Snap, Crackle, Pop... a Scene is Born
In last week’s Pop Rocks & Poetry post, I explored the essential differences between embodied intelligence and its machine imitator. Although a large language model (LLM) cannot physically experience the zing of a pop rock bursting in its mouth, it can still do some interesting things. Today, I will examine a relatively successful ChatGPT encounter, one where the model responded creatively and unexpectedly.
We start with the 4th edition of Understanding Poetry, written by Cleanth Brooks and published in 1976. Although this text is dated, the writing is beautiful and clear. A lot of mid-century (50’s, 60’s, 70’s) writing was like that. I don’t know, but I’ve wondered if the slow process of composing with paper and typewriters makes for clearer thinking. Anyway, in Chapter 1, Brooks presents a poem called Sir Patrick Spence. It's the story of a noble sailor named Sir Patrick Spence who receives a command from the Scottish king to execute an expedition. But winter has come, and it is not a time to sail the high seas. Given that this is a tragic Scottish poem, you probably know how the story ends. The ship encounters a storm and sinks with all hands onboard. Yet, as Brooks points out, the poem has no storm! He writes, “We are forced to create the storm scene in our imaginations, and that very fact perhaps makes the storm more vivid” (p. 25). All the poem gives us is a single image of hats floating on the sea. But that’s all we need. With just that, our minds can imagine the storm, the struggle to save the ship, and the sinking. I just felt a little zing in my mouth. Did you? Pop Rocks strike again!
The mind’s ability to create something from practically nothing is truly unique. Large language models (LLMs) cannot fully mimic this ability, though – as we will see – they come pretty close. I decided to see how ChatGPT fared with Sir Patrick Spence. I started the conversation by telling the model that I would give it two parts of a poem and then ask it to fill in the details of what happened in between. To make the challenge easier, I updated the English. Here’s how the conversation went:
Me
ChatGPT
Me
ChatGPT
After reading ChatGPT’s response, I thought, “This is pretty good.” The model even generated three rhyming lines. Maybe not great poetry, but at least they give the reader a sense of how facing an inevitable fate on a doomed ship feels.
How might we account for this success, given that the model makes inferences quite differently from humans? Here, it appears that a critical factor was ChatGPT’s ability to identify the poem. From there, it was in an excellent position to summarize the content, which LLMs do well.
When I replaced “Sir Patrick Henry '' with my name in a different conversation, the model inferred that a storm sank the ship. It even generated two bridging stanzas of 4 lines each, though its output's lyrical and evocative quality was subpar.
In summary, ChatGPT performed well on this simple test. However, it’s certainly not a poet, at least not a human one.