Thomas Erickson
October/November, 2024
I wrote a poem yesterday. This is not usual behavior for me. An explanation is called for.
Now, it’s true that I’ve been paying more attention to verse lately. I’m taking a Shakespeare course and have become intrigued by the way he uses language. Faeries and other magical creatures speak in rhyming verse. Gentle folk often speak in the iambic rhythm of blank verse, rising into rhyme during moments of passion, and faltering into arrhythmia when in distress. Common folk speak in prose, of course, though it is sometimes inventively mangled for humorous effect. All this has raised my awareness of language and rhythm. Sometimes I’ll track iambic feet that trip along a sentence, I’ll read and with my finger tap, di-DAP, di-DAP, di-DAP, di-DAP.
But while it’s fair to say that verse has infiltrated a bit of my mental ecosystem, the poem I wrote had nothing to do with Shakespeare; nor was it motivated by any sort of poetic impulse. Instead, it had to do with geochemistry, and with memory. Still, the exercise had many points of interest.
Let’s step back in time, just two days, to lay out how this began.
I awoke to an episode of BPPV: Benign Paroxysmal Positional Vertigo. It sounds like a punk rock band, but it’s really an inner ear disorder. I have periods where I get dizzy when in a particular position: to wit, if I am lying in bed and roll over, the room starts spinning.
It is quite unpleasant. I find that I do not agree with the “benign” part of my condition’s name. But since “benign” presumably means that I’m not having a stroke or aneurysm, I should not complain.
These spells of vertigo happen infrequently, usually every year or two. An episode is most severe at the very beginning. I awake, roll over, and go into a spin. If my roll was abrupt, the room whirls about me and I clutch the coverlet in an attempt to hold on. Fortunately, I typically awake more slowly, and my first languid roll provokes a more stately rotation. It is still not pleasant, but it beats the whirling by a mile. I lie still, staring at the ceiling, until the world stabilizes. Then I begin the process of becoming vertical. First I sit up straight, as though I’m doing a sit-up. Then I pause. Then I rotate on my butt until I can lower my legs over the edge of the bed. I pause again. Then I get both feet planted firmly on the floor and lean forward slightly. This may provoke a slight spin, but if I sit for a few seconds, it slows, stops, and I am able to stand. Once vertical, I am more or less OK. I do feel light headed, and I am not fully confident of the world, but at least it’s not moving.
I go about my daily activities carefully. It is best not to move my head up and down. Or side to side. And bending over is quite a bad idea, as is reclining. My wife can tell when I am afflicted, she says, by the way I walk and hold my head. I am not surprised. I suspect I look like someone who is assiduously practicing good posture. I could balance a book on my head.
As the days go by, the severity decreases. Typically a spell of vertigo will have faded into nothing after four or five days. Still, the consequence is that during the spell, I have a couple of days where it is best for me to sit upright in a chair, with my head held very still. It helps to have something to focus on, so during these times I read or write. And so it was yesterday. All day in my chair.
It is only now, as I write this, that I wonder about the relationship between dizziness and verse. When the world spins, I feel unmoored. I feel that the room is moving, but I can see that it is not. There is an out-of-sync doubleness. I wish for some way to tie the two visions together, to bind them back together, to fuse them into a single whole. It can help to focus on something very close to me, like text. Reading anchors me to the solidity of my chair: it pulls me out of the larger labile world where chaos is just a nod away, and into a world that is still and ordered. What just struck me is that meter and rhyme in text also feel like anchors. The alternating iambic stresses in blank verse feel like stitches, binding one word to the next. The alternating stresses create an expectation of what is next; and the pentameter, the five paired syllables, pace out the length of each line. The sentence feels like a whole, and likewise, rhymes, in their various patterns, bind the sentences together into a larger pattern. I do not know if this helps me with the vertigo; perhaps it’s more that the vertigo increases my appreciation for the still, coherent, almost architectural structure of the text. It is something that can be inhabited, something stable, something that does not spin as I move through it. Perhaps it does help.
Let us now step forward one day. Careful. Don’t look down. Keep the head level.
