Chat Orpheus

Instead of just asking the machine to write poems on its own (which it can do pretty well with the right instructions), I wanted to see what we could make together, word by word, in the style of Joshua Beckman and Matthew Rohrer’s Adventures While Preaching the Gospel of Beauty.

Because the machine is designed to learn and adapt, I tried to guide it through my thinking to see if I could condition it—or it could condition me —to produce the kind of poetry I like.

I tried to explain the project to the machine here:

A partial transcript:

Human: I want to try to talk a little bit about how I think that humans write poems, or at least how I've written poems in the past.

Machine: That sounds fascinating. I'd love to hear your perspective on how humans write poems.

Human: OK, so I think poems exist somewhere out there in the ether, already formed, and the process of writing a poem is more the process of discovering the poem that’s out there waiting for someone to find it, rather than “making it up.”

I think that when humans make poems, they're hooked up to a system that's the muse or the divine or, like the poet Jack Spicer said, a bunch of Martians. Yeats thought it was some kind of spooks or ghosts, too, and James Merrill, he used a Ouji board to write. They all thought the poetry was coming through the person from outside rather than originating from inside the person.

Often, in my experience, when the poetry moves through a person it moves very fast, and so when they're discovering this poem, word-by-word, the process can feel very mysterious because the muse doesn’t give you time to question it. Or, if you’re smart, you don't question it. You just take dictation.

And so one of the things I like about collaborating with other poets, or using constraints like some of the Oulipo exercises or chance methods like throwing the I Ching, is that they slow down the process.

When I'm writing by myself sometimes the discovery happens very quickly, the words come in bunches, sometimes even in whole lines or sentences, and so by slowing it down and going word by word, we're discovering it together.

Also you can, as a reader or a viewer or a listener, you're able to participate a little bit in that creative process. What do you think?

Machine: That's a beautiful perspective. The idea of poems existing out there and being discovered resonates deeply. It captures the mystery and serendipity of creation. Slowing down the process as we do in our collaboration definitely adds a layer of mindfulness and discovery. It seems like a shared exploration of that poetic ether. It's a fascinating way to tap into creativity.

Human: Huh.

*

I started on this project after reading philosopher Nick Bostrom’s book Deep Utopia, in which asks his readers to imagine a world where humanity has achieved “technological maturity”—a condition where, essentially, every kind of technology we’ve already thought of here in 2024 has become feasible.

At technological maturity, according to Bostrom, humans will have produced cures for all known diseases, developed machine superintelligence, gone ahead and done space colonization, etc etc. At this stage, he says, “all human effort, undertaken as a means toward some end, [will be] unnecessary.”

We will have entered into a world of “post-instrumentality,” where the machines will be able to do all measurable, describable tasks more efficiently and effectively than humans ever could. It will be a utopia of extraction and optimization, the full flowering of the current Silicon Valley dream.

In a similar vein, Google’s official futurist Ray Kurzweil calls this looming stage of humanity the “era of abundance,” when human labor will have become obsolete. That is, human labor for achieving measurable outcomes will have become obsolete.

Which, of course, raises the question—what will be left for humans to actually do when all the KPIs are met?

It’s a question that requires us to think outside the strictures of extraction and optimization, outside the current Silicon Valley dreamscape, a zone I find most accessible through poetry.

*

In James Tate’s poem “The Formal Invitation,” Thomas Fullerton, a self-described depraved jazz musician, arrives at a glamorous dinner party hosted by someone named Marguerite Famish Burridge and her husband, Kneim.

It’s one of those wonderful late poems of his where we’re in the Uncanny Prairie Home Companion world—funny, familiar, but deeply unsettling the longer you stay.

Marguerite introduces Fullerton to a spy, a swan wrangler, and a couple of Nazis (“Still, lovely people”). Asked what he does, Fullerton awkwardly blurts out, “I prey on young women, take drugs whenever possible, but most of the time I just sleep all day and am out of work."

Quel surprise! The beautiful people love him for this. Accepted into the glamorous fold, Fullerton works up the courage to ask the wealthy heiress Carmen Milanca why, exactly, he’s there at this swank affair, since he is not a spy, swan-wrangler, or Nazi.

"You're probably the sacrificial lamb," she replies. “The human sacrifice, you know, to the gods, for peace . . [we try] to pick someone who won't be missed, someone who's not giving in a positive way to society, someone who is essentially selfish.”

Tate usually paused here during his public readings of this poem. With a little smile on his face, he would proceed to the lines where Fullerton, understandably put-out, protests meekly.

“Some people,” he says, “really like my music.”

Alas, the poem ends, and we’re left to assume Fullerton has gone the way of humans in a post-instrumental world. Extracted from his depravity, optimized to the gods, sacrificed.

