The Economist’s language columnist Johnson joins the discussion of AI that’s been roiling the online world:
More and more people are using simple, free tools, not only to decode text but also to speak. With these apps’ conversation mode, you talk into a phone and a spoken translation is heard moments later; the app can also listen for another language and produce a translation in yours.
You may still get a surprise or two. Google Translate may be the best-known name in machine translation, but it often blunders. Take “my wife is gluten-free,” the kind of thing you might say at a restaurant abroad. In French or Italian, Google Translate renders this as “my wife is without gluten”—true to the words rather than the meaning. DeepL, a rival, does better, offering various options, most of them along the correct lines.
The best tool may not be a translation app at all. Though not marketed for the purpose, ChatGPT, a generative AI system that churns out prose according to users’ prompts, is multilingual. Rather than entering an exact text to translate, users can tell ChatGPT to “write a message in Spanish to a waiter that my wife and I would like the tasting menu, but that she is gluten-free, so we would like substitutions for anything that has gluten.” And out pops a perfect paragraph, including the way Spanish-speakers actually say “my wife is gluten-free”: mi esposa es celíaca. It is a paraphrase rather than a translation, more like having a native-speaking dinner companion than an automated interpreter.
There’s an idiotic anecdote from the frequently annoying Douglas Hofstadter, who “has argued that something profound will vanish when people talk through machines. He describes giving a halting, difficult speech in Mandarin, which required a lot of work but offered a sense of accomplishment at the end. Who would boast of taking a helicopter to the top of Mount Everest?” Yes, and when he was a boy he had to walk six miles to school in the snow, uphill each way. What’s the point of being in class at all if you’re not sweaty and exhausted when you get there? Then we get this sensible take:
As AI translation becomes an even more popular labour-saving tool, people will split into two groups. There will be those who want to stretch their minds, immerse themselves in other cultures or force their thinking into new pathways. This lot will still take on language study, often aided by technology. Others will look at learning a new language with a mix of admiration and puzzlement, as they might with extreme endurance sports: “Good for you, if that’s your thing, but a bit painful for my taste.”
And some people will go from one group to the other, and then back! People are different, one size does not fit all, and better, easier translation is a good thing. Thanks, cuchuflete!
I’ve used the Microsoft Translator app to gain some insight into the Chinese and Japanese texts in manuals for many years, especially when it clearly says something different (i.e., much longer) or completely incomprehensible gibberish in “English”.
I doubt learning another language or two will disappear that much. The reason virtually everyone in Brussels seems to speak Dutch to me (making it surprisingly hard to practice one’s French even there) isn’t going to disappear. Or it might, but not because of AI, anyway. A delay of at least a second or two is simply unavoidable — it has to listen before it can convey what it heard, after all.
Cars have not eliminated walking in general, but there are certainly people who will always, if they can, take a car to avoid a five-minute walk.
Relying on machine translation is not nearly as harmful as that.
A lot of Douglas Hofstadter’s peculiarities, both positive and negative, seem to be related to the household environment in which he grew up. In many respects, it was an extremely intellectually intense environment, but Hofstadter also had a sister who was severely autistic and nonverbal, which also affected him quite a bit.
Even Google Translate is undeniably useful, depite its severe limitations. It woud be stupid to deny that LLMs can be useful to actual intelligent agents wishing to translate.
But (yet again) calling them “AI” is a pure marketing gimmick (where it’s not simple confusion or intellectual idleness.) It’s just like calling powered flight “antigravity.” People need to stop doing this. (There is hope: I haven’t seen the offspring of IVF called “test tube babies” for a long while now.)
Hofstadter is a very clever man but he’s always been full of it. He’s persuasive because he truly believes it all himself.
From the article: “Nor do most people learn languages for the purpose of humanising themselves or training their brains. On their holiday, they just want a beer and the spaghetti carbonara without incident”.
Reading that, I can well imagine the author sitting irritated at a table in Milan thinking “why don’t these damn foreigners speak English?” Perhaps she would have been better off visiting Disneyland, where they have clean, safe English-speaking versions of foreign places, and no requirement for humanizing oneself.
Yes, it’s an essentially subhuman understanding of translation: as such, quite possibly amenable to emulation by systems with no understanding of … anything at all.
Hofstadter is certainly clever enough to see this: I presume he’s pretending not to, so as to be hip and happening and all.
I think I would find a program that identifies languages spoken around you quite useful.
(unrelated to anything in this thread, but it has the right title)
Also “identify a song” – as long as it is able to identify something in a moderately obscure (as obscure as possible to still be on the internet/youtube) language.
Someone listened to music in a language I can’t identify right under my window.
