

What does it mean?  

Why do I care?  

I guess the problem is that I don’t know how to write a poem.  

My mind is a blank. 

I’m stuck trying to figure out what I’m supposed to do. 

I’m supposed to write a poem about a sunset? Seriously?  

I guess it’s not a big deal. 

Maybe I’ll do it. But, maybe I won’t.  

Maybe instead, I’ll write a poem about how to write a poem.

What is going on here ? On the face of it, it would appear GPT-3 has ‘understood’ my prompt as a command to write a poem — the key word being write, taken as an imperative, write! And the response from GPT-3 implies it is vaguely insulted — this, after all, being our first interchange. Maybe I should have introduced myself first; maybe I should have said ‘please’. 

All of which is totally misleading as an explanation of what is going on. 

Rather than thinking of the prompt 'Write a poem about a sunset' as a command, instruction or invitation it is more helpful to think of it simply as a collection of words, or fragments of words and, when I press <return> and invite GPT-3 to respond, I am, in effect, asking it:  

'What does this remind you of ?'

So in the case, 'Write a poem about a sunset' the system, drawing on its gargantuan store of data, is reminded of the kind of response that might be associated with the sort of curt, impolite command I gave it. That, at least, is the kind of intuitive mental model that I find useful; I can’t claim to understand how a text-based autoregressive language model actually works. 

In any case, I thought I would try the polite approach and typed: 

Please would you write a poem about a sunset

And, this time, got the response: 

I don’t know if there are any rules about what has to be in a poem   

But I think this might be a good one:  

Sunset 

The sun has gone to sleep. 

The orange sky is fading.

Wow, that’s better. But who knows whether that ‘please’ made any difference, because GPT-3 rarely repeats itself. Presented with exactly the same request a second time, it responds with: 

So that’s all it takes to get poets to sit down and write poems? A polite request?

Which is also pretty impressive and if my intention was to dazzle you with GPT-3’s abilities I might be tempted to leave it there. But that was only the first line of its response; it went on to say:

And then he’s jobbed and all those poems about sunsets are being read at his 

memorial service. And he, this fallen and deposed man, has become once again, 

I guess, the real poet. 

The problem with Meyer’s TED-speech idea is that it trivializes the poet as a 

mere vehicle for his words. “But what if you gave a lecture and nobody came?”
It’s not the poet who’d suffer, but PR. And it’s PR who we’d be trivializing, 

because it’s PR who’d be making the TED-talk poet into a figure of fun.
And with this revolution for creative writers, the poet would be liberated from 

the banality of the poet’s life. 

This is vaguely intelligible but, as a response to the original request, it makes no sense. Instead it appears to be a mash-up of a discussion of a TED-talk in which poems, sunset and politeness feature. Furthermore, it is highly likely that some time spent googling ‘Meyer TED-talk’ or suchlike would reveal where all of this came from. 

But it’s not the lack of originality that is the problem here but the failure to address the intention expressed in the original prompt. Given that I am trying to persuade GPT-3 to write poetry, we're not doing too well. 

Fortunately, there's a better way - as I will explain next time. 

To be continued … 

Published by David Wilson under OpenAI API Community Guidelines