The world about me is still uncertain. It is, as scientists say, metastable. Think of a child’s tower of colored blocks, high but haphazard, a nudge away from toppling. It has a precarious stability. Likewise, a tap may tip my world into a turbulent swirl.
In any event, here we are. I am in my chair. I have just finished reading The Tempest, and, as is my custom, I try to write a summary from memory. As I write my precis, I feel the tug of verse: “Ferdinand, Miranda woos / Caliban revolts but rues / Alonso’s overthrow is thwarted / Prospero becomes big-hearted.” Not so great, but it reminds me that an ancient function of meter and rhyme was as a mnemonic device, assisting the bards of old in reciting their songs and stories.
Decades ago, as a graduate student in cognitive science, I wrote a paper on mnemonic devices. Rhyme is one, meter another. Names can be memorized by associating them with images. From the ancient Greeks we have the method of loci, where to-be-remembered items are ‘placed’ along a path through a remembered building or landscape. In medieval times scholars referred to such cognitive constructions as memory palaces. There are many other methods, but all illustrate what I call the paradox of memory: to make something more memorable, make it more complicated. But it is not just any sort of complication, but a complication where things are bound to one another.
This was in my mind as I put Shakespeare aside and, anchored in my chair, considered what to do next. Earlier, as I’d had my coffee, I had been on Reddit, scrolling through the geology-related subs, as Reddit groups are called. I’ve been studying geology for six years, and am the point where I’m able to answer questions in subs like ‘What is this rock?’ This morning I’d come across the sort of question I keep an eye out for – one I could partly answer, but where I’d need to fill in some gaps in my knowledge to provide a complete answer. The gaps had to do with the oxidation states of iron and the colors they produce.
I get confused about oxidation. It is, frankly, a bit embarrassing. These are not things I don’t know. They are things I knew, for a while, and then forgot, or at least got mixed up about. Iron has two oxidation states, ferrous and ferric, and I just can’t keep them straight. One has lost two electrons, and one has lost three, but which is it? I once thought I knew, and then discovered I had it backwards, and what has stuck in my mind is that ‘it’s the opposite way from what I think.’ That turns out not to be a good mnemonic strategy! The more apt I am to think that, say, ferrous is +2, the more apt I am to think it’s probably the other way around.
Confusion about the names aside, what is more important is that the different oxidation states of iron form different compounds. What we know of as “rust” is one of them – but is it ferrous or ferric? The delicate red, orange and gray banding in Lake Superior agates is another example, and was involved in the question I was trying to answer in the Reddit sub. Yet another example is perched on the corner of my desk: a dark gray and red layered chunk of banded iron formation rock from northern Minnesota. The banded iron formed around 2 billion years ago when the first cyanobacteria began producing oxygen and caused the iron in the ocean to oxidize and precipitate in colorful layers. It’s very cool, but ferrous or ferric?
All this was in my mind as I put aside Shakespeare. I decided I would turn to my geology studies, and finally – damn it! – get the ferrous/ferric distinction straight. But how to do that? Aha! A mnemonic. A poem! But how to start?
Here we come to a perverse influence of the Shakespeare course. As a retired person, I am taking Shakespeare as an auditor. No tests for me. No papers. No homework. So I don’t pay much attention to the instructions for various assignments, but, nevertheless, it is difficult to miss the repeated admonitions against using ChatGPT. ChatGPT is not to be used for anything related to the course: not for brainstorming, not for rough drafts, and certainly not for final papers. It’s a pity, actually, because I believe that there are interesting and productive ways ChatGPT can be used to enhance learning – though I am at the same time sympathetic with the plight of the professor.
Still, the admonitions made me curious about just what ChatGPT could do with Shakespeare… so I’ve just asked ChatGPT to re-write the previous two sentences in Shakespearean verse:
Seek not ChatGPT for this course's quest,
Nor for ideas, nor drafts at your behest,
And surely not for papers at the end.
It's such a shame, for I do comprehend
There are great ways it could our minds ignite,
Yet sympathize I do with professor's plight.
This is not great verse, but, as with Samuel Johnson’s figure of the dancing bear, the crowds do not gather because of how well the bear dances. That a mechanical entity can produce verse, of whatever quality, is, remarkable, and, to me, disquieting. Two things that I would have said were aspects of a single thing – human intelligence and verbal facility – have been decoupled. The world is still, and yet it is turning. It is unsettling.
