How this came about

Volume One: Neruda

I had no intention of doing any of this.

In the mid-2000s I wandered into a bookstore on vacation. I had been trying to learn Spanish and The Poetry of Pablo Neruda caught my eye. Months later I actually started reading it, and loved the whole (very long) book. The poems were already translated into English, but ninety or so were printed alongside the original Spanish, so I decided to practice my Spanish by translating those.

It did not go well. I gave up. Translation is extremely tedious and difficult.

But I decided not to be so hard on myself. I started writing whatever came to mind as I read the Spanish. I knew enough of the language that some meanings were impossible to avoid; the rest of the words were only sounds and rhythms, and a sound can trigger any word you know, filtered through whatever else you happen to be thinking about at the time. As a cognitive scientist I found myself watching my brain as it worked.

Then my brother suddenly died.

My whole world and worldview changed. The grief was immense and lasted many years. The process became both a distraction from it and an extraction of it, because the filter it ran through was now an entirely different filter.

Only after that did the thing get a name. I had been calling it mistranslating, on the model of mishearing — but a mishearing is an accident, and none of this was accidental. I settled on transbliteration: neither a translation nor an obliteration, but both.

The Numbers of Nature / The Nature of Numbers

after Gautama Cristo

Nature
    In particular
Numbers her representations
    Using the same
        1
    For Jesus and Christ
        According to recorded word of mouth

She assigns them
    7 to sadists
    8 to eaters
    10 to tentacles
    80 to ears
    40 to rivers
    20 to life
Like empty shells
    173
Awaiting hermits
    357

    Without asking
She touches these natural numbers
    Ignoring the irrational ones
        Like petals too-well-placed
And swims with them in the ocean of rationality
    That divides the Ordinary from the Cardinals
And by adding nothing
She makes them whole

While Nature provides these numbers
    For the best and the poor
        The clean and the dirty
    For Whites and Blacks
    For bloody assassins
    For enriched victims
        And those of sky-sent strikes
She leaves to the imagination
    How to represent the death of a brother

    As fewer and fewer transcendentals arrived on the shore
Men began
    To study the colors
        Of a drip of honey
        Of the lips of the anus
    To search with skepticism
    To hope for the knowledge
        Of what matters and what doesn’t
        Of a technique for laughter
        Of how to go farther
                Unlimited
            Without sleep

We find distasteful what the ages believed to be sacred
Our humanity succumbs to the cynicism of mortality
We are incapable of ignoring the mirage
    That we can safely subtract Nature’s numbers
    And cube them for our amusement

Our progeny will question us
    Confront us with a lyric
        Uniting our voices
    Confront us with the fractions
            Left by our living
        Their roughness stuck in our throats
            Such that we cannot swallow
                The Oneness of our Nature

I was filled with nostalgia and memories, and trying to understand my brother better. For example, there is another poem for him called Ode to George Carlin, after Neruda’s Oda a Walt Whitman.

Transbliterations, Volume One: Neruda runs to 170 pages. I shared it with a few friends and that was all. It got me through my grief. If you can’t turn tragedy into art, then tragedy wins.

I did build an auto-tweeter that posted the entire book to Twitter. The hard part was cutting the poems into sensible 140-character pieces. The solution cemented the style of unpunctuated tabs that had emerged from the transbliteration process of going line by line, not knowing where things would end up. That shape could also be used to let a machine break them into meaningful chunks automatically. I deleted all of those posts when I started using the account again. Maybe someone has an archive.

Volume Two: Shakespeare

Three months after I finished Neruda, I wanted the practice back, and needed something to point it at. Shorter poems would be easier. Horace’s odes? Shakespeare’s love sonnets? But early modern English would not have enough noise. So I made it Spanish first. Each sonnet went through Google Translate as it was then, came back mostly nonsense, and I ran the Neruda process on the nonsense.

Then my wife suddenly left me.

My whole world and worldview changed. The grief was immense and lasted many years. The process was once again both a distraction from it and an extraction of it, because the filter was once again an entirely different filter.

So these poems were written in real time, over 1,818 days: from 841 days before the day I learned my life was a lie, to 976 days after it. Just under five years. Revisions went on for another 99 days after the last poem was begun. I have wondered since whether some part of me already sensed it, and that’s why I chose love sonnets? In any case, many now came through my brain as hate sonnets, and after that as the sound of someone trying to get a self back after losing the two most important people in his life, and then of meeting people again (I had not dated in this century), and eventually of finding love again.

The twenty poems on this site are that arc in miniature. The book’s six Parts (Sensing, Knowing, Recovering, Loving, Rebelling, Accepting) are all represented, and reading the twenty in order is reading it happen.

The pictures

All 154 poems were finished long before the AI image models arrived. As with Neruda, the work had gone to a handful of people and no further, and I had only vaguely played with the idea of publishing. But now I could see a way to illustrate the poems that might be more interesting to somebody else than just reading my therapy sessions.

It took years of trying. Disco Diffusion in April 2022 was not good enough, and neither were Latent Diffusion, JAX CLIP guided diffusion, simplified Disco Diffusion, many versions of Stable Diffusion through Deforum, nor SDXL Lightning. What finally worked was FLUX. But there was so much still to explore. What would be the medium of the image generated? And how would it evolve over the course of a poem? As part of the experimentation, I had every medium as a possibility. Early on, I liked woodblock prints, but when images with “sculpture” were produced, suddenly a whole new world opened up. Why stop at two dimensions? I could imagine a real thing in the world, a sculpture, standing in a garden you could walk through. Walking through the garden would be like walking through the poem.

