Currents Magazine
    

Issues

Events & ventures

Residencies
 
About



 



.•.·:*¨ ¨*:·...••°°°°••..°


© poopopop












web design by 
Ronin Reeves 
&
Adrian Wong
“Why is it the snake’s language that we’re all using to build the future?”
Lili Newberry interviews Zoë Hitzig





Zoë Hitzig is the author of two books of poetry, Mezzanine (Ecco 2020) and Not Us Now (Changes 2024). She holds a PhD in economics, works as a Research Scientist at OpenAI, and currently serves as poetry editor of The Drift.

I was first introduced to Hitzig’s work when she read from Not Us Now at The Grolier in 2024. In a series of inventive forms, the collection renders algorithmic life through procedures of compression, dissection, and ventriloquy. Hitzig extracts vestiges of human life from an ever-growing mass of data. Before her reading, I had never seen a poem encode its own set of variables—at least, not explicitly. It’s entirely possible The Procedure began spinning into shape that evening.


Lili Newberry: In the foreword to your second collection, Not Us Now, Srikanth Reddy writes that the poems “both court and interrogate the logic of algorithm.” How do you see this happening?


Zoë Hitzig: What I see in that statement is something that’s been a theme in my life – you have to get up close and intimate with the things you’d like to criticize. In order to deeply understand what algorithmic systems are doing to us as people, as a society, as bodies in space, I’ve had to get really close and almost flirt with them. What I think I was trying to do with the algorithm poems in Not Us Now, these dramatic monologues spoken from the perspective of specific algorithms, was to try my best to understand where they come from. If they could talk, what kind of a pathos would they have? Part of what that does is it reminds us how they were built. It reminds us that they were created by humans, humans who had feelings and who had experiences and had a way of looking at the world that then they infused in these systems. 


LN: Is this coming out of a particular experience?

ZH: Not really, though through my work in and around computer science, I became obsessed with the language that is used to describe these systems. Being a poet, I couldn’t ignore the actual words they use—like regret in learning algorithms. Regret in this context has a simple definition: how much better could you have done if you had known everything you came to know at the end. It’s simple, but why’s it called regret, as opposed to ex-post loss or something?


LN: I mean, even loss is very charged…  


ZH: Yeah, even loss is a very loaded term. I’m just endlessly fascinated by the actual language of code. In the book, I was focused mainly on the language of learning algorithms and fundamental computational heuristics. But the linguistic landscape of computers is really wild. Why are we writing in “Python”? Why is it the snake’s language that we’re all using to build the future?


LN: So you’re a poet, economist and Research Scientist at OpenAI. How do you see your professions and fields of study bleeding into one another?

ZH:  I’ve always been obsessed with what happens when you take something very human, like an experience, a feeling, a thought, a goal, and you try to codify that human muck. What happens when you distill it down into code in the sense that you can put it on paper or in a computer and give it to someone else and they can understand it. That’s a certain kind of compression that happens through language—or different languages. It’s a compression that happens when you build a mathematical model of the world and of society, which is what economics is, or at least the kind that I do: microeconomic theory.

It’s also what happens in algorithmic systems. You take in a whole bunch of data, and you try to find the patterns in it and distill it so that you get something portable. At a really abstract level, there’s something in that sort of compression into a code that I find really beautiful and really terrifying. I think it’s terrifying in economic theory and in algorithmic systems because a lot of the time, the people who are affected by that compression don’t understand the simplifications and what could have been lost in translation, so to speak.

In my work in economics, I’ve talked about the algorithms that are used to allocate kids to public schools in Boston. There’ve been major policy changes driven by the mathematical properties of these systems—properties that economists know how to analyze and solve for. It’s impressive that we can take a messy, complex reality and describe it in precise mathematical terms. In a sense, we’ve tamed reality. But at the same time, the stakeholders who are impacted by these decisions made based on compressions, often cannot fully understand or interrogate those compressions. 

I think something similar happens with big machine learning algorithms. It’s even more extreme—a bot starts talking to someone and they immediately feel like this is an oracle. But it’s not—it’s just a statistical model trained on a lot of data. There’s something about this compression into code that’s kind of frightening, socially. What happens when people aren’t attending to those compressions? In poetry though, compression feels different –


LN: It doesn’t really affect our choices or policy. 

