|
|
|
|
|
by NitpickLawyer
196 days ago
|
|
> It makes me less eager to create which is quite unfortunate. It's been the exact opposite for me. Coding assistance is a great boon towards productivity simply because otherwise I wouldn't work on any of my old ideas stashed in numerous note taking apps. It's way easier today to go from 0 to something like an MVP, and see if there's something there. If there isn't, not much is lost. But without these tools it would be 0 all across the board. |
|
So, how hard can it be to translate them automatically? Tried a few ready-made tools and they either just plain didn't work or made stupid mistakes due to bad prompting.
I fired up Claude and had an MVP that could 1) rip english subtitles out of a mkv file 2) shove them to OpenAI API for translation in batches within a few hours
v2 took two evenings of me watching TV and bouncing around between Claude+Codex+Crush(GLM-4.6)
1) Get subtitles from original
2) Feed full subtitle file to GPT-4o for first pass analysis. It finds names for people, locations etc and decides a singular translation so they don't vary and gives a general context for the tone/genre of the movie for translation
3) Give gpt-4o-mini the first pass context file and a section of the subtitles in a loop
4) save them as .srt
The results are SO MUCH better than most of the "real" translation we've encountered. I think a bunch of them were done by a zero-shot AI or a human who didn't give a fuck about quality and had to meet a quota.
And none of this would've been done without AI, there's so much crappy boilerplate on top of the few actually interesting bits I would've just tolerated the bad ready-made solutions instead.