And if I may complain some more… I don’t know why it is, but Japanese people are terrible at learning languages and understanding how language in general works. It’s the biggest reason why I have a career in the first place. But it’s also hampered money-making at times, as Japanese companies will settle for garbage translation and/or interpretation without the first clue as to how that might be negatively impacting the company.
Case in point: I had a discussion with a recruiter not too long ago about a potential industrial translation job back in December in BFE Georgia. It sounded good at first, but then the recruiter didn’t get back to me (recruiters suck, by the way), so I called him up, and he told me that the client had decided that someone with no experience at all was fine for the job, so someone expensive like me was not needed. And I knew that this was just ***ing idiocy, that their experience using me would be absolutely, completely different than using a rank beginner (typically a kid right out of school who doesn’t even really know the language). But what can you do? Japan is often cheap and stupid.
Right, and this is a key point. Using the phone app that can read Korean and tell you where the bi bim bap is on the menu is a real help.
I’m not my wife (shocking) but I’d think you’d save a substantial amount of time in just not having to manually rewrite the document, much less the mental labor of rewriting it correctly in another language. But everyone has their own workflow and perhaps you personally wouldn’t find it helpful. The service she uses is DeepL which also might have some workflow benefits vs other general purpose LLMs. She also primarily translates for Spanish which I assume has benefits in both similarity to English and wealth of examples to train a model on.
I think that part of this, at least, is people gravitating to the same models so you see a lot of the same sort of output, especially if the model isn’t really trained specifically for that sort of thing. To use image generation again (just because I’m most familiar), I can easily tell most AI images immediately because most of them are people using Bing and ChatGPT and it all has the same glossy cartoonish look to it. By that basis, I could easily say that all AI image gen is “slop” and trash, etc. But there’s other models able to generate some extremely convincing photorealism or artistic-looking works that are nothing remotely like “Prawn Jesus on Facebook”. They often require some sort of cost on the part of the user so the vast majority of images are through the free services and give the impression that all AI image gen is the same.
I suspect the same goes for the LLMs as well – every AI written bit sounds like every other AI written bit because they all came from the same models kind of like the bit about nearly every top-charting pop song being written by the same couple of Swedish guys.
And the question is also, sure, a lot of AI content is easy to spot a mile away. The models matter and also how you prompt it and sculpt it to give an answer matters, too. More importantly, is how do you even know how much AI content you’ve completely missed? Obvious AI is obvious. But what about non-obvious AI? I’m sure there’s AI content I’ve missed. And there’s human content that I’ve mistaken for AI. The line can be fuzzy, especially if you’re the structured type of writer that AI out-of-the-box likes to be.
Right. If it’s just “translate so as to be accurate and understandable,” then I wouldn’t rewrite it. But I don’t have all that much experience in translating such documents (e.g., long contracts, medical texts, etc.). It’s either been marketing, technical (to the point where the AI probably doesn’t have the correct specialized vocabulary, which must be teased out from other texts by the client), or both at the same time. E.g., I did a marketing piece for Honda’s AWD system back in 2023, which had to use Honda’s terms and not just generically correct terms, and sound like both technical and marketing language.
Yes, and tools have been developed for certain types of translations as well, that I don’t know about or at least haven’t used.
I know also from a friend getting into offering translation online that there are some people out there charging an insanely low amount for massive translations. I have to assume that, in such cases, the human input is minimal.
Yes. Another issue, one that I imagine is hard to correct, is that the generated texts can be almost too good without seeming great or particularly personalized. For example, one tell for me in the YouTube video scripts is the use of long, Ciceronian periods with a bunch of appositives, participle clauses, etc. A bit of an exaggeration, but I’m sure you know what I mean. These are tools that very good writers should have in their toolbox, but while great writers know how to use them judiciously and with the proper frequency, AI does not.
Good point. The thing is, I encounter the same problem with AI-generated art: it’s a bit too good and facile. There’s a YouTube channel called the Daily Stoic, and this dude was putting a frig-ton of pretty damn good AI art in his videos: Like a bronze statue of Marcus Aurelius wearing like business casual and tastefully animated with a pie chart behind him in an office. Stuff like that. Over and over. And it’s just visually and mentally fatiguing to watch and… just bad, in its own way.
And I think that’s a danger of AI that is as yet underappreciated. You are starting to hear, “That looks/sounds like it’s AI-generated!” all the time. Sure, a lot of this has a spoken (or unspoken) “because it sucks!” appended to it, but more and more it’s as if just getting “busted” for using AI is enough to engender disrespect. Because it’s not real, organic, etc. There are many, many contexts in which a serviceable output and nothing more is totally fine. But there are many others for which I think humans tend to feel that some sort of effort or intention is necessary for it to qualify.
For my part, I don’t think AI can write a “great” novel–yet. I also feel that I don’t particularly want it to be capable of that, but I definitely feel that I don’t want to read such novels at all were they to exist.
Absolutely agree. It can be good in helping a writer with ideas and themes, but to be “truly creative” – whatever that means – it doesn’t seem to quite have that “spark” yet that makes good human creative writing still quite distinguishable from an AI simulacrum. That’s one area where it’s lags quite a bit, IMHO, behind humans. It can create “original” ideas, but not ones I’ve really seen that are really interesting to a human or as an extended piece of prose. (And it’s even worse at poetry.)