ZH: Yeah, the stakes are pretty low in a certain sense. But also, poetic compression is explicitly an invitation to interpret, to question and to disagree, rather than something we’re asked to accept as a neutral, authoritative system. 

How people communicate with each other is one of the wildest things to think about. How do you stay faithful to a message? There’s no message that can be perfectly represented as it’s transmitting from one person to another. Short of literally merging our brains, I don’t think there’s any way around that gap. And if you just sit with that for a moment, it’s really astonishing. On a social level, those failures to communicate start to compound. Some people know how to take advantage of them—by marshalling specialized forms of knowledge, or by manipulating the terms of the conversation through jargon, euphemism, strategic ambiguity... It’s kind of mind blowing that we’ve managed to get along as well as we do. 

There is something magical in poetry, though: through compression, through reliance on the magic of verse and breath and rhythm, you can actually say more, not less, dissolving for a moment those boundaries between brains and between hearts.


LN:  Does that feel like compression too, or is it something else? 

ZH: Yeah, I feel like it has to be something else.


LN: Yeah. Because in a way, poetry bypasses a lot of the expected avenues of communication. 


ZH: I have a pretty mystical view of poetry. Whenever I’ve written a good poem, it doesn’t feel wholly authored by me. It feels like something that existed before and I happened to be there to write it down. Not every poem I’ve written was written like that, but my better poems definitely were.  

Poetry is such a funny thing, because there are so few of us who actually care about it at this point. And I often find myself in this difficult situation where people are like, what makes a poem good? And I have a response I can give that sounds like what a critic is supposed to say. But my honest answer is obnoxious and absurd—you just know when it’s good because having read it you are completely changed, as if by an unshakeable dream or memory.


LN: How does procedure factor into your poetry?

ZH: Procedure is a scary word, isn’t it? I’m now thinking about process versus procedure, two words that have the same root… I sometimes stumble into formal constraints that give me a procedure to write into. For example, the long poem at the end of Not Us Now has a bizarre structure where most of it is written in lines that are separated by dashes, and words are broken up by hyphens. There are six “feet” in each line, where each foot is at most two syllables. I don’t know how this structure came to me, but once I had it, I found the formula or procedure so generative. I wrote so much I thought that poem might be its own novel (I cut most of it out).

I had a similar experience writing with wild momentum into a procedure with one of the poems in my first book, Mezzanine. Generally, Mezzanine is a lot more normal than Not Us Now, but there’s a pretty whacko poem in there called  “Fragments from the Imagined Epic: The Island of Stone Money.” I stumbled on an idea to write a long series in the shape of these ancient stones that have a hole in the middle that were used as currency on the island of Yap in Micronesia. I wrote like dozens and dozens of pages of stone money poems – what ended up in the book is just a tiny sliver of what came out of that procedure. So maybe that’s actually where the nightmare of the procedure comes back—once it’s in place, it can be enabling, but then maybe it’s too enabling, and you can’t stop. 


LN: What you’re saying now is making me realize there’s a passivity in the word as compared to process, right? Something done to you, that you’re not actively choosing. 


ZH: Definitely. I think a lot of my poetry, because of my obsession with codification and all that, is sort of obsessed with procedure, and where procedures go wrong. In Mezzanine, I was thinking a lot about the law. What a terrifying set of procedures which so often lead to something other than what it was designed for. I also wrote about economic procedures – like auctions—what does this procedure for selling an object mean, what kind of spectacle is it? I also wrote about statistical procedures. I’m thinking of one poem called “Generalized Method of Moments,” which takes its title from a statistical procedure. 

LN: In the case of “Exit Museum” in Not Us Now, the long poem with all the dashes, you mentioned writing along with the form. I think that’s what most people do. But looking back at the poem, what do you think that procedure is doing to the language?


ZH: It slows the reader down. And in particular, it calls attention to how experience gets codified—it’s essentially a stream of consciousness. The text on its own is not that interesting, plot-wise. But the format calls attention to how any kind of experience gets codified—what happens when experience is recorded, in one format versus another, or in formats that are kind of alien. On one hand the poem is deeply human with its fixations on the needs of the body – and also ancient, because the six-foot lines are hexameter-ish.  But at the same time, the poem’s text is corrupted in a way that makes it feel alien and futuristic. The syllables are broken up in ways that don’t make sense to a human reader. So what it does, even though I didn’t necessarily set out to do this when I was writing it, is that it forces the reader to wobble through an otherwise normal text, and through that encounter confront the oddness of language and the imperfection of how we codify our deepest experiences, with or without intention.