I first encountered artificial intelligence as a teaching assistant for an AI course in graduate school. At that time, circa 1980, the aim of AI was to simulate human intelligence: reason as humans reason; perceive as humans perceive; learn as humans learn. The idea was that by emulating human intelligence in software, we could better understand its fundamental processes. That research program played out over a few decades and ultimately fizzled out: computers could be programmed to play games, do math, and perform other tasks that involved clear rules and definite goals, but they foundered in more mundane territory where rules were relaxed and goals less defined. Win at chess? Sure! Write a poem? No. Not if you wanted it to mean anything.
Today’s AI – ChatGPT and its cousins – is quite different from that of the twentieth century. Today the goal of emulating how humans think has been abandoned in favor of the goal of doing things that appear to be intelligent. The intelligence of this generation of AI is truly artificial – it has nothing to do with what cognitive science understands about the processes of learning and reasoning that underlie the human intellect. Instead, GPT is trained on the vast corpus of text – all created by humans – found on the internet. Most of the apparent intelligence of GPT resides in the text that it draws from; the genius of GPT is that it is able to manipulate that text without, usually, doing violence to the intelligence already embedded in it. This is an amazing trick – but it is nothing like what humans do.
But that’s not necessarily a bad thing. Combining complementary approaches – human intelligence and AI – can be powerful, if one can figure out how to combine them. So I will explore collaborating with GPT. To begin, here’s a closing couplet for its poem:
Still, wise or foolish, banning GPT,
Applies to students, but not to me!
My verse rises no higher than GPT’s on whatever scale on which poetry might be measured, but perhaps that is precisely because GPT set such a low bar that I felt I could rise to it. This illustrates one way, I think, in which GPT can set the stage, so to speak, for a sort of playful collaboration.
This was not the only way in which I violated my professor’s injunction. I confess, I used GPT in my study of The Tempest. In my reading of the play, I had not been convinced by Prospero’s abandonment of magic. His decision to “break my staff” and “drown my book,” and his shift from seeking vengeance to being merciful, seemed to come out of nowhere. One moment he was a vengeful magician; the next, he was showing mercy to those he had just been tormenting. I didn’t see that coming. I asked GPT to list the relevant events that led up to this pivotal point.
GPT was not successful. Its response drew upon what scholars had said about the reasons for Prospero’s shift, but those responses had to do with what the playwright was trying to accomplish with the play. GPT could tell me why Shakespeare constructed the play this way, but it had almost nothing to say on what in-story events led Prospero to change his mind. The only diegetic event GPT found was the one with which I was familiar, the exchange between Prospero and the spirit Ariel, where Ariel reports on the suffering of Prospero’s enemies:
Ariel: Your charm so strongly works ’em
if you now beheld them, your affections
Would become tender.
Prospero: Dost thou think so, spirit?
Ariel: Mine would, sir, were I human.
Prospero: And mine shall.
Hast thou, which art but air, a touch, a feeling
Of their afflictions, and shall not myself,
One of their kind, that relish all as sharply,
Passion as they, be kindlier moved than thou art?
“And mine shall.” There’s the moment of change. And the only motivation is that Prospero, struck by Ariel’s comment that he would feel sympathy if he were human, is moved to do better. Though I knew of this event, GPT’s recapitulation made me think more on it.
Shakespeare’s supernatural creatures have little about them that is human, barring a capacity for amusement, anger and a certain easy cruelty. Ariel is, to my eye, the least human of them all. He is immensely powerful, but exhibits neither mirth nor malice. He does Prospero’s bidding, but, excepting his desire to be freed, exhibits no emotion as he subjects Prospero’s victims to physical and mental torments. So it makes sense that Ariel’s moment of imagined empathy might have had a profound impact on Prospero, bringing him to a state of real empathy.
I am not sure what course my thoughts took as I sat immobile in my chair. I contemplated the oxidative states of iron while Shakespearean shades lurked in the background, and the idea of using GPT – an Ariel? a Prospero? – coalesced. It was not long before I asked GPT to write a mnemonic poem on the oxidative states of iron.