“People need to stop doing this.”
DE, please, suggest a better name such that my interlocutors would be able to understand what I’m speaking about.
I call them AI not because everyone does.
But (yet again) calling them “AI” is a pure marketing gimmick
Yet again, no.
something profound will vanish when people talk through machines
Like when people started to talking to each through the telephone rather than sending letters or attaching messages to pigeons?
(I’m being facetious but not entirely).
It is not entirely clear what the word of science-fiction-style translators will give us. Maybe a lot, maybe so much that what we lose is simply negligible. Or maybe not. I just don’t know.
Thus far, machine translators hardly make my own life more interesting.
And one more thing: if I want to visit Yunnan, I need Yunnan where almost no one understands me, not Yunnan where everyone speaks English.
Seriously, English is your “machine translation”, when you come to a country where everyone knows English, you understand people, people understand you and you still have their exotic culture.
I know what it feels like and no, I’m quite curious about native English speakers but knowing that everyone knows English or Russian or … as L2 in some country does NOT make it more attractive to me:)
“Yes, and when he was a boy he had to walk six miles to school in the snow, uphill each way. What’s the point of being in class at all if you’re not sweaty and exhausted when you get there? ”
Can we also say that virgin forests with roads and McDonalds are better than just virgin forests?
I have been asking chatGPT to do some writing in my L2. Perhaps this is my imagination but it seems that the output accuracy has gotten significantly worse recently.
Since the model acquires content from the internet and so much more is written by chatGPT and published as-is these days, I have heard it posited that we are already getting into a recursive model degradation feedback loop.
@davidL
I am rather uncomfortable having difficult conversations over the telephone (this may be because I lack emotional intelligence, at least compared to many of my interlocutors, and depend on nonverbal signals to avoid “putting my foot in it”). So I would rather ask for a meeting or at least a videocall.
This has been noted in the LLM community and studied. It is generally believed to be a direct side-effect of the steps they take after training to keep the models from going off on antisemitic sexist homophobic rants and such. The hackers running models on their own machines all want to get the unfiltered ones for this and other reasons.
The model does not continue to acquire content from the internet. We don’t yet have models that can be continually trained. The chatbots I use only have data up to 2021 which is when the training data was collected. They are trained once at huge expense. There is some fine tuning and other things that can be done after training that I don’t understand the details of though. Because it’s so expensive they take some time between new versions when they tweak the code and massively increase the parameters, the training data size, and the compute-hours spend on training.
A couple of chatbots can do internet searches and incorporate the search result into their answer. You.com and Bing I think can both do that. But they can’t incorporate the results of those searches into their model.
We haven’t gotten into a big feedback loop yet with the major models, but we surely will. Maybe some of the smaller ones that are still scraping the web since GPT came out are already suffering from this problem though.
For anybody interested in how they really work, there are two quite good quite technical but still quite comprehensible write-ups that don’t dumb it down to junk phrases like “fancy markov chains” or “pattern matching on steroids”:
What Is ChatGPT Doing … and Why Does It Work?
A jargon-free explanation of how AI large language models work
The phone is such a suboptimal option that many people have, shall we say, made the call to avoid it.
Hofstadter is certainly clever enough to see this: I presume he’s pretending not to, so as to be hip and happening and all.
An intellectual given unlimited space will probably produce a range of nuanced statements of the on-the-one-hand/on-the-other-hand class.
A journalist has limited space, and in processing such lengthy output may select one of the statements and rely on other sources or authorial commentary to provide nuance, falsely implying the quoted intellectual has none.
A public intellectual asked for a soundbite may sometimes save the journalist time by making this selection themselves.
jargon-free explanation of how AI LLMs work
I was surprised to learn that they use animals:
#
The training process happens in two steps. First there’s a “forward pass,” where the water is turned on and you check if it comes out the right faucet. Then the water is turned off and there’s a “backward pass” where the squirrels race along each pipe tightening and loosening valves.
#
Well, not that surprised. I had already noticed that squirrely people get excited about it.
DE, please, suggest a better name such that my interlocutors would be able to understand what I’m speaking about
How about “Large Language Models”? Happily, this actually does seem to be catching on with sensible commentators.
(“LLM” has the further advantage for Welsh speakers of evoking llwm “bare, barren, naked, threadbare.”)
llwm
Cognate with Latvian lupt ‘to peel; eat’ and thus clearly with Hausa loma ‘mouthful; piece of food.’
https://dil.ie/search?q=lomm&search_in=headword
Did we get it from the Welsh?
