Why haven't we auto-translated all AI alignment content?
5NicholasKees
3Nition
3NicholasKees
2Algon
4AnthonyC
5kaiwilliams
5Algon
2AnthonyC
2Algon
2AnthonyC
2[comment deleted]
New Answer
New Comment
Or a newsletter which was natively multi-lingual (e.g. Rohin Shah's Newsletter was always translated to Chinese, though not by AI). Or a forum where people can discuss AI in whatever language they prefer, and things are automatically translated between users?
It seems like there are a lot of ways cheap translation could broaden the conversation to include people not in the Anglosphere. The cost is that AI translation will often make mistakes (even human translation is imperfect), but I'm not sure why that cost isn't worth paying. Currently most people outside the Anglosphere need to rely on local elites to decide what ideas are worth taking seriously (e.g. a local could report on AI 2027, like this Dutch summary. Also apparently the NYT translated their coverage of AI 2027 into Spanish, which seems cool.)
Or a forum where people can discuss AI in whatever language they prefer, and things are automatically translated between users?
Cool idea. I've never seen this done, yet it sounds very achievable with today's tech. I presume you'd set your desired language to whatever you prefer, and every post would now appear in that language, either natively or auto-translated. Put a little "translated from x, click here to view original" on each auto-translated post. Write your own posts in whatever language you like and everyone would see them in their preferred language.
When I order food on UberEats this already happens automatically when I chat with a delivery person who doesn't speak English. Similar thing for reviews on several websites.
1Or a newsletter which was natively multi-lingual (e.g. Rohin Shah's Newsletter was always translated to Chinese, though not by AI). Or a forum where people can discuss AI in whatever language they prefer, and things are automatically translated between users?
Sounds good.
I'm currently at like 45% that this is a good idea, and people just aren't doing it, and at 35% that people are doing it, but I just don't know about it.
I think this seems like a good idea, but even among native English speakers a lot of readers seem to ignore, disbelieve, or misunderstand all the places people warn about the importance and nuance of subtle details (e.g. why you should assume all the obvious thoughts that come to mind about how to solve or do X won't work by default), which are just the kinds of things that are likely to get somewhat garbled in translation. (Note: I'm not asserting that all such warnings are true, but I do think they are important). Not sure what kinds of disclaimers or warnings should be included with this kind of translated content, but I would definitely hope there are some.
One cheap test might be to translate into Dutch and then back to English with the same prompt twice. How garbled does the output end up? Does it elide important nuances?
(Though this might overestimate the quality given that the original pieces in English are likely in the training data. If you're fluent in another language, I'd be quite curious about the results in the target language).
That is a sensible test! However, in my testing, I just asked a fluent-in-Dutch co-worker about the results, and he was satisfied. There were some problems, but it was basically OK.
That's good news!
Depending on how high a bar you want to meet, it may be worthwhile to also try with a language farther from English than Dutch is. If I really wanted to be thorough, I might try to take two bilingual people, give one the English and one the non-English, and ask them to discuss the results with one another and/or with you. Or, even better, give it to people who don't speak English, and don't know it was translated from English, and therefore can't subconsciously use knowledge of English to correct translation problems. But for a lot of use cases I agree that these would be unnecessary.
As in, the warnings will get garbled, or the subtle details will? I agree the subtle details are more likely to be garbled. Which would suck for e.g. a young alignment researcher trying to deeply understand some translated article, say. Still seems plausibly worth it? Even more so if the person reading doesn't intend to be a researcher, or gain a deep understanding.
(I hope this doesn't show up as an answer to my own question...)