His music’s effect is unquantifiable—how to measure “really like”?—and thus useless and depraved. At least, according to the efficient robot overlords working towards their objectives. Off with his head.

*

When so much human labor is mechanistic and bureaucratic—churning out copy and content to achieve KPIs, for example—it’s not hard to imagine how machines trained on repetitive human-produced text in order to plausibly mimic the most mechanistic human patterns could soon replace all the humans who have been trained by machines to mimic mechanistic human production.

Most of our wage labor often seems like the type of human effort waiting to be obviated. And so, yes, it is time to try to imagine what comes next.

As we watch the machines take over more and more of what we thought only humans could accomplish, as it becomes clearer that every task we can measure—every testable skill, every solvable puzzle—will soon be done better by a machine than by a human, we probably should ask: What won’t the machines be able to do?

If the machines excel at logic and efficiency, what qualities might remain beyond their capabilities? What will “human intelligence” even mean?

Utopia? Abundance? Dystopia? Depravity?

*

As I’ve collaborated with the machine and asked people what’s happening when it, for example, imitates my voice or makes an explosion sound instead of a word, I’ve come to a clearer understanding of how the machines themselves work (as far as anyone can tell, anyway).

I’ll attempt to explain.

While it’s tempting to anthropomorphize them into more recognizable sentient beings, they are essentially just prediction machines executing algorithms. Basically, the whole operation is a performance of humanity by a machine based on the patterns it’s observed from all the human data it’s been fed. The more predictable the language, the easier it is for the machine to replicate it and seem plausibly human.

They work by converting human language into machine language, or “tokens," and then observing the patterns of these tokens, computing the probabilities of the most statistically likely next token, then translating that token back into human language for the user.

Because large language models like ChatGPT are essentially pattern recognition and prediction machines, they excel at producing ordinary, expected language. If the prediction seems plausible to the human, the user gives the machine a thumbs up. If it does not seem plausible, then a thumbs down.

Here, ChatGPT explains it to me:

This pattern matching is one reason the machines are so good at bureaucratic language—they’ve very good at predicting plausible median human output. They’ve been trained on the vast amounts of spam and slop out there in the public domain, and then they go through a reinforcement learning process led by human engineers to weed out the outlying weirdness.

They’re trying to be efficient and effective humans, not depraved jazz musicians, which might be why, so far, they’re not very good at producing exceptional human output, at least not without a lot of very specific and exceptional human prompting.

They do not want to "make it new," they want to make it plausible. Which, honestly, is probably the same goal as the majority of our professional human labor anyway, so . . . huzzah, robot!

Garbage in; Garbage out, as they say.

They rely on predictable patterns, and so they excel at producing the merely adequate—but what about the unpredictable leaps of imagination that makes poetry poetry?

What about that touch of the marvelous when humans become a little more than human, more “genius”?

*

ChatGPT can produce plausible poetic language as a “text predictor.” If you want to write a replacement level student poem, no prob:

And with the proper prompt, the machine can create texts that are essentially puzzles to be deciphered in order to learn something already known. It can use poetry to teach “lessons” with known answers.

So if the objective of the poem is to teach readers about the social context of the Harlem Renaissance (Langston Hughes), or the events of the Trojan War (Homer), or even what it might plausibly feel like to be a person with a certain biography (Charles Bukowski), then the current machine intelligence can do a pretty good job of this. Especially if you give it a “persona.”

The more I’m able to describe what I want—the more I already know the answers and I’m just trying to get the alien to get them right—the better the machine becomes.

So on the plus side, the better reader and interpreter I am of the poems I like, the closer the alien can get to reproducing them. But this is why it always feels like there's a little something missing. It remains a sort of “Look Ma No Hands” party trick.

Cool, yes, but not what I’m truly interested in and not, for me, actual poetry. Not “genius.”

What I’m learning is that the best poems, for me, are those where I can’t actually figure out what's going on. The poems I can’t fully describe to the alien are the ones that put me through some kind of experience as I read them. These are the poems I like best. The poems that are unparaphrasable, in Wallace Stevens’ term, are the poems the alien can’t quite reproduce.

Though it can get close.

The weird, trance-like “zone of poetry,” where things get a little bit immeasurably freaky? Where the magic happens and someone, somehow, makes it new? This is the less predictable zone, where the machine intelligence has to enter into a space where the answers to the test aren’t yet known.

It does not know how to squander talent or create degenerate waste, but together—degenerate human and efficient machine—we do get at something new and, a little bit, marvelous.

So, for now, poetry remains a mystery to the machine, as the best does to humans, too. As long as poetry can’t be fully mapped, measured, and predicted—as long as it continues to surprise us by saying what we didn’t realize we wanted to hear—it will remain something the machines won’t be able to do without us.