I think a lot about what will happen to all of our stories in the long run. Where will all of our data be in fifty years, in a hundred years? There’s so much of any one of us on so many servers all over the world. And where’s it going to go? Who will have access to it? What if someone or something stumbles on it and tells your story in this strange, corrupted way? What would that feel like?  Maybe it would feel a bit like “Exit Museum”—though not exactly since that’s a stream of consciousness, and I don’t think anyone has access to raw streams of data from our brains right now. 


LN: But soon.

ZH: Yeah! Soon. And in that poem, there’s a lot of digital communication. It’s like, “text – Xo” “text – Jo” “text – from – Ry” “Li – texts – a – phot-o” “log – on – share – screen” …  that’s literally the poem. And so even if the actual stream of consciousness isn’t the thing that would be stored and then rediscovered in some number of years in a corrupted format, all of those communications which are a kind of mosaic of the narrator’s day, exist somewhere. And if someone one day had all the power in the world to control all the servers in the world, they would be able to piece together pretty much what the narrator was doing on that day, and a good deal, perhaps, of what they were thinking, too. 


LN: I really like how that procedure draws so much attention to the syllable. It creates kind of a strain on communication, points to the mechanics we don’t think about. 


ZH: Totally. And what’s funny is, even though I wasn’t thinking about language models, the syllables in “Exit Museum” kind of look like tokens, which are the basis of language models. In language models, text is broken into tokens, which are small groups of characters including some whole words like “hello” and “world,” but also sub-words like “un” and “re” and “ology” and punctuation—GPT-3 for instance had a “vocabulary” of roughly 50,000 of these tokens. This means that language models are generating text at the level of these sub-words – they see a sub-word like “un” and might predict “able”—they’re staggering through and sounding out sentences in a pretty inhuman way. In the early days there were also some tokens that are known to like, send the models off the rails... Some of these “glitch tokens” were inconspicuous and some of them aren’t—if I recall correctly a lot of them were random usernames like, “SolidGoldMagikarp.” 

Anyway, I wasn’t thinking about tokens when I wrote. I wasn’t really thinking about language models then. I saw GPT-2 in 2019 and was only mildly interested in it.

LN: It was still silly back then.


ZH: It was super silly. A friend of mine back then fine-tuned GPT-2 on Ted Hughes’ work and generated hundreds of imitation poems. Hughes is in some ways a perfect poet for this exercise because his corpus is so huge and he has such a distinctive style—a decent chunk of the model’s output could probably be slipped into a Collected and most readers wouldn’t notice. But is the machine good at poetry? No, it’s good at copying recognizable style—and a lot of Ted Hughes poems were like… just OK.   


LN: You mentioned the Boston Public Schools study, where you used the phrase “algorithmic life.” Could you speak to what that means and how that manifests? 


ZH: One of the things that freaks me out is that the more we let algorithmic systems we don’t fully understand control our lives, the more we become like them without knowing it. There’s a kind of slow, scary drift that can happen when we allow these procedures—whether designed by experts or corporations—to shape our choices and our society’s choices.
We can’t avoid the need for some kind of procedure or algorithm for assigning kids to public schools. What troubles me is when only a small group of experts really understands how the system works, and they end up dominating the conversation about it. That has real consequences for everyone else’s ability to weigh in on how the allocation should work.

In that article about Boston Public Schools, the story is especially stark. Boston was at various points in history at the center of some of the most inflammatory battles over how to allocate children to schools. In the 1970s, Boston was  under court orders to desegregate its schools by busing black kids and white kids to different neighborhoods. Now contrast that with a world where all of these choices are made by a small set of people who can say “We ran the numbers, this is the optimal outcome,” and  what can parents do other than say, “Oh, okay, I don’t like it but I guess I just don’t understand why it’s optimal.” So even if they feel like their kid is not getting the chance they deserve, it’s hard to know how to challenge it. That’s one way technocratic expertise becomes incredibly powerful once algorithms make so many social decisions. 

That’s the public policy side. There are also corporate algorithms that dictate our choices even more invasively and with less accountability on a micro and macro level every hour of every day. We vaguely seem to understand the stakes and social consequences—everyone’s always talking about misinformation, addiction, teen depression… We can point to these symptoms but for the most part we don’t really have the vocabulary or the access to say “This system is literally optimizing for something that is in direct conflict with my own interest.” In the “algorithmic life,” a tiny handful of people are writingalgorithms that define our choices and society’s choices, while the rest of us lack the power or information that would allow us to understand what they’re optimizing for. 


LN: Yeah. How is the layman going to interrogate that? 


ZH: Yeah, exactly. Some people say, “Regulation will catch up! We’ll fix this with policy,” and blah, blah, blah… maybe in Europe, but not here in the US, not now. Especially now that I’ve worked in industry for a bit, I just don’t see how regulators are ever going to have the information they’d need to judge whether these systems are operating in people’s interests—and certainly not fast enough to stop them if they’re not.  

One concrete example some economist friends have worked on—kind of mundane but perfect for showing the information symmetry—is Amazon’s “buy box.” When you want to buy a vacuum cleaner, you’ve got the Buy-from-Amazon option or you can buy from a third-party seller. But then Amazon also chooses via some proprietary algorithm which seller to promote in the “buy box” when there are twenty sellers selling the same vacuum cleaner—who gets the coveted “Add to cart” position? Winning that box is life or death for sellers, because almost all Amazon sales go through the buy box. Regulators have been worried that Amazon might be favoring its own products with that box, which would of course be an issue from an antitrust perspective. Amazon says, “No, no, this is what people actually want, we’re showing people the best offer in terms of price, delivery speed, reliability and their preferences—we’re maximizing consumer welfare.” 

I just love this example because it’s so absurd. As outsiders, it’s like, how are we supposed to argue with that? We can try to scrape together some third-party data and build fancy econometric models with loads of assumptions to figure out if Amazon was self-preferencing. But it’s pretty damn hard to tell what’s going on. It’s like trying to audit a casino by standing in the parking lot with binoculars. 
That’s why I think there’s a deeper political problem here, not just a technical problem or an economic problem. It has to go beyond antitrust. We keep trying to fight the tech companies with frameworks that were built for railroads and oil companies. It’s a bit like arguing about 19th century contract law while factories are dumping toxic sludge into rivers, i.e. completely missing the point. Eventually people had to invent an entirely new category—environmental law—when the Cuyahoga River caught fire. Only then did we get the Clean Water Act, where people were like “Hmm, yeah actually it’s kind of a problem that the river is on fire, not because of some other law but because the river is on fire.” Before, they didn’t have the right grounds to litigate on these things. I can only hope that we’ll have the right laws and moral vocabulary for digital platforms in the future—one that allows us to say that even if prices are low and people keep coming back, it may be a problem that we’re all engaging in a system that invisibly optimizes against our long-term interests. 


LN: Can you please explain a simplex algorithm to the Currents readership?


ZH: The simplex algorithm is a method for solving a linear programming problem. It also has some deep historical importance—it was one of the first optimization algorithms used at scale, famously during the Berlin Airlift. It was used to figure out how to get as much food and fuel as possible through limited runways, planes, crews and hours. Its inventor, George Dantzig, was working for the U.S. military at the time. 
I was struck by the social life of this algorithm, but also by its geometry and what it’s doing on a mathematical level—the procedure it describes.  Geometrically, you can imagine there’s a hill you’re trying to climb: say a smooth curve shaped roughly like a parabola, representing how “good” each choice is. Now take a bunch of straight-line constraints (like “you can’t spend more than this,” “you need at least that much of something”) and draw them as lines in the plane. The region where all those constraints are satisfied is a polygon. In higher dimensions, that polygon becomes a many-faced shape called a polytope.

The simplex algorithm lives on the corners of that shape. It starts at one corner (one feasible solution), then looks at the corners it’s connected to by edges, and moves to whichever neighbor makes the objective function go up—like taking a step uphill. Then it does the same thing again from the new corner, and again, and again, snaking its way around the polytope. When it finally lands on a corner where none of the neighboring corners are higher, that’s the optimum for your linear problem. Conceptually it’s super simple. It’s not a fancy algorithm by any stretch. But that jagged slithering from vertex to vertex is what helped optimize the delivery of goods during the airlift.

And so that poem, “Simplex Algorithm,” has one syllable per line, it looks like this snake. And the poem is kind of biblical, about the fall. And the fall is the Berlin airlift. 

LN: Thank you. It’s interesting how snakes recur…


ZH: Isn’t it? Maybe the next book is just called Python. Python for beginners.