Right. Another issue is that I think, when it comes to creative endeavors in general (novels, poetry, music, visual art, etc.) most of the low-hanging fruit has been picked, so artists really have to dig to come up with something that captures people’s attention at all. (I think this is a big and underappreciated reason why so many movies are flopping these days.) This is also where, it seems to me, AI is weakest, rendering it more or less incapable of generating affecting art.
Also, in the case of using AI to generate ideas, that is a combination of artificial intelligence and human intelligence to comb through the ideas and find something workable. In fact, such human intelligence is a big factor in most of the AI-generated art we see. If a human goes through 1,000 songs or 1,000 images to find that masterpiece, then the human intelligence factor in the resulting end user experience is indispensable.
I’m definitely not going to underplay the importance of human curation in the process. It’s part of what I would consider successfully using AI entails. It’s a bit like brainstorming: get the room to throw out 15 slogan ideas for you and, as director, sift through the ideas, see which ones have potential, work from there. You can get some reasonable results in one-pass for straightforward prompts, but for anything requiring creativity, usually human input/judgment of some type is necessary.
As I said earlier, I was involved in the translation industry in the early 90, and machine translation sucked. A couple of companies tried to get us to use their services, but it seemed that simply throwing the translation out and starting over by person was as fast. It actually was more annoying because of the stupidity of the word choices.
IMHO, this really isn’t a fair criticism of AI, as it’s already better than a good percentage of traditional human translators.
When I was in the translation field, I also did copywriting and saw all the issues you have described. Marketing is a different creature and requires different skills and knowledge.
However, LLM translations, with minimal rewriting, are as good now as a lot of the work I was involved with back in the 90s.
Not all restaurants are Michelin class. Not all translation requires the greatest skilled translators.
I think most people don’t understand the limitations of AI translation and consequently have too great of expectations. For those who don’t, it can be a good tool.
I did not know machine translation existed in the early 1990s. My first experience of it was Google translate in the late 1990s or early 2000s, and I saw the results of the specialized medical translation software in Japan in 2001. It was pretty bad.
In the text of mine you quoted, I was not criticizing AI translation per se but simply stating my opinion about its capabilities wrt higher-end translation. I agree that it’s better than a pretty large percentage of human translators, inasmuch as most human translators are pretty bad. The reason for this being fairly basic: it’s really hard to know two languages well enough, as well as to understand one’s specific strengths and weaknesses in them, to be a great translator. Put another way: How many people do you know who are truly great writers of English? Not many, I bet. But unless you are a very good writer in English, you will not be a very good translator of a specific language into English, and vice versa.
But this fact is also instructive about the weakness of AI wrt translation: AI tends to be a generically “pretty good” but not great writer of English. By its very nature, AI writing tends to be risk-averse, mean-reverting, and stylistically bland in its approach. It needs to fulfill the prompts and not be too crazy. This doesn’t work for marketing materials, video scripts, and fiction. In order to work, very perceptive, creative, and aggressive AGI would be necessary.
I can still remember a book I read (fiction) about teaching machines to translate. The expression “out of sight; out of mind” was translated as “invisible idiot.” I liked that.
Short answer: They claim to have created the first topological qubit (they’ve been working on it for about 15 years), but the evidence is not conclusive. And topologically protected states have been proposed for many years, but they aren’t “a new state of matter”.
So I held off on commenting on this, first of all because Scott already covered most of the important points, but also because I simply haven’t had the time to look into this in great detail. But I think it’s worth pointing out this caveat hidden in the (publicly accessible) referee report on the paper (scroll down to the bottom to download the ‘peer review file’):
The editorial team wishes to point out that the results in this manuscript do not represent evidence for the presence of Majorana zero modes in the reported devices. The work is published for introducing a device architecture that might enable fusion experiments using future Majorana zero modes.
So, hype notwithstanding, Microsoft has not demonstrated the presence of topological modes in its experiment, much less shown that they can do any actual quantum computation with them.
Here’s a recent Fortune article going into some more detail regarding the likely death of the pure scaling approach:
OpenAI has quietly made an about-face on its core strategy. After years of preaching a bigger-is-better approach that calls for pre-training models with ever more data, Altman effectively acknowledged the scaling technique was no longer producing a big enough performance boost.
Huh. I’m surprised this is that recent. I could have sworn people realized pure scaling was not sustainable a while ago, like at least a year ago. And I don’t see that as a bad thing or a necessary impediment to the development of AI.
There were certainly earlier claims and indications, but at least outwardly, many key figures in AI (mostly those with a sizeable stake in the matter) have held fast to the sort of ‘scaling is all you need’-picture.
(I, certainly not a ‘key figure’ in anything, on the other hand was skeptical of this already more than four years ago, in terms of a metaphor also employed by the Fortune article:
What GPT-3 lacks is the corrective facility System 2 lends to System 1’s free-wheeling association. There is no consistency check to disabuse it of the notion that Emily’s father’s three daughters are named April, May, and June: it sees the pattern, and completes it in a sensible way; but it does not understand how the concepts the words map to interrelate.