The hypothetical traveler’s wife is presumably made out of human flesh, which is indeed gluten-free. Gluten-avoiding cannibals have nothing to worry about. It seems like an odd thing to want to tell a stranger, though, whether or not the stranger understands English.
@J.W. Brewer: It sounds very much like a foreign custom that might have been described by Montaigne: announcing one’s palatability to people with various dietary restrictions, for the benefit of any local cannibals who one might up nourishing. “My wife is gluten free, but she isn’t kosher, because she has a clubfoot.”
The reason virtually everyone in Brussels seems to speak Dutch to me […] isn’t going to disappear.
Since you mention it, why is that? I should think there were few Francophone cities where people address strangers in Nonfrench (a language related to Undeutsch, an old name for Latvian).
Hofstadter also had a sister who was severely autistic and nonverbal, which also affected him quite a bit.
Good point. It makes me wonder how my younger grandchildren will grow up if their older brother (not Dorian) remains nonverbal. A speech therapist I know swears he can be taught to speak, it’s just that nobody has consistently tried.
spaghetti carbonara without incident
Surely that should be without incidenti, an Italian condiment of some sort.
Like when people started to talking to each through the telephone rather than sending letters
Indeed, the decay of letter-writing was a multigenerational thing, only ended (for a while) by the rise of email.
So I would rather ask for a meeting or at least a videocall.
For me, it’s the contrary: I find in-person meetings exhausting compared to phone calls.
The phone is such a suboptimal option that many people have, shall we say, made the call to avoid it.
I think that has more to do with the phone’s bell (i.e. its ability to demand attention) than with anything about the phone as a medium. Many people wish to be able to communicate with someone at their will and regardless of the other person’s will, so it suits them to use it, but it will be a long time before people are wiling to text or email emergency services (fire, police, ambulance) There are already jokes about people dialing 911/999 and being told “All our operators are busy helping other victims, but your call is very important to us” etc. etc.
How about “Large Language Models”?
I tend to hear that as “(Large Language) Models”, and I also often see it as LMM, though perhaps these issues will go away.
spaghetti carbonara without incident
Surely that should be without incidenti, an Italian condiment of some sort.
Molto divertente. Spaghetti carbonara without mishaps (incidents), or without side orders (incidentals), or without any bits that get left between your teeth (interdentals, or according to some authorities “inside-dentals”). Not be confused with spaghetti alla puttanesca, senza nulla di indecente.
It is 9 p.m. and your wife is still alcohol-free!
@John Cowan:
> Since you mention it, why is that? I should think there were few Francophone cities where people address strangers in Nonfrench (a language related to Undeutsch, an old name for Latvian).
A not dissimilar reason as why Brussels switched from Dutch to French in the late 19th century, i.e., mainly economics, but I don’t expect we’ll be able to witness a reversal. The economics were coupled with discriminatory practices by the Belgian state and elites back in the 19th century.
> I think that has more to do with the phone’s bell
Indeed, the way to go is to ask if it currently fits to call on Slack/Discord/Skype/SMS.
This has been noted in the LLM community and studied. It is generally believed to be a direct side-effect of the steps they take after training to keep the models from going off on … rants and such.
Where would the training data have come from that recommends (for example) tasting a mushroom in the wild to see if it’s an edible variety? Pro tip: if you drop dead, it isn’t.
@AntC For the most part those systems are happy to oblige nonsense questions with made up answers that are a reasonably plausible description of how things would be if the proposition weren’t nonsense. Of course with this specific example being in the news the model will probably be whipped into shape to respond “safely” to queries about wild mushrooms, but for now I’ve got this:
That is funny. Especially the “tú” and the “Querido”. William would conclude that Philip had lost his mind.
ChatGPT did as well, in completely neglecting the mushrooms. But it seems to be aware of it.
.
The system seems to have been improved (or neutered?) to be much less imaginative since I tried it half a year ago. But perhaps that’s for the best, as long as you adjust your prompt slightly.
.
Proust, ChatGPT and the case of the forgotten quote [Grauniad/Elif Batuman]
And it wasn’t: ChatGPT just made up quotes from Proust (in atrocious French). The next day, she went back for more punishment; asked the same questions; got entirely different replies …
She seems incredibly naive. Is it genuine naïveté or faux-naïveté? Hmm, maybe I’ll ask ChatGPT…
The first example of this kind of thing I ran across was when a usually reasonable political writer was doing an article that involved heights of American presidential candidates, and he explained that when he couldn’t find a height for a person he asked ChatGPT. That to me seems an obviously ridiculous thing to do, but when Google and Microsoft are promoting LLMs to replace search engines it’s hard to blame people who don’t know much about LLMs for believing them.