Should we enable automatic natural language translation here?

Meta Discussion
21

Is it one million credits per day, or per month?

EDIT: Per here and here, it’s per day.

22

Maybe I am missing something, but I find it difficult that I cannot really endorse a translation. While my native language version remains the global source of truth, I cannot really judge the quality of all possible translations.

I would thus greatly prefer a translation to some lingua franca, which I understand and where I can confirm: Yes, that is what I want to say.

23

The post author (along with mods and admins) can actually edit the translations! In fact, we could turn on multilingual support independently of the automatic LLM-based translator, wherein all such translations are hand-written.

I asked this question on meta: Cómo se presentan las traducciones editadas manualmente - Translations - Discourse Meta… but I did it in Spanish and then edited the English. It doesn’t look like you can see that edit history or the fact that I’ve “blessed” the English, however.

image
image1610×1874 402 KB

24

So either you run the risk of overspending, or you risk a degraded user experience for those that rely on translation? Also who’s going to pay for this anyway. 1 million tokens doesn’t sound like a lot.

25

It’s a million tokens a day, and we’re posting tens of thousands of words per day. Of course words are not tokens, but it seems like we could keep up with translations at this budget. Backfilling the translations of the previously-posted 60 million words would take quite some time at that rate, however.

26

But anyone browsing the forum will consume tokens right? Since all of that will be translated as well. So you should look at the number of words read per day, not posted? On top of that, each post has to be individually translated for every requested langauge. So for sure the minimum is N_words * N_languages

27

No, the translations are translated once and then stored — that’s how you can edit them manually. They’re not translated as they’re read… it’s as they’re posted. And then the system can “backfill” (or not!) to some particular date(s) in batches.

28

Time to talk entirely in Julia, the solution to the 7000+ language problem.