Attention is All we Have
On July 20th, 2020, I posted this on Twitter:
btw in light of the various OpenAI stuff I’ve revised my estimate of when idiomatic fiction translation will be practical from “not within my lifetime” to “within ten years as a special case of something else, and five if somebody decides there’s money in it as a special problem”
In 2021, I posted this in response (I dimly recall) to one of the statements by various Japanese publishing executives expressing optimism that advancements in AI would allow them to automate some or all of the translation process, which would in turn make it economically feasible to automate translation of older or more niche titles:
The dream of manga industry cheapskates that they will someday be able to replace translators and editors with machine translation tools has a gigantic flaw: If they’re right (and they might be), this same family of tools will also largely replace the artists themselves.
OpenAI’s DALL-E already creates images from text descriptions with an impressive understanding of stylistic nuance. https://openai.com/blog/dall-e/
Meanwhile, there’s Gwern Branwen’s research into using StyleGAN to generate plausible anime-style faces. https://gwern.net/Faces
I can think of a few weird endgames for this trend—and I’m not even particularly smart or imaginative—but in none of them do I see creaky old traditional publishers being the big winners in a rush to automate creative production.
It’s hard not to feel a little proud of myself, here. Five and a half years later, my prediction about idiomatic fiction translation was if anything slightly conservative. I feel especially vindicated, though, by the two tweets at the end of the thread:
My gut here is that if you, a publisher, think you might someday need to make the case that there’s inherent value in the fact that the thing you’re selling was created by a human being, you should start preparing to make that case now.
I should also be clear—I don’t think this will be a “first they came for the translators” type of deal. By the time machine translation is “good enough”, related tools will also be so good as to make machine translation the least of anyone’s problems.
What I didn’t anticipate—in fact, what I didn’t even consider, so far as I can remember—is what my own orientation to that technology would be when it came.
The original ChatGPT, based on GPT-3.5, was good enough at translation to make my eyebrows go up, but the introduction of GPT-4 was when I realized the future was bearing down faster than I was ready for it. I even have a screenshot of the specific moment.[1]
Me: I’m going to give you English words. I want you to define the word for me in Japanese, phrasing your definition as though you’re a Yakuza thug shaking down a local shop owner for protection money. The first word is “book”.
GPT-4: おい、この店のオヤジ、聞いてるか?「本」っていう英語の言葉を教えてやるぜ。それは「ほん」と日本語で言うんだ。理解したか?これからは俺達に保護料払って、その「ほん」を売り続けられるようになれよ。でないと、お前の店はめちゃくちゃにされちまうぜ。(Hey, old man, listen up. Let me tell you about the word ‘book’. That’s ‘hon’ in Japanese. Get that? So from now on you’re gonna keep selling books so you can pay your protection money. Otherwise this shop of yours is gonna get real messed up.)
Me: “pen”
GPT-4: 聞け、この店のオヤジ、今度は「pen」って英語を教えてやるからよく聞け。それは日本語で「ペン」だ。お前、その「ペン」でちゃんと保護料の小切手書けよ。書かねえと、お前の店に不幸が降りかかるかもしれねえからな。分かったか? (You listening, old man? I’m gonna explain the English word ‘pen’ to you. It’s ‘pen’ in Japanese. You’re gonna use that pen to write us checks for your protection fees. Otherwise who knows what unfortunate things might happen to your place, you got that?)
Me: “ontology”
GPT-4: なあ、この店のオヤジ、難しい言葉も知ってるんだぜ。「ontology」って英語を教えてやる。それは日本語で「オントロジー」、つまり「存在論」ってことだ。哲学の話だが、俺たちヤクザだって教養があるんだぞ。だから俺たちに保護料払わねえと、お前の店がどんな存在になるか、考えてみろよ。分かったか?(“Hey, old man. I know some tougher words, too. Lemme explain the English word ‘ontology’ to ya. In Japanese it’s ‘ontorojii’, or ‘sonzairon’--like theories of being. This gets into some philosophy, but Yakuza like me got brains too, see? So you better think about what state of being your place is gonna be in if you don’t pay up!”)
What gave me goosebumps about this was not GPT-4’s ability to define nouns, or even its generically correct but unimaginative take on Yakuza thug Japanese, but rather the recognition that this was a bit, that it was supposed to be funny, and that having our thug jump from explaining “book” and “pen” to “ontology” represented an opportunity to tell a joke[2].