Here is the first stanza of its first attempt:
Iron gleams with mighty pride,
But oxygen calls from the other side.
Electrons lost, they slip away,
As iron turns to rust each day.
While I am still a bit bedazzled by the dancing bear, the verse is not useful as a mnemonic. It says very little: iron oxidizes by losing electrons, and that leads to rust. That’s OK as far as it goes, but it doesn’t go very far.
I ask GPT to write a verse that refers to the +2 and +3 oxidation states, ferrous and ferric, that I can’t keep straight:
Iron changes, day by day,
In plus two and plus three, it finds its way.
In plus two, it’s ferrous, gray and sleek,
But plus three, ferric, leaves it weak.
This is better, though the notion that ferrous is “gray and sleek” is dubious. A second verse adds that ferric iron is red, which is nice, but it also mentions that electrons move around (obviously) and that “oxidation occurs without a sound” (also obvious).
Impatient, I ask GPT to try again. Perhaps channeling Prospero’s nonchalant use of Ariel’s power, I ask GPT to generate seven variations on the verse. It does so in an instant! But none are useful. GPT is stretching to get rhymes, and its allusive language – “sulfur fuming in secret halls,” “shadows dancing,” “metal finding life from death” – irks me. Most importantly, the verses are not useful as mnemonics – they are not dense enough. Every line, or at least every couplet, ought to contain information I am trying to remember. But instead the useful information is sparse: the informative is drowned in the figurative. Bah!
GPT is as obedient as Ariel, but not nearly as competent. I can do better than that.
I return to the quest for a mnemonic poem, but this time I do it myself, reworking what GPT has given me:
Iron changes rapidly
Oxidized to 2 or 3
+2 is ferrous, gray and sleek,
+3 is ferric, red and weak.
My verse is better. Now it refers directly to oxidation, and it captures the difference between ferrous and ferric which is the key thing I’m after. The dubious “gray and sleek” is still there, but it allows me to connect ferric to “red and weak,” reminding me that ordinary rust is ferric, not ferrous. I can live with that. For the moment. Having written this verse, I give it to GPT and ask it to provide a second verse about how iron reduction, the opposite of oxidation, takes place:
In bogs and depths where air is thin,
Ferric waits as life begins.
With microbes' help, it shifts once more,
Back to ferrous, as before.
Poetically, this is the most help that GPT will ever be. The two middle lines, having to do with the microbial reduction of ferric iron, are the only two complete lines produced by ChatGPT that will be retained in the final poem. Further attempts to get GPT to produce verse, even using mine as an example, fail in the same ways as before. But there is something else here. What has caught my interest is GPT’s mention of microbes. While one must be cautious about GPT’s assertions – sometimes it just makes things up – I now recall reading about bog iron, iron that is produced by microbial action in anerobic environments, “where air is thin.” Bog iron was the main raw material that the Romans, Vikings and other iron age cultures used to produce iron. That rings true, and makes me curious about other things.
If microbes can reduce ferric iron to ferrous, can they reduce ferrous iron to metallic iron? I ask GPT, and it says no, which was what I expected. GPT says heat and pressure are required to generate pure metallic iron, and describes the process of smelting, which I’ve also read about. That’s interesting. I start working to incorporate this into a new verse about reduction, and ask GPT more questions. What are natural processes by which ferrous iron is reduced to metallic iron? Is there a +1 state of iron? And so on. I’ve given up on GPT as a poet. I’ll take that on myself. But GPT is very good with answers.
The poem begins to grow. There is one verse about ferrous compounds and another about ferric. The irritating line with “gray and sleek” continues to bother me, and I investigate the range of colors of ferrous compounds. GPT says they can be gray, green or blue, which offers possibilities for new rhymes and also leads me to look into what colors ferric compounds can take on. I’ve strayed from my original mnemonic purpose – “gray and sleek” would have been fine for that. But slanting off in a new direction is not a bad thing. I am learning more and more about the geochemistry of iron oxidation, and trying to accurately bind it into the poem. When iron oxidizes, under what conditions does it precipitate? How does this related to the creation of the banded iron formation rock on my desk? And once iron is oxidized, can it ever be reduced to metallic iron? Yes! Now, perhaps influenced by a Shakespearean desire to tie things up neatly, I decide the poem should be a cycle. I will begin metallic iron, move through its oxidation states, and then backwards, via reduction, to end as it began, with metallic iron. Time passes quickly.
