The A.I. Thread of OMG We're Being Replaced


Like it’s tautological that translation has multiple dimensions for it, and that human translation can accidentally or intentionally alter nuance.

But machine translations destroy nuance constantly. In focusing on a direct word for word translation major errors of syntax or meaning can occur. It’s why things like ‘take this song, run it through Google Translate a dozen times through different languages, and see what comes back’ can be so amusing.

Which isn’t to say there isn’t a use for machine translations. If you’re reading a board at a train station, or a sign, or one of the many other pieces of function first communication mediums we experience, that can be a perfectly serviceable and good use. Sure it can get things wrong, but if you’re trying to read a map without anyone speaking your language, it certainly beats the alternative.

However for anything like literature or more prosaic communications a human translator is infinitely superior. That isn’t to say all translators are equally good, or that a specific translator could create a result that is worse than the machine translation (even possibly intentionally), any half decent translator is going to be able to use nuance, euphemisms, vernacular, and tone in ways that provide a far better outcome.

And of course there are multiple valid translation objectives. Yes a direct word to word meaning, with minimal accounting for time and place specific changes (this word may translate directly as shoe, but in the time it was written would more correctly be conveyed as step stool in our modern terminology), is a valid choice, as can be one where literal meanings lose the intent so translating for intended message/meaning. Or you could try and convey style, or tone, or emotion, but doing so requires inserting or changing the words in more drastic ways. Because we aren’t 5th century BCE Greeks hearing Homer or something.

If you’ve ever studied the nuances of how classics and ancient literature are translated, or wondered why there are so many variant translations of old works (we already have translated Thucydides, why is someone making a new one?) then you would know how such things are products of much scholarly debate and effort, precisely because there is no one perfect way to translate things, and different methods can produce different and yet valid results.

Yes, a famous example is this:

Around 1830, Paris and Washington entered into heated talks over an indemnity, and Jackson proposed extraordinary measures to Congress. A message sent to the White House by France began as follows: “Le gouvernement français demand” which a secretary translated as “The French Government demands.” Jackson swiftly and forcefully replied that if the French government dared to “demand” anything whatsoever from the United States, it would obtain nothing.

Demande means ask in French, not demand.

But this is exactly why these new systems are so much better at translation, because they aren’t just doing a word-by-word translation. They are taking input in one language, turning it into an abstract embedding that represents the “meaning”, and then decoding that meaning into the target language.

So, @CraigM basically said what I would answer, but yes - about x1000 more than I would trust the current crop of AIs.

And they still destroy nuance. The big difference is that now they mostly seem competent, so when they do trip you up, you probably will never notice, because it will be totally confident that it the translation is X…

Sorry, you are right that this is not the right translation. What it should actually say is Y.

Sorry, you are right that this is an inaccurate translation. What it should actually say is Z.

I apologize for that mistake. What it actually does say is X…

Wars have been fought over the nuance of the meaning of one word translations.

And no, I am not being hyperbolic. One could look at the entire history of the European wars of Religion, culminating in the 30 years war, as the consequence of translational choice and accuracy, where extremely minor deviations led to significant theological shifts, with deadly consequence.

Obviously nobody is going to war over the meaning of a specific phrase from Sun Tzu or Chaucer, but small errors can have meaningful impacts. And when we are talking about any matter of significance, be it diplomatic or literary, my response to AI being the arbiter is an emphatic Fuck No.

Absolutely. With human translators.

What I’m saying is that there’s not necessarily a basis for the assumption that AI translation is going to be worse than human translation.

Certainly that’s been the case in the past, but I don’t know how long that will hold, or if it even holds with the current state of the art. I definitely don’t think that it will always be the case.

Sometimes, a translation is not an optimal mapping of word in language X to work in language Y. There isn’t a single function that optimally performs this, but a (potentially infinite) family of functions. This discussion reminds me of a great story I heard on Radiolab, where they discussed translation of one single French poem from the 15th century. (The person being interviewed wrote Godel, Escher, Bach, so you might consider checking it out.)

The poem was written by Clément Marot in the early 1500s, to send well wishes and a “get better” message to the daughter of the Queen, who was sick.

DOUGLAS HOFSTADTER: A sick young lady. Ma mignonne, je vous donne le bon jour; le séjour, c’est prison. Guérison recouvrez, puis ouvrez votre porte et qu’on sorte vitement, car Clément le vous mande. Va, friande de ta bouche, qui se couche en danger pour manger confitures; si tu dures trop malade, couleur fade tu prendras, et perdras L’embonpoint. Dieu te doint santé bonne, ma mignonne.