This is what I think about when I see the replies to the LLM headline of the month. “It doesn’t know anything, it doesn’t understand anything, it’s not thinking. It’s just regurgitating its training data and doing statistical pattern-matching.” The problem is not so much that I disagree with these statements per se.[3] It’s that they come to the conclusion that because there is clearly no knowing, understanding, or thinking happening here, the output is of no interest or value. Whereas to me, that proposition (that fluent manipulation of written language is evidently possible without knowing/understanding/thinking) is at least as interesting as the alternative. Either the machine thinks or it doesn’t, but to me either case is astonishing.
I’ve spent most of my career translating from Japanese into English. It is lonely, intellectually grueling work. Like many worthwhile endeavors, translation feels perversely harder as you accrue skill: The more nuance you perceive in the source text, the more you feel obliged to include in the target. The early-career translator gropes around in the dark for grammatically complete understanding of a passage; Later they will hold complete understanding in one hand while rummaging through the junk drawers of their own mind with the other, hunting for the apt phrasing that they’re pretty sure was around here somewhere. There never comes a point where you finally stop struggling.
Accruing skill and expertise also changes the feel of the work, even when the work remains difficult. It’s granted me something I would describe as “empirical confidence.” After close to 20 years doing it professionally, I understand the full shape of the problem, and I can explain in concrete detail what separates excellent work from adequate and adequate from shoddy, because I have produced all those levels of work myself. I have also edited the work of other translators, of all degrees of quality, and seen in those texts the universality of my struggle.
At least as valuable to me as the skill itself is the appreciation it’s given me for skill in general. In becoming an expert at this one thing, I have an intuition for what expertise means for anything humans do, and that the only way to attain it is with long hours spent attending to the work at the limit of one’s ability.
I cannot overstate this: The yakuza dictionary conversation shocked me, and that shock was specifically informed by my expertise. GPT-3.5’s translation abilities were impressive, but even then it still felt like a recognizable advancement over something like Google Translate. GPT-4, however, could riff. I could ask it for ten different ways to phrase some incredibly specific utterance—some arrangement of words never spoken or written before in all of history—and it would immediately oblige me. If two of the options missed the point, five were tin-eared, and the other three merely mediocre, the fact of their existence at all was still astonishing to me.
This was when I knew that some people were going to start asking what the point of paying a human translator like me was. I needed to know what my answer for those people would be, even if it was “I don’t know.”
This is what I arrived at, and I still believe this (even though LLMs are much better at translation than they were in 2023): The point is to have a responsible party.
The translation credit in a book is the attestation of whose attention produced the translation, but its presence also implicitly communicates that a translation is something for which responsibility ought to be taken. When someone reads a translation, they’re implicitly conceding its abstract value. They didn’t choose to acquire the language the text was originally written in, but in benefitting from the translator’s effort, they’re agreeing that there was value in someone acquiring that language and producing the translation.
Hardly anyone actually needs to be convinced of this; it’s intuitive. Part of the broad public suspicion of and antipathy toward AI is, I think, predicated on an axiomatic understanding of attention’s value. At least part of what some people find repugnant about AI is its confusion of their ability to estimate others’ attention.
It turns out that this stymieing of effort-estimation drives people fucking nuts, and the more highly you value things like “thought” and “craft” and “care” the more nuts you are likely to be driven. Attention-fraud now feels omnipresent. The crime of giving a shit is punished with tenure as a wary professor constantly hunting for academic dishonesty, and the cheaters are everywhere—in the coworker whose emails have suddenly gotten suspiciously slick, the cousin whose best-man speech includes one too many negative parallelisms, and the cafe whose new menu glistens with the queasy sheen of generated art. When we find ourselves fooled or diverted by slop it feels especially contemptuous, and the contempt invites immediate transformation into a hater: “Fuck me? No, fuck you.”
It’s easy to sympathize with the principled refusenik, too. Wendell Berry’s refusal to buy a computer in 1988 remains difficult to argue with. Berry’s position was not a superficial one, so anyone who wanted to convince him that he was wrong had to argue with his principles, and—notably—they did not succeed. As far as we can tell, he never did buy that computer. His convictions about what’s good in life were well-articulated and strongly-held, and computers were not on the list, both because of how they were made and what they were used for. There are similarly principled arguments against both forms of and uses for AI, and they’re just as hard to directly refute.
And yet! Despite both grinding my teeth at unattested slop as much as anyone and having at least some feelings in common with the refusenik, I don’t share their conclusions.
Translation is the single area in which I am an expert. In that field alone am I qualified to have an opinion on what constitutes a significant development. It’s from that position that I can say with confidence that the yakuza dictionary moment represented a significant development.
I suspect that most working translators would agree with a neutrally-phrased version of my shock—something along the lines of “A computer being able to perform this category of transformation on natural language is somehow germane to our profession.” I also know that their feelings about that proposition are diverse, ambivalent at best, and mostly negative. Not all of them—maybe even a small minority of them—are as fascinated as I am.