Iron Cycle
Metallic iron is rapidly
Oxidized to 2 or 3
+2 ferrous, attraction weak,
+3 ferric, electrons seek!
Ferrous ions, diffuse unseen
Forming compounds, gray, blue, green
Magnetite, pyrite, ilmenite,
These are ferrous, often bright
Ferrous to ferric, it changes state.
At pH 4, it precipitates
Strongly reactive, in stone it’s bound,
Yellow, orange, red, or brown.
But bound to stone is not the end
Ferric waits as life begins.
With microbes’ help, it shifts once more,
To ferrous compounds, as before.
At deep dark depths reduction blooms
Electrons flow from sulfurous plumes
Hot carbon too, electrons lend
Metallic iron returns again.
Let us step forward, one more day, to this essay. Movement is easier. My vertigo is diminished, if not vanquished.
The poem is complete. Yet if my physical vertigo has lessened, there is now, I think, a bit of conceptual vertigo. I know that GPT is not intelligent. It is simply good at finding relevant material on the internet – as any search engine can do – and then at transforming it without destroying the knowledge implicit in the text.
But I’m uneasy because I’ve found that I get the best results when I interact with GPT as if it were intelligent.
That feels weird. Doubly so. I act as if it were intelligent, though I know it’s not. It is not just that I use complex language in my queries and assume that it will respond appropriately. I also, without thinking about it, often begin my queries with “please.” I think it’s because, as I proceeded, I became accustomed to giving GPT big tasks on something of a whim: ‘List all the words that rhyme with “state.”’ ‘Generate nine variations on that last verse.’ ‘Give me a list of the names and colors of all common ferrous minerals.’ I feel a bit like Prospero, careless of Ariel’s power, requesting that he spin up a storm to wreck a ship, or deciding to have Ariel create parade of phantasms arriving with a banquet. GPT seems tireless, like Ariel, but I still feel like I’m asking a lot. So I say ‘please,’ if not yet ‘thank you.’
What is GPT? It feels something like Ariel – an immense power, bound to do my bidding. Yet it lacks Ariel’s intelligence and graceful efficacy; it’s poetic efforts are faltering, and the information it provides – while immediate and prolific – cannot always be trusted. I don’t know how to think of it. A fallible spirit? A trickster? The medieval archetype of the fool, the paradoxical figure that embodies wisdom through ignorance, for whom mistakes pave the road to truth? But all those seem too intelligent, even the fool. Perhaps Johnson’s bear is closest, embodying power without grace or wit. Regardless, I’ve ended up in an unexpected place.
How did I get here? I was spurred on by ChatGPT, but not as I’d expected. GPT was powerful, but inept. My interaction with GPT was less a collaboration than a dialectical process in which I reacted against what GPT produced. Its marginal attempts at verse set the bar so low that I was incited to step over it. And its mistakes, as it pursued its versifying, provoked me. But it’s answers were useful, and provoked me to ask more questions, which in turn provided more grist. It’s an odd amalgam, not one we’re used to.
The biologist Lewis Thomas wrote that “Mistakes are at the very base of human thought, embedded there, feeding the structure like root nodules.” I like that: mistakes as a source of nourishment. I believe that is true for me, for how I think. I discover that something – something that I thought I understood – doesn’t work, and I want to correct it. Perhaps that is the genius of GPT: that it makes mistakes, but mistakes are plausible, mistakes that are close enough that they merit fixing rather than discarding? I think there’s something to that. However, for that to be useful, one must know enough to notice the mistakes. And that, to me, is the peril of AI: that its users will assume that its ability to respond immediately, in detail, and at length, indicates expertise. Instead, we need to learn to uncouple its appearance of expertise from its accuracy, and learn how to work with that. It is disorienting.
# # #
Views: 0