Comment by throwaway2037
9 days ago
At the very bottom of the article, I see this notice:
> This article was originally written in Korean and translated by a bilingual reporter with the help of generative AI tools. It was then edited by a native English-speaking editor. All AI-assisted translations are reviewed and refined by our newsroom.
I like that. It is direct and honest. I'm fine with people using LLMs for natural language related work, as long as they are transparent about it.
Especially since LLM tech was originally developed for translation. That’s the original reason so much work was done to create a model that could handle context and it turned out that was helpful in more areas than just translation.
While LLM usage is just spinning up in other areas, for translation they have been doing this job well for over 5 years now.
Specifically, GNMT came out in 2016, which is 9 years ago.
GNMT used seq2seq with attention to do translations. GNMT plus some RNN and attention lead to transformers, and here we are today.
> While LLM usage is just spinning up in other areas,
Oh?
This is how I’ve done translation for a number of years, even pre-LLM, between the languages I speak natively - machine translation is good enough that it’s faster for me to fix its problems than for me to do it from scratch.
(Whether machine translation uses LLMs or not doesn’t seem especially relevant to the workflow.)
My partner is a pro-democracy fighter for her country of origin (she went to prison for it). She used to translate english articles of interest to her native language for all the fellow-exiles from her country. I showed her Google translate and it blew her mind how much work it did for her. All she had to do was review it and clean it up.
The AI hype train is bs, but there're real and concrete uses for it if you don't expect it to become a super-intelligence.
I agree 100% with this sentiment. Another good use case: Ask an LLM to summarize a large document. Again, not super-intelligence, but can be a big timesaver to reduce "intern work". I have heard some people have a LLM plug-in to their Microsoft Outlook (Exchange) that allows them to summarize an email thread. Again, not perfect, but helps to reduce cognitive load. Another practical example: Using an LLM with conference calls to transcribe meeting notes and provide a summary. Then you can review the summary, fix any obvious errors, and send by email to participants.
> The AI hype train is bs, but there're real and concrete uses for it
When you consider that there are real and concrete uses for it across a wide variety of domains, the hype starts to make more sense.
Obviously Sam “we’ll build a Dyson sphere with it” is off in hype lala land somewhere while he tries to raise a trillion dollars to burn through fossil fuels as fast as possible, but that’s a kind of symptom of the real underlying capabilities and promise here.
That footnote does make me question the bilingual reporter's skills in both languages though. If the reporter needs an LLM to help translate they could easily be missing subtle mistranslations.
The final note that all AI-assisted translations are reviewed by the newsroom is also interesting. If they are going to take the time to review it and have enough experience in both languages to verify the translation, why use the LLM for it at all?
> That footnote does make me question the bilingual reporter's skills in both languages though. If the reporter needs an LLM to help translate they could easily be missing subtle mistranslations.
I've done my fair share of translating as a bilingual person and having an LLM to do a first pass at translation saves TON of time. I don't "need" LLM, but it's definitely a helpful tool.
> If they are going to take the time to review it and have enough experience in both languages to verify the translation, why use the LLM for it at all?
People generally read (and make minor edits if necessary) much faster than they can write.
If using LLM can shorten the time reporter needs to rewrite the whole article again in the language the reporter is fluent but take effort to write, why not?
This will give the reporter more time to work on more articles, and we as a foreigner to Korea, getting more authentic Korean news that is reviewed by Korean and not be Google Translate.
You raise an interesting point about "missing subtle mistranslations". Consider the stakes for this article: This highly factual news reporting. There are unlikely to be complex or subtle grammar. However, if translating an interview, this stakes are higher, as people use many idiomatic expressions when speaking their native language. Thinking deeper: The highest stakes (culturally) that I can think of is translating novels. They are full of subtle meanings.
The reporter does not need the LLM, but it's often faster to review/edit a machine translation than doing the whole translation by yourself
> It was then edited by a native English-speaking editor.
Two different editors.
But as others mentioned, this is helpful even for the same editor to do.
As long as the LLM doesn't hallucinate stuff when translating, by generating text that is inaccurate or even completely fabricated.
why would you not be fine about it?
You probably don’t want to read news websites which are nothing but LLM output without a journalist reviewing the articles. Unless you’re a fan of conspiracy theories or ultra-aligned content.
Case in point:
A New Gaza Rage Machine–With Polish Origins - https://news.ycombinator.com/item?id=45453533
So just a blanket message at the bottom of the page "anything and everyone you read here might be total bullshit"
FWIW that happens sometimes with traditional reporting to. At the end of the day, it's just a matter of degree, and to be truly informed you need to need to be willing to question the accuracy of your sources. As the parent comment said, at least they're being transparent, which isn't even always the case for traditional reporting
> I'm fine with people using LLMs for natural language related work
phew I'm relieved you're okay with people using modern tools to get their job done
It's still worse than useless :
https://www.bloodinthemachine.com/p/ai-killed-my-job-transla...
I really don't get this take where people try to downplay AI the most where it is obviously having the most impact. Sure. A billion people are supposed to go back to awful machine translation so that a few tens of thousands can have jobs that were already commodity.
I have sympathy for those affected but this article is disingenuous. I speak Spanish and have just gone to 3 or 4 Spanish news sites, and passed their articles through to ChatGPT to translate "faithfully and literally, maintaining everything including the original tone."
First it gave a "verbatim, literal English translation" and then asked me if I would like "a version that reads naturally in English (but still faithful to the tone and details), or do you want to keep this purely literal one?"
Honestly, the English translation was perfect. I know Spanish, I knew the topic of the article and had read about it in the NYTimes and other English sources, and I am a native English speaker. It's sad, but you can't put the toothpaste back in the tube. LLMs can translate well, and the article saying otherwise is just not being intellectually honest.
What isn't tested here, and what I can't test myself as a mono-linguist, is how well english is translated to other languages. I'm sure it's passable, but I absolutely expect it to be less sufficient because most of the people working on this live in the USA / speak english and work the most on that.
I want to know how it holds up translating Spanish to Farsi, for example.
2 replies →