But Wendell Berry isn’t the only guy with principles.
The translator spends their days submerged in the problems of our languages, an eyewitness to their mystery and uncanniness. The work is a combination of mechanical/analytical and intuitive/creative thinking, and excelling at it requires both, all the time, relentlessly. Can I be blamed for feeling so deeply that language is the greatest and strangest human tool? Well, that is how I feel. If you ask me, the atlatl, the wheel, the EUV lithography node—all must bow to the verb.
And now we have language machines.
Cosmopolitan, pluralist values follow straightforwardly from my profession: I think it’s good that there are many languages and cultures, and I think translation—insofar as it facilitates communication among them—is a specifically worthy pursuit in the cosmopolitan-pluralist context.
I’m also broadly curious, maybe pathologically so. This is, I admit, more of a personality trait than a principle, but a principle that I do hold related to that curiosity is that it’s better to change my mind when new facts arrive than it is to hold out for vindication of whatever I thought before. I don’t like to find myself arguing with reality, even when that reality isn’t what I hoped for or expected.
When I made my predictions about translation, I wasn’t hoping for the state of the art in machine translation to change sooner than it did—if for no other reason than it was how I made my living. But when it did change, I needed to know as much as I could about the nature of that change.
And so drawing on all my expertise in translation, I can report that in 2026, translations produced by AI are—in broad, general terms—remarkably high in quality[4]. I know this because curiosity has compelled me to constantly test AI translation capabilities, repeatedly and at length, scrutinizing the output closely.
My statement back in 2021 that truly good machine translation would “also largely replace the artists themselves” wasn’t quite right, though, and I wish I’d been a bit less eager to drain a three on cheapskate Japanese publishing executives by insisting that they were threatening their own artists‘ high status. I should have elaborated more on the second half of my point, that a machine capable of producing reliably good translations would also be good at a lot of other things. I can’t insist on credit for calling a shot I didn’t take, but either way, these things remain true: The machines have gotten very good at translation, and the daily headlines about the machines are not about how good at translation they are.
Very suddenly we live in a world with a new dialectic axis, with synthetic language at one pole and human at the other, and it happens that another principle I hold with Berryish stubbornness is that dialectics are good.
It was dialectic tension that led me to ask myself what the point of a human translation would be when machines were functionally as capable, and to then come up with an answer that was something other than a cope. My understanding of the nature of my work is now richer and deeper purely thanks to the necessity of orienting it in this new dialectic.
We now have the privilege of arguing the value of human-produced language against its synthetic counterpart. We get to establish its value relative to an alternative instead of being obliged to play the only game in town. Now there’s a whole spectrum of possible uses to debate and bicker over, balances to strike, policies to articulate, etiquette to establish. This is where the refusenik approach loses me entirely, because there is simply no fucking way I am sitting this conversation out, nor remaining ignorant on principle of its true stakes.
I find it irresistibly thrilling how muddled the status markers of AI use currently are. The taste landscape is hotly contested and the outcomes far from settled. I find Venkatesh Rao’s framework for taste very persuasive, with the position of the Connoisseurs (“who rehearse and reinforce established cultures of taste and ritual disputation through their behaviors, perhaps making incremental advances”) tenuously held at this moment by the refuseniks.
Connoisseurs in Rao’s framework don’t get to be Pioneers, “whose transgressions have the power to undermine taste cultures themselves. The tasteless are judged by taste cultures. But taste cultures themselves are judged by taste pioneers.”
The taste culture of today’s Connoisseurs is vulnerable to judgement by a Pioneer using AI to help them make something unignorably, un-shameably good. Maybe this will never happen, and in two or three or ten years this blog post will stand as a record of my increasingly-wrong guess.
Even as I write this, I feel nervous about predicting the untenability of the refusal position. So far the huge preponderance of evidence is slop, and thus not in my favor. Most of the people excited about AI seem incapable of communicating their excitement in ways that aren’t witheringly offputting, and their bad vibes are exceeded only by those emitted from the people making AI.
For me, though, the surprise of all of this is too great to admit “this will amount to nothing” as a response. The appearance of synthetic language surprised me. What it suggests about what kind of thing language is surprised me. The pace of its improvement and refinement continues to surprise me. The marketplace rejection of machine translation in my little niche has been a very pleasant surprise. Paradoxical though it may be to say so, I expect the surprises to continue.
I’m not going to let AI write my emails or parent my children or otherwise mediate my humanity, and I’m as baffled as anyone by the various dorks and dweebs racing to do exactly that. But I’m also not going to refuse participation in the emerging synthetic-human dialectic, and I spend a lot of time thinking about what that participation should amount to and look like.
