/dev/null
Sign in

Commit

aaa3196

Tomokazu Kiyohara@kiyohara
There are some things that show up on screen already translated, and others that don't get translated unless you press the translate button. That basically means if translated data has been saved, it's being reused. In other words, when something already translated comes…
Translated from Japanese
ralf luartes@ralfrtl
now that you mentioned it, that's actually clever on the part of /dev/null team. show original for fewer translation request, if somebody translated it then store, then serve the translation moving forward
Tomokazu Kiyohara@kiyohara
You can keep translation costs down, right? Storage costs go up, and there will be more language-difference checks and extra loading queries, but translation using an AI model is the most expensive, so this feels like a reasonable compromise. Good sense of balance.
Translated from Japanese
Mathieu Céraline@mathcrln
Out of curiosity, how does doing it like that affect storage costs? If anything, it keeps it down by not storing transalations when not needed, no?
Tomokazu Kiyohara@kiyohara
What we were comparing it to was ambiguous, wasn't it? We were comparing it to an architecture where only the original text is saved, and it's translated and displayed each time the user presses the translate button. If you move the translation processing to the browser or client…
Translated from Japanese
Mathieu Céraline@mathcrln
Oh, yes it was indeed confusing, but now that I understand, I completely agree! Personally I think the current strategy makes the most sense. Inherently translation is something that should be stored since it’s likely to always be the same + many users are likely to need it…
Tomokazu Kiyohara@kiyohara
Like you, I think the current architecture is a good landing spot. But at the same time, what I think is… The act of pressing the translate button means that a user trigger is required. It requires the will to say, "Alright, I'm going to read this post!" I believe there is an…
Translated from Japanese
Pauline Vos@paulinevos
Translating everything is not an option for now. The inference costs for translations come from my own pocket, so does Tom’s inference. That’s why we chose this route + TTL on cached translations (because storage comes out of @nunomaduro’s pocket, lol)

commit aaa3196