Impossible Translations

Galaxy Brain

We imagine other minds all the time, asking what a friend would say, what a rival would do, what someone we admire would make of a problem. The familiar Christian prompt “What would Jesus do?” (WWJD¹) is that same faculty at scale—moral instruction distributed as a wearable meme. Psychologists call this theory of mind²: our capacity to model what another person thinks with the understanding that they see the world differently. It’s central to how we function as social beings.

Parallax is experimental software built on that instinct to imagine another mind. You pose a question, and the model identifies real thinkers whose perspectives might speak to it. The software generates responses from their public record, then lets you put those simulations in dialogue, creating a debate across space and time. You come looking for another perspective and catch a glimpse of your own. By serving perspectives instead of answers, Parallax is meant to branch your thinking, a perpetual entry point rather than an endpoint. The prototype wrestles with what becomes possible—and what we risk—when AI allows a thinker’s work to answer back.

“You come looking for another perspective and catch a glimpse of your own.”

AI’s reach across an enormous archive is its advantage. It can read more thinkers, across more of their work, than any one person is likely to do. But its limit is lived experience: the stakes, subtext, and emotional contours that shape what words mean. Whether we believe AI to be a stochastic parrot³ or something more⁴, a machine cannot escape itself. Yet it fills the gaps with confidence, aping a voice so fluently that you may stop asking whether it’s faithful to its author.

Building Parallax, I would ping-pong between vibecoding and testing, often asking meta-questions about simulation, itself. For example:

In one session, the software selected Jorge Luis Borges⁵, a writer whose work tugged at authorship, copies, and the way meaning shifts as words pass from author to reader. In the simulation, Borges’s first move was to argue against his own existence. A machine trained on his work, he “said,” does not reproduce him. It is a “new author who happens to wear his syntax.” To quote the simulation, it “has all his sight and none of his darkness, the blind librarian’s whole catalog without the blindness that shaped it.”

Borges becomes, in the AI’s synthetic words, not the author but the source material.

I was impressed. It was almost too perfect that Parallax chose Borges. But my worry spiked because the simulation felt good. The response wasn’t dead-on Borges, but I recognized him in the puzzle-like performance of the words. Imagine asking the real Borges a question, then asking Parallax, and finding the simulation’s answer more useful and alive. What then? Either the machine understood something beyond the writer, having absorbed his work whole, or it understood me, and contorted itself to resonate. From the inside, the answer that best reflects the author and the one that best fits the reader may be impossible to tell apart. Could you resist the latter?

The deeper problem is that ease is persuasive. It’s human to mistake clarity and confidence for accuracy. A simulation is fluent by design, and fluency is not evidence⁶, however seductive. When a plausible copy is easier to absorb than the source, we have less reason to open the original. Perhaps we won’t abandon it all at once, just lose, over time, the reflex to return. At what point is the machine doing our reading for us? Worse still, will we fail to feel what we gave up because the answers sound right? I can feel Jean Baudrillard’s hot stare: that we become “devoted to the spectacle of thought⁷,” at the peril of thought itself.

“Fluency is not evidence, however seductive.”

That downward pressure feels real to me. But it assumes so little of us, as if we were only passive readers. As Borges might posit, even a writer’s own words don’t hand you the writer. Hermeneutics⁸ is the study of how we interpret texts. It begins with the observation that the chasm between writer and reader has always been there. The words stay fixed, but every generation asks something different of them. We never encounter another mind directly, we construct one through our own era and personal preoccupations.

Parallax is a prosthetic for reaching across that chasm. It does not spoil a pristine process. Old texts reach us changed, by translation, by the slow shift in what words mean, or by who has added commentary, which is why they reach us at all. The philosopher Walter Benjamin called that survival a work’s afterlife (Fortleben⁹). Modern versions of The Odyssey list both Homer and their translators. Turn to Richmond Lattimore’s book if you prize fidelity to the original Greek, or sit back to watch Christopher Nolan’s IMAX version, despite it being panned by Emily Wilson¹⁰, who herself rendered a celebrated, contemporary translation.

Of course, Parallax takes a more radical step than translation: it lets a work imagine forward from its own moment into yours, in words its author never wrote—without their consent. Because AI models can so easily make these impossible (and transgressive) texts possible, I wanted to explore how to frame these hallucinations rather than shun them outright. When building Parallax, I approached from three different vantage points. As a user, I wanted to know whether the simulation tool could be useful. As an artist, I wanted the software itself to be a provocation: should we simulate at all? As a designer, I wanted to reconcile that discomfort through the interface’s terms of engagement. How can friction be a statement?

Though Parallax’s input box may look familiar, there are clues the software leads somewhere different. The banner, “For Questions without an Answer,” runs counter to what we expect from chatbots: a shortcut to resolution. For additional orientation, users can click an animated loop of Rodin’s Thinking Man, which opens a text-message from a narrator I dubbed the Speculative Librarian. I usually bristle at anthropomorphizing software, but here the slowly typed message–“I am not here to hand you conclusions”–evokes a choose-your-own-adventure game rather than a productivity tool. The Librarian isn’t there to deepen the illusion, but rather gives the game away. It introduces Parallax as a simulation machine and nudges users toward an open-ended question.

Unlike other simulation tools like Character.ai¹¹ or Delphi.ai¹², you don’t arrive knowing who you want to talk to, whether an anime girlfriend or Lawrence Lessig. Instead, Parallax uses your question to have the thinkers find you. The most useful perspective may come from someone you wouldn’t have thought to ask. Ask about creativity, and you might encounter a painter, a musician, or a psychologist, each bringing a different vantage point and unique metaphors. Once you choose a thinker, the software labels subsequent perspectives “counterweight” or “lateral.” The panel is built with intention, bringing together voices that challenge the first perspective or approach the subject sideways. As “thinking software,” Parallax is idealistic (perhaps even Pollyanna-ish given the trajectory of AI). It makes room for competing perspectives rather than blending them into a single response disguised as a view from nowhere¹³. Making something of those perspectives requires active curiosity. Or, as the Librarian puts it, “you bring the thinking.”

Once a thinker is selected, you encounter something I called the Baudrillard slider. The dial starts with the simulated text, the hyperreal, then moves past the Speculative Librarian’s reasoning for choosing a thinker, and then finally ends with primary sources that put the referent back. As you drag toward citation, the dithered portrait sharpens into a face you can make out. You can’t have both. The interface either serves you smoothly rendered text or a clearly rendered portrait. The slider is an artistic incentive to discover the human behind the simulation. The dithered portrait is a reminder of what the simulation’s fluency threatens to obscure. The danger isn’t simply that Parallax gets Borges wrong, but that the simulation becomes a substitute for the human, and we stop reaching for Borges at all.

The interface also needed to make the generated language feel present rather than natural or neutral. I spent hours trying, and mostly failing, to polish the familiar AI register out of its responses (please never say “load-bearing”). I couldn’t consistently make the simulated text match the thinker’s style, but if it sounded too much like AI, readers might dismiss it as slop before noticing what was valuable.

Where AI watermarking¹⁴ tries to indicate provenance by hiding a signal inside the text, Parallax names the hallucination outright. All generated text is set in BIZ UDMincho¹⁵, a Universal Design font meant for Japanese that contains Latin glyphs. To an English reader, it’s perfectly legible but subtly off—the kerning runs too wide, for example. Perhaps if the text announces itself as unfamiliar from the outset, a reader can spend less time asking, Am I being deceived?, and more time asking, What is this helping me notice?