DH became really interested in translating this - why not translate this into english? What would that look like? If you were more literal:

Hello my dear. I’m sorry you’re sick. Being sick is like prison.
I, Clément, wish you to open your doors, get out into the world.
Get out of bed. Eat some jam.
So you don’t look so pale
and lose your plump shape.

But that loses entire spirit of the poem. So DH started translating it other ways:

My sweet dear. I send cheer. All the best. Your forced rest is like jail. So don’t ail very long. Just get strong. Go outside. Take a ride. Do it quick. Stay not sick. Ban your ache for my sake. Buttered bread while in bed makes a mess. So unless you would choose that bad news, I suggest that you’d best soon arise. So your eyes will not glaze. Douglas prays health be near, my sweet dear.

(what? Words are changing? It’s not a one for one translation anymore? But the spirit is retained, and now it’s more playful, which was lost in the first attempt.)

Someone else’s take:

Fairest friend, let me send my embrace. Quit this place, its dark halls and dank walls. In soft stealth, regain health. Dress and flee off with me, Clément who calls for you.

Eventually, DH had 700 pages of various translations, all of one poem! His favorite:

Pal petite, gal so sweet. Hug from Doug. Some dumb bug dragged you down. Zap that frown. Feel the urge, bugs to purge. From the scourge, you’ll emerge in a trice. Sound advice from – ahem – Doug slash Clem. So smash flu. Come. Yoo-hoo. Come you who live to chew. Sheets eschew. Sweets let’s chew. Pop a tart. Make your heart palpitate. Clem’s mandate. Sure hope God cures your bod head to feet. Pal petite


Hi toots. Get well. Hospital’s prison and prison’s hell. Get well. Flee your cell. Clément’s orders in a nutshell. Go pig out. Open wide your mouth. Keep those sweetmeats going south. Unless you’re hail, you’ll turn pale. Lose ooh-la-la that wiggles your tail. God restore good health to you my little flower, mon petite shoe.

Even in this incredibly short passage, there are an almost infinite number of translations.

Translation of any literary work isn’t a function you apply, but an intentional re-interpretation of the text, to maintain both the words and the spirit of the message. Translators become co-authors for writers, as they add all of their own nuance to a translation.

(Seriously, go listen to the start of that podcast, it’s freaking delightful.)

There is no one perfect Platonic translation.

Even today you will find different high quality translations with different focus points. And a good translation will have scholar notes on the translation explaining what this focus is.

For example I recently was rereading my copy of Tolkien’s translation of Sir Gaiwan and the Green Knight. As part of the author preamble, he explained what his intent was for the translation. Because the language used was atypical of the contemporary London style more akin to Shakespeare. It was a dark and stilted literary style drawing on an older midlands poetic tradition. As such Tolkien had a choice in how to translate. He chose to do so using a method to preserve the emotional impact of the original literary form, one that was intentionally a bit dark and harder to read. Because he felt that it was valuable to preserve that emotional feeling that would have been original to the author.

Likewise other translators have and can choose to make it flow more naturally for modern readers, in order to help capture the epic feel of the story.

Neither method is inherently superior. Both represent deliberate choice by the translator, and both yield similar, but substantively different, works. The story remains the same, but the feelings will vary.

Now to your average casual reader it is a highly nuanced difference with only minor notable terrain. But from a scholarly one, it is significant.

AI simply does not have that kind of intentionality, that considered forethought. It lacks the broader nuance and understanding of historical time and place. It can’t make a choice to deliberately make a translation with a specific emphasis. It simply does whatever the median sentiment analysis of the words says.

It may make a translation that seems entirely adequate and serviceable for the average person, who is unaware of the intent and nuance of an original work. There are things lost due to the AI translation they would never be aware of. But that does not mean the translation is just as good, because it isn’t. It just means most people lack the basis to judge the translation as inferior, as they don’t have the specialist knowledge to detect that. They are losing something, but have no idea they are.

So even the best AI translation will always be inferior, because it lacks that intentionality and forethought into the deeper aspects of translation.

Thanks for that, a delightful application of what I described. Unfortunately I don’t have my copy of Gaiwan so couldn’t transcribe the translation notes directly, but your post perfectly encapsulates what I was expressing.

