In my ideal undergraduate course in literary criticism, the first semester would include a brisk introduction to large language models. This is less absurd than it sounds. Recall that large language models are essentially machines for choosing combinations of words. And if we are being crudely reductionistic—which is to say, if we are following the spirit of our age—poetry is likewise just a matter of choosing such combinations. To date, I have yet to see ChatGPT “generate” any masterpieces, only doggerel of truly awesome inanity. Despite this, you can learn a surprising amount about what poetry is by thinking of it as a problem for the new language machines to solve.
One of the more reliable guides to the inner workings of large language models is Stephen Wolfram, physicist and creator of the Mathematica computing system. According to Wolfram, these models are designed to solve one simple problem: predicting what the next word in a sentence is. They do this by breaking down text into sub-semantic units known as tokens. The tokens are converted into vectors, which give us an array of numbers, which in turn can be used to estimate the probabilities associated with each word as the machine moves along its sentence. Through judicious use of engineering voodoo, the system is then directed to find what Wolfram calls “the rational continuation” of the sentence. The system thus generates meaningful text by simply asking over and over what the next “expected word” is. And thus Peter Lax’s dictum is shown once more to be right: “If you can reduce a mathematical problem to a problem in linear algebra, you can most likely solve it.”
What does any of this have to do with poetry? ChatGPT is so interesting because the process Wolfram describes is basically the opposite of what poets do. Consider a remark made by the excellent but neglected war poet Sidney Keyes. According to Michael Meyer, the editor of Keyes’s Collected Poems, the young Keyes would judge his poetry according to “Yeats’s demand for ‘the intellectually surprising word which is also the correct word.’” I was introduced to this passage from Keyes by Geoffrey Hill during his tenure as professor of poetry at Oxford, and like him I have been unable to find the text in Yeats that Keyes refers to. But I would contrast that phrase—“the intellectually surprising word which is also the right word”—with Wolfram’s “rational continuation,” and propose that in this contrast we discover something important about what poetry is, and how it differs from other ways of using language.
Keyes does not give us much to work with when we try to think about how to choose the “right word” so as to pass the Yeatsian test. This is where we can call upon help from a very strange but brilliant little book by Simone Weil, The Need for Roots. Here is the relevant passage:
Simultaneous composition on several planes at once is the law of artistic creation, and wherein, in fact, lies its difficulty. A poet, in the arrangement of words and the choice of each word, must simultaneously bear in mind matters on at least five or six different planes of composition. The rules of versification—number of syllables and rhymes—in the poetic form he has chosen; the grammatical sequence of words; their logical sequence from the point of view of the development of his thought; the purely musical sequence of sounds contained in the syllables; the so-to-speak material rhythm formed by pauses, stops, duration of each syllable and of each group of syllables; the atmosphere with which each word is surrounded by the possibilities of suggestion it contains, and the transition from each atmosphere to the other; the psychological rhythm produced by the duration of words corresponding to such and such an atmosphere or such and such a movement of thought; the effects of repetition and novelty; doubtless other things besides; and finally a unique intuition for beauty, which gives all this a unity.
In a provocative mood, I would say that this quotation, qualified by Keyes’s Yeatsian criterion, is a skeleton key for all poetry. Even if Weil concedes that her list is arbitrary and incomplete, the choice of each word in a poem will be in some way determined by these planes or functions. A whole book could be written on the development of this passage into a fully elaborated poetics.
Returning to where we started, with the reductionist account of poetry as a game of word combinations, we can say that the poet selects his words to create the maximum amount of meaning along all of Weil’s planes of composition, and to do so in such a way that we can read the poem along these different planes simultaneously. The weight the poet gives to each plane depends in part on his sensibility, and in part on the formal constraints of the poem he is writing. A short lyric, for example, may primarily be constructed around the potentialities of meaning in what Weil calls the “atmosphere” of words—the way their lexical fields can become entangled—whereas the longer narrative poem requires a more open texture, but one that is also more metrically conditioned and in which deviations from metrical expectations carry a greater weight of meaning.
In discussing planes of composition, Weil uses a mathematical metaphor, one that forces our minds back to the language machines and to the terrible question: Is there a mechanism for creating poetry? I think not, though the recent success of AI systems has made anyone with a brain open to the idea that these models can get close to most forms of human cognition. What are the reasons for thinking that machines don’t really do poetry, that poetry creation cannot be formalized in a mechanistic way? For one, there is a sort of Wittgenstinian hunch. Just as we would be surprised if a lion took on speech and started to write sonnets of fourteen lines subdivided into octave and sestet, we would be equally surprised if an assembly of logic switches did the same thing. And of course, as Wittgenstein said, we would not expect to understand the poetry of either lions or logic switches.
But the real mystery, the double miracle of poetry, is not that we can make it—that is, search out this very specific “intellectually surprising” language structure—but that the fully achieved poem, the great poem, is not just structure but vision. Returning to Weil, we find that the final plane in her list is what she calls the “unique intuition for beauty, which gives all [the planes] a unity”—that is, something that is not confined to any one plane but traverses the whole. It is this intuition or vision that we experience when, as readers, we want to speak about the “truth” of the poem. To borrow from the philosopher-mathematician Gian Carlo Rota, the characteristic experience of reading great poetry is a sense of “enlightenment” wherein enlightenment is not reducible to logical verification. He likens it to a lichtung, a clearing in the woods. I prefer Wallace Stevens at the end of his late long poem An Ordinary Evening in New Haven: “It is not in the premise that reality / Is a solid. It may be a shade that traverses / A dust, a force that traverses a shade.”
