[casey2]

>よつば

>よっば

Just the fact that it misrecognizes this character means at least 10% of it's translations are going to be nonsense. That the creator puts this video up as a demo doesn't instill any confidence.

For any real translation app you will need something like deep research with a much longer context window and a much lower hallucination rate, better FIR and a slate with the input words and your confidence in them. Arguably a conscious robot is required.

Until OCR is a solved problem manga won't be a solved problem, and even then because I guarantee there is more wacky shit in comic books than in government and corporate documents, it still won't be 90% and that's just OCR

If you actually understood the scope of the problem, and were acting rationally, you would just go learn japanese. Of course if you compare current AI to bad localization you can say it's a million times faster and more accurate even though in reality it's worth than any bilingual person reading it and reporting the meaning. As I ramble perhaps pretraining with "this is the meaning" at various clause, sentence, paragraph and page levels for every popular or useful media will make Gemini or ChatGPT seem like magic.

[adibzaini]

Thank you for your feedback! I will try to improve the OCR and also the problem with understanding context.

[ryao]

  よつば = yotsuba
  よっば = yobba

As someone who has tried learning Japanese, it took me longer than I am willing to admit that to realize the two were different and things made much more sense after I realized the difference.

That said, learning Japanese is an extremely slow process and even if one does learn enough to understand a novel, one’s reading speed will be extremely slow, even if one is lucky enough to remember enough that it is not necessary to use a dictionary every other kanji. Even upon learning Japanese, it is quite tempting to dump the text into a LLM to have it translated.