Rebecca Kuang’s novel Babel is exactly about this tension between translation and authorship. She locates (literal) magic in the shades of nuance between even single words with direct translations into other languages.

Google bills this as “let Google do the googling for you.” Rather than searching the web yourself, you’ll delegate this task to Google. Hidden in this pitch is a tacit admission that Google is no longer a convenient or reliable way to retrieve information, drowning as it is in AI-generated spam, poorly labeled ads, and SEO garbage:

AI-powered search promises to fix this, not by making Google search results better, but by having a bot sort through the search results and discard the nonsense that Google will continue to serve up, and summarize the high quality results.

Now, there are plenty of obvious objections to this plan. For starters, why wouldn’t Google just make its search results better? Rather than building a LLM for the sole purpose of sorting through the garbage Google is either paid or tricked into serving up, why not just stop serving up garbage? We know that’s possible, because other search engines serve really good results by paying for access to Google’s back-end and then filtering the results:

Google makes a lot of money selling placement with search results. You think they’re just going to give that up? Nope, they’re going to sell placement with AI searches as well, and it’ll be even worse since it’ll look organic. You think advertisers won’t be able to buy the ability to be the resort chosen when someone asks Google to plan a vacation?

Enshitification is 100% guaranteed.

Google’s been fronting AI results for a few months now and there’s not a single time I’ve looked at the AI summary or thought it was useful. It’s just more cruft to scroll past.

Absolutely, the same goes for pretty much every interesting book with a broad audience. Like the Tao te Ching, there are some translations that attempt to be super literal, and some that attempt to better incorporate what the commonly understood meaning to contemporary readers would have been at the time of its creation. And both have value, depending on what you want to get or of it.

To me, this is actually what makes some of the most recent language models so amazingly interesting, because of their ability to perform different tasks, or the same task in different ways, based on modification of their input.

As an example, I was using a model to generate bridge to bridge chatter for simulated ships. One of the interesting things that I was able to do, very easily, was to tell it things like, “say this (English) in (other language),” and it would do a very serviceable job. Cool.

But what’s really cool, was that you could also say things like, “you only have a rudimentary understanding of the language,” and it would translate the English into a kind of simplified, crappy translation. And then in that particular case, it’s actually a more realistic effect, because you are trying to simulate a poor speaker, and the model grasps the language well enough to do it.

Now, my desire to have a crappy translation is a pretty narrow use case, but it highlights some cool things that these models can do. You can ask it for super literal translations. You can ask it to explain stuff to you like you are a child. You can ask it to explain stuff to you like IT is the child! And to me, that’s is just a super cool thing that has an infinite multitude of uses.

So what I’m trying to say here, is that some of the stuff you are talking about, these systems actually can already do, to varing degrees.

Microsoft’s producing a lot more planet-heating pollution now than it did when it made a bold climate pledge back in 2020. Its greenhouse gas emissions were actually around 30 percent higher in fiscal year 2023, showing how hard it could be for the company to meet climate goals as it simultaneously races to be a leader in AI.

Yeah, I view the Google AI search results the same way as the sponsored results that I already scroll past.

As far as Google AI enhancing search? They already poisoned the well with all the SEO nonsense and sponsored results, why trust they can fix this?

Glad we are boiling the planet to fix the search they enshittified.

There is a sizable Klingon population in the Louisville area.

Microsoft is supposed to be working on nuclear reactors to power it’s AI ambitions. If we can’t get the government to do it, might as well be a megacorp, right?

Imagine, 20 years from now, Microsoft TerraPower plants in every county and your GamePass subscription includes your electric bill!

Didn’t the elves who sailed west at the end of LotR settle in Kentucky?

So, is this people leaving OpenAI over concerns of safety and not wanting to be involved, or something much simpler and stupider.

These guys didn’t leave OpenAI because they were worried about safety issues and not being listened to, but that OpenAI is never going to create anything that would need a superalignment team to exist anyway. OpenAI is going to focus on making LLM related toys and things. OpenAI is just not going to be focusing on any sort of cutting edge AI that could ever be any sort of threat. Like, we are way too overqualified for what this company is doing now, and what is in the 5 year plan.

I mean, look at the twee videos they are putting out trying to make “Her” a reality. When you have a PHD in AI and ethics, this does not seem like a good use of your time, or something mentally engaging.