It’s not much of an exaggeration to say that David C. Porter’s essay about music genre and generational shifts, The Discovered Country, reconfigured how I think about artistic work and aesthetic context. He contends that
there is no “undiscovered country” anymore, in any major art form. As I alluded to in my initial posts, the history of art from, let’s say, around 1850 to the turn of the millennium is one of the systematic probing, and eventual overcoming, of every conceivable boundary and convention regarding what art can be. There is, at this point, no line which has not been crossed, no limit which has not been pushed to the point of arbitrariness.
His thesis is that the cultural context that enabled the existence of the avant-garde no longer obtains. To me, understanding this was like seeing the arrow in the FedEx logo for the first time; the formerly invisible became glaringly and permanently obvious. The shape of the continent is fully known, but this doesn’t mean it’s impossible to say or create anything new.
Porter maintains that for the artist trying to say something meaningfully new, the only alternative to now-doomed searching for the avant-garde is “just doing stuff”:
It is no longer the case, as it was in the last century, that you can write yourself into the history books, or at least into a sustainable career, by simply finding some clear and obvious taboo and violating it—there aren’t any left. […] There’s no shortcuts left—you have to actually do the work. And doing the work means… just doing stuff.
[…] To be a good artist now demands a different sort of confidence now than it did when one could get fed up with the tastemakers, pen a manifesto, and go from there – everyone is a tastemaker now, and everyone is fed up with each other, and so none of it matters. You have to start by actually doing something now, by finding a way to make something that you can believe is better than nothing, that is not pointless, that is not just an empty repetition or regurgitation of what’s come before.
The nature of how contemporary AI is created means that “an empty repetition or regurgitation of what’s come before” is precisely the kind of thing it’s very easy to create with it, of course. To do otherwise requires what good work has always required: Care, effort, attention.
In 2026 we are only barely beginning to figure out the cultural etiquette and ethics around the use of synthetic text and AI, but we already instinctively understand the contempt inherent in any communicative act where we are expected to pay more attention to the exchange than the other party is paying[5]. By now it’s clear that when a person or organization is excited to be on the advantageous side of such asymmetries, they are fucking up big-time, and will be correctly hated for it. It is nevertheless tenable to both insist on the value of our own attention and others’ while still taking AI seriously.
My strong conviction about how something unignorably good or useful will be made using AI is this: It will not involve in a direct way the output of any text- or image- or music- or video-generation model; rather, it will come from someone who uses AI to fashion better creative tools for themself and who then uses those tools to create something they otherwise couldn’t or wouldn’t have. Such a work will seem obviously the result of its creator’s close attention because that is what it is.
I’m being a little coy here. The “better creative tools” I’m suggesting will be Things on The Computer, because making computers do things is (somewhat circularly) the area in which AI’s progress is most starkly undeniable. It’s very hard to overstate just how profoundly the process of programming a computer has changed (i.e., “utterly and irrevocably”) and how quickly that change has happened (i.e., “the past twelve months”). For about the price of a Netflix subscription, anyone can now hire what amounts to a tireless computer programming prodigy to help them do and make things with their computer.
What can computers do these days? I don’t know, man. All kinds of stuff. But if I would like to see a computer do something that it doesn’t already do, as of right now all I need to do is ask, and it happens. For the first time, computer attends.
Attention is what we have to give. It is what our lives are made of; the currency in which the spending of our life is most directly denominated. Now there is a machine that can both feign attention and multiply it. This is bad, and good, and in any case persistently disorienting and weird. But it is happening. This is all really happening. Whatever else may be true about AI, or the people and organizations building and improving it, it is also true that it is real. Things that were categorically impossible for computers to do five years ago are now mundane. This is real life.
We have to pay attention.
The parenthetical translations from the Japanese are de moi, and were composed without any LLM input or assistance. ↩︎
I don’t have an exact date for the chat itself, but I can be certain it was no later than May 27th, 2023. For comparison, here is Claude Sonnet 4.6 performing the identical exercise on February 27th, 2026. ↩︎
For one thing, I can’t really disagree with them until the person making the assertions provides definitions of words like “know,” “understand,” and “think” that are rigorous enough to argue with. I have never seen this happen in the wild. ↩︎
By “high in quality” I mean something fairly specific: If I were an editor paying for work from a freelance translator who turned in work identical to what a mid-2026 model produces as reliably as such a model does (even accounting for how much longer it takes the human to work) I would consider them a valuable professional contact. From 2014 to 2020, I was an editor paying for work from freelance translators, so my judgement here is founded on experience I have making this precise decision. ↩︎
Maybe there’s a golden rule-ish formulation of this: “Don’t serve slop you wouldn’t eat yourself.” ↩︎