What should the sculptures look like? I had no idea. I am at best a photographer when it comes to art, and certainly no sculptor. But I could search the entire space of sculpture, including regions it would be impossible to sculpt in the real world. So I built a system that generated thousands of candidates, weighted the properties of the prompts by which ones I chose, and generated thousands more. I call these Aesthetic Experiments, and you can run your own in a Google Colab if you like.

In June 2025 JBR Visuals interviewed me for their Unfiltered Conversations series. Part of it is a walk through the public version of the Colab notebook on lines from Sonnet 154: the random prompts, the voting, the seeds. It is followed by an earlier iteration of the sculptures and voice for Sonnet I. Near the end I explain transbliteration and the plan to turn the sonnets into a sculpture garden you could walk through, before any of them were 3D.

Diffusion

Most AI image generators work by diffusion. An image of pure noise is made, then run through a neural network in the context of a text prompt to produce a slightly less noisy version. That version is fed through again for a less noisy one, and so on, until what is left is not noise at all but something like what the prompt intended.

When these techniques came out I thought immediately: that is exactly the transbliteration process. I give myself a noisy text and my own natural neural network tries to make it less noisy, over and over, until I have the poem my emotions were prompting.

Diffusion also depends utterly on the noise it starts from. That first image is random, made from a seed that sets the values of every pixel — a different seed is a different starting noise, and so a different final image. On top of a nearly infinite search through prompts, then, there is another nearly infinite set of images for any one prompt. That is why there are so many versions of styles here, and versions of versions.

Every image in a set on this site was generated from the same seed. That is deliberate: holding the noise still is how I find the styles that are most flexible, most reactive to the line they are given. It is also why you will often see the same pose, or the same spacing of objects, from one line to the next. They all started from the same noise.

I have always put some art in my science and some science in my art. Working on this I feel like a nineteenth-century explorer, finding art inside these models that has never been seen. Every day I see something that astounds me and want it to be real — to make a reality-twin of the digital, when nearly all of this type of work runs the other way.

Into three dimensions

A digital object gets me halfway to the real thing, to a garden you can physically walk through. An image cannot be walked around.

In the summer of 2025 there was an open call on The HUG, using Illust.Space, for an augmented-reality artwalk in Vancouver, British Columbia. That sounded intriguing, and it made sense to do things in AR before trying to 3D print anything. So I went looking and found the image-to-3D models — TripoSG and Hunyuan3D, and later TRELLIS. I was hooked. Some of the results have been shown by the Balanced Forces team, to whom I am very grateful.

The sculptures on this site are those models. What they are standing in is a maquette; the colophon sets out the terms of the real garden, which does not exist yet.

Zen Materialist

The sculptures are shown under that name. It is the handle I took on social media a long time ago and it has become my nome d’arte by use rather than by decision.

It came about decades ago, when I was into Zen. I traveled around the world in my early twenties and saw how much needed changing — how much there is that you should not be Zen about, and how the material world we have is a choice that can be changed for the better of everyone.

As a personal religion it deals, like most of them, with desire. I summarize it as: it’s okay to want. It’s also okay not to have. A friend of mine summarized that as “So, easy come, easy go?”

Is this art?

People like to compare AI images to the arrival of photography. Instead of sitting for hours while an oil painter does your portrait, along comes a guy, almost always a guy, who makes your image in a second.

I think it is closer to the arrival of the synthesizer. All music is waveforms, and all of a sudden waveforms could be made digitally. I am not a drummer, but I can make waveforms that sound like drums, or a violin, or a guitar, or a saxophone — and waveforms no physical system could ever produce. In my basement. Endlessly.

That did not take work away from violinists or guitarists or saxophonists. It let people like Tangerine Dream make new music (tracks, as they say, not songs) that others enjoy. A new genre, enabled by a technology and powered by people.

Every generated image is likewise just pixels, just numbers, and numbers can be modeled. A musician sits down at a company’s synthesizer and works every key, button and knob to find out what it will do — prompting the system and seeing what comes out. This is the same: play with the input, look hard at the output, and follow the leads worth following.

Afterwards

Only afterwards did I find out that other poets had been doing things like this for a long time, and I did not know about any of them when I started.

Celia and Louis Zukofsky’s Catullus (1969) follows the sound of the Latin as well as its sense. In The Practice of Poetry (1992), a book of exercises from poets who teach, Charles Bernstein suggests translating a poem by what it sounds like, without a dictionary, relying on “what your ears hear,” both from a language you know and from one you don’t. That is close to where I started with Neruda. Erín Moure called her versions of Pessoa transelations (2001). Jen Bervin erased Shakespeare’s sonnets down to new poems in Nets (2004), and in The Sonnets (2012) 154 poets each rewrote one of them. An essay on Bervin’s work and others’, that I only recently read, even mentions her “not-entirely-obliterating hand in her collection of sonnet erasures.” People were also already making poems and games out of Google Translate’s mistakes by the time I started on Shakespeare.

Robert Lowell’s Imitations (1961) is the nearest published relative. I had never heard of it while I was writing. It reached me long after the sonnets were finished, when I was describing what I had done and the person I was telling said it sounded like that book.

There is one thing I did know, though I didn’t consciously think about it at the time. I had studied Eugène Ionesco about 25 years earlier. In La Cantatrice chauve, Ionesco copied sentences from a primer for learning English to make an absurdist play. I did not think of him once while I was writing, and only thought about it again recently. Maybe it was in the back of my mind. Maybe it was what let me be absurdist and surrealist without concern. But it was not a conscious connection at the time, and every other name here I met after the poems were done.

So my process was not totally unique. It was useful to me at the time, and I was curious where what I made lands next to previous work. As a data scientist I was interested in the numbers, so I measured it. The